Datasets:

memray

/

nus

like 0

191

The Potential of the Cell Processor for Scientific Computing

The slowing pace of commodity microprocessor performance improvements combined with ever-increasing chip power demands has become of utmost concern to computational scientists . As a result, the high performance computing community is examining alternative architectures that address the limitations of modern cache-based designs. In this work, we examine the potential of using the forthcoming STI Cell processor as a building block for future high-end computing systems. Our work contains several novel contributions. First, we introduce a performance model for Cell and apply it to several key scientific computing kernels: dense matrix multiply, sparse matrix vector multiply, stencil computations , and 1D/2D FFTs. The difficulty of programming Cell, which requires assembly level intrinsics for the best performance, makes this model useful as an initial step in algorithm design and evaluation. Next, we validate the accuracy of our model by comparing results against published hardware results, as well as our own implementations on the Cell full system simulator. Additionally, we compare Cell performance to benchmarks run on leading superscalar (AMD Opteron), VLIW (Intel Itanium2), and vector (Cray X1E) architectures. Our work also explores several different mappings of the kernels and demonstrates a simple and effective programming model for Cell's unique architecture . Finally, we propose modest microarchitectural modifications that could significantly increase the efficiency of double-precision calculations. Overall results demonstrate the tremendous potential of the Cell architecture for scientific computations in terms of both raw performance and power efficiency.

INTRODUCTION Over the last decade the HPC community has moved towards machines composed of commodity microprocessors as a strategy for tracking the tremendous growth in processor performance in that market. As frequency scaling slows, and the power requirements of these mainstream processors continues to grow, the HPC community is looking for alternative architectures that provide high performance on scientific applications, yet have a healthy market outside the scientific community. In this work, we examine the potential of the forthcoming STI Cell processor as a building block for future high-end computing systems, by investigating performance across several key scientific computing kernels: dense matrix multiply, sparse matrix vector multiply, stencil computations on regular grids, as well as 1D and 2D FFTs. Cell combines the considerable floating point resources required for demanding numerical algorithms with a power-efficient software-controlled memory hierarchy. Despite its radical departure from previous mainstream/commodity processor designs, Cell is particularly compelling because it will be produced at such high volumes that it will be cost-competitive with commodity CPUs. The current implementation of Cell is most often noted for its extremely high performance single-precision (SP) arithmetic, which is widely considered insufficient for the majority of scientific applications . Although Cell's peak double precision performance is still impressive relative to its commodity peers (~14.6 Gflop/s@3.2GHz), we explore how modest hardware changes could significantly improve performance for computationally intensive DP applications. This paper presents several novel results. We present quantitative performance data for scientific kernels that compares Cell performance to leading superscalar (AMD Opteron), VLIW (Intel Itanium2), and vector (Cray X1E) architectures . We believe this study examines the broadest array of scientific algorithms to date on Cell. We developed both analytical models and lightweight simulators to predict kernel performance that we demonstrated to be accurate when compared against published Cell hardware result, as well as our own implementations on the Cell full system simulator. Our work also explores the complexity of mapping several important scientific algorithms onto the Cell's unique architecture in order to leverage the large number of available functional units and the software-controlled memory. Additionally , we propose modest microarchitectural modifications that could increase the efficiency of double-precision arithmetic calculations, and demonstrate significant performance improvements compared with the current Cell implementation . Overall results demonstrate the tremendous potential of the Cell architecture for scientific computations in terms of both raw performance and power efficiency. We also conclude that Cell's heterogeneous multi-core implementation is inherently better suited to the HPC environment than homogeneous commodity multicore processors. RELATED WORK One of the key limiting factors for computational performance is off-chip memory bandwidth. Since increasing the off-chip bandwidth is prohibitively expensive, many architects are considering ways of using available bandwidth more efficiently. Examples include hardware multithreading or more efficient alternatives to conventional cache-based architectures such as software-controlled memories. Software-controlled memories can potentially improve memory subsystem performance by supporting finely controlled prefetch-ing and more efficient cache-utilization policies that take advantage of application-level information -- but do so with far less architectural complexity than conventional cache architectures . While placing data movement under explicit software control increases the complexity of the programming model, prior research has demonstrated that this approach can be more effective for hiding memory latencies (including cache misses and TLB misses) -- requiring far smaller cache sizes to match the performance of conventional cache implementations [17, 19]. The performance of software-controlled memory is more predictable, thereby making it popular for real-time embedded applications where guaranteed response rates are essential. Over the last five years, a plethora of alternatives to conventional cache-based architectures have been suggested including scratchpad memories [9,16,30], paged on-chip memories [12, 17], and explicit three-level memory architectures [18, 19]. Until recently, few of these architectural concepts made it into mainstream processor designs, but the increasingly stringent power/performance requirements for embedded systems have resulted in a number of recent implementations that have adopted these concepts. Chips like the Sony Emotion Engine [20, 23, 29] and Intel's MXP5800 both achieved high performance at low power by adopting three levels (registers, local memory, external DRAM) of software-managed memory. More recently, the STI Cell processor has adopted a similar approach where data movement between SPE 256 KB PPC 512 KB memo ry con troller I/O I/O EIB 4 rings, 8bytes/ core cycle 25.6 GB/s SPE 256 KB SPE 256 KB SPE 256 KB SPE 256 KB SPE 256 KB SPE 256 KB SPE 256 KB Figure 1: Overview of the Cell processor these three address spaces is explicitly controlled by the application . For predictable data access patterns the local store approach is highly advantageous as it can be very efficiently utilized through explicit software-controlled scheduling . Improved bandwidth utilization through deep pipelining of memory requests requires less power, and has a faster access time, than a large cache due in part to its lower complexity . If however, the data access pattern lacks predictabil-ity , then the advantages of software-managed memory are lost. This more aggressive approach to memory architecture was adopted to meet the demanding cost/performance and real-time responsiveness requirements of Sony's upcoming video game console. However, to date, an in-depth study to evaluate the potential of utilizing the Cell architecture in the context of scientific computations does not appear in the literature. CELL BACKGROUND Cell [8,27] was designed by a partnership of Sony, Toshiba, and IBM (STI) to be the heart of Sony's forthcoming PlayStation3 gaming system. Cell takes a radical departure from conventional multiprocessor or multi-core architectures. Instead of using identical cooperating commodity processors, it uses a conventional high performance PowerPC core that controls eight simple SIMD cores, called synergistic processing elements (SPEs), where each SPE contains a synergistic processing unit (SPU), a local memory, and a memory flow controller. An overview of Cell is provided in Figure 1. Access to external memory is handled via a 25.6GB/s XDR memory controller. The cache coherent PowerPC core, the eight SPEs, the DRAM controller, and I/O controllers are all connected via 4 data rings, collectively known as the EIB. The ring interface within each unit allows 8 bytes/cycle to be read or written. Simultaneous transfers on the same ring are possible. All transfers are orchestrated by the PowerPC core. Each SPE includes four single precision (SP) 6-cycle pipelined FMA datapaths and one double precision (DP) half-pumped (the double precision operations within a SIMD operation must be serialized) 9-cycle pipelined FMA datapath with 4 cycles of overhead for data movement [22]. Cell has a 7 cycle in-order execution pipeline and forwarding network [8]. IBM appears to have solved the problem of inserting a 13 (9+4) cycle DP pipeline into a 7 stage in-order machine by choosing the minimum effort/performance/power solution of simply stalling for 6 cycles after issuing a DP 10 instruction. The SPE's DP throughput [14] of one DP instruction every 7 (1 issue + 6 stall) cycles coincides perfectly with this reasoning. Thus for computationally intense algorithms like dense matrix multiply (GEMM), we expect SP implementations to run near peak whereas DP versions would drop to approximately one fourteenth the peak SP flop rate [10]. Similarly, for bandwidth intensive applications such as sparse matrix vector multiplication (SpMV) we expect SP versions to be between 1.5x and 4x as fast as DP, depending on density and uniformity. Unlike a typical coprocessor, each SPE has its own local memory from which it fetches code and reads and writes data. All loads and stores issued from the SPE can only access the SPE's local memory. The Cell processor depends on explicit DMA operations to move data from main memory to the local store of the SPE. The limited scope of loads and stores allows one to view the SPE as having a two-level register file. The first level is a 128 x 128b single cycle register file, where the second is a 16K x 128b six cycle register file. Data must be moved into the first level before it can be operated on by instructions. Dedicated DMA engines allow multiple concurrent DMA loads to run concurrently with the SIMD execution unit, thereby mitigating memory latency overhead via double-buffered DMA loads and stores. The selectable length DMA operations supported by the SPE are much like a traditional unit stride vector load. We exploit these similarities to existing HPC platforms to select programming models that are both familiar and tractable for scientific application developers. PROGRAMMING MODELS The Cell architecture poses several challenges to programming : an explicitly controlled memory hierarchy, explicit parallelism between the 8 SPEs and the PowerPC, and a quadword based ISA. Our goal is to select the programming paradigm that offers the simplest possible expression of an algorithm while being capable of fully utilizing the hardware resources of the Cell processor. The memory hierarchy is programmed using explicit DMA intrinsics with the option of user programmed double buffering to overlap data movement with computation on the SPEs. Moving from a hardware managed memory hierarchy to one controlled explicitly by the application significantly complicates the programming model, and pushes it towards a one sided communication model. Unlike MPI, the intrinsics are very low level and map to half a dozen instructions. This allows for very low software overhead and good performance , but requires the user to be capable and either ensure correct usage or provide an interface or abstraction. For programming the parallelism on Cell, we considered three possible programming models: task parallelism with independent tasks scheduled on each SPE; pipelined parallelism where large data blocks are passed from one SPE to the next; and data parallelism, where the processors perform identical computations on distinct data. For simplicity, we do not consider parallelism between the PowerPC and the SPEs, so we can treat this as a homogeneous parallel machine . Data pipelining may be suitable for certain classes of algorithms and will be the focus of future investigation. We adopt the data-parallel programming model, which is a good match to many scientific applications and offers the simplest and most direct method of decomposing the problem . Data-parallel programming is quite similar to loop-level parallelization afforded by OpenMP or the vector-like multistreaming on the Cray X1E and the Hitachi SR-8000. The focus of this paper is Cell architecture and performance ; we do not explore the efficacy of the IBM SPE XLC compiler. Thus, we heavily rely on SIMD intrinsics and do not investigate if appropriate SIMD instructions are gener-ated by the compiler. Although the produced Cell code may appear verbose -- due to the use of intrinsics instead of C operators -- it delivers readily understandable performance. Our first Cell implementation, SpMV, required about a month of learning the programming model, the architecture, the compiler, the tools, and deciding on a final algorithmic strategy. The final implementation required about 600 lines of code. The next code development examined two flavors of double precision stencil-based algorithms. These implementations required one week of work and are each about 250 lines, with an additional 200 lines of common code. The programming overhead of these kernels on Cell required significantly more effort than the scalar version's 15 lines, due mainly to loop unrolling and intrinsics use. Although the stencils are a simpler kernel, the SpMV learning experience accelerated the coding process. Having become experienced Cell programmers, the single precision time skewed stencil -- although virtually a complete rewrite from the double precision single step version -- required only a single day to code, debug, benchmark, and attain spectacular results of over 65 Gflop/s. This implementation consists of about 450 lines, due once again to unrolling and the heavy use of intrinsics. SIMULATION METHODOLOGY The simplicity of the SPEs and the deterministic behavior of the explicitly controlled memory hierarchy make Cell amenable to performance prediction using a simple analytic model. Using this approach, one can easily explore multiple variations of an algorithm without the effort of programming each variation and running on either a fully cycle-accurate simulator or hardware. With the newly released cycle accurate simulator (Mambo), we have succesfully validated our performance model for SGEMM, SpMV, and Stencil Computations , as will be shown in the subsequent sections. Our modeling approach is broken into two steps commensurate with the two phase double buffered computational model. The kernels were first segmented into code-snippets that operate only on data present in the local store of the SPE. We sketched the code snippets in SPE assembly and performed static timing analysis. The latency of each operation , issue width limitations, and the operand alignment requirements of the SIMD/quadword SPE execution pipeline determined the number of cycles required. The in-order nature and fixed local store memory latency of the SPEs makes the analysis deterministic and thus more tractable than on cache-based, out-of-order microprocessors. In the second step, we construct a model that tabulates the time required for DMA loads and stores of the operands required by the code snippets. The model accurately reflects the constraints imposed by resource conflicts in the memory subsystem. For instance, concurrent DMAs issued by multiple SPEs must be serialized, as there is only a single DRAM controller. The model also presumes a conservative fixed DMA initiation latency of 1000 cycles. The model computes the total time by adding all the per-11 Cell X1E AMD64 IA64 SPE Chip (MSP) SIMD Multi-Multi Super VLIW Architecture core chip scalar SIMD Vector Clock (GHz) 3.2 3.2 1.13 2.2 1.4 DRAM (GB/s) 25.6 25.6 34 6.4 6.4 SP Gflop/s 25.6 204.8 36 8.8 5.6 DP Gflop/s 1.83 14.63 18 4.4 5.6 Local Store 256KB 2MB -L2 Cache -512KB 2MB 1MB 256KB L3 Cache -3MB Power (W) 3 ~40 120 89 130 Year -2006 2005 2004 2003 Table 1: Architectural overview of STI Cell, Cray X1E MSP, AMD Opteron, and Intel Itanium2. Es-timated total Cell power and peak Gflop/s are based on the active SPEs/idle PowerPC programming model. iteration (outer loop) times, which are themselves computed by taking the maximum of the snippet and DMA transfer times. In some cases, the per-iteration times are constant across iterations, but in others it varies between iterations and is input-dependent. For example, in a sparse matrix, the memory access pattern depends on the nonzero structure of the matrix, which varies across iterations. Some algorithms may also require separate stages which have different execution times; e.g., the FFT has stages for loading data, loading constants, local computation, transpose, local computation, bit reversal, and storing the results. For simplicity we chose to model a 3.2GHz, 8 SPE version of Cell with 25.6GB/s of memory bandwidth. This version of Cell is likely to be used in the first release of the Sony PlayStation3 [28]. The lower frequency had the simplifying benefit that both the EIB and DRAM controller could deliver two SP words per cycle. The maximum flop rate of such a machine would be 204.8 Gflop/s, with a computational intensity of 32 FLOPs/word. For comparison, we ran these kernels on actual hardware of several leading processor designs: the vector Cray X1E MSP, superscalar AMD Opteron 248 and VLIW Intel Itanium2. The key architectural characteristics are detailed in Table 1. 5.1 Cell+ Architectural Exploration The Double Precision (DP) pipeline in Cell is obviously an afterthought as video games have limited need for DP arithmetic. Certainly a redesigned pipeline would rectify the performance limitations, but would do so at a cost of additional design complexity and power consumption. We offer a more modest alternative that can reuse most of the existing circuitry. Based on our experience designing the VI-RAM vector processor-in-memory chip [12], we believe these "Cell+" design modifications are considerably less complex than a redesigned pipeline, consume very little additional surface area on the chip, but show significant DP performance improvements for scientific kernels. In order to explore the limitations of Cell's DP issue bandwidth , we propose an alternate design with a longer forwarding network to eliminate the all but one of the stall cycles -- recall the factors that limit DP throughput as described in Section 3. In this hypothetical implementation, called Cell+, each SPE would still have the single DP datapath, but would be able to dispatch one DP SIMD instruction every other cycle instead of one every 7 cycles. The Cell+ design would not stall issuing other instructions and would achieve 3.5x the DP throughput of the Cell (51.2 Gflop/s) by fully utilizing the existing DP datapath; however, it would maintain the same SP throughput, frequency, bandwidth, and power as the Cell. DENSE MATRIX-MATRIX MULTIPLY We begin by examining the performance of dense matrix-matrix multiplication, or GEMM. This kernel is character-ized by high computational intensity and regular memory access patterns, making it an extremely well suited for the Cell architecture. We explored two storage formats: column major and block data layout [26] (BDL). BDL is a two-stage addressing scheme (block row/column, element sub row/column). 6.1 Algorithm Considerations For GEMM, we adopt what is in essence an outer loop parallelization approach. Each matrix is broken into 8n x n element tiles designed to fit into the memory available on the Cell chip, which are in turn split into eight n x n element tiles that can fit into the 8 SPE local stores. For the column layout, the matrix will be accessed via a number of short DMAs equal to the dimension of the tile -- e.g. 64 DMAs of length 64. BDL, on the other hand, will require a single long DMA of length 16KB. Since the local store is only 256KB, and must contain both the program and stack, program data in the local store is limited to about 56K words. The tiles, when double buffered, require 6n 2 words of local store (one from each matrix) -- thus making 96 2 the maximum square tiles in SP. Additionally, in column layout, there is added pressure on the maximum tile size for large matrices, as each column within a tile will be on a different page resulting in TLB misses. The minimum size of a tile is determined by the FLOPs to word ratio of the processor. In the middle, there is a tile-size "sweet spot" that delivers peak performance. The loop order was therefore chosen to minimize the average number of pages touched per phase for a column major storage format. The BDL approach, as TLB misses are of little concern, allows us to structure the loop order to minimize memory bandwidth requirements. A possible alternate approach is to adapt Cannon's algorithm [3] for parallel machines. Although this strategy could reduce the DRAM bandwidth requirements by transferring blocks via the EIB, for a column major layout, it could significantly increase the number of pages touched. This will be the subject of future work. Note that for small matrix sizes, it is most likely advantageous to choose an algorithm that minimizes the number of DMAs. One such solution would be to broadcast a copy of the first matrix to all SPEs. 6.2 Single Precision GEMM Results The Cell performance of GEMM based on our performance model (referred to as Cell pm ) for large matrices is presented in Table 2. SGEMM simulation data show that 32 2 blocks do not achieve sufficient computational intensity to fully utilize the processor. The choice of loop order 12 Cell pm + Cell pm X1E AMD64 IA64 DP (Gflop/s) 51.1 14.6 16.9 4.0 5.4 SP (Gflop/s) -204 .7 29.5 7.8 3.0 Table 2: GEMM performance (in Gflop/s) for large square matrices on Cell, X1E, Opteron, and Itanium2 . Only the best performing numbers are shown. Cell data based on our performance model is referred to as Cell pm . and the resulting increase in memory traffic prevents column major 64 2 blocks from achieving a large fraction of peak (over 90%) for large matrices. Only 96 2 block sizes provide enough computational intensity to overcome the additional block loads and stores, and thus achieving near-peak performance -- over 200Gflop/s. For BDL, however, 64 2 blocks effectively achieve peak performance. Whereas we assume a 1000 cycle DMA startup latency in our simulations, if the DMA latency were only 100 cycles, then the 64 2 column major performance would reach parity with BDL. At 3.2GHz, each SPE requires about 3W [8]. Thus with a nearly idle PPC and L2, Cell pm achieves over 200 Gflop/s for approximately 40W of power -- nearly 5 Gflop/s/Watt. Clearly, for well-suited applications, Cell is extremely power efficient. 6.3 Double Precision GEMM Results A similar set of strategies and simulations were performed for DGEMM. Although the time to load a DP 64 2 block is twice that of the SP version, the time required to compute on a 64 2 DP block is about 14x as long as the SP counterpart (due to the limitations of the DP issue logic). Thus it is far easier for DP to reach its peak performance. -- a mere 14.6 Gflop/s. However, when using our proposed Cell+ hardware variant, DGEMM performance jumps to an impressive 51 Gflop/s. 6.4 Performance Comparison Table 2 shows a performance comparison of GEMM between Cell pm and the set of modern processors evaluated in our study. Note the impressive performance characteristics of the Cell processors, achieving 69x, 26x, and 7x speed up for SGEMM compared with the Itanium2, Opteron, and X1E respectively. For DGEMM, the default Cell processor is 2.7x and 3.7x faster than the Itanium2 and Opteron. In terms of power, the Cell performance is even more impressive , achieving over 200x the efficiency of the Itanium2 for SGEMM! Our Cell pm + exploration architecture is capable, for large tiles, of fully exploiting the DP pipeline and achieving over 50 Gflop/s. In DP, the Cell+ architecture would be nearly 10 times faster than the Itanium2 and nearly 30 times more power efficient. Additionally, traditional micros (Itanium2, Opteron, etc) in multi-core configurations would require either enormous power saving innovations or dramatic reductions in performance, and thus would show even poorer performance/power compared with the Cell technology. Com-pared to the X1E, Cell+ would be 3 times as fast and 9 times more power efficient. The decoupling of main memory data access from the computational kernel guarantees constant memory access latency since there will be no cache misses, and all TLB accesses are resolved in the communication phase. Matrix multiplication is perhaps the best benchmark to demonstrate Cell's computational capabilities, as it achieves high performance by buffering large blocks on chip before computing on them. 6.5 Model Validation IBM recently released their in-house performance evaluation of their prototype hardware [4]. On SGEMM, they achieve about 201 Gflop/s, which is within 2% of our pred-icated performance. SPARSE MATRIX VECTOR MULTIPLY At first glance, SpMV would seem to be a poor application choice for the Cell since the SPEs have neither caches nor word-granularity gather/scatter support. Furthermore, SpMV has a relatively low O(1) computational intensity. However, these considerations are perhaps less important than the Cell's low functional unit and local store latency (<2ns), the task parallelism afforded by the SPEs, the eight independent load store units, and the ability to stream nonzeros via DMAs. 7.1 Algorithmic Considerations Two storage formats are presented in this paper: Compressed Sparse Row (CSR) and Blocked Compressed Sparse Row (BCSR). Only square BCSR was explored, and only 2x2 BCSR numbers will be presented here. Future Cell SpMV work will examine the entire BCSR space. Because of the quadword nature of the SPEs, all rows within a CSR tile are padded to a multiple of 4. This greatly simplifies the programming model at the expense of increasing memory traffic. Note that this is very different than 1x4 BCSR.. To perform a stanza gather operation the Cell utilizes the MFC "get list" command, where a list of addresses/lengths is created in local store. The MFC then gathers these stanzas from the global store and packs them into the local store. It is possible to make every stanza a single quadword, however , without an accurate performance model of the MFC "get list" command, one must resort to tiling to provide a reasonable estimate for performance. For simplicity all benchmarks were run using square tiles. The data structure required to store the entire matrix is a 2D array of tiles, where each block stores its nonzeros and row pointers as if it were an entire matrix. We chose not to buffer the source and destination vector tiles as this would result in a smaller block size. These tradeoffs will be examined in future work. Collectively the blocks are chosen to be no larger than ~36K words in SP (half that in DP). The inner loop of CSR SpMV either requires significant software pipelining, hefty loop unrolling, or an approach al-gorithmically analogous to a segmented scan [1]. As there are no conditional stores in the SPE assembly language, we chose to partially implement a segmented scan, where the gather operations are decoupled from the dot products. This decoupled gather operation can be unrolled and software pipelined, thereby completing in close to three cycles per element (the ISA is not particularly gather friendly). It is important to note that since the local store is not a write back cache, it is possible to overwrite its contents without fear of consuming DRAM bandwidth or corrupting the actual arrays. As the nonzeros are stored contiguously in arrays, it is 13 # Name N NNZ Comments 15 Vavasis 40K 1.6M 2D PDE Problem 17 FEM 22K 1M Fluid Mechanics Problem 18 Memory 17K 125K MotorolaMemory Circuit 36 CFD 75K 325K Navier-Stokes, viscous flow 06 FEM Crystal 14K 490K FEM stiffness matrix 09 3D Tube 45K 1.6M 3D pressure Tube 25 Portfolio 74K 335K Financial Portfolio 27 NASA 36K 180K PWT NASA Matrix 28 Vibroacoustic 12K 177K Flexible box structure 40 Linear Prog. 31K 1M AA T -7pt 64 256K 1.8M 64 3 7pt stencil Table 3: Suite of matrices used to evaluate SpMV performance. Matrix numbers as defined in the SPARSITY suite are shown in the first column. straightforward to stream them in via DMA. Here, unlike the source and destination vectors, it is essential to double buffer in order to maximize the SPEs computational throughput. Using buffers of 16KB for SP allows for 2K values and 2K indices for CSR, and 1K tiles for 2x2 BCSR. Note that for each phase -- loading nonzeros and indices -- there is the omnipresent 1000 cycle DMA latency overhead in addition to the startup and finalize penalties (as in traditional pipelining). To partition the work among the SPEs, we implemented a cooperative blocking model. By forcing all SPEs to work on the same block, it is possible to broadcast the blocked source vector and row pointers to minimize memory traffic. One approach, referred to as PrivateY, was to divide work among SPEs within a block by distributing the nonzeros as evenly as possible. This strategy necessitates that each SPE contains a private copy of the destination vector, and requires an inter-SPE reduction at the end of each blocked row. The alternate method, referred to as PartitionedY, partitions the destination vector evenly among the SPEs. However there is no longer any guarantee that the SPEs' computations will remain balanced, causing the execution time of the entire tile to be limited by the most heavily loaded SPE. Thus for load balanced blocks, the PartitionedY approach is generally advantageous; however, for matrices exhibiting irregular (uneven) nonzero patterns, we expect higher performance using PrivateY. Note that there is a potential performance benefit by writing a kernel specifically optimized for symmetric matrices. For these types of matrices, the number of operations can effectively double relative to the memory traffic. However, the algorithm must block two tiles at a time -- thus the symmetric matrix kernel divides memory allocated for blocking the vector evenly among the two submatrices, and performs a dot product and SAXPY for each row in the lower triangle. 7.2 Evaluation Matrices In order to effectively evaluate SpMV performance, we examine a synthetic stencil matrix, as well as ten real matrices used in numerical calculations from the BeBop SPARSITY suite [11, 31] (four nonsymmetric and six symmetric). Table 3 presents an overview of the evaluated matrices. 7.3 Single Precision SpMV Results Single and double precision tuned SpMV results for the SPARSITY matrices are show in Tables 4 and 5. Surpris-ingly , given Cell's inherent SpMV limitations, the SPARSITY nonsymmetric matrices average over 4 Gflop/s, while the symmetric matrices average nearly 8 Gflop/s. Unfortunately , many of these matrices are so small that they utilize only a fraction of the default tile size. Unlike the synthetic matrices, the real matrices, which contain dense sub-blocks, can exploit BCSR without unnecessarily wasting memory bandwidth on zeros. As memory traffic is key, storing BCSR blocks in a compressed format (the zeros are neither stored nor loaded) would allow for significantly higher performance if there is sufficient support within the ISA to either decompress these blocks on the fly, or compute on compressed blocks. This is an area of future research. Overall results show that the PrivateY approach is generally a superior partitioning strategy compared with PartitionedY . In most cases, the matrices are sufficiently unbalanced that the uniform partitioning of the nonzeros coupled with a reduction requires less time than the performing a load imbalanced calculation. When using the PartionedY approach, the symmetric kernel is extremely unbalanced for blocks along the diagonal. Thus, for matrices approximately the size of a single block, the imbalance between SPEs can severely impair the performance -- even if the matrix is uniform. In fact, symmetric optimizations show only about 50% performance improvement when running the nonsymmetric kernel on the symmetric matrices. Once again DMA latency plays a relatively small role in this algorithm. In fact, reducing the DMA latency by a factor of ten results in only a 5% increase in performance. This is actually a good result. It means than the memory bandwidth is highly utilized and the majority of bus cycles are used for transferring data rather than stalls. On the whole, clock frequency also plays a small part in the overall performance. Solely increasing the clock frequency by a factor of 2 (to 6.4GHz) provides only a 1% increase in performance on the SPARSITY nonsymmetric matrix suite. Similarly, cutting the frequency in half (to 1.6GHz) results in only a 20% decrease in performance. Simply put, for the common case, more time is used in transferring nonzeros and the vectors rather than computing on them. 7.4 Double Precision SpMV Results Results from our performance estimator show that single precision SPMV is almost twice as fast as double precision, even though the nonzero memory traffic only increases by 50%. This discrepancy is due to the reduction in the number of values contained in a tile, where twice as many blocked rows are present. For example, when using 16K 2 SP tiles on a 128K 2 matrix, the 512KB source vector must be loaded 8 times. However, in DP, the tiles are only 8K 2 -- causing the 1MB source vector to be loaded 16 times, and thus resulting in a much higher volume of memory traffic. Future work will investigate caching mega blocks across SPEs to reduce total memory traffic. 7.5 Performance Comparison Table 4 compares Cell's estimated performance (the best partitioning and blocking combination) for SpMV with re-14 SPARSITY nonsymmetric matrix suite Double Precision (Gflop/s) Single Precision (Gflop/s) Matrix Cell F SS Cell pm + Cell pm X1E AMD64 IA64 Cell pm AMD64 IA64 Vavasis 3.79 3.17 3.06 0.84 0.44 0.46 6.06 0.70 0.49 FEM 4.28 3.44 3.39 1.55 0.42 0.49 5.14 0.59 0.62 Mem 2.21 1.69 1.46 0.57 0.30 0.27 2.79 0.45 0.31 CFD 1.87 1.52 1.44 1.61 0.28 0.21 2.33 0.38 0.23 Average 3.04 2.46 2.34 1.14 0.36 0.36 4.08 0.53 0.41 SPARSITY symmetric matrix suite Double Precision (Gflop/s) Single Precision (Gflop/s) Matrix Cell F SS Cell pm + Cell pm X1E AMD64 IA64 Cell pm AMD64 IA64 FEM -6 .79 6.32 3.12 0.93 1.14 12.37 1.46 1.37 3D Tube -6 .48 6.06 2.62 0.86 1.16 11.66 1.36 1.31 Portfolio -1 .83 1.60 2.99 0.37 0.24 3.26 0.42 0.32 NASA -1 .92 1.66 3.30 0.42 0.32 3.17 0.46 0.40 Vibro -3 .90 3.47 2.54 0.57 0.56 7.08 0.56 0.64 LP -5 .17 4.87 1.27 0.47 0.63 8.54 0.55 0.92 Average -4 .35 4.00 2.64 0.60 0.67 7.68 0.80 0.83 Synthetic Matrices Double Precision (Gflop/s) Single Precision (Gflop/s) Matrix Cell F SS Cell pm + Cell pm X1E AMD64 IA64 Cell pm AMD64 IA64 7pt 64 Stencil 2.20 1.44 1.29 -0 .30 0.29 2.61 0.51 0.32 Table 4: SpMV performance in single and double precision on the SPARSITY (top) nonsymmetric and (bottom) symmetric matrix suites. Note: Cell F SS represents the actual implementation and runs on the cycle accurate full system simulator sults from the Itanium2 and Opteron using SPARSITY, a highly tuned sparse matrix numerical library, on nonsymmetric (top) and symmetric matrix suites. X1E results where gathered using a high-performance X1-specific SpMV implementation [6]. Considering that the Itanium2 and Opteron each have a 6.4GB/s bus compared to the Cell's 25.6GB/s DRAM bandwidth -- one may expect that a memory bound application such as SpMV would perform only four times better on the Cell. Nonetheless, on average, Cell pm is more than 6x faster in DP and 10x faster in SP. This is because in order to achieve maximum performance, the Itanium2 must rely on the BCSR storage format, and thus waste memory bandwidth loading unnecessary zeros. However, the Cell's high FLOP to byte ratio ensures that the regularity of BCSR is unnecessary allowing it to avoid loading many of the superfluous zeros. For example, in matrix #17, Cell uses more than 50% of its bandwidth loading just the DP nonzero values , while the Itanium2 utilizes only 33% of its bandwidth. The rest of Itanium2's bandwidth is used for zeros and meta data. It should be noted that where simulations on Cell involve a cold start to the local store, the Itanium2's have the additional advantage of a warm cache. Cell's use of on-chip memory as a buffer is advantageous in both power and area compared with a traditional cache. In fact, Cell is 20 times more power efficient than the Itanium2 and 15 times more efficient than the Opteron for SpMV. For a memory bound application such as this, multicore commodity processors will see little performance improvement unless they also scale memory bandwidth. Comparing results with an X1E MSP is far more difficult . For unsymmetric matrices, the Cell pm performance on average is twice that of the X1E. For symmetric matrices, Cell pm performs somewhere between half and triple the performance of the X1E, but on average is 50% faster. The fact that the X1E consumes about three times the power of Cell guarantees Cell, in double precision, is at least as power efficient as the X1E 7.6 Model Validation Some might claim that matrix-matrix multiplication performance can be easily predictable. Most, however, would agree that SpMV is very difficult to predict. As seen in Table 4, we tested our implementation of the DP SpMV kernel on the cycle accurate IBM full system simulator, referred to as Cell F SS . The actual implementation makes dynamic blocking and partitioning decisions at run time, based on the lessons learned while exploring optimization strategies for the performance model; however, the current version but does not include the BCSR approach, and only pads rows to the nearest even number. The cycle accurate simulations with a superior implementation proved to be about 30% faster than the initial performance estimate, and averages impressive results of more than 3 Gflop/s for nonsymmetric matrices. The 30% discrepancy disappears when static partitioning and blocking strategies used. We can clearly see how the actual implemen-tation's run time search for structure boosted performance of the heat equation from about 1.3 Gflop/s to 2.2 Gflop/s -achieving a 7x speedup over the Itanium2. Cell F SS , for double precision nonsymmetric matrices, is more than 8 times faster than the Itanium2, and 27 times more power efficient. These results confirm our performance model's predictive 15 X next [i, j, k, t+ 1] = X[i - 1, j, k, t]+ X[i+ 1, j, k, t]+ X[i, j - 1, k, t]+ X[i, j+ 1, k, t]+ X[i, j, k - 1, t]+ X[i, j, k+ 1, t]+ X[i, j, k, t] X[i, j, k, t+ 1] = dt 2 dx 2 (X[i - 1, j, k, t]+ X[i+ 1, j, k, t])+ dt 2 dy 2 (X[i, j - 1, k, t]+ X[i, j+ 1, k, t])+ dt 2 dz 2 (X[i, j, k - 1, t]+ X[i, j, k+ 1, t])+ X[i, j, k, t] - X[i, j, k, t- 1] Figure 2: Stencil kernels used in evaluation. Top: Chombo heattut equation requires only the previous time step. Bottom: Cactus WaveToy equation requires both two previous time steps. abilities on complex kernels, and clearly demonstrate Cell's performance superiority when compared with leading microarchitectural approaches. STENCIL COMPUTATIONS Stencil-based computations on regular grids are at the core of a wide range of important scientific applications. In these applications, each point in a multidimensional grid is updated with contributions from a subset of its neighbors. The numerical operations are then used to build solvers that range from simple Jacobi iterations to complex multigrid and block structured adaptive methods. In this work we examine two flavors of stencil computations derived from the numerical kernels of the Chombo [5] and Cactus [2] toolkits. Chombo is a framework for computing solutions of partial differential equations (PDEs) using finite difference methods on adaptively refined meshes. Here we examine a stencil computation based on Chombo's demo application, heattut, which solves a simple heat equation without adaptivity. Cactus is modular open source framework for computational science, successfully used in many areas of astrophysics. Our work examines the stencil kernel of the Cactus demo, WaveToy, which solves a 3D hyperbolic PDE by finite differencing. The heattut and WaveToy equations are shown in Figure 2. Notice that both kernels solve 7 point stencils in 3D for each point. However, the heattut equation only utilizes values from the previous time step, while WaveToy requires values from the two previous timesteps.. Additionally, WaveToy has a higher computational intensity, and can more readily exploit the FMA pipeline. 8.1 Algorithmic Considerations The algorithm used on Cell is virtually identical to that used on traditional architectures except that the ISA forces main memory loads and stores to be explicit, rather than caused by cache misses and evictions. The basic algorithmic approach to update the 3D cubic data array is to sweep across the domain, updating one plane at a time. Since a stencil requires both the next and previous plane, a minimum of 4 planes must be present in the local stores: (z-1,t), Z+2 Z+1 Z Z-1 Z-1 Z Time t+1 Time t Z+2 Z+1 Z Z-1 Z-2 Z-1 Time t+2 Time t Z Z-1 Z-2 Time t+1 Figure 3: Flow Diagram for Heat equation flow diagram . Left: Queues implemented within each SPE perform only one time step. Right: Time skewing version requires an additional circular queue to hold intermediate results. (z,t), (z+1,t), and (z,t+1). Additionally, bus utilization can be maximized by double buffering the previous output plane (z-1,t+1) with the next input plane (z+2,t). In order to parallelize across SPEs, each plane of the 3D domain is partitioned into eight overlapping blocks. Due to the finite size of the local store memory, a straightforward stencil calculation is limited to planes of 256 2 elements plus ghost regions. Thus each SPE updates the core 256x32 points from a 258x34 slab (as slabs also contain ghost regions ). To improve performance of stencil computations on cache-based architectures, previous research has shown multiple time steps can be combined to increase performance [13, 21, 32]. This concept of time skewing can also be effectively leveraged in our Cell implementation. By keeping multiple planes from multiple time steps in the SPE simul-taneously , it is possible to double or triple the number of stencils performed with almost no increase in memory traffic ; thus increasing computational intensity and improving overall performance. Figure 3 details a flow diagram for the heat equation, showing both the simple and time skewed implementations. Note that the neighbor communication required by stencils is not well suited for the aligned quadword load requirements of the SPE ISA - i.e. unaligned loads must be emu-lated with permute instructions. In fact, for SP stencils with extensive unrolling, after memory bandwidth, the permute datapath is the limiting factor in performance -- not the FPU. This lack of support for unaligned accesses highlights a potential bottleneck of the Cell architecture; however we can partially obviate this problem for the stencil kernel via data padding. 8.2 Stencil Kernel Results The performance estimation for the heattut and WaveToy stencil kernels is shown in Table 5. Results show that as the number of time steps increases, a corresponding decrease in the grid size is required due to the limited memory footprint of the local store. In SP, the heat equation on the Cell pm is effectively computationally bound with two steps of time skewing, resulting in over 41 Gflop/s. More specifically , the permute unit becomes fully utilized as discussed 16 Double Precision (Gflop/s) Cell F SS Cell pm + Cell pm + Cell pm X1E AMD64 IA64 Stencil (2 step) Heat 7.25 21.1 10.6 8.2 3.91 0.57 1.19 WaveToy 9.68 16.7 11.1 10.8 4.99 0.68 2.05 Single Precision (Gflop/s) Cell F SS Cell pm Cell pm X1E AMD64 IA64 Stencil (4 step) (2 step) Heat 65.8 41.9 21.2 3.26 1.07 1.97 WaveToy -33 .4 22.3 5.13 1.53 3.11 Table 5: Performance for the Heat equation and WaveToy stencils. X1E and Itanium2 experiments use 256 3 grids. The Opteron uses a 128 3 . Cell uses the largest grid that would fit within the local stores. The (n steps) versions denote a time skewed version where n time steps are computed. in Section 8.1. In DP, however, the heat equation is truly computationally bound for only a single time step, achieving 8.2 Gflop/s. Analysis also shows that in the Cell+ approach, the heat equation is memory bound when using a single time step attaining 10.6 Gflop/s; for time skewing, performance of Cell+ DP jumps to over 21 Gflops/s. We believe the temporal recurrence in the CACTUS WaveToy example will allow more time skewing in single precision at the expense of far more complicated code, and will be the subject of future investigation. 8.3 Performance Comparison Table 5 presents a performance comparison of the stencil computations across our evaluated set of leading processors. Note that stencil performance has been optimized for the cache-based platforms as described in [15]. In single precision, for this memory bound computation, even without time skewing, Cell pm achieves 6.5x, 11x, and 20x speedup compared with the X1E, the Itanium2 and the Opteron respectively. Recall that the Cell has only four times the memory bandwidth the scalar machines, and 75% the bandwidth of the X1E indicating that Cells potential to perform this class of computations in a much more efficient manner is due to the advantages of software controlled memory for algorithms exhibiting predictable memory accesses. In double precision, with 1/14 th the floating point throughput , Cell pm achieves a 2x, 7x, and 14x speedup compared to the X1E, the Itanium2, and the Opteron for the heat equation -- a truly impressive result. Additionally, unlike the Opteron and Itanium2, simple time skewing has the potential to at least double the performance in either SP (either version of Cell) or in DP on the Cell+ variant. Finally, recall that in Section 7 we examined Cell SpMV performance using 7-point stencil matrices. We can now compare those results with the structured grid approach presented here, as the numerical computations are equivalent in both cases. Results show that for two time step calculations , the single precision structured grid approach achieves a 23x advantage compared with the sparse matrix method. This impressive speedup is attained through the regularity of memory accesses, reduction of memory traffic (constants are encoded in the equation rather than the matrix), the ability to time skew (increased computational intensity), and that stencils on a structured grid dont require multiplications by 1.0 like a sparse matrix would. For double precision, the stencil algorithm advantage is diminished to approximately 12x, due mainly to the lack of time skewing. 8.4 Model Validation As with SpMV, we implemented an actual double precision kernel on the full system simulator, with Cell F SS results shown in Table 5. At first, we were surprised that measured performance fell short of our prediction by 13%. However, upon closer examination it was discovered that the actual Cell implementation prohibits dual issuing of DP instructions with loads or permutes, even though it allows SP with loads or permutes to be dual issued. Thus for kernels with streaming behavior, it is realistic to assume that one double precision SIMD instruction can be executed every 8 cycles -- instead of every 7 as we had predicted previously. This discrepancy results in a 14% architectural performance reduction , which corresponding very well to the 13% difference observed in Table 5 between the predicted (Cellpm) and simulated (Cell F SS ) DP data. Nonetheless, the actual DP Cell F SS implementation of our evaluated stencil kernel is about 13x faster, and nearly 30x more power efficient than the Opteron. We also developed a SP version of the heat equation that allowed four time-skewed stencil steps. (Our original performance estimation assumed one or two time steps.) Results show spectacular SP Cell F SS performance of nearly 66 Gflop/s -- more than 60x faster and 136x power efficient compared with the Opteron, even though Cell has only four times the bandwidth and 20 times the single precision throughput. FAST FOURIER TRANSFORMS The FFT presents us with an interesting challenge: its computational intensity is much less than matrix-matrix multiplication and standard algorithms require a non-trivial amount of data movement. Extensive work has been performed on optimizing this kernel for both vector [24] and cache-based [7] machines. In addition, implementations for varying precisions appear in many embedded devices using both general and special purpose hardware. In this section we evaluate the implementation of a standard FFT algorithm on the Cell processor. 9.1 Methods We examine both the 1D FFT cooperatively executed across the SPEs, and a 2D FFT whose 1D FFTs are each run on a single SPE. In all cases the data appears in a single array of complex numbers. Internally (within the local stores) the data is unpacked into separate arrays, and a table lookup is used for the roots of unity so that no runtime computation of roots is required. As such, our results include the time needed to load this table. Additionally, all results are presented to the FFT algorithm and returned in natural order (i.e. a bit reversal was required to unwind the permutation process in all cases). Note that these requirements have the potential to severely impact performance. For simplicity we evaluated a naive FFT algorithm (no double buffering and with barriers around computational segments) for the single 1D FFT. The data blocks are dis-tributed cyclically to SPEs, 3 stages of local work are performed , the data is transposed (basically the reverse of the 17 cyclic allocation), and then 9 to 13 stages of local computation is performed (depending on the FFT size). At that point the indices of the data on chip are bit-reversed to unwind the permutation process and the naturally ordered result copied back into main memory. Once again, we presume a large DMA initiation overhead of 1000 cycles. However, a Cell implementation where the DMA initiation overhead is smaller, would allow the possibility of much larger FFT calculations (including out of core FFTs) using smaller block transfers, with little or no slowdown using double buffering to hide the DMA latency. Before exploring the 2D FFT, we briefly discuss simultaneous FFTs. For sufficiently small FFTs (<4K points in SP) it is possible to both double buffer and round robin allocate a large number of independent FFTs to the 8 SPEs. Although there is lower computational intensity, the sheer parallelism, and double buffering allow for extremely high performance (up to 76 Gflop/s). Simultaneous FFTs form the core of the 2D FFT. In order to ensure long DMAs, and thus validate our assumptions on effective memory bandwidth, we adopted an approach that requires two full element transposes. First, N 1D N-point FFTs are performed for the rows storing the data back to DRAM. Second, the data stored in DRAM is transposed (columns become rows) and stored back to DRAM. Third the 1D FFTs are performed on the columns, whose elements are now sequential (because of the transpose). Finally a second transpose is applied to the data to return it to its original layout. Instead of performing an N point bit reversal for every FFT, entire transformed rows (not the elements of the rows) are stored in bit-reversed order (in effect, bit reversing the elements of the columns). After the first transpose, a decimation in frequency FFT is applied to the columns. The columns are stored back in bit-reversed order -- in doing so, the row elements are bit reversed. With a final transpose, the data is stored back to memory in natural order and layout in less time. 9.2 Single Precision FFT Performance Table 6 presents performance results for the Cell 1D and 2D FFT. For the 1D case, more than half of the total time is spent just loading and storing points and roots of unity from DRAM. If completely memory bound, peak performance is approximately (25.6GB/s/8Bytes) 5NlogN/3N cycles or approximately 5.3logN Gflop/s. This means performance is limited to 64 Gflop/s for a 4K point SP FFT regardless of CPU frequency. A clear area for future exploration is hiding computation within the communication and the minimiza-tion of the overhead involved with the loading of the roots of unity. Unfortunately the two full element transposes, used in the 2D FFT to guarantee long sequential accesses, consume nearly 50% of the time. Thus, although 8K simultaneous 4K point FFTs achieve 76 Gflop/s (after optimizing away the loading of roots of unity), a 4K 2 2D FFT only reaches 46 Gflop/s -- an impressive figure nonetheless. Without the bit reversal approach, the performance would have further dropped to about 40 Gflop/s. The smaller FFT's shown in the table show even poorer performance. 9.3 Double Precision FFT Performance When DP is employed, the balance between memory and computation is changed by a factor of 7. This pushes a Double Precision (Gflop/s) N Cell pm + Cell pm X1E AMD64 IA64 4K 12.6 5.6 2.92 1.88 3.51 1D 16K 14.2 6.1 6.13 1.34 1.88 64K -7 .56 0.90 1.57 1K 2 15.9 6.6 6.99 1.19 0.52 2D 2K 2 16.5 6.7 7.10 0.19 0.11 Single Precision (Gflop/s) N Cell pm + Cell pm X1E AMD64 IA64 4K -29 .9 3.11 4.24 1.68 1D 16K -37 .4 7.48 2.24 1.75 64K -41 .8 11.2 1.81 1.48 1K 2 -35 .9 7.59 2.30 0.69 2D 2K 2 -40 .5 8.27 0.34 0.15 Table 6: Performance of 1D and 2D FFT in DP (top) and SP (bottom). For large FFTs, Cell is more than 10 times faster in SP than either the Opteron or Itanium2. The Gflop/s number is calculated based on a naive radix-2 FFT algorithm. For 2D FFTs the naive algorithm computes 2N N-point FFTs. slightly memory bound application strongly into the computationally bound domain. The SP simultaneous FFT is 10 times faster than the DP version. On the upside, the transposes required in the 2D FFT are now less than 20% of the total time, compared with 50% for the SP case. Cell pm + finds a middle ground between the 4x reduction in computational throughput and the 2x increase in memory traffic -increasing performance by almost 2.5x compared with the Cell for all problem sizes. 9.4 Performance Comparison The peak Cell FFT performance is compared to a number of other processors in the Table 6. These results are conservative given the naive 1D FFT implementation we used on Cell whereas the other systems in the comparison used highly tuned FFTW [7] or vendor-tuned FFT implementations [25]. Nonetheless, in DP, Cell pm is at least 12x faster than the Itanium2 for a 1D FFT, and Cell pm + could be as much as 30x faster for a large 2D FFT. Cell+ more than doubles the DP FFT performance of Cell for all problem sizes. Cell performance is nearly at parity with the X1E in double precision; however, we believe considerable headroom remains for more sophisticated Cell FFT implementations. In single precision, Cell is unparalleled. Note that FFT performance on Cell improves as the number of points increases, so long as the points fit within the local store. In comparison, the performance on cache-based machines typically reach peak at a problem size that is far smaller than the on-chip cache-size, and then drops precip-itously once the associativity of the cache is exhausted and cache lines start getting evicted due to aliasing. Elimination of cache evictions requires extensive algorithmic changes for the power-of-two problem sizes required by the FFT algorithm , but such evictions will not occur on Cells software-managed local store. Furthermore, we believe that even for problems that are larger than local store, 1D FFTs will con X1E FFT numbers provided by Cray's Bracy Elton and Adrian Tate. 18 tinue to scale much better on Cell than typical cache-based superscalar processors with set-associative caches since local store provides all of the benefits as a fully associative cache. The FFT performance clearly underscores the advantages of software-controlled three-level memory architecture over conventional cache-based architectures. CONCLUSIONS AND FUTURE WORK The Cell processor offers an innovative architectural approach that will be produced in large enough volumes to be cost-competitive with commodity CPUs. This work presents the broadest quantitative study Cell's performance on scientific kernels and directly compares its performance to tuned kernels running on leading superscalar (Opteron), VLIW (Itanium2), and vector (X1E) architectures. We developed an analytic framework to predict Cell performance on dense and sparse matrix operations, stencil computations, and 1D and 2D FFTs. Using this approach allowed us to explore numerous algorithmic approaches without the effort of implementing each variation. We believe this analytical model is especially important given the relatively immature software environment makes Cell time-consuming to program currently; the model proves to be quite accurate, because the programmer has explicit control over parallelism and features of the memory system. Furthermore, we propose Cell+, a modest architectural variant to the Cell architecture designed to improve DP behavior . Overall results demonstrate the tremendous potential of the Cell architecture for scientific computations in terms of both raw DP and SP performance and power efficiency . In addition, we show that Cell+ significantly out-performs Cell for most of our evaluated DP kernels, while requiring minimal microarchitectural modifications to the existing design. Analysis shows that Cell's three level software-controlled memory architecture, which completely decouples main memory load/store from computation, provides several advantages over mainstream cache-based architectures. First, kernel performance can be extremely predictable as the load time from local store is constant. Second, long block transfers can achieve a much higher percentage of memory bandwidth than individual loads in much the same way a hardware stream prefetch engine, once engaged, can fully consume memory bandwidth. Finally, for predictable memory access patterns, communication and computation can be overlapped more effectively than conventional cache-based approaches. Increasing the size of the local store or reducing the DMA startup overhead on future Cell implementations may further enhance the scheduling efficiency by enabling more effective overlap of communication and computation. There are also disadvantages to the Cell architecture for kernels such as SpMV. With its lack of unaligned load support , Cell must issue additional instructions simply to permute data, yet still manages to outperform conventional scalar processor architectures. Even memory bandwidth may be wasted since SpMV is constrained to use tiling to remove the indirectly indexed accesses to the source vector . The ability, however, to perform a decoupled gather, to stream nonzeros, and Cell's low functional unit latency, tends to hide this deficiency. Additionally, we see stencil computations as an example of an algorithm that is heavily influenced by the performance of the permute pipeline. Here, the lack of support for an unaligned load instruction Speedup vs. Power Efficiency vs. Cell+ X1E AMD64 IA64 X1E AMD64 IA64 GEMM 3x 12.7x 9.5x 9x 28.3x 30.9x SpMV >2.7x >8.4x >8.4x >8.0x >18.7x >27.3x Stencil 5.4x 37.0x 17.7x 16.2x 82.4x 57.5x 1D FFT 2.3x 10.6x 7.6x 6.9x 23.6x 24.7x 2D FFT 2.3x 13.4x 30.6x 6.9x 29.8x 99.5x Speedup vs. Power Efficiency vs. Cell X1E AMD64 IA64 X1E AMD64 IA64 GEMM 0.8x 3.7x 2.7x 2.4x 8.2x 8.78x SpMV 2.7x 8.4x 8.4x 8.0x 18.7x 27.3x Stencil 1.9x 12.7x 6.1x 5.7x 28.3x 19.8x 1D FFT 1.0x 4.6x 3.2x 3.0x 10.2x 10.4x 2D FFT 0.9x 5.5x 12.7x 2.7x 12.2x 41.3x Table 7: Double precision speedup and increase in power efficiency of (Top) Cell+ and (Bottom) Cell, relative to the X1E, Opteron, and Itanium2 for our evaluated suite of scientific kernels. Results show an impressive improvement in performance and power efficiency. is a more significant performance bottleneck than either the SP execution rate or the memory bandwidth. For dense matrix operations, it is essential to maximize computational intensity and thereby fully utilize the local store. However, if not done properly, the resulting TLB misses adversely affect performance. For example, in the GEMM kernel we observe that the BDL data storage format, either created on the fly or before hand, can ensure that TLB misses remain a small issue even as on-chip memories increase in size. Table 7 compares the advantage of Cell and Cell+ based on the better of performance model or actual implementation (where available) in terms of DP performance and power efficiency for our suite of evaluated kernels and architectural platforms. Observe that Cell+ has the potential to greatly increase the already impressive performance characteristics of Cell. By using the insight gained in the development of our estimation model, we developed an optimized SpMV version that outperformed our initial predictions by 25% 70%. If a full system simulator could model the modest improvements of our Cell+ variant, we feel confident that we could demonstrate comparable improvements to DP performance as well. We also note that DP stencil performance fell short of our model by 13% due to previously unknown microarchitectural limitations. However, time skewing showed a huge benefit in SP and we believe a similar benefit would be present in DP on Cell+ variant. It is important to consider these performance differences in the context of increasingly prevalent multi-core commodity processors. The first generation of this technology will instantiate at most two cores per chip, and thus will deliver less than twice the performance of today's existing architectures . This factor of 2x is trivial compared with Cell+'s potential of 10-20x improvement. While peak Cell DP performance is impressive relative to its commodity peers, a fully utilizable pipelined DP floating point unit would boost Cell (i.e. Cell+) performance and efficiency significantly. 19 Acknowledgments This work was supported by the Director, Office of Science, of the U.S. Department of Energy under Contract No. DE-AC02 -05CH11231. The authors gratefully thank Bracy Elton and Adrian Tate for their assistance in obtaining X1E FFT performance data, and Eduardo D'Azevedo for providing us with an optimized X1E SpMV implementation. REFERENCES [1] G. Blelloch, M. Heroux, and M. Zagha. Segmented operations for sparse matrix computation on vector multiprocessors. Technical Report CMU-CS-93-173, CMU, 1993. [2] Cactus homepage. http://www.cactuscode.org. [3] L. Cannon. A Cellular Computer to Implement the Kalman Filter Algorithm. PhD thesis, Montana State University, 1969. [4] Cell broadband engine architecture and its first implementation. http://www-128.ibm.com/ developerworks/power/library/pa-cellperf/. [5] Chombo homepage. http://seesar.lbl.gov/anag/chombo. [6] E. D'Azevedo, M. R. Fahey, and R. T. Mills. Vectorized sparse matrix multiply for compressed row storage format. In International Conference on Computational Science (ICCS), pages 99106, 2005. [7] FFTW speed tests. http://www.fftw.org. [8] B. Flachs, S. Asano, S. Dhong, et al. A streaming processor unit for a cell processor. ISSCC Dig. Tech. Papers, pages 134135, February 2005. [9] P. Francesco, P. Marchal, D. Atienzaothers, et al. An integrated hardware/software approach for run-time scratchpad management. In Proceedings of the 41st Design Automation Conference, June 2004. [10] Ibm cell specifications. http://www.research.ibm.com/cell/home.html. [11] E.-J. Im, K. Yelick, and R. Vuduc. Sparsity: Optimization framework for sparse matrix kernels. International Journal of High Performance Computing Applications, 2004. [12] The Berkeley Intelligent RAM (IRAM) Project. http://iram.cs.berkeley.edu. [13] G. Jin, J. Mellor-Crummey, and R. Fowlerothers. Increasing temporal locality with skewing and recursive blocking. In Proc. SC2001, 2001. [14] J. Kahle, M. Day, H. Hofstee, et al. Introduction to the cell multiprocessor. IBM Journal of R&D, 49(4), 2005. [15] S. Kamil, P. Husbands, L. Oliker, et al. Impact of modern memory subsystems on cache optimizations for stencil computations. In ACM Workshop on Memory System Performance, June 2005. [16] M. Kandemir, J. Ramanujam, M. Irwin, et al. Dynamic management of scratch-pad memory space. In Proceedings of the Design Automation Conference, June 2001. [17] P. Keltcher, S. Richardson, S. Siu, et al. An equal area comparison of embedded dram and sram memory architectures for a chip multiprocessor. Technical report, HP Laboratories, April 2000. [18] B. Khailany, W. Dally, S. Rixner, et al. Imagine: Media processing with streams. IEEE Micro, 21(2), March-April 2001. [19] M. Kondo, H. Okawara, H. Nakamura, et al. Scima: A novel processor architecture for high performance comp uting. In 4th International Conference on High Performance Computing in the Asia Pacific Region, volume 1, May 2000. [20] A. Kunimatsu, N. Ide, T. Sato, et al. Vector unit architecture for emotion synthesis. IEEE Micro, 20(2), March 2000. [21] Z. Li and Y. Song. Automatic tiling of iterative stencil loops. ACM Transactions on Programming Language Systems, 26(6), 2004. [22] S. Mueller, C. Jacobi, C. Hwa-Joon, et al. The vector floating-point unit in a synergistic processor element of a cell processor. In 17th IEEE Annual Symposium on Computer Arithmetic (ISCA), June 2005. [23] M. Oka and M. Suzuoki. Designing and programming the emotion engine. IEEE Micro, 19(6), November 1999. [24] L. Oliker, R. Biswas, J. Borrill, et al. A performance evaluation of the Cray X1 for scientific applications. In Proc. 6th International Meeting on High Performance Computing for Computational Science, 2004. [25] Ornl cray x1 evaluation. http://www.csm.ornl.gov/ dunigan/cray. [26] N. Park, B. Hong, and V. Prasanna. Analysis of memory hierarchy performance of block data layout. In International Conference on Parallel Processing (ICPP), August 2002. [27] D. Pham, S. Asano, M. Bollier, et al. The design and implementation of a first-generation cell processor. ISSCC Dig. Tech. Papers, pages 184185, February 2005. [28] Sony press release. http://www.scei.co.jp/ corporate/release/pdf/050517e.pdf. [29] M. Suzuoki et al. A microprocessor with a 128-bit cpu, ten floating point macs, four floating-point dividers, and an mpeg-2 decoder. IEEE Solid State Circuits, 34(1), November 1999. [30] S. Tomar, S. Kim, N. Vijaykrishnan, et al. Use of local memory for efficient java execution. In Proceedings of the International Conference on Computer Design, September 2001. [31] R. Vuduc. Automatic Performance Tuning of Sparse Matrix Kernels. PhD thesis, University of California at Berkeley, 2003. [32] D. Wonnacott. Using time skewing to eliminate idle time due to memory bandwidth and network limitations. In International Parallel and Distributed Processing Symposium (IPDPS), 2000. 20

GEMM;FFT;Cell processor;three level memory;SpMV;Stencil;sparse matrix

192

The Use of Mediation and Ontology Technologies for Software Component Information Retrieval

Component Based Development aims at constructing software through the inter-relationship between pre-existing components. However, these components should be bound to a specific application domain in order to be effectively reused. Reusable domain components and their related documentation are usually stored in a great variety of data sources. Thus, a possible solution for accessing this information is to use a software layer that integrates different component information sources. We present a component information integration data layer, based on mediators. Through mediators, domain ontology acts as a technique/formalism for specifying ontological commitments or agreements between component users and providers, enabling more accurate software component information search.

INTRODUCTION Component Based Development (CBD) [1] aims at constructing software through the inter-relationship between pre-existing components, thus reducing the complexity, as well as the cost of software development, through the reuse of exhaustively tested components. Building new solutions by combining components should improve quality and support rapid development, leading to a shorter time-to-market. At the same time, nimble adaptation to changing requirements can be achieved by investing only in key changes of a component-based solution, rather than undertaking a major release change. For these reasons, component technology is expected by many to be the cornerstone of software production in the years to come. According to Jacobson, Griss and Jonsson [1], the effectiveness of component reuse depends on the connectiveness among them and their binding to specific application domains. The connectiveness is one of the most discussed problems in CBD [6, 12]. Approaches that deal with component interfaces (one of premises for connection between components) focus on their capability to provide and request services. Although this interface aspect is important in a CBD approach, other problems arise when trying to connect components. The connectiveness also depends on the execution environment, the heterogeneity of components, the distance between them and the architecture that controls their connections [1, 5, 12]. The architecture that governs the connections between components also depends on the application domain. Therefore, reuse possibilities increase when components are bound to domain concepts. As stated by Krueger [1], while retrieving reusable software components, it is advantageous to use a terminology that is familiar to the domain. This approach diverges from other software component retrieval proposals, such as the Agora System [6] that bases the search only on the component interfaces, covering solely the component connectiveness problem, and the RIG initiative [10] that presents an approach for domain repository integration with minor user transparency and without web access. Suppose that, in a typical component retrieval scenario, a software developer wants to find software components to use in the construction of an application under development. If he does not know any other specialized service that provides information about components, the natural search space will be the Internet. Now, consider that this developer has no knowledge about the available components. Thus, the following actions are necessary to discover software components that satisfy his needs: 1. To locate information about components that can be stored in distributed repositories. This could be typically done through an Internet search engine that takes as input a few keywords and returns a list of relevant resource descriptions that might be available in these repositories. The success of this task directly depends on the interest of the repository administrators in publicizing their data and the precision of the user while providing the keywords. 2. To determine usability of search results. Due to the complexity in the analysis of component usefulness (i.e., considering the component domain, functionality and connection possibilities based on architecture decisions), the An interface of a component can be seen as a component's part that defines its access points. These points allow clients (components themselves) to access services provided by the component [12]. Thus, a naive Internet approach will not cope with the complexity of software component retrieval. The previous actions require an engine that combines the following three characteristics: (i) Distribution and Heterogeneity software components can be distributed and use different kinds of storage; (ii) Domain Ontology - to organize component repositories within a domain in order to ease its search; (iii) Software Component Information Evolution to insert new information (including legacy information). However, as stated before, current software component retrieval proposals either lack from heterogeneity or domain ontology. On the other hand, many database projects [3,4,7,8,9] are particularly concerned with distribution, heterogeneity (and ontology) found in legacy databases. These projects are known as "multi-database" or Heterogeneous and Distributed Data Base Systems (HDDS) [4]. One solution found in HDDS is the use of mediators [2] combined with ontology [14] to integrate, identify and retrieve related legacy databases. Ontology in this context can be defined as a vocabulary of terms and the relationship between them. For each term, a definition must be created, using an informal description, some concrete examples in the domain, and also a formal specification of the relationships between terms, thus forming a semantic network. We believe that the HDDS technologies can be adapted to handle software component repositories in the place of legacy databases. Mediators can represent and integrate domain information repositories (distributed and/or heterogeneous). Metadata found in mediators can describe the repositories of components, presenting the domain, their semantics, software architecture and interfaces. Usually a query engine is available in HDDS and therefore ad hoc queries over this metadata can be used to analyze the available components. The organization of mediators with ontology drives the user search along heterogeneous vocabulary. Therefore, our main objective is to present a software component information retrieval engine named Odyssey Mediation Layer (OML) that combines the connectiveness of components and the domain concept approach. We address both issues through the adoption of mediation [2] and ontology [14] technologies, respectively. Our approach is motivated by a project that is being conducted in the Legislative House domain. There are several applications that can benefit from reusable information within this domain and from other related ones, such as justice domain, criminal domain, among others. Our users are not specialists in the latter domains, only in the Legislative domain. However, it is important that relevant reusable information from all related domains can be presented to them, particularly when they are not aware of its existence. Most components of legislative process applications can be reused from the legislative domain (e.g., Proposal Creation, Legislature Evaluation, Council Members referee, among others), but sometimes it is worth looking at components from other related domains such as justice domain. Our retrieval engine is able to identify and suggest components from other related 2 According to Wiederhold [2], mediators are modules that encompass layers of mediation services, connecting bases of heterogeneous and distributed data (producers) to information systems (consumers). domains in the same way as it suggests components from the Legislative domain. With our retrieval engine, the user search can rely on a controlled vocabulary (ontology) composed of domain terms that are familiar to him. Thus, the search is more focused and executed over relevant available component information repositories. Besides, the usefulness of the retrieved components is supported by the bindings between ontology terms and related components. This binding is accomplished by domain specialists together with domain engineers during a domain engineering process [5, 17] thus enforcing the biding precision. In order to address these issues, the retrieval engine organizes component information repositories within domain ontologies while preserving its original characteristics of distribution and heterogeneity, all in a flexible way. The main contribution of our proposal is to provide an approach for accessing software components through the use of ontologies and mediators. Our innovative aspect is to provide flexibility, transparency and accuracy in software component information retrieval. In order to present our approach, the paper is organized as follows: Section 2 discusses the novelties of our proposal with related works; Section 3 details the architecture of Odyssey Mediation Layer; Section 4 shows a component retrieval example; and Section 5 presents our concluding remarks. RELATED WORKS In a broader approach, several information retrieval systems use semantic brokering with respect to their resources. These systems include SIMS [3], TSIMMIS [4], InfoMaster [7], and Information Manifold [8]. These systems work in the definition of some sort of common vocabulary (similar to an ontology) to define objects in their domain. Individual information sources that contain these objects describe constraints on objects that they can provide, in terms of this common vocabulary. The broker then uses these constraints to determine how to process queries. These projects deal with generic designs for database retrieval. Our approach combines the mediation technology with specific domain ontology [2] to integrate different software components data sources. The main difference between these projects and ours is that our approach is particularly concerned with software components. Thus, we use an ontology, which is specifically constructed for that. This ontology is specified during a Domain Engineering process [5] tailored for this purpose. Hence, the ontology accuracy is more efficient, and consequently the usefulness of the retrieved components. Another work worth mentioning is the InfoSleuth system[9]. It is a large project, conducted by MCC, which uses an ontology approach to retrieve information from distributed and heterogeneous databases. InfoSleuth can be seen as a framework that can be tailored for a given purpose. One interesting application is the EDEN project in the environmental domain [9]. In this case, some tools of InfoSleuth were customized for this project. Our project adopts a similar approach, where our constructs are specific for software component information retrieval. The Agora System [6] describes a search engine for retrieving reusable code components, such as JavaBeans and CORBA components. Agora uses an introspection mechanism for 20 registering code components, through its interface. As a result, this information may not be available at a certain time because the repository is not running, or the information cannot be located. In each of these cases, the interface information cannot be successfully retrieved and indexed, and the component is not registered in the AGORA index database. In our proposal, the use of mediators provides the flexibility to access remote component repositories. There is no need for an index phase, and the mediator is able to capture updates that may occur in a remote repository (i.e., the repository, using the translator 3 services, sends a message with these updates to the mediator). Moreover, the mediator metadata manager has access to all the ontological terms of a given domain, facilitating the identification of the existence of a component, even if its repository is out of service. In this case, the user knows that the component exists and that he can retrieve it latter. Moreover, new information is always associated to domain terms within a given domain ontology, improving its accessibility and reuse. Ye and Fischer [18] presents an approach that provides an active repository for components. The work emphasizes active delivery of reuse information, helping on the identification of components that developers did not even know that existed. Regarding this last aspect, it is similar to our approach. Our component information retrieval system also provides this functionality too, once it accesses components from other domains based on semantic similarity. The active repository functionality, although not described in this paper, is also part of our work [16]. One aspect that is different in our proposal is the retrieval of distributed information, which is not mentioned in Ye and Fischer's work. Another important work to mention is the RIG initiative [9], which describes a reuse library interoperability approach. The idea of the asset library interoperability is based on the storage of domain information in several databases. These databases are static and based on a unique global model. The integration requires that information is stored according to this unique model. Therefore, if any reuse database is to be integrated, it has to be translated to the RIG model. Alternatively, the mediation approach creates a new level of abstraction above the database model, allowing the insertion and/or removal of repositories from the mediation structure without the need for updates on the whole structure. Moreover, RIG lacks a more effective search engine that provides searches based on domain concepts and filtering of relevant information, including Internet access, as we do in our work. SPECIALIZING A MEDIATION ARCHITECTURE TO A COMPONENT INFORMATION RETRIEVAL ENGINE The effectiveness of a software component information retrieval engine is associated to its capacity to handle the distribution and heterogeneity of software components, to organize component repositories within a domain, and to enable software component evolution. These requirements can be accomplished through a software layer that is seen as a particular case of HDDS. As stated before, mediators are modules that encompass layers of mediation services, connecting bases of heterogeneous and distributed data (producers) to information systems (consumers). Hence, in order to be really useful in software component information search this solution has to be tailored to component 3 A translator in this context is the same as a wrapper or adapter. information retrieval, considering the component domain, its semantics, architecture, and interfaces. In a mediation architecture, as new sources of information are aggregated to the mediation structure, the amount of information to be modeled increases, frequently generating inconsistencies, ambiguities, and conflicts in the represented information. One way to deal with this problem is to partition the consumers and producers' models and the structure of mediation by domain. The description of these models, partitioned by domains, forms the so-called domain ontology [14]. In the context of component information retrieval, the use of mediators allows the information access to be carried independently of the format and the operational platform where it is stored. Therefore, the structure of a retrieval engine as a whole can be flexible, since existing component data sources can be added to the architecture in an easy way, with no need to convert from the original information format (format form of data/information source) to the format used by the reuse environment. Another interesting feature of mediators within a component information retrieval engine is that reusable information is naturally organized by domain (Figure 1), which facilitates the search for domain concepts, since specific domain data is accessed in a search. Moreover, the use of mediators allows the aggregation of information already stored in legacy databases, without the need to transform the original database format. In order to help on the correct choice of mediators for a given domain, the mediator layer provides a specific ontology for each domain. Therefore, this ontology must be specified by domain specialists, facilitating the search for specific components, since the ontology definition is directly connected to software components within the domain. The use of this layer in the Legislative Domain is particularly interesting, since in Brazil as in other countries [15], there are some legislative houses that are more up to date with software technology than others. The former represents a reference source of software components to several Legislative Houses. Without this kind of layer, there exist some barriers for reusing components among these houses, such as the distance, scarce financial resources, and semantic conflicts among components (a common component functionality can be identified differently in each legislative house). Figure 1 presents an example of a mediation layer configuration for this specific application domain. Several mediators are presented as sub-domains, such as State Legislative (SL) and Municipal Legislative (ML) domains. The SL Mediator is aggregated (P1) to ML, generating a more generic mediator that combines the two domains. The latter can be used in cases where information concerning the two domains is necessary. Each mediator is connected to the related domain data sources that contain reuse component information. The Justice Domain Mediator may be accessed in cases where the user wants components related to the Justice domain. In order to provide an architecture that is able to handle the requirements of component information search and retrieval, we specified and implemented OML (Odyssey Mediation Layer) was specified and implemented, based on mediation and ontology technologies. OML is derived from a HDDS mediator, the HIMPAR architecture [12], adding to it more precision and semantics, using ontologies tailored to software component information retrieval. 21 Retrieval Interface Legislative State Domain Mediator Justice Domain Mediator Legislative Domain Mediator Legislative Municipal Domain Mediator P1 P2 ORB Component Repository Translator Component Repository Component Repository Translator P1 : aggregation P2: association Translator Component Repository Translator Figure 1 - An example of a mediation layer for the Legislative domain User Interface Service Manager SM Query Translator Query Packer Query Decomposition Query Manager(QM) Ontology Model OntologyManager Metadata Manager (MM) Mediator Translator Translator Service Manager Metadata Model ORB Bus Component Repository Component Repository ORB Bus Reuse Tools Figure 2- Architecture of the Odyssey Mediation Layer The OML engine is part of a reuse environment, named Odyssey, that deals with component based development of applications within a given domain. Specifically, OML is part of a system of agents that helps users in their search for reusable components [16]. It uses intelligent mechanisms such as learning techniques and user preferences in order to present in advance components that the user does not even know that exist. The agent system infers this information and presents the components to the user. It is important to notice that although OML was built in the context of the Odyssey project, it can be used standalone or integrated to other tools, since OML offers a user interface (as seen in figures 3 through 8) and a CORBA IDL interface for its communication with other tools. Figure 2 presents OML, which comprises four levels: Interface, Mediation Layer, ORB bus, and Translators. The Interface level is implemented by the Service Manager (SM), which stores metadata about available mediators, and is capable of creating ontological bindings between related ontologies in order to query several mediation layers. Also, SM is responsible for the creation and modification of mediators. The Mediation Layer provides the management of each mediator through the Metadata Manager 22 (MM), and provides access to mediators through the Query Manager. At the ORB level, communication between the mediation layer and translators is established through CORBA standard services. Finally, the Translator level provides one translator for each component repository in such a way that it can participate in the Mediation Layer integration model. 3.1 Service Manager The Service Manager (SM) stores metadata about mediators, translators and data sources availability, and deals with ontological commitments between related mediators (domains). Schema 1 presents an overview of SM Metadata in ODL notation, and Schema 2 provides the IDL interface to access mediators through the CORBA bus. class Object_Himpar ( extent Himpares) { attribute string Name; } class Mediator extends Object_Himpar ( extent Mediadores) { relationship list<Container> AssociatedDataSources inverse DataSource::Medis; attribute string Description; attribute string KeyWords; relationship list<Mediador> Super inverse Mediador::*; relationship list<Mediador> Spec inverse Mediador::*; relationship list<Mediador> Assoc inverse Mediador::*; attribute string BaseName; relationship list<OntologyTerm> TermRel inverse OntologyTerm::MediadorRel; attribute String password; } class Wrapper extends Object_Himpar ( extent Wrappers) { attribute string description; attribute string type; relationship list<DataSource> Repositories inverse DataSource::Trad; } class DataSource extends Object_Himpar ( extent DataSources) { attribute string owner; relationship list<Mapping> Structure inverse Mapping::Cont; attribute string AbstractionLevel; relationship Wrapper Trad inverse Wrapper::Repositories; relationship list<Mediator> Medis. inverse Mediator:: AssociatedDataSources; attribute String password; } class Mapping { attribute string DataSourceName; attribute string map; relationship DataSource Cont inverse DataSource::Structure; } class OntologyTerm ( extent Terms) { attribute string Name; relationship list<OntologyTerm> Synonym inverse OntologyTerm::*; relationship list< OntologyTerm > Hipernym inverse OntologyTerm::*; relationship list< OntologyTerm > Hiponym inverse OntologyTerm::*; relationship Mediator MediadorRel inverse Mediator: TermRel; } class Component ( extent Components) { attribute string type; } Schema 1 Metadata of Service Manager The ontological commitments between related mediators are all done at SM level. SM provides the necessary metadata for this, using the Mediator, DataSource, Mapping, OntologyTerm and Component classes described in Schema2. The decision to concentrate all ontological commitments at SM level was mainly based on the SM available information. It knows the availability of all OML components and thus is able to indicate which domain the user could access. In order to use a given mediator within OML, the administrator has to register the mediator, its related data sources, and translators used by these data sources into SM. Figures 3 and 4 present examples of the interfaces for doing this. module Mediator { interface Access { struct OntologyTerm { string Name; string Description; }; struct object { string type; string definition; }; typedef sequence<Ontology> ListOntology; typedef sequence<object> ListObjects; // Functions for the management of bases string open_base (in string basename); void close_base (in string basename); // Functions for the management of ontologies ListOntology retrieve_Ontology (in string mediator-name); // Functions for retrieve components ListObjects queryMediator(in string query); }; }; Schema 2 SM IDL interface to access mediators Figure 3 shows some information about the Municipal Legislative Mediator within SM. Some basic metadata are: i) the mediator name, ii) the executable file that has to be loaded (if it is not already loaded) by ORB, in order to respond to some request to this specific mediator, iii) the keywords related to the mediator 23 (this information provides a fast and limited knowledge about the contents of the mediator), iv) the password required by the mediator to attend the request (if necessary), among others. Figure 4 presents a data source registration, associating a specific data source, File System2, to the Municipal Legislative Mediator. We may also register the available types of components in order to know to which phase of the application development the component belongs (analysis, architectural or implementation see Figure 8). One important characteristic of OML is to use domain ontology to search for domain terms and its ontological relationship, within or among various domains at different levels of abstraction. Thus, SM has to capture the ontological model 4 of each mediator and associate terms among them. For capturing each ontological model, SM uses the ORB bus, through IDL retrieve_Ontology() interface method (Schema 2), to access the specific mediator, retrieving its ontological terms. Therefore, the ontological model provides the main structure for dealing with domain ontology relationships. Relationships involve semantic links such as Hypernyms, Hyponyms, and Synonyms. A Synonym link associates ontological terms in several domains that represent synonyms for a particular ontological term. Hypernyms and Hyponyms links relate ontological terms from various domains that can be either more general or more specific than the current one. Thus, it is possible to associate ontological terms from multiple domains, providing accessibility for domain information. Figure 5 presents the interface for the association of ontological terms. In a query formulation, SM accesses and retrieves all related mediators, searching for component information that fulfills the query semantics. The ontological information about requested 4 Each ontological term is specified as a domain term and a detailed description of it, and relationships with other ontological terms, at different levels of abstraction, are created (see section 3.2) domains is transferred by the Broker (ORB) to the proper domain (mediator), using the method queryMediator (in string query), and correct ontological domain terms. In the example shown, SM will query all mediators that are related to the Municipal Legislative mediator. Thus, SM will transmit the query to the Legislative Domain Mediator and Justice Domain Mediator (see Figure 2). The retrieved components and their corresponding match levels are shown, if each retrieved component exactly matches the query or if the component Figure 3 A Mediator Registration Example Figure 4 A data source registration example 24 partially attends the request. OML registers this information and presents how well the component attends the request (total or some percentage in the latter case). In section 4, we present a more concrete example of this kind of query. 3.2 Mediator Manager Each mediator has its own metadata. This metadata represents the ontological model of the domain. The Mediator manager (MM) also stores relationships among the ontological terms and components stored in data sources. This metadata provides the capability to retrieve components related to this mediator (domain). In order to provide this feature, MM also stores the ontology metadata related to this domain. For each ontological domain term, it is necessary to register its name, its type (in this specific case a term that represents a functionality in the domain), its importance within the domain and the related domain terms. These terms permit the expansion of the query range within the domain, i.e., if there is no component information on data sources related to this specific ontological term, OML could query related ontological terms in the same domain. Of course, this "shift" must be reported to the user. In order to relate each ontological term with its counterparts in data sources, MM retrieves related information of data sources from SM, using the ORB bus. The retrieved information is used by MM to locate and retrieve software components from data sources. Thus, we can associate each ontological term with its related components, with the help of a specific translator. A RETRIEVAL EXAMPLE USING OML Consider a user who is developing an application to handle new legislative proposals, among other characteristics, in the legislative domain. He wants to know if he can use pre-existing software components in his application. Thus, he can use the OML user interface to know about the availability of this kind of components, and to retrieve some candidates. In our example (Figure 6), data source 2 has a binary software component called "New Subject", and data source 1 has a Java package (set of related classes) named "Proposal Creation". Both data sources were mapped into the Legislative Municipal Domain Mediator. Therefore, the Justice Domain mediator has an ontology term named Justice Code that is mapped to a component named Code Database. These components are made available to the Legislative Municipal Domain through the ontological term in the mediator called "Creation of New Proposals", and are mapped to the above components in data sources, i.e., data sources 1 and 2. During the creation of new proposals within a Municipal Legislative House, there are some cases when it is necessary to consult justice database rules. This justice database can impose some restrictions on a new proposal creation. Figure 5 Multiple Ontology Association 25 When the Municipal Legislative Mediator was registered, the SM administrator associated this mediator with the Justice Mediator. This Justice Mediator provides software components used for the development of applications in the Justice domain. Thus, when our user accesses the SM interface in order to retrieve components related to the creation of new proposals, he can choose to access information from all related mediators, i.e., generic mediators, specific mediators, associated mediators or all of them. Suppose our user decides to retrieve information from the Legislative Mediator and associated mediators, then he will access components from the Legislative Mediator and Justice Mediator (see Figure 2). The formulation of the query (Figure 7), selecting the type of component to be retrieved (components belong to analysis, architectural, codification or all phases of development), and the result of this query is presented in Figure 8. Note that for each component, a description of the retrieval is presented (in Data Source 1 Data Source 2 Mediator Manager ORB Justice Rules Proposal Creation New Subject Data Source 3 Rule Database Mediator Manager Creation of New Proposals Service Manager Creation of New Proposals Hyponym Justice Rules Figure 6 Retrieval Schema Example Figure 7 Query Formulation 26 Figure 8, the description presented is related to the Rule Database Access component) and the user can select one or more components to retrieve. Through the mediation structure, OML users can search for components in a transparent and uniform way. In the above example, users of OML do not have to know where components are stored. Moreover, users do not have to query all component repositories, using each specific repository query language format (when a query language exists) to find where the needed components are stored. They do not even have do know how to access data sources. The complexity for dealing with these heterogeneous repositories is treated by OML. Without this layer, users would have to handle these repositories individually, increasing the complexity of the query and access. By using mediators, users can query specifically the mediation metadata, using one single model. The mappings between mediator metadata and translators redirect and decompose the query to data sources 1, 2, and 3. Also, the identification of components of the same domain that are in different repositories can be detected at the time of their registration in the mediation layer. Afterwards this is all transparent to users. CONCLUSIONS This work addresses the interoperability problem between component information repositories. An integration layer was developed to help searching and identifying suitable reuse components. This layer is based on mediators and ontologies to provide the binding of different components to their domain concepts. To assist the identification of related components and their appropriate domain organization, each mediator encloses one domain ontology and provides the mapping to their respective repository of components. Mediators provide a uniform view of the available components organized in domain taxonomy. Domain ontologies are used to help searching for reusable components information through the representation of domain semantic concepts. Therefore, this mediation layer promotes domain information integration and provides mechanisms to translate component requests across ontologies. The important aspect of our proposal is the use of domain ontologies, for reusable component retrieval, in a concrete situation, allowing users to express component requests at a higher level of abstraction when compared to keyword based access or component interface based access used in other proposals. Without OML, users would have to access directly various repositories, dealing with specific characteristics of each repository. Therefore, the main contribution of this paper is to show the potential of the technology of mediators, together with ontology models, for dealing with components repositories complexities, and organizing the manipulation of different components within a domain ontology. Although, the mediation technology is quite popular within HDDS, its adaptation using domain ontologies for component information retrieval is innovative. OML is an operational interoperability architecture based on the use of mediators, translators, and a CORBA communication protocol, which is responsible for the connection among translators and mediators in a distributed and heterogeneous environment. It was constructed using the C++ language together with the Visibroker ORB for C++. Currently, OML is being extended in order to publish and search for components on the Internet [19], based on XML standard. References [1] Jacobson, I.; Griss, M.; Jonsson, P. : "Software Reuse: Architecture, Process and Organization for Business Success;" Addison Wesley Longman, May 1997. Figure 8 Example of component information retrieval in OML 27 [2] Wiederhold, Gio; Jannink, Jan: "Composing Diverse Ontologies;" 8th Working Conference on Database Semantics (DS-8), Rotorua, New Zealand (DS-8) January 1999 (Final version to be published by IFIP/Kluwer/Chapman&Hall). [3] Arens Y., Knoblock C.A., and Shen W.: Query reformulation for dynamic information integration. Journal of Intelligent Information Systems, 6(2):99130, 1996. [4] Molina, Garcia and et.al. :The TSIMMIS approach to mediation: Data models and languages. Journal of Intelligent Information System, 8(2), 1997. [5] Braga, R.; Mattoso, M.; Werner, C.: "The Use of Mediators for Component Retrieval in a Reuse Environment," In: Proc. Technology of Object-Oriented Languages and Systems Conference (TOOLS-30 USA'99), IEEE CS Press, Santa Barbara, pp.542-546, August 1999. [6] Seacord, R.; Hissan, S.; Wallnau, K,: "Agora: A Search Engine for Software Components," Technical Report CMU/SEI-98-TR-011, August 1998. [7] Genesereth M.R., Keller A., and Duschka O.M.: Infomaster: An Information Integration System. In SIGMOD RECORD, Proceedings of the 97 ACM SIGMOD International Conference on Management of Data, pp. 539542, Tucson-Arizona, 1997. [8] Levy, Alon Y., Rajaraman, Anand, and Ordille, Joann J.: Querying heterogeneous information sources using source descriptions. In Proceedings of the 22nd VLDB Conference, pp. 251262, Mumbai (Bombay), India, 1996. [9] Fowler, Jerry, Perry, Brad, Nodine, Marian, and Bargmeyer, Bruce: Agent-Based Semantic Interoperability in InfoSleuth , SIGMOD Record 28(1): pp. 60-67, 1999. [10] RIG; "Reusable Library Interoperability Group" at http://www.asset.com/rig/, 1996. [11] Pires, P.; Mattoso, M.: "A CORBA based architecture for heterogeneous information source interoperability;" Proceedings of Technology of Object-Oriented Languages and Systems - TOOLS'25, IEEE CS Press, pp.33-49, November 1997. [12] Szyperski, C.: Component Software: Beyond Object Oriented Programming, Addison Wesley, 1998 [13] Ram, S.: "Guest Editor's Introduction: Heterogeneous Distributed Database Systems,"; IEEE Computer, Vol. 24 No.12, December 1991. [14] Nieto, E. M.: OBSERVER: An Aproach for Query Processing in Global Information Systems based on Interoperation across Pre-existing Ontologies, Doctoral Thesis, Universidade de Zaragoza, November 1998. [15] Weinstein, P. C.: Ontology-based Metadata: Transforming the MARC Legacy, Proceedings of the 1998 ACM 7 th Internacional Conference on Information and Knowledge Management, pp. 52-59, 1998. [16] Braga, R.; Mattoso, M.; Werner, C.: "Using Ontologies for Domain Information Retrieval," in DEXA 2000 DomE Workshop, pp.100-104, September 2000. [17] Braga, R.; Werner, C.; Mattoso, M.: "Odyssey: A Reuse Environment based on Domain Models"; In: Proceedings of IEEE Symposium on Application-Specific Systems and Software Engineering Technology(ASSET'99), IEEE CS Press, Richardson, Texas, pp.50-57, March 1999. [18] Ye, Y.; Fischer, G.: "Promoting Reuse with Active Reuse Repository Systems," IEEE ICSR 2000, Vienna, pp.302-317 , June 2000 [19] Pinheiro, R.; Costa, M.; Braga, R.; Mattoso, M; Werner, C.; "Software Components Reuse Through Web Search and Retrieval", Proceedings of the International Workshop on Information Integration on the Web - Technologies and Applications, Rio de Janeiro, Brazil, 2001 (to appear). 5 Each ontological term is specified as a domain term and a detailed description of it, and relationships with other ontological terms, at different levels of abstraction, are created (see section 3.2) 28

Domain Engineering;Software Classification and Identification;Component Repositories;Component Based Engineering

193

Through Different Eyes Assessing Multiple Conceptual Views for Querying Web Services

We present enhancements for UDDI / DAML-S registries allowing cooperative discovery and selection of Web services with a focus on personalization. To find the most useful service in each instance of a request, not only explicit parameters of the request have to be matched against the service offers. Also user preferences or implicit assumptions of a user with respect to common knowledge in a certain domain have to be considered to improve the quality of service provisioning. In the area of Web services the notion of service ontologies together with cooperative answering techniques can take a lot of this responsibility. However, without quality assessments for the relaxation of service requests and queries a personalized service discovery and selection is virtually impossible. This paper focuses on assessing the semantic meaning of query relaxation plans over multiple conceptual views of the service ontology, each one representing a soft query constraint of the user request. Our focus is on the question what constitutes a minimum amount of necessary relaxation to answer each individual request in a cooperative manner. Incorporating such assessments as early as possible we propose to integrate ontology-based discovery directly into UDDI directories or query facilities in service provisioning portals. Using the quality assessments presented here, this integration promises to propel today's Web services towards an intuitive user-centered service provisioning. Categories and Subject Descriptors

INTRODUCTION Web services are expected to provide an open platform not only for electronic B2B interaction, but also for the provisioning of so-called user-centered services, i.e. B2C services that can provide useful information and a variety of service offers to support users in a modern mobile lifestyle. Though the capabilities of such services are still relatively simple, their sophistication will grow with the improvement of (wireless) networks, bandwidths, and client device capabilities. However, finding the adequate service for subsequent use of each individual user becomes a more and more demanding problem. Given the convergence of networks in forthcoming (mobile) environments and the evolving innovative business models for third party service deployment (e.g. NTT DoCoMo's i-mode service certification/licensing model for mobile service portals<A href="193.html#10"> [19]) the variety of services is even expected to grow. Making an informed choice of the `right' service will therefore include matching individual users' preferences or dislikes against the concepts and capabilities of the services offered. Usually the interaction process for Web services consists of three distinct phases: a discovery of possible services, the selection of the most useful, and the subsequent execution. In understanding what a service actually offers the first two phases are crucial and the general acceptance of user-centered services will depend on the solutions of still demanding problems in interaction like cooperative querying. As sh<A href="193.html#9">own in [4] and [5]<A href="193.html#10"> the discovery and selection processes of user-centered Web services involves a high degree of respect for user preferences to be flexible enough for real world use. In that respect providing user-centered services strongly differs from the well-defined capabilities of traditional B2B services. As a running example of a typical user-centered service we will use an extension of the cooperative restaurant booking Web service <A href="193.html#9">presented in [4]: restaurant booking services subscribe to the least general applicable node along a complex service ontology for a number of characteristics. A service request can then be performed including a choice of various individual categories. However, the individual services offered will usually only more or less match all the user's expectations. Ranking services with respect to requests is thus an ongoing challenge, as is also evident from the research areas of IR or Web search engines for information provisioning. Service providers almost always can anticipate some typical interactions with their services. For our example typical tasks are for instance booking a certain restaurant for a specific evening, finding a suitable restaurant in the vicinity for lunch, etc. The characteristics and input parameters for Web services for restaurant booking thus usually contain a number of general input values that can be specified in a service request/query: the name of the restaurant, its location, its specific address, the type of cuisine, the date and time for a booking, its price range or even third party content like recommendations (e.g. the Zagat reviews). However, from a service provisioning point of view the nature of these parameters strongly differs. A user expecting to book a certain restaurant on a specific evening will expect that the request may be granted or may fail depending on current reservations of that restaurant for the given date, but relaxing the constraints of the date given or booking a different restaurant for the evening might simply not do. In contrast a user simply wishing for a close-by restaurant to have lunch will rarely provide such fixed terms as a restaurants name, but rather use descriptive terms like a preferred cuisine and an approximate location. Figure 1: Concept of enhanced UDDI service registries Distinguishing such query stereotypes like `book a table at the `Chez Panisse' for the 12/3/03 8:00 pm' and `give me the name and address of a Chinese restaurant in the commercial district of San Francisco with medium price range' and the subsequent personalization of service provisioning also needs different types of input parameters. Whereas simple variables like the restaurant's name or a certain category in a clear request can be handled in an exact match fashion, more fuzzy attributes in a somewhat tentative request like an approximate location or the choice of cuisine have to be understood as a user`s preferences with respect to certain concepts (soft constraints). In the area of the Semantic Web the management of such concepts is usually done by the very powerful tool of ontologies that describe a generalization hierarchy of such concepts. In the course of this paper we will show how to open up service provisioning to the better understanding, adequate handling and quality assessment of each individual user's intentions and preferences. The contribution of the paper thus is twofold: On one hand we relate the use of ontologies and the handling of conceptual views like given by the Semantic Web to co-operatively evaluating preferences for each specific user On the other hand we show how to effectively deal with the problem of relaxing multiple conceptual views for more complex queries and give quality measures to assess the most useful results for each specific user. Both contributions can be expected to improve the service provisioning of user-centered Web services and help to boost their usability and thus subsequently their acceptance. SEMANTIC REGISTRY ENHANCE-MENTS Today Web services are usually provided via an Internet wide network of services registries given by the Universal Description Discovery<A href="193.html#10"> and Integration (UDDI) [21]. UDDI builds on the Web Service Definition Language WSDL<A href="193.html#10"> [7] which features basic information about providers of a service and technical service invocation details. Even though UDDI has become the de facto standard in the field it suffers from a major shortcoming: the information offered on individual services is rather limited. A yellow -page-style lookup mechanism provides the service interface together with a short verbal description of what task the service performs. Mainly targeted a human Web service experts and developers , advanced query capabilities and cooperative matchmaking , however, are still lacking. Research in the area of the Semantic Web seeks a solution to this unsatisfy<A href="193.html#10">ing situation, e.g. [20][4]<A href="193.html#9">. Generally speaking, the Semantic Web fosters a population of the Web with content and services having formal semantics and rich service descriptions. Several semantic frameworks for Web services are currently emerging <A href="193.html#9">with DAML-S [2] and W3C's recently established <A href="193.html#10">OWL-S [9], [10] initiative as the most prominent approaches. We have built our previous work on DAML-S as a relatively mature ontology-based approach to the description of Web services that tries to provide a common ontology of services for the Semantic <A href="193.html#10">Web. Building on top of DAML+OIL [8] the Web service representations in DAML-S consist of a service profile for advertising and discovering services, a process model giving a detailed description of a service's operation and a service grounding providing details on how to interoperate with services via message exchange . Figure 1 shows a schematic view of our semantically enriched Web service provisioning concept. A user states personal needs and preferences in an enhanced service request. The enhanced UDDI registry matches this request against the descriptions of all registered services. The actual matching can be carried out using cooperative database technology like sh<A href="193.html#9">own in [4]. The query is split in hard and soft constraints where the hard constraints are processed as filter conditions, whereas the soft conditions can be relaxed if necessary. If no user-specific preferences are given with the service request, the relaxation follows the domain-specific conceptual views of the service ontology given by the service providers or portal operators. To distinguish between several possible relaxations the quality assessment, which is the main aspect of this paper, will evaluate the degree of match for each service with respect to the original user query and offer all best matches. After a certain implementation has been chosen, the service provider will execute the service and deliver the result. From a service provisioning viewpoint centralized and publicly available service ontologies can be understood as a default service conceptualization or the most common service concept hierarchy, i.e. encoding common and widely accepted knowledge or world/domain knowledge. Due to the hierarchical nature of ontologies a user asking for a specific service will also be served with any more specific service concept subsumed by his request. On the other hand, in the case where a best match to his initial request is not available he/she might also be satisfied with more general services from a super-class of the requested one. This is determined by a relaxation step in the service ontology, i.e. a generalization of concepts along the lines of the ontology. ONTOLOGY-BASED WEB SERVICE DISCOVERY AND SELECTION In this section we provide a brief overview of the use of ontologies for relaxation of soft query constraints and show how common knowledge serves as a default for cooperative retrieval. 197 American California Seafood Shopping Commercial Cajun Texmex Organic Fusion Center Suburb City Cultural ... ... ... Location Thing Restaurant Akasaka Chez Panisse Avocado Garden Cesar The Walrus ... Se rvi c e s S e rvi c e O n t o lo gy ,,Cuisine" Figure 2: Service Ontology for restaurant booking. 3.1 Service Ontologies, User Preferences and Usage Patterns The purpose of a service ontology is to describe the kinds of entities available in a service repository and how they are related. To this end service ontologies may include descriptions of service classes, properties and their instances (the actual services that are eventually selected for execution). A basic service ontology is depicted in Figure 2. Here restaurant booking services are classified according to their cuisine and their location in a city. For instance the restaurant `Chez Panisse' serves `Organic' food with `Organic' being a specialization of the `Californian' cuisine (as well as of `American'). Furthermore the restaurant `Chez Panisse' is located in the `Shopping' district which itself is part of the city `Center'. We have used W3C's Web Ontology Language (OWL) and its predecessor DAML+OIL to enrich DAML-S service pro-<A href="193.html#9">files in Web service repositories [4][5]<A href="193.html#10">. Modeled in OWL the most general `Restaurant' and `Location' concepts are anchored in `owl:Thing' the most common concept of any ontology. A restaurant booking service using the service ontology from Figure 2 might on one hand assume that a user asking for `a restaurant featuring American cuisine' will be well served by all restaurants with e.g. Cajun, Californian or Texmex cuisine, since they are all instantiations of American cuisines. On the other hand, if a user asks for `a Californian fusion cuisine restaurant' and no such service should be registered, implicitly relaxing the query to all Californian restaurants and offering restaurants with organic cuisine or Californian seafood to our user will be more helpful than just stating the empty result. And even if a user has different conceptions (e.g. cuisines being related based on their flavors) or explicit preferences (a Chinese restaurant rather than an Italian one), an ontology-based discovery/selection model is <A href="193.html#9">still useful. [4] shows in detail how to deal with these cases by overwriting the default ontology with an explicitly provided (or implicitily derived) generalization hierarchy of a user somewhat similar to the view definitions proposed by<A href="193.html#10"> [14]. Since for our assessment framework here the exact kind and classes/values of an ontology do matter less than its actual structure in each instance , such overwritings by user specified conceptions are always possible to facilitate. We advocate the use of service ontologies together with a proprietary notion of basic user preferences and typical service usage patterns for the stepwise refinement of service requests in a cooperative service provisioning environment. While the basic approach and combination of ontologies, preferences and patterns is <A href="193.html#10">published elsewhere [5] we will now concentrate on enhancements to the relaxation along the lines of ontologies alone: unlike the specialization of a request, a generalization can result in severe changes of the initial query semantics. This is especially true, if several relaxation steps have to be performed until a match can be found. At the point of relaxing a constraint to the root of an ontology, the respective constraint can even be considered as entirely dropped. But nevertheless, since an ontology resembles common knowledge (and thus implicit preferences), for high quality service provisioning offering somewhat related features is usually still a better default than just returning an empty result set. 3.2 Multiple Conceptual Views Individual users might have quite specific ideas about differing domain concepts (conceptions) or very clear expectations how to be served differing from the usual domain assumptions (explicit preferences), but also implicit preferences play an important part. Consider for instance location-based services, e.g. for restaurant booking. If a user asks to book `a Chinese restaurant for dinner', the common domain knowledge tells us that this restaurant should be in the vicinity (e.g. a 30 miles area) of his current or usual whereabouts and we can add this information as an implicit constraint for better provisioning quality. A user in San Francisco would usually be annoyed by offers of Chinese restaurants in Hong Kong no matter how good their actual quality or rating is. If this general assumption would not hold, however, (e.g. if a user wants to fly to Hong Kong in the morning and then have dinner there) he or she would have stated this unusual detail already within the query and had asked for a `Chinese restaurant in Hong Kong for dinner'. Such explicit information within a service request is provided due to the psychological notion that though users want a service to know what is sensible (like they expect to be served in human-human interaction), no user expects a service to be clairvoyant. Thus, not only having further (explicit) knowledge of a user, but also assuming typical behavior, concept hier-198 archies given by ontologies can be used as good default relaxation hierarchies for user preferences. Should, however, some preferences or a specific conception be given, the underlying ontology has to be exchanged against the user-provided terms or concepts. We introduce the notion of conceptual views on a service ontology to account for all the different interests a user wants to express in a service request. Such a conceptual view is modeled as a clipping from the full service ontology that starts with the most general concept associable with a specific user interest. For this paper we will for the ease of understanding assume conceptual views to be non-overlapping, tree-shaped clippings from the full service ontology where each service is registered with the node that describes its value with respect to the most specific characteristics . An example conceptual view is indicated as a grey shaping in figure 2: the view named `Cuisine' is basically a sub-ontology only concerned with the classification of restaurants according to the offered type of food. Whereas the restaurants `The Walrus', `Chez Panisse', `Avocado Garden' and Cesar are classified as being `Fusion' or `Organic' places the restaurant `Akasaka' is not reachable in this view as it is only classified as located in the cultural district. As we will discuss in the remainder of this paper multiple conceptual views can be used to account for different interests a user wants to express in a service request and the relative importance between them. In the case of the restaurant booking example it is conceivable that a user values the fulfillment of location constraints over cuisine constraints if he/she is only up for a quick work lunch. However, for the ambitious `hobby gour-met' this might be just the other way round on the weekend. Thus we will need ways to assess the respective quality of different relaxation schemes to allow users to make an informed choice. RELAXING MULTIPLE ONTOLOGIES Let us now focus on problems that arise when multiple soft constraints have to be relaxed over the service ontology. We will first look at our sample scenario, investigate the relaxation of conceptual views and then discuss some quality considerations. 4.1 Relaxation Plans for Multiple Selection Predicates Let us consider our restaurant booking service from above. A typical query would be "Find a Californian fusion cuisine restaurant in the commercial district of Berkeley". Here we will have to deal with two soft constraints: the type of cuisine and the location. Assuming that we do not have more specific information about the user's preferences both constraints could be relaxed along the two default conceptual views of the service ontology of figure 2. `The Walrus' fusion `Cesar' `Chez Panisse' californian organic `Avocado Garden' `Cucina Calabrese' seafood `The Mediterraneum' Figure 3: The cuisine ontology with respective instances Figures 3 and 4 show the full first two concept levels of the respective conceptual views with some instances. So in figure 3 we can for instance see that there is a service for a restaurant called `The Walrus' which is classified as offering fusion cuisine and as such, also offering Californian cuisine. The relaxation of query predicates over such a view is straightforward. If the query predicate specifies `fusion cuisine' restaurants, we would have the choice between the respective instances, here `The Walrus' and `Cesar'. If for some reason there would be no fusion cuisine restaurants registered, or the instances cannot satisfy some other constraints (like booking for a certain date), we will relax along the ontology to the more general concept of Californian cuisine and can also consider the restaurants that are registered under the `organic' and `seafood' characterizations of California cuisine. The technical problem of how to relax single conceptual views with adequate query languages over cooperative database systems <A href="193.html#9">is in detail addressed e.g. in [4] and [5]<A href="193.html#10">. The beauty of the design is that all details of a UDDI or DAML-S style description for each service together with some more characteristics (e.g. taken from RDF statements of a restaurants homepage) can be stored in a classic relational database by the service provider and can be searched using a declarative query language extended by preference constructors like sh<A href="193.html#10">own in e.g. [13]. Thus an added-value service using semantically meaningful content can be provided quite easily. `Tom's Grill' commercial `Sushi and Soul' `Avocado Garden' city center cultural `The Mediterraneum' `Akasaka' `Chez Panisse' shopping `Cesar' `Pizza My Heart' Figure 4: The location ontology with respective instances A more serious problem arises when several soft query constraints over various conceptual views have to be evaluated. Consider for instance the query on a `fusion cuisine restaurant in the commercial district'. We can easily verify that in our example even though `The Walrus' and `Cesar' are fusion cuisine, they are not registered in the commercial district. Likewise `Tom's Grill', `Sushi and Soul' and `Avocado Garden' are in the commercial district, but they do not offer fusion cuisine. Aiming at a cooperative retrieval behavior we are left with three choices: relaxing the constraint on the cuisine, relaxing the constraint on the location or relaxing both constraints. When relaxing the cuisine to `califor-nian' we would retrieve the service of `Avocado Garden' the only Californian cuisine restaurant within the commercial district. Relaxing the location-based ontology would result only in the service of `Cesar' the only fusion cuisine restaurant in the city center . Finally relaxing both constraints would result in all Californian cuisine restaurant within the city center, i.e. `Avocado Garden' , `The Mediterraneum', `Chez Panisse' and `Cesar'. Thus 199 different kinds of relaxation will usually result in essentially differing answer sets. This problem will remain if more relaxation steps have to be taken. If we have relaxed only one constraint we could e.g. decide to relax this property even further or relax another property of our service during the next step. Let us first define the task of finding the `best' service under relaxation The Problem of Best-Matching Service Provisioning: Given various characteristics that describe all services in a service ontology, a concept hierarchy (conceptual view) for each characteristic and a user request stating a number of hard and soft constraints , the best-matching service is given by all services that: - fulfill all the hard constraints - fulfill all soft constraints with a minimum amount of relaxation of characteristics with respect to a suitable quality measure So given that each service is registered in one concept node for each characteristic given by the respective conceptual view, and all services not fulfilling the hard constraints have been filtered, the problem of selection over multiple constraints comes down to deciding what is `a minimum amount of relaxation'. Obviously the basic task is finding a service that is registered with all concepts (or any of their respective sub-concepts) as specified by the query, a `perfect match'. But if there is no such service registered, the decision what soft constraints to relax and how far they are relaxed, is paramount for the quality of provisioning. 4.2 Basic Service Quality Considerations Since the decision about the relaxation scheme is important for the output, some way of considering which scheme to follow is needed. Usually ontologies are of a qualitative nature. A superclass / subclass taxonomy is established, but there is no knowledge of the `degree' or the relative distance between different concepts. However, such knowledge could crucially change the utility of certain relaxation plans. Relaxing more refined ontologies or views will generally hurt the user preferences less than relaxing already coarse views or ontologies. The less general the concept, the more refined are the sets of objects that will be offered to the user, and flooding the user with too much too general content is avoided. Let us first take a closer look at merely qualitative views on ontologies of comparable granularity, etc. (i.e. relaxing one constraint is introducing the same amount of generalization as relaxing any other) and then investigate ways to deal with quantitative measures in the following section. For scenarios of merely qualitative preferences and their relaxation for the restricted class of ceteris paribus preferences <A href="193.html#10">[16] proposes a scheme of ordering different objects in the result set according to the count of necessary relaxation steps from the top or the bottom of the hierarchy or simply the relative distance to all violated query constraints. Since we assumed symmetrical views on ontologies, also in our relaxation problem a similar concept will help us to understand what should be relaxed preferably and what this means for the objects in the result set. Let us first show the approach of simply counting the relaxation steps from the violated constraints. We will label the services we found in each step by the number of necessary steps to find them. But first we need to define the necessary concept of relaxation paths. Given tree-shaped conceptual views of a service ontology that arranges concepts or values with respect to certain service characteristics using a generalization semantics, a relaxation path is a path along the edges of each conceptual view that leads from a base concept to the respective root of the view. Usually this base concept is specified in a service request or user query and relaxing along the relaxation path leads to an increasing generalization of this concept. Assuming that all services have been assigned to the node of their first appearance along the relaxation path (i.e. they are registered to any of the respective node's sub-trees of concepts , but not to an earlier node of the relaxation path) we get a chain of concepts with all services registered under the aspect of least necessary level of generalization. An example for a relaxation path can be easily derived from figures 2 and 4. If a user is primarily interested in the commercial district, the appropriate nodes of the relaxation path would be `commercial district', `city center', `city' and `location'. The services registered in the nodes are e.g. `Tom's Grill', `Sushi and Soul' and `Avocado Garden' for `commercial district'. The node `city center' would also contain all services registered to its sub-concepts , (i.e. `The Mediterraneum', `Akasaka', `Chez Panisse', `Cesar' and `Pizza My Heart') and so on. Figure 5 shows the first two steps of respective relaxation paths for both conceptual views in figures 3 and 4 focusing only on the services registered in both views. As we pointed out we will always assume tree-shape views for the course of this paper. Please note that all the concepts easily can be transferred to the case where sub-concepts can have multiple parent nodes. In this case the node along the relaxation path would consist of the intersection of the different parent concepts , or the intersection of their registered services respectively. This generalization has already been successfully employed in a similar fashion for mapping queries between differing ontologies by<A href="193.html#10"> [18]. Distances then can simply be measured by the minimum distance, if multiple paths for relaxation should be available. Let us now see, how relaxation can be done using an unlabeled relaxation graph (see figures 3 and 4). If we begin by either relaxing the cuisine or the location constraint of our query we would have to assign a quality value of 1 to both the `Avocado Garden' and `Cesar'. Thus in terms of quality they are incomparable, which closely resembles our missing knowledge of what kind of relaxation the individual user would prefer. If there should be more ways to relax constraints to encounter a service, we will always count the minimum number of relaxation steps necessary. Relaxing both constraints again leads to quality values of 1 for `Avocado Garden' and `Cesar' and a value of 2 for `The Mediterraneum' and `Chez Panisse', because they have only been seen by relaxing both constraints and thus are probably less desirable for the user. However, relaxing the cuisine ontology two steps (i.e. to all American cuisines) and sticking to the commercial center constraint might result in `Tom's Grill' turning up also with a value of 2. This is because in the nave model the semantic difference between `deep' relaxation and `broad' relaxation is considered the same. To be sensitive to the differing implications of broad and deep relaxation with respect to generality we will use our concept of relaxation paths and show how to use labels along this path to get to a more sophisticated relaxation paradigm. fusion californian american city center commercial city 1 2 1 2 fusion californian american city center commercial city 1 2 1 2 Avocado Garden Tom`s Grill Cesar Cesar Avocado Garden Tom`s Grill Chez Panisse Mediterraneum Chez Panisse Mediterraneum Figure 5: Relaxation paths and assigned services 200 Usually the generalization throughout ontologies will become quickly rather unspecific with decreasing distance to the root. Hence a broad relaxation strategy (breadth first relaxation) often is preferred to deep relaxation steps. Summing up the distances like before, but weighing each relaxation step with the relative distance to the original query term can implement this. Consider our query for `fusion cuisine' restaurants in the `commercial district' . So for example the `Avocado Garden' and `Cesar' need each only one relaxation step with a distance of one to the original constraint (cf. labels in figure 5), so their quality value is 1. `Chez Panisse' and `The Mediterraneum' both need two relaxation steps with a distance of 1 each resulting in a value of 2. In contrast `Tom's Grill' also needs only two relaxation steps, but whereas the first step has a distance of 1, the second step already shows a distance of 2. Thus the final value for `Tom's Grill' is 3 (i.e. 1+2*1) and we can now effectively distinguish between deep and broad relaxation. Depending on the nature and granularities of the ontology or views we can of course also use higher weightings for the deep relaxations, for instance 10 (distance-1) . So the first step will be weighted by 1, the second deep step by 10, the third deep step by 100, and so on. Since the broad relaxation steps are still simply added up, this will `punish' deep relaxation and avoid too broad generalizations of constraints. If we always want to punish deep steps symmetrically until all constraints in turn are relaxed at least to the same level a factor of (number_of_ontologies) (distance-1) will be adequate as shown in the following lemma. Lemma 1: Weightings to Foster Broad Ontology Relaxation Given n soft query constraints with their respective relaxation hierarchies. To always prefer a broad relaxation scheme, label each object by summing up the numbers of edges relaxed to find this object in each hierarchy and weigh every edge by n (d 1) using the number of soft constraints n and the relative distance d to the original query constraint. Proof: Since within each depth all weightings are the same, it is obvious that within a certain depth of the hierarchy any object seen with less relaxation steps has a smaller label than an object that needs more steps. Thus if we e.g. have to relax two constraints within a level this object will always be labeled with a higher weight than an object that needed only one relaxation independently of which constraints have been relaxed. We still have to show that if we do a step with a deeper distance, an object O encountered there always gets a higher label than any object P encountered in all hierarchies only with relaxations up to a lower distance. We will do that by showing that the minimum label for object O is higher than the maximum label for object P. Let us assume that in order to encounter object O we have to relax at least one constraint to a distance of k. The minimum label for O thus is given by relaxing only a single constraint to distance k and not having to relax any other constraint. Hence object O's label is given by (n 0 +n 1 +...+n (k-2) +n (k-1) ). The maximum possible label for object P on the other hand is given by having to relax every of the n constraints (k-1)-times, i.e. the maximum distance smaller than k. Thus the maximum label for P is n*(n 0 +n 1 +...+n (k-2) ) = (n 1 +...+n (k-2) +n (k-1) ) and thus P's label is -even relaxing all constraints to a maximum- at least by 1 smaller than the best possible label for O. Thus in relaxing constraints for equally important conceptual views an adequate algorithm would be the processing of decreasing levels of quality, i.e. finding services with increasing weightings . Starting with the minimum possible relaxation the algorithm will always work over an entire sequence of services with the same quality index and return all the discovered services of the lowest level found, together with their quality estimation. This is important for having to restart the algorithm at the previous point of termination, if the services discovered so far should not have been sufficient and the evaluation of lesser quality levels becomes necessary. Similarly, knowing the labeling technique and the views involved a user can also specify a maximum quality value up to which he/she is willing to accept more general services. For our algorithm we will assume a declarative query mechanism on UDDI directories enhanced by all the feature characteristics described by their respective conceptual views like pres<A href="193.html#9">ented in [4]. However, symmetrically relaxing just the same number of steps even when assuming views of the same granularity and importance with respect to the user query, will generally not lead to our desired broad relaxation scheme with as little generalization as possible. Imagine a query that specifies two soft predicates of which one is a leaf node of a view, whereas the other predicate specifies a direct descendant node of the root in the respective view. If no perfect match should be given, our strategy would result in relaxing either of the two constraints a single step. But whereas the relaxation by a single step from our leaf node usually leads to a slight generalization allowing a few more services for selection, the relaxation step in our second constraint would relax to the root node and thus offer the total number of services registered in the entire ontology for the respective characteristic, i.e. entirely drop the second constraint. Obviously that would not be a sensible behavior. So we do not only have to punish deep relaxation steps but even more severely refrain from relaxations the closer we are to the root. This concept will be implemented in our relaxation algorithm by letting the longest possible relaxation path determine the weightings for all constraints (again assuming a comparable level of detail throughout our ontology). The weightings for edges along the relaxation path will then be assigned in each conceptual view in descending order starting from the root down to the concept specified by the query. Thus the relaxation of all concepts at least to the same level of generalization is enforced before having to relax already more general concepts. In our example from above the relaxation path for a leaf node concept would be assigned weightings as given by the height of the conceptual view, whereas the relaxation path for a concept node right below the root would be assigned the highest possible weighting. Thus (following lemma 1), our leaf node concept will have been relaxed to a generalization level of the respective concept right below the root before the second constraint is relaxed for the first time. Now we are ready to present an algorithm for relaxation of symmetrical constraints incorporation a breadth first paradigm and a minimum level of generalization strategy. Algorithm: Symmetrical Constraints Breadth First Paradigm 1. Pose the query containing only the hard constraints against the enhanced UDDI / DAML-S directory. 1.1. If an empty result should be returned, terminate the algorithm outputting the empty result set. 1.2. Repose the query with all soft constraints included. 1.3. If a non-empty result should be returned for the expanded query, terminate the algorithm and output the respective services as perfect matches with a relaxation level of 0. 201 2. Given n the number of soft constraints we have to label each edge along the relaxation path from the concept specified by the query to the root node in every conceptual view (cf. lemma 1). 2.1. Among the n views find the longest possible relaxation path and set maxdepth as the maximum depth of all ontologies relative to the class specified in the service request 2.2. For every view label the relaxation path starting from the root by n d down to the concept specified by the query (d := (maxdepth 1) to 0 descending) 3. For i = 1 to n j (1 j maxdepth) 3.1. Start with the query including all hard and soft constraints and build statements containing any possible relaxation with a weighting of i, i.e. relax in turn all conceptual views by one step up to the point of reaching the desired weighting. Due to construction of the weighting this will result in a breadth first strategy. 3.2. If any of the statements produced in 3.1. retrieves a non-empty result set, collect results of all possibilities and terminate the algorithm. In this algorithm step 3 can efficiently be implemented using an A*-Algorithm that successively explores the differently weighted tree edges finding all possible combinations for each quality weighting. However, please note that not every weighting is possible to reach by relaxing constraints. We will now exemplify the above algorithm through different examples: consider the three conceptual views X, Y and Z given in figure 6 and assume a user has posed a query requesting the characteristics X3, Y2 and Z6 as soft constraints. Let us assume all hard filter conditions have already been satisfied, but the basic query for our soft constraints fails, i.e. no service with these capabilities is registered. Step two of our algorithm will now label the relaxation paths like shown in figure 6 (bold edges and shaded vertexes). Starting with all services having quality values of one, concept Z6 is generalized to Z3 and the query is reposed with constraints X3, Y2 and Z3. If we still should have no matching services we have to go on relaxing. A query for a value of two is not possible, but for a value of 3 we can relax X3 to X2 and repose the query with constraints X2, Y2 and Z3. Let us assume we still have not found a result the next quality value would be 4. For this we have two possibilities and have to unite the results of the query on X2, Y2, Z3 and the query X3, Y2, Z2. The next possible quality value would be 7 with a query on X2, Y2, Z2. Please note that we indeed have relaxed all constraints to same level until we relax any constraint to the top level (e.g. in the view Y) for the first time. The algorithm would terminate at the latest after relaxing all views to the top level with a quality value of 34. Z2 Z1 Z4 Y5 Y4 Y3 X1 X4 X5 X6 Z6 1 9 3 9 3 9 X3 X2 Y2 Z3 Y1 Figure 6: Conceptual views with labeled relaxation paths 4.3 Quantitative Service Quality Measures In the last section we have seen an effective scheme for the case of symmetrical relaxations under the assumption of equal useful-ness , i.e. one broad step was as useful as any other broad step of a comparable level. But in real world applications query constraints are not always only of a qualitative, incomparable nature. Deeper knowledge of individual user's preferences or knowledge of stereotypical usage can become interesting parameters in assessing the quality of different relaxation schemes. Tuning the factor to a certain ratio for each application (x broad steps equal 1 deep step) will express the desired semantics in each instance. The exact coefficient used for the discrimination of deep relaxation in each instance, however, will typically strongly depend on the domain, the total number of soft query constraints and the respective granularity of the views / ontology (i.e. the semantic level of detail) used. Views that are modeled with a very fine granularity can be relaxed introducing a smaller degree of generalization to the service request results than would be introduced by those views that are modeled rather coarsely anyway. Hence, a deep relaxation step in a very detailed ontology might be worth only three or four broad relaxations of other user constraints, whereas a deep step in a coarser ontology used within the same query might add up to the worth of ten broad relaxation steps or even several deep relaxation steps of other constraints. Likewise very flat hierarchies with many subclasses to each node are not suited too well for deep relaxation. So the discrimination will in each application depend on: The relative semantic importance of a view with respect to the user request The maximum depth of each conceptual view, Its total number of (sub-) concepts and The (average) number of instances in each concept. The relative semantic importance of a view can usually only be determined by directly consulting the user. But in the following we will give an overview of techniques that will generally help to deal with problems of relaxing views with different granularities. As a rule of thumb we can state that the relaxation of views having a low maximum depth and rather high numbers of services attached to each node should be delayed as long as possible. Our technique of starting the view graph labeling from the root node already helps facilitating this rule. If a shallow conceptual view is used (usually an indication for coarse modeling) together with a more detailed view, even the edges to leaf nodes in the shallow ontology will be assigned rather high weightings unlike the leaf nodes in ontologies with a rather high depth. This behavior is, however, not always the best choice. If unlike in figure 6 the respective depths of conceptual views differ by a considerable amount, we should not simply delay the relaxation of shallow views, but have to insert several intermediate steps in the more detailed views before relaxing the next step in a coarse view. Figure 7 shows pairs of views X, Y and X', Y'. In both cases the depth of X, X' is only two whereas the depth of Y, Y' is four. That means that in terms of relaxation we can assume that for some reason the more shallow ontologies X and X' are modeled rather coarsely. On the left hand side in figure 7 we can see the labeling scheme from our algorithm. Ontology Y would be relaxed to Y3 or even Y2 before a single step in X would be relaxed . On the right hand side we can see a better labeling scheme with interleaved relaxation steps (two steps in Y' for a step in X'). 202 Y5 8 4 1 8 4 2 Y5 12 3 1 8 4 2 Y4 Y2 Y3 Y1 Y4 Y2 Y3 Y1 X1 X2 X3 X1 X2 X3 Figure 7: Differently labeled relaxation paths Since the interleaving of relaxation steps with respect to the maximum depth generally seems a fairer approach, we will incorporate this behavior into our algorithm. If the maximum depth of some conceptual views should severely differ, we will assume a coarser level of detail and find out how many steps in the most detailed view represent a single step in the coarser view. We then re-label the coarse view beginning from the root by adding so many of the appropriate sequence of weightings, as steps in the detailed view are necessary. For instance in figure 7 we can easily see that a single step in X (depth 2) represents two steps in Y (depth 4) and thus we would have to re-label the first edge by the sum of the two highest weights of Y (8+4) and its second edge by the sum of the next two weights (2+1). Following this scheme we can gain the more suitable relaxation weights of view X'. In terms of our quality assessment algorithm that means that we have to reconsider the labeling of the relaxation paths in step two and replace the respective section by the following: 2.1. Among the n conceptual views find the longest possible relaxation path multiplying possible steps in each view relative to the class specified in the service request by the view's respective factor of q, where q is given as the integer part of the result of dividing the maximum view depth by the maximum depth of the current view (i.e. q steps in the most detailed view represent one step in this view). Set the maximum value for maxdepth. 2.2. For every conceptual view label the relaxation path starting from the root by n d + n d-1 +...+ n d-q+1 (i.e. the sum of the first q weights in terms of the most detailed relaxation path) down to n d + n d-1 +...+ n 0 with d:= (maxdepth 1). A second possibility to control the relaxation properties of multiple conceptual views is the incorporation of user preferences giving a preferred relaxation order. Incorporating such preferences into the weights along the ontology, however, is a very difficult problem in its own rights and therefore beyond the scope of this paper. For the case that a simple ordering of relaxation for the views is given by the user, we can use double-labeled relaxation paths like <A href="193.html#9">for the tree patterns in [1]. The second label for each node is e.g. the respective rank of the view in a specified relaxation ordering or the number of the node's respective sub-concepts. In the case that for the execution of step 3 in our algorithm more than one query should be possible, the relaxations can then be executed minimizing the sum of second labels and retrieval can be terminated whenever a result occurs. Also for the case that a prioritization of relaxations is give<A href="193.html#10">n (cf. [13]), the respective relaxation scheme is straightforward: The prioritized view is successively relaxed until a first result set is retrieved. Then the second , third, etc. views are used to break ties. However, when it comes to integrating weights from preferences into relaxation weights deeper research is still needed. RELATED WORK While in the above discussions we assumed the existence of different conceptual views to a given ontology, the actual creation of these views is beyond the scope of this paper. Multiple views as abstractions of data sources are a well understand concept in classical databases systems. Yet the concept of such views has only recently been addressed in the context of the Semantic Web through the proposal of the view definition language RVL for the low level ontology<A href="193.html#10"> language RDFS [14]. RVL uses a declarative query language for the creation of virtual schemas which in turn serve as views on existing complex ontologies. Please note that although we merely focused on tree-shaped clippings from OWL ontologies as simple relaxation hierarchies in our examples, the presented concepts are general enough to be used with only the slightest adaptations together with other types of ontologies and more complex views, e.g. virtual RVL views. Choosing the `right' Web service for execution has been considered in several ways. For legal or economical points of view especially assurance structures guaranteeing that a service performs <A href="193.html#10">the desired task like [11] or [17] have been addressed. However, when negotiating about execution guarantees or costs the semantic content of a service has to be understood and its specific capabilities have already to be agreed on. Taking a more user-centered view the notion of services' reputation for subsequent selection <A href="193.html#10">[15] or the quality assessment for the negotiation of service level agreements<A href="193.html#9"> [3] have been proposed. However, these approaches focus on conceptual designs omitting algorithms how to assess the quality in each instance. The most complete framework with respect to heterogeneous environments featuring multiple ontologies is given by<A href="193.html#10"> [18] where the notion of information loss for query reformulation is defined. Unlike our work presented here, where multiple conceptual views of a service ontology occur in a single request, this work, however, deals with the loss of information when a query has to be translated from one into another ontology (e.g. in order to pose it to a different data source). Thus it is rather concerned with the problem of ontology mappings. The area of service request relaxation over ontologies also shows some similarities with database query relaxation frameworks like given <A href="193.html#10">in [6] or [13] and especially recent work on querying semi-structured data like in XML databases. In the case of XML the DTD of a document defines its structure together with the (semantic ) type of data within each node. The main focus of querying in that area is on building queries without perfect knowledge of documents structure or the exact data it contains. Exploiting the set of labels given by a XML document's DTD as ontology in <A href="193.html#10">term of the documents' structure [12] uses the result sets of queries to define the semantic equivalence of alternative query expressions , however without relaxing concepts within queries. The area of relaxation for tree-shaped queries not only on a structural level (`relax to any descendant node instead of child node'), but also on a limited semantic level (`find author of document instead of book') is in <A href="193.html#9">detail addressed in [1]. Here also weightings along the edges of trees comparable to our user preference-driven quality assessments in section 4.3 are discussed. Our work differs mainly in that we can rely on fine-granular concept ontologies that are custom made by domain experts and used in central service provisioning portals or UDDI / DAML-S directories. Not only are we able to exploit semantics by a far larger extent than previous work, but we also provide means to derive sensible weightings for edges within conceptual views based on general user preferences and a fair relaxation paradigm. 203 The general area of enhancement of UDDI goes back to describing the capabilities of Web services on a more detailed level by using ontology languages <A href="193.html#10">like DAML+OIL [8]. An example for the efficient mapping of DAML+OIL capability descriptions onto <A href="193.html#10">UDDI records is given in [20]. Our approach for result quality assessment here is facilitated by the database-based approach for UDDI <A href="193.html#9">enhancement featured in [4]. Using cooperative database technology and an extended declarative query language like the one <A href="193.html#10">given in [13] this framework features a good implementation framework for our quality assessment. Queries using hard and soft constraints can be automatically rewritten using the relaxed concepts and posed against a database of service descriptions using suitable relaxation ontologies. However, in terms of quality these languages feature only a qualitative result set under the notion of Pareto-optimality. Quantitative quality constraints that limit the flood of incomparable results delivered by the exponential growth of Pareto sets with the number of soft constraints in the request are not considered. A first study of such quality measures is given <A href="193.html#10">in [16] for the restricted class of ceteris paribus preferences like discussed in section 4.2. SUMMARY AND OUTLOOK In this paper we presented a framework for the discovery and selection of Web Services based on personalized quality assessments for individual service users. Starting with a set of conceptual views over a service ontology that express a generalization hierarchy of concepts (conceptual views) for different services' capabilities and characteristics, we proposed to enhance today's UDDI / DAML-S registries by a matching component that will not only perform a filtering of services according to user specified terms, but also allows for cooperative matchmaking between service descriptions and the individual user's preferences. Focusing on the quality assessment component of such an enhancement we described in detail how to deal with different kinds and multiple instances of conceptual views. The views in our framework contain the domain-specific understanding of concepts or the common knowledge that users typically will expect when trying to find an adequate service for execution. If no service that offers all required capabilities can be found, relaxing along these views will step by step generalize the services' requested features until a best possible match can be found. Thus cooperative behavior can be introduced for improved service provisioning. We focused on controlling these relaxation steps implementing a breadth first strategy control flow to delay far-reaching generalizations for as long as possible. We also discussed the influence of relaxation plans for various views, which may differ in their granularity or accuracy of discrimination and the influence of individual user preferences for relaxation orders. For the case of views with differing level of details we gave an adequate scheme to balance the control flow. Nevertheless, the exact instantiation of the views and the adequate weightings chosen still will usually differ between application areas. Anticipating stereotype interaction patterns or experiences of past interactions, however, the service provider usually is able to also provide some suitable default ontologies for cooperative matchmaking within UDDI / DAML-S registries or managed service portals. In any case the provisioning of suitable facilities for the assessment of Web service quality can be expected to become a central part of service provisioning and will essentially influence the future acceptance of Web service offers by individual clients. In this paper we have restricted conceptual views to tree-shaped clippings of ontologies with view elements being exclusively related through `is-a' relationships (stating an explicit generalization of concepts in a superclass/subclass fashion). Of course also other relationships within ontologies might be available for relaxation tasks in service requests; on the other hand their semantic meaning will usually be somewhat more difficult. Since complex ontologies commonly contain named relationships between entities that might be used to make views more flexible and query relaxation more meaningful, an important future work item will be to break down this restriction on relationships of the `is-a' type. With our algorithms' focus on relaxation paths as an abstraction of the underlying views, our general framework for quality assessment can be expected to be extensible also to these new types of views in a straightforward manner independently of the exact type of view the relaxation path was derived from. If a relaxation of a constraint along a certain relationship is backed by sensible relaxation semantics (i.e. is meaningful), however, has to be checked in each individual instance. Furthermore, our future work will focus on a tighter integration of individual user preferences into the quality assessment process like addressed in section 4.3. Choosing the adequate weightings does not have an obvious semantics. The meaning of `relaxing one constraint is two-times better than relaxing another constraint' can only be guessed, what about three-times, etc.? The area between quantitative quality assessments like e.g. re-weighting techniques or relevance feedback as known from the area of IR, and the purely qualitative approaches like Pareto optimality of solu-<A href="193.html#10">tions like given in [13] offers a vast variety of possibilities to explore for real world applications. Also here the notion of stereotypical usage of services and the grouping of users with similar intensions might lead to improved service provisioning. We believe that using and extending our framework is a vital step towards getting a better understanding of these topics. In any case assessing quality of service request results in a semantically sensible way promises to pave the road to cooperative provisioning for user-centered services. ACKNOWLEDGMENTS We would like to thank Achim Leubner and Anthony Tarlano for helpful comments and suggestions. This work was partially funded by an Emmy-Noether-Grant of the German Research Foundation (DFG). REFERENCES [1] S. Amer-Yahia, S. Cho, D. Srivastava. Tree Pattern Relaxation . In Proc. of the Int. Conf. on Extending Database Technology (EDBT'02), Prague, Czech Republic, 2002. [2] A. Ankolenkar, M. Burstein, J. Hobbs, et. al. DAML-S: Web Service Description for the Semantic Web. In Proc. of the Int. Semantic Web Conf. (ISWC'02), Sardinia, Italy, LNCS 2342, Springer, 2002. [3] W.-T. Balke, A. Badii. Assessing Web Services Quality for Call-by-Call Outsourcing. In Proc. of the Int Workshop on Web Services Quality (WQW'03), Rome, Italy, 2003. [4] W.-T. Balke, M. Wagner. Cooperative Discovery for User-centered Web Service Provisioning. In Proceedings of the First International Conference on Web Services (ICWS'03), Las Vegas, USA, 2003. 204 [5] W.-T. Balke, M. Wagner. Towards Personalized Selection of Web Services. In Proceedings of the 12th International World Wide Web Conference (WWW 2003) Alternate Track on Web Services, Budapest, Hungary, 2003. [6] S. Chaudhuri. Generalization and a Framework for Query Modification. In Proc. of the Int. Conf. on Data Engineering (ICDE'90), Los Angeles, USA, 1990. [7] E. Christensen, F. Curbera, G. Meredith, S. Weerawarana. Web Services Description Language (WSDL) 1.1. http://www.w3.org/TR/2001/NOTE-wsdl-20010315, 2001. [8] D. Connolly et al. DAML+OIL Reference Description. W3C Note, December 2001. [9] DAML. OWL-S: Semantic Markup for Web Services. http://www.daml.org/services/owl-s/1.0/owl-s.html#foot29 [10] DAML. OWL-S 1.0 Release. http://www.daml.org/services/owl-s/1.0/ [11] M. Jakobsson, M. Yung. On Assurance Structures for WWW Commerce. In Proc. of Int. Conf. on Financial Cryptography (FC'98), Springer LNCS 1465, Anguilla, British West Indies, 1998 [12] Y. Kanza, Y. Sagiv. Flexible Queries over Semistructured Data. In Proc. of the ACM Symp. on Principles of Database Systems (PODS'02), Santa Barbara, USA, 2001. [13] W. Kieling, G. Kstler. Preference SQL - Design, Imple-mentation , Experiences. In Proc. of the Int. Conf. on Very Large Databases (VLDB'02), Hong Kong, China, 2002. [14] A. Magkanaraki, V. Tannen, V. Christophides, D. Plexousakis. Viewing the Semantic Web Through RVL Lenses. In Proc. of the Int. Semantic Web Conf. (ISWC'03), LNCS 2870, Sanibel Island, USA, 2003. [15] E. M. Maximilien, M.Singh. Conceptual Model of Web Service Reputation. In SIGMOD Records 31(4), 2002. [16] M. McGeachie, J. Doyle. Efficient Utility Functions for Ce-teris Paribus Preferences. In Proc. of Conf. on Artificial Intelligence and Conf. on Innovative Applications of Artificial Intelligence (AAAI/IAAI'02), Edmonton, Canada, 2002. [17] G. Medvinsky, C. Lai, B. Neuman. Endorsements, Licensing , and Insurance for Distributed System Services. In Proc. of the ACM Conf. on Computer and Communications Security , Fairfax, USA, 1994 [18] E. Mena, V. Kashyap, A. Illarramendi, A. Sheth. Imprecise Answers in Distributed Environments: Estimation of Information Loss for Multi-Ontology based Query Processing. In International Journal of Cooperative Information Systems (IJCIS), 9 (4), 2000. [19] NTT DoCoMo home page. http://www.nttdocomo.com/home.html, 2003. [20] M. Paolucci, T. Kawamura, T. Payne, K. Sycara. Importing the Semantic Web in UDDI. In Proc. of the Int. Workshop on Web Services, e-Business and the Semantic Web (WES'02), Toronto, Canada, 2002 [21] UDDI. The UDDI Technical White Paper. http://www.uddi.org. 205

selection of the most useful;Web Service Definition Language;Web services;Tree-shaped clipping of ontologies;subsequent execution;Semantic Web;user profiling;The generalization throughout ontologies;ontology resembles common knowledge;Universal Description Discovery and Integration;discovery of possible services;generalization hierarchy of concepts;cooperative service discovery;personalization;preference-based service provisioning;Domain-specific understanding of concepts;Relaxing multiple ontologies

194

Topic Modeling in Fringe Word Prediction for AAC

Word prediction can be used for enhancing the communication ability of persons with speech and language impair-ments . In this work, we explore two methods of adapting a language model to the topic of conversation, and apply these methods to the prediction of fringe words.

INTRODUCTION Alternative and Augmentative Communication (AAC) is the field of research concerned with finding ways to help those with speech difficulties communicate more easily and completely. Today there are approximately 2 million people in the United States with some form of communication difficulty. One means to help ease communication is the use of an electronic communication device, which may have synthetic speech as output. However, one issue in using an AAC device is communication rate. Whereas speaking rate is estimated at 180 words per minute (wpm), many AAC users' communication rates are lower than 15 wpm [3, 7, 16]. Thus one goal of developers is to find ways to increase the rate of communication, by making AAC devices easier to use and more intelligent. Some researchers have attempted to speed communication rate by providing quick access to the core vocabulary the relatively small set of frequently used words. Methods for doing this include abbreviation expansion and iconic methods such as semantic compaction [1]. In contrast, in this work we attempt to speed access to the much larger set of words often called fringe vocabulary. This set is of interest because although each individual word occurs less frequently, the set of fringe words on the whole is very significant . Suppose that the user wants to enter "I want a home in the country." After typing, "I want a h", they might see something like shown below. The system has created a prediction window containing the five words that it thinks the user may be trying to type. In this example, the user can press F5 to complete the word "home" and the system will enter the word with a space afterwards. So in this example, the user needed 2 keystrokes to enter what would normally take 5 keystrokes. It is difficult to judge how much word prediction can speed communication rate. Much of this determination is dependent on the accuracy of the prediction method, the characteristics of the user, such as their physical and cognitive abilities, and the characteristics of the user interface, such as where the prediction list is displayed and how a word in the list is selected. Here, the prediction method is evaluated separately from the rest of a word prediction system by simulating what a user would type in a conversation if he/she were taking full advantage of the prediction list. This theoretical evaluation measures the percentage of keystrokes that were saved by word prediction over typing out every character. In this paper we first describe related work and give some background in statistical approaches to word prediction. We present approaches to topic modeling and compare the results of topic modeling to a baseline method. For a more thorough account of this work, visit http://www.cis.udel.edu/fringe/. RELATED WORK Several previous researchers have used n-gram models in word prediction for AAC [4, 5, 12, 18]. For example, Lesher et al. [12] show how increasing training set size and unigrams to bigrams (going from 47% to 54.7%) to trigrams (another .8%). These evaluations used a window size of 6. Other researchers have integrated grammatical information into n-gram word prediction systems. Garay-Vitoria and Gonzalez-Abascal [10] integrated a statistical chart parser, while Fazly and Hirst [8] and Copestake [7] used part-of-speech (POS) tagging. These yielded improvements of 1-5% keystroke savings. There have been several attempts at topic modeling in the language modeling community, particularly for speech recognition [2, 14, 17, 6, 9, 13]. Some of the evaluations of topic modeling have found different variants of it to be very beneficial [2, 14, 9]. Lesher and Rinkus [13] is an attempt at topic modeling for word prediction, but does not use dynamic topic modeling like [9, 2] and this work. Table 1: The keystroke savings of topic modeling is shown compared to a bigram and trigram baseline. METHODS Like several of the aforementioned word prediction researchers , we use n-gram methods for language modeling. Our baseline word prediction methods use bigram and trigram-based n-gram models with backoff with Good-Turing smoothing , the current best practice in statistical language modeling according to Manning and Sch utze [15]. Additionally, we incorporate a special unigram model for the first word of each sentence. In word prediction, these language models our used to rank all the words that the user could possibly be typing. The top W words are presented to the user, where W is the prediction window size. Statistical approaches require a collection of text to construct a language model. Ideally, our corpus would be a large collection of conversations involving one or more people using an AAC system. Such a corpus is unavailable, so we follow [13] in using the Switchboard corpus, which is a collection of telephone conversations and their transcriptions. 1 The training section contains a randomly pre-selected 2217 conversations and the testing section contains the remaining 221 conversations. We perform preprocessing to remove some speech repairs in accordance with Hindle [11]. These editing rules bring the Switchboard conversations closer to what we envision an AAC user would type. 3.1 Evaluation We compare the number of keystrokes required for a user taking full advantage of our word prediction system to the number of keystrokes required to enter each character of the conversation. We use immediate prediction for our evaluations , which allows use of the prediction list before the first character of a word has been entered. We assume that one keystroke is required to "speak" each turn of input and that a space is automatically inserted after a word is selected from the prediction list. KS = keys normal - keys withprediction keys normal 100% Because we are interested in the prediction of fringe words, our evaluations are measured on fringe words only. Core words are excluded from the list of predictions. The particular core vocabulary we chose is available from the AAC Centers at the University of Nebraska at Lincoln, available from http://aac.unl.edu/. We used the "Young Adult Conversation Regular" core vocabulary list, as it is the most similar to the type of conversations in the Switchboard corpus . 1 The Switchboard transcriptions were available from http://www.isip.msstate.edu/projects/switchboard/ TOPIC MODELING The goal of topic modeling is to identify the current topic of conversation, then increase the probability of related words and decrease the probability of unrelated words. Some words will be unaffected by topic modeling, such as function words, which are used similarly in all topics. It is for this reason that we chose to improve fringe word prediction with topic modeling: we feel that topic modeling specifically improves fringe word prediction. Researchers are consistent in representing a topic by creating a collection of representative text of the topic. However, researchers differ on the best way to organize a collection of topics. Some researchers have created a hierarchical collection of topics [9], while others have created a disjoint set of topics [14, 2, 17]. We feel that the primary lure of a hierarchical approach, the ability to generalize, can be captured in the set approach as well, by giving varying weight to all topics and not just the most likely topic. For this reason, we represent topics as disjoint sets of conversations. The current topic of conversation must be identified from the part of the conversation that has taken place so far, and updated periodically in the conversation. Thus, we must devise a representation for a partial conversation for assessing the similarity of the conversation to each topic. In representing the conversation so far, we choose to implement an exponentially decayed cache, like [2], using TF-IDF values rather than raw frequencies. This follows the work of Mahajan et. al. [14] in considering the inverse document frequency of a word to proportional to its utility in identifying the current topic. Because our approach is for topic identification, we ignore words that occur in 85% or more of the topics, with the intuition that such words are irrelevant to selection of topic. As a step to convert our model of the current conversation to a model of the current topic, we compute the document similarity between the cache and the unigram model for each topic. We chose to use the cosine metric, following [9]. Given that we have computed a similarity score between each topic and the current conversation, there are two main variations on how to construct a new language model. Mahajan et. al. [14] implemented a k-nearest solution, constructing the topic model from the most similar k topics. Each topic's language model was weighted equally for their experiments. Instead, we chose to follow Florian and Yarowsky's approach [9]. They expand the probability for a word (w) given a history (h) as follows: P (w | h) = X ttopics P (t | h) P (w | t, h) P (w | t, h) is simply the probability of w taken from the language model constructed for topic t. The probability of the topic is estimated as follows: P (t | h) S (t, h) P t topics S (t , h ) where S(t, h) is the cosine similarity of the topic to the current part of the conversation. 4.1 Method A Our first method of topic modeling is most similar in spirit to the work of Mahajan et. al. [14] and Florian and Yarowsky [9]. In training, a bigram model is computed for 277 each topic in Switchboard. In testing, the cache representation of the current conversation is compared against the unigram representation of each topic and similarity scores are computed. The similarity scores are then used to weight the frequencies obtained from each topic in a linear interpolation . Then this interpolated bigram model is used to compute the probabilities used for word prediction. Topic modeling shows a sizable improvement over the the bigram baseline: 1.6% 1.7%. We've included the comparison to a bigram baseline because it is the most natural baseline in terms of language understanding. However, a trigram baseline is also a natural comparison when considering that it can run with the same or less computational resources than topic modeling. When compared against the trigram baseline, the topic model gives 0.8% 1.5% improvement. 4.2 Method B Our second method of topic modeling is more similar to the work of Bellegarda [2]. Like Bellegarda, we compute topic-dependent unigram probabilities. These topic-dependent probabilities are multiplied with probabilities from a trigram backoff model. Additionally, we weight the topic component with a tuning parameter. After manual tuning on a two conversations, we found that = .15 worked well. Method B is an improvement over a trigram baseline, but only a minor improvement. We feel that the problem is that a low value was necessary to avoid overriding the word preference that is due to context, but that it also reduced the ability of the overall model to adapt to a particular topic. 4.3 Comparison Method A offers an additional 1% or more keystroke savings over Method B for most window sizes. This is due to the low weight of the tuning parameter for Method B. However , as previously mentioned, the low weight was necessary. Additionally, notice that Method A becomes comparatively better as the window size is increased. The trigram model component in Method B can be thought of as a stronger source of knowledge than the interpolated bigram model of Method A. Because of this, when the trigram history exists in the language model, Method B's predictions are more accurate . However, because the trigram model is sparse, it can only contribute to the top few predictions. Thus, it has a much greater effect on the top few window sizes. For real world systems, however, absolute performance is not the only factor. The computational demands of each approach are often considered when selecting a practical solution . The trigram baseline processed at 1,325 words per minute (wpm). Method A processed conversations in testing at 32 wpm and Method B processed 1,267 words per minute. Method B uses barely more processing time than the trigram baseline model. CONCLUSIONS Topic modeling can be implemented in many different ways. We've demonstrated two such methods for topic modeling : one for computationally limited devices and another for computationally rich devices. Both methods show a clear improvement over a trigram model with backoff. Before the advent of word prediction, a user would've pressed 6.4 keys per fringe word on average. Now, with topic modeling for word prediction, only 2.5 keys per word are required. ACKNOWLEDGMENTS We would like to thank the US Department of Education for funding this research under grant H113G040051, under the National Institute on Disability and Rehabilitation Research program. We would also like to thank Dr. Gregory Lesher for correspondence regarding his work and Dr. David Saunders for lending us a compute server. REFERENCES [1] B. Baker. Minspeak. Byte, pages 186202, 1982. [2] J. Bellegarda. Large vocabulary speech recognition with multispan language models. IEEE Trans. On Speech and Audio Processing , 8(1), 2000. [3] D. R. Beukelman and P. Mirenda. Augmentative and alternative communication: Management of severe communication disorders in children and adults . P.H. Brookes Pub. Co., 1998. [4] L. Boggess. Two simple prediction algorithms to facilitate text production. In Proceedings of the second conference on Applied natural language processing , pages 3340, Morristown, NJ, USA, 1988. Association for Computational Linguistics. [5] A. Carlberger, J. Carlberger, T. Magnuson, M. S. Hunnicutt, S. Palazuelos-Cagigas, and S. A. Navarro. Profet, a new generation of word prediction: An evaluation study. In Proceedings of Natural Language Processing for Communication Aids , 1997. [6] S. Chen, K. Seymore, and R. Rosenfeld. Topic adaptation for language modeling using unnormalized exponential models. In Proc. Int'l Conf. on Acoustics, Speech and Signal Processing , 1998. [7] A. Copestake. Augmented and alternative nlp techniques for augmentative and alternative communication. In Proceedings of the ACL workshop on Natural Language Processing for Communication Aids , 1997. [8] A. Fazly and G. Hirst. Testing the efficacy of part-of-speech information in word completion. In Proceedings of the 10th Conference of the European Chapter of the Association for Computational Linguistics , 2003. [9] R. Florian and D. Yarowsky. Dynamic nonlocal language modeling via hierarchical topic-based adaptation. In Proceedings of ACL'99 , pages 167174, 1999. [10] N. Garay-Vitoria and J. Gonz alez-Abascal. Intelligent word-prediction to enhance text input rate. In Proceedings of the second international conference on Intelligent User Interfaces , 1997. [11] D. Hindle. Deterministic parsing of syntactic non-fluencies. In Proceedings of the 21st Annual Meeting of the Association for Computational Linguistics , 1983. [12] G. Lesher, B. Moulton, and J. Higgonbotham. Effects of ngram order and training text size on word prediction. In Proceedings of the RESNA '99 Annual Conference , 1999. [13] G. Lesher and G. Rinkus. Domain-specific word prediction for augmentative communication. In Proceedings of the RESNA '01 Annual Conference , 2001. [14] M. Mahajan, D. Beeferman, and X. D. Huang. Improved topic-dependent language modeling using information retrieval techniques. In Proceedings of the International Conference on Acoustics, Speech, and Signal Processing , 1999. [15] C. Manning and H. Sch utze. Foundations of Statistical Natural Language Processing . MIT Press, 2000. [16] A. Newell, S. Langer, and M. Hickey. The r^ ole of natural language processing in alternative and augmentative communication. Natural Language Engineering, 4(1):116, 1996. [17] K. Seymore and R. Rosenfeld. Using story topics for language model adaptation. In Proceedings of Eurospeech '97, pages 19871990, Rhodes, Greece, 1997. [18] A. L. Swiffin, J. A. Pickering, J. L. Arnott, and A. F. Newell. Pal: An effort efficient portable communication aid and keyboard emulator. In Proceedings of the 8th Annual Conference on Rehabilitation Techonology , pages 197199, 1985. 278

core vocabulary;identify current topic of conversation;AAC;language modeling;accuracy of prediction method;fringe vocabulary;prediction of fringe words;conversations in the Switchboard corpus;Word prediction;immediate prediction;decrease probability of unrelated words;increase probability of related words;prediction window size;communication rate;construct a new language model;topic modeling

195

Topic Transition Detection Using Hierarchical Hidden Markov and Semi-Markov Models

In this paper we introduce a probabilistic framework to exploit hierarchy, structure sharing and duration information for topic transition detection in videos. Our probabilistic detection framework is a combination of a shot classification step and a detection phase using hierarchical probabilistic models. We consider two models in this paper: the extended Hierarchical Hidden Markov Model (HHMM) and the Coxian Switching Hidden semi-Markov Model (S-HSMM) because they allow the natural decomposition of semantics in videos, including shared structures, to be modeled directly, and thus enabling efficient inference and reducing the sample complexity in learning. Additionally, the S-HSMM allows the duration information to be incorporated, consequently the modeling of long-term dependencies in videos is enriched through both hierarchical and duration modeling . Furthermore, the use of the Coxian distribution in the S-HSMM makes it tractable to deal with long sequences in video. Our experimentation of the proposed framework on twelve educational and training videos shows that both models outperform the baseline cases (flat HMM and HSMM) and performances reported in earlier work in topic detection . The superior performance of the S-HSMM over the HHMM verifies our belief that duration information is an important factor in video content modeling.

INTRODUCTION The ultimate goal of the video segmentation problem is to characterize the temporal dynamics of the video whereby it can be segmented into coherent units, possibly at different levels of abstraction. Seeking abstract units to move beyond the shots has thus been an active topic of much recent research. While the problem of shot transition is largely solved at a satisfactory level [7], the `abstract units' or scene detection problem is much harder, partially due to the following three challenges identified in [29]: (a) the variety in directional styles, (b) the semantic relationship of neighbouring scenes, and (c) the knowledge of the viewer about the world. While the last aspect is beyond the scope of this work, the first two clearly imply that effective modeling of high-level semantics requires the domain knowledge (directional style) and the modeling of long-term, multiple-scale correlations of the video dynamics (neighboring semantic relationship). Modeling temporal correlations over a long period is generally a challenging problem. As we shall review in the subsequent section, this problem is usually solved in a specific domain setting so that the expert knowledge about the domain can be utilised. While organization of content in generic videos (e.g., movies) is too diverse to be fully char-acterized by statistical models, the hierarchy of semantic structure in the class of education-oriented videos is more defined, exposing strong temporal correlation in time, and thus make it more desirable to probabilistic modeling. In this paper, we concentrate on this video genre and develop an effective framework to segment these videos into topi-cally correlated units. This problem is an important step to enable abstraction, summarization, and browsing of educational content a rich class of film genre that has an increasing important role in building e-services for learning and training. Probabilistic modeling of temporal correlations in video data is however a difficult problem. It is complicated because the underlying semantics naturally possess a hierarchical decomposition with possible existence of tight structure sharing between high-level semantics. In addition, the typical duration for these structures usually varies for each of its higher semantic. As an example, assisted narration a section that involves the narrator talking to the audience is usually used in both the introduction and in the main body of a topic in an educational video. However while one, or rarely two, shots of assisted narration (AN) are considered sufficient for the introduction, the body typically requires many AN shots. Thus it is important to exploit and fuse hierarchical decomposition, structure sharing and duration information in a unified framework to effectively address the problem of topic transition detection. The most widely used pobabilistic model is the hidden Markov model (HMM). However, in many cases, the HMM is unsuitable for video analysis since the strong Markov assumption makes it too restrictive to capture correlations 11 over long periods. This limitation is usually overcome in the literature by the use of a series of HMMs in a hierarchic manner . The underlying problem in these approaches still is the manual combination of HMMs at the higher levels which results in the excessive expense of preparing the training data and, more importantly, the interaction across higher semantic levels is not incorporated during model training. One rigorous approach to overcome this limitation is the use of the Hierarchical Hidden Markov Model (HHMM), first introduced in [6] and later extended to handle structure sharing in [3]. The sophisticated model in [3] allows natural hierarchical organization of the videos, including any existing structure sharing, to be modeled rigorously. Practically this will result in computational savings and a reduction in sample complexity for learning. Given its advantages, we use this model in this paper to model educational video content for topic transition detection. It is natural to see that durative properties play an important role in human perception. An excessively long lecture would bore the students. As such, education-oriented videos (e.g., news, documentaries, lectures, training videos, etc.) exhibit strong duration information in their content. We thus propose an alternative approach towards handling temporal dependencies over long periods through the explicit modeling of duration information captured in semi-Markov models. In these models, a state is assumed to remain un-changed for some duration of time before it transits to a new state, and thus it addresses the violation of the strong Markov assumption from having states whose duration distributions are non-geometric. Existing semi-Markov models commonly model duration distributions as multinomials. Video data is however typically very long, thus making a multinomial semi-Markov model unsuitable for video analysis since it would result in both excessive computation and the number of parameters required. Continuous modeling of duration does exist such as in the use of the Gamma distribution, or more generally the exponential family, described in [12, 16] to provide more compact parameterization. However, there are still two limitations applied to these models for video analysis: (a) learning these distributions requires numerical optimiza-tion and the time complexity still depends on the maximum duration length, and (b) no hierarchical modeling has been attempted. Fortunately, in [5], a Switching Hidden Semi-Markov Model (S-HSMM) is introduced in which the duration is modeled as a discrete M -phase Coxian distribution . This model is particularly interesting for video analysis since: (1) it can model hierarchical decomposition, and (2) the Coxian duration modeling results in fast learning and inference, the number of parameters is small and close-formed estimation exists. Parameterizing long-term temporal correlations existing in video is thus enriched by both the hierarchical architecture and the duration modeling at the bottom level of the S-HSMM. To model video content, we argue that it is beneficial to exploit both the hierarchical organization of the videos, their semantically shared substructures and typical durations of important semantics. These aspects are all addressed in this paper in a unified and coherent probabilistic framework . We use the HHMM and the S-HSMM and propose a two-phase architecture for detecting topical transition in educational videos. In the first phase, shots are classified into meaningful labels. Using classified shot labels, the second phase trains a hierarchical probabilistic model (HHMM or S-HSMM) which is then used at a later stage for segmentation and annotation. Prior knowledge about the domain, including shared structures, is incorporated into the topo-logical structure during training. Our cross-validation on a dataset including a mix of twelve videos demonstrates promising results. The performances from the baseline cases (HMM and HSMM) have shown that they are too restrictive and unsuitable in our detection scheme, proving the validity of hierarchical modeling. The performances of the hierarchical models, including the HHMM and S-HSMM, are shown to surpass all results reported in earlier work in topic detection [23, 20, 4]. The superior performance of the S-HSMM over the HHMM has also demonstrated our belief that duration information is indeed an important element in the segmentation problem. Exploiting the hierarchy, structure sharing and duration in a unified probabilistic framework, our contributions are twofold: (1) we provide a coherent hierarchical probabilistic framework for topic detection. Although the current report concentrates on the educational genre, this framework can clearly generalize to other genres such as news and documentaries , and (2) to our knowledge we are the first to investigate duration and hierarchical modeling for video segmentation 1 in a unified framework. The remainder of this paper is organized as follows. In the next section, we provide related background to this work. This is followed by a detailed section on the detection framework including the description of the HHMM and S-HSMM. We detail the shot classification phase in Section 4. Experimental results are then reported in Section 5. Finally, the conclusion follows in Section 6. RELATED BACKGROUND Seeking high-level semantics to move beyond the shots has been the central theme of much recent research. Attempts towards this problem have resulted in a fast growing body of work, and depending on the investigating domain, the abstracting units appear under different names such as scene, story, episode for motion pictures; topic, subtopic, macro segments, story units for information-oriented videos (news, documentaries, training and educational videos), or general term like logical story units used in [8, 32]. Otherwise stated, we shall the term `scene' in this section to mean all of the aforementioned names. Early attempts have targeted extracting scene-level concepts in broadcast programs, in particular news videos (e.g., [9, 14, 26]). In these attempts, the semantic extraction problem is usually cast as the classification problem. The authors in [26], for example, combine a number of visual and aural low-level features together with shot syntax presented in news videos to group shots into different narrative structures and label them as anchor-shot, voice-over, or inter-1 Since topic change coincides with a shot transition, the shot boundary provides crucial information in detecting topic transitions, therefore the term `duration' in this work is calculated in terms of the number of shots. This drastically simplifies the modeling process. An alternative way of modeling duration is to uniformly replicate a shot label based on its length. However, doing this would require an extra modeling of shot transition knowledge. In this work, we avoid this complication and concentrate on duration information based on the shot counts. 12 view. Liu et al. [14] propose a video/audio fusion approach to segment news reports from other categories in broadcast programs with different types of classifiers (simple threshold method, Gaussian mixture classifier, and support vector machine ). Ide et al. [9] propose an automatic indexing scheme for news video where shots are indexed based on the image content and keywords into five categories: speech/report, anchor, walking, gathering, and computer graphics. Caption text information is then used with classified shots to build the indices. Segmentation of the news story is the second major theme explored in the broadcast domain. The common underlying approach used in these works is the use of explicit `rules' about the structure of news programs to locate the transitions of a news story. Commonly accepted heuristics are for example: a news story often starts and finishes with anchor-person shots [31]; the start of a news story is usually coupled with music [2]; or a relative long silence period is the indication of the boundary between two news stories [33]. More complicated rules via temporal analysis are also exploited such as the work of [37] which utilises detection results of anchor-persons and captions to form a richer set of rules (i.e., if the same text caption appears in two consecutive anchor-person shots, then they belong to the same news story). There is also a body of work which casts the segmentation problem of news story in a HMM framework [10, 4]. The authors in [10], for example, propose the news segmentation problem as problem of decoding the maximum state sequence of a trained HMM whose transition matrix is tailored by explicit rules about the news program. A somewhat similar approach to the work in this paper is [4] (whose results came first in the TRECVID2003 story segmentation benchmark). Shots in [4] are first classified into a set common labels in news (e.g., anchor, 2anchor, text-scene, etc.). These labels are then input to a HMM for the segmentation task. They report best performances of 74.9% recall and 80.2% precision for the TRECVID dataset. The work of [4] however remains limited due to the flat structure HMM, and it is not clear how the set of `transition' states were chosen. In an effort to move beyond flat structure, the authors of [4] have raised the need for high-order statistical techniques, which will be addressed in this paper through the HHMM and S-HSMM. More recent approaches towards scene extraction have shifted to motion pictures (e.g., [30, 34, 1, 31]). Detecting scenes in motion pictures is in general a challenging problem and there are three main existing approaches as outlined in [31]: temporal clustering-based, rule-based and memory-based detection. In the clustering-based approach, shots are grouped into scenes based on visual similarity and temporal closeness (e.g., [8, 13]). Scene breaks in the rule-based detection approach are determined based on the semantic and syntactic analysis of audiovisual characteristics and in some cases further enhanced with more rigorous grammars from film theory (e.g., [34, 1]). The authors in [30] propose a memory-based scene detection framework. Visual shot similarity in these works is determined based on the consistency in color chromaticality, and the soundtrack is partitioned into `audio scenes'. Visual and aural data are then fused within a framework of memory and attention span model to find likely scene breaks or singleton events. Further related background on scene detection can be found in many good surveys (e.g., [30, 28, 31]). Existing HMM-based approaches towards modeling long-term temporal dependencies typically use pre-segmented training data at multiple levels, and hierarchically train a pool of HMMs, in which HMMs at the lower levels are used as input to the HMMs at the upper levels. In principle, some fundamental units are recognised by a sequence of HMMs, and then likelihood values (or labels) obtained from these HMMs are combined to form a hierarchy of HMMs 2 to capture the interactions at higher semantic levels (e.g., [11, 18]). Analysing sports videos, Kijak et al. [11] propose a two-tiered classification of tennis videos using two layers of HMMs. At the bottom level, four HMMs are used to model four shot classes (`first missed serve',`rally', `replay', and `break'). Each HMM is trained separately and subse-quently topped up by another HMM at the top level which represents the syntax of the tennis video with three states of the game: `sets', `games', and `points'. Parameters for the top HMM are, however, all manually specified. In [18], a generic two-level hierarchy of HMMs is proposed to detect recurrent events in movies and talk shows. Their idea is to use an ergodic HMM at the top level, in which each state is another (non-ergodic) sub-HMM representing a type of signal stationary properties. For the case of movies, the top HMM has six states, and each is in turn another three-state non-ergodic HMM. The observations are modelled as a mixture of Gaussians. After training, the authors claim that interesting events can be detected such as `explosion', `male speech', and so on. While being able to overcome the limitation of the flat HMM in modeling long-term dependencies , approaches that use HMMs at multiple levels still suffer from two major problems: (1) pre-segmented and an-notated data are needed at all levels for training, and (2) in most existing work parameterization at higher levels has to be manually specified. In many cases, preparing training data at multiple levels is extremely tedious and at worst, may not be possible. With respect to the second problem, since each semantic level has to be modeled separately, the underlying problem is that the interactions across semantic layers are not modeled and thus do not contribute to the learning process. One framework that integrates the semantics across layers is the Hierarchical Hidden Markov Model (HHMM) proposed recently in [6]. The hierarchical HMM extends the standard HMM in a hierarchic manner to allow each state to be recursively generalised as another sub-HMM, and thus enabling the ability to handle hierarchical modeling of complex dynamic processes, in particular "the ability to infer correlated observations over long periods in the observation sequence via the higher levels of hierarchy" [6]. The original motivation in [6] was to seek better modeling of different stochastic levels and length scales presented in language (e.g., speech, handwriting, or text). However, the model introduced in [6] considers only state hierarchies that have tree structures, disallowing the sharing of substructures among the high-level states. Recognizing this need, the authors in [3] have extended the strict tree-form topology in the original HHMMs of [6] and allowed it to be a general lattice structure. The extension thus permits a state at any arbitrary level of the HHMMs to be shared by more than one parental state at its higher level (i.e., resulting in a compact form of parameter typing at multiple levels). This extended 2 Not to be confused with the Hierarchical HMMs. 13 form is very attractive for video content modeling since it allows the natural organization of the video content to be modeled not only in terms of multiple scales but also in terms of shared substructures existing in the decomposition. Further details on the HHMM are provided in Section 3.1. Early application of the HHMM for video analysis is found in [36] and later extended in [35]. In these works, the authors use the HHMM to detect the events of `play' and `break' in soccer videos. For inference and learning, the HHMM is `collapsed' into a flat HMM with a very large product state space, which can then be used in conjunction with the standard forward/backward passes as in a normal HMM. Four methods are compared in [36] to detect `play' and `break': (1) supervised HMMs, in which each category is trained with a separate HMM, (2) supervised HHMMs, in which bottom level HMMs are learned separately and parameters for the upper levels are manually specified, (3) unsupervised HHMMs without model adaptation, and (4) supervised HHMMs with model adaptation. In (3) and (4), two-level HHMMs are used. Their results have shown a very close match between unsupervised and supervised methods in which the completely unsupervised method with model adaptation performs marginally better. These figures are 75.5%, 75.0%, 75.0% and 75.7% respectively for those four methods. While presenting a novel contribution to the feature selection and model selection procedure, the application of the HHMMs in this work is still limited both for learning and for exploitation of the hierarchical structure. Flattening a HHMM into a flat HMM as done in [36, 35] suffers from many drawbacks as criticized in [17]: (a) it cannot provide multi-scale interpretation, (b) it loses modularity since the parameters for the flat HMM get constructed in a complex manner, and (c) it may introduce more parameters, and most importantly it does not have the ability to reuse parameters , in other words parameters for the shared sub-models are not `tied' during the learning, but have to be replicated and thus lose the inherent strength of hierarchical modeling. Being able to model shared structures, the extended HHMMs of [3] allows us to build more compact models, which facilitates more efficient inference and reduces the sample complexity in learning. This model is applied in [20] and [22] for the problem of topic transition detection and video structure discovery respectively. The authors in [20] use a three-level HHMM for the detection of topic transitions in educational videos. Differing from our experiments in this paper , the HHMM in [20] is modified to operate directly with continuous-valued observed data via the use of Gaussian mixture models as the emission probabilities. Each shot-based observed vector consists of seven features extracted from visual and audio streams. They report a 77.3% recall rate and 70.7% precision for the detection task. In another application, with the help of prior knowledge about educational videos, a topology for a three-level HHMM is used in [22] to automatically discover meaningful narrative units in the educational genre. Their experiments have shown encouraging results in which many meaningful structures are hierarchically discovered such as `on-screen narration with texts', `expressive linkage', `expressive voice-over', etc. The work of [22] is somewhat similar to that of [18] reviewed earlier in this section, except the model in [22] allows more domain knowledge to be encoded and the parameters are all learned automatically. THE PROBABILISTIC TOPIC DETECTION FRAMEWORK Our topic detection framework consists of two phases. The first phase performs shot detection and low level feature extraction and then classifies a shot in a meaningful label set . This phase is described in Section 4. In the next phase, we train a HHMM or S-HSMM over the alphabet space from the training data and then use it in conjunction with the Viterbi to perform segmentation and annotation. The architecture of the framework is depicted in Figure-1. 1 2 F E A T U R E E X T R A C T I O N SHOT DETECTION AND CLASSIFICATION Direct Narration Assisted Narration Voice-Over Expressive Linkage Functional Linkage M-phase Coxian M-phase Coxian M-phase Coxian M-phase Coxian M-phase Coxian END `Intro' `main body' 1 2 3 4 5 Video & Audio Signals Figure 1: The architecture for topic detection framework. The two-level HHMM and the S-HSMM (whose topology is shown on the top of Figure-1) are special cases of the hierarchical model with two layers. For the S-HSMM (HHMM), the top layer is a Markov sequence of switching variables, while the bottom layer is a sequence of concate-nated HSMMs (HMMs) whose parameters are determined by the switching variables at the top. Thus, the dynamics and duration parameters of the HSMM (HMM) at the bottom layer are not time invariant, but are `switched' from time to time, similar to the way the linear Gaussian dynamics are switched in a switching Kalman filter. When mapping to the topic modeling problem, the bottom layer is used to capture `atomic' semantics such as voice-over, expressive linkage or assisted narration. Combinations of these atomic semantics then form higher-level semantics, each of which is represented by a hidden state at the top layer in our model. 3.1 The Hierarchical HMM With the assumed knowledge of the flat HMM (e.g., see [24]), we shall now briefly describe the HHMMs. A hierarchical HMM is formally defined by a three-turple , , : a topo-logical structure parameterized by and an emission alphabet space . The topology specifies the model depth D, the state space S d available at each level d, and the parent-children relationship between two consecutive levels. For example, the two-level topology shown on the top of 14 Figure-1 specifies the children set at the bottom level for state 1 is {1, 2, 5} and for state 2 is {2, 3, 4}. Here, state 2 at the bottom level has been `shared' by both state 1 and 2 at the top level. Given , the parameter of the HHMM is specified in the following way. For d &lt; D, p S d and i, j S d+1 are the children of p: d,p i is the initial probability of i given p; A d,p i,j is the transition probability from i to j given the parent p; and A d,p i,end is the probability that state i going to end-state (i.e., returns control to its parent) given the parent is p. Finally, for each state i at the lowest level D and an alphabet v : B v|i is the emission probability of observing v given the current state at the lowest level is i. The whole parameter set is written compactly as: = {, A, A end , B}, where: = [ 1d&lt;D pSd n d,p : 1 M o , B : |S d | || A = [ 1d&lt;D pSd nA d,p : M M o , A end = [ 1d&lt;D pSd nA d,p end : 1 M o where in each each M is implicitly meant the number of children of p and |.| is the cardinality operator. Stochastic constraints require: P i d,p i = 1, P v B v|i = 1 and P j A d,p i,j + A d,p i,end = 1. An intuitive way to view the set is to consider the subset { d,p , A d,p , A d,p end } as the parameter of the p-initiated Markov chain at level d. This chain is terminated when one of the children i of p reaches the end-state with the probability of A d,p i,end . For inference and learning, the HHMM is represented as a dynamic Bayesian network (DBN) and can be learned by the Asymmetric Inside-Outside algorithm in [3] or by the forward/backward passes in [17]. Figure-3 shows on its left the DBN representation of the HHMM with two levels, i.e., D = 2. We refer readers to [6, 17, 3] for further information on the HHMMs. 3.2 The Switching-Hidden Semi Markov Model To provide an intuitive view to the S-HSMM, starting from the description of the HHMMs from the previous section , let us consider the case of a two-layer HHMM (D = 2) defined as follows. The state space is divided into the set of states at the top level Q = S 1 = {1, . . . , |Q |} and states at the bottom level Q = S 2 = {1, . . . , |Q|}. This model is parameterized by = { , A , , A, A end , B}. At the top level, p and A pq are respectively the initial probability and the transition matrix of a Markov chain defined over the state space Q . For each p Q , ch(p) Q is used to denote the set of children of p. As in the case of the extended HHMM in [3], it is possible that different parent states may share common children, i.e., ch(p) ch(q) = for p, q Q . A transition to p at the top level Markov chain will initiate a sub-Markov chain at the lower level over the state space ch(p) parameterized by { p , A p , A p end } where q i and A p ij are the initial and transition probabilities as in the normal HMM setting, and A p i,end is the probability that this chain will terminate after a transition to i. At each time point t, a discrete symbol y t is generated with a probability of B v|i where i is the current state at the bottom level. In the description of this two-level HHMM, the duration d for which a bottom state i remains the same clearly has a geometric distribution parameterized by its non-self-transition probability (1 - A p ii ), i.e., d Geom(1 - A p ii ). In many cases, the geometric distributions are often too restricted to model realistic data. The Switching Hidden Semi-Markov Models (S-HSMMs) proposed in [5] overcomes this restriction and allows the duration d of state i at the bottom level to follow a more general discrete distribution d D p,i d . More precisely, the p-initiated chain at the bottom level is now a semi-Markov sequence parameterized by { p i , A p ij , D p,i d } as opposed to the normal Markov chain in the HHMM case. The authors in [5] consider two families of distributions for modeling the duration: the multinomial and the Coxian. However for the multinomial case, the complexity of the learning algorithm is proportional to the maximum duration length, thus making it unsuitable for the problem of modeling video data which is usually very long in nature. Apart from the disadvantage of assuming a maximum duration , our empirical testing on the multinomial case with the maximum length of 50 has also shown that it is about 20 times slower than its Coxian counterpart reported in this paper, thus making it impractical in our settings. We will therefore omit the multinomial case and will consider exclu-sively the Coxian parameterization in this paper. A discrete M -phase Coxian distribution Cox(; ), parameterized by = { 1 , . . . , M } (P i i = 1) and = { 1 , . . . , M }, is defined as a mixture of P M i=1 i S i where S i (X i + . . . + X M ), in which X i are independent random variables having geometric distributions X i Geom( i ). This distribution is a member of the phase-type distribution family and has the following very appealing interpretation . Let us construct a Markov chain with M + 1 states numbered sequentially with the self transition parameter A ii = 1 i as shown in Figure-2. The first M states rep-1 absorbing state 2 M 1 M 2 M 1 M 2 1 Figure 2: The phase diagram of an M -phase Coxian. resent M phases, while the last is the absorbing state which acts like an end state. The duration of each individual state (phase) i is X i Geom( i ). If we start from state i, the duration of Markov chain before the end state reached is S i = X i + . . . + X M . Thus, Cox(, ) is indeed the distribution of the duration of this constructed Markov chain with as the initial state (phase) distribution. The discrete Coxian is much more flexible than the geometric distribution: its probability mass function is no longer monotonically decreasing and it can have more than one mode. Using the Coxian distribution, the duration for the states at the bottom level in the S-HSMM is modeled as follows. For each p-initiated semi-Markov sequence, the duration of a child state i is distributed according to D p,i d = Cox(d; p,i , p,i ). The parameter p,i and p,i are M -dimensional vectors where M is a fixed number representing the number of phases in the discrete Coxian. It is easy to verify that for M = 1, the model reduces identically to a two-layer HHMM. 15 3.3 Inference and Learning in the S-HSMM For inference and learning, the S-HSMM is represented as a dynamic Bayesian network as shown in Figure-3 and then forward/backward passes are applied to compute the filtering and smoothing distributions required for EM learning. t +1 t z t z t +1 m t m t +1 y t y t +1 x t +1 x t e t +1 e t z t z t +1 y t y t +1 x t +1 x t t t +1 e t e t +1 Figure 3: Two-slice DBN representation of a two-level HHMM (left) and the (Coxian) S-HSMM (right). At each time-slice t, an amalgamated hidden state S t = {z t , t , x t , e t , m t } together with the observation y t are maintained . The top level state is updated via z t and t is a boolean-valued variable set to 1 when the z t -initiated semi-Markov sequence ends at t. At the bottom level, x t is the current child state in the z t -initiated chain, m t represents the current phase of x t and e t is a boolean-valued variable set to 1 when x t reaches the end of its duration. The forward and backward procedures in the general DBN are then used to compute the filtering distribution Pr(S t |y 1:t ) and two smoothing distributions Pr(S t |y 1:T ) and Pr(S t , S t+1 |y 1:T ). With these smoothing distributions, it is sufficient to derive all expected sufficient statistics required during EM learning . The overall complexity for the forward pass (and also for the EM) is O(|Q| 2 |Q | 2 M T ). Further information can be found in [5]. 3.4 Viterbi decoding for segmentation To compute the best sequence state, that is to find: S 1:T = argmax S 1:T Pr(S 1:T |y 1:T ) Viterbi decoding algorithms for the HHMM and S-HSMM are developed. These algorithms are similar to the one used in the standard HMM outlined in [24] except we replace the normal state in the HMM setting by our amalgamated state S t which {z t , x t , t , m t , e t } for the S-HSMM and {z t , x t , t , e t } for the HHMM (cf. Figure-3). SHOT-BASED SEMANTIC CLASSIFICATION In this section, we detail the first phase in the detection framework. This includes the formulation of an alphabet set for shot labeling, low-level feature extraction and shot classification. 4.1 Shot labels set: Existing work on the educational videos analysis (e.g., [21, 19]) has studied the nature of this genre carefully. As noted in [21], the axiomatic distinction of the educational genre is in its purpose of teaching and training; and as such a well-crafted segment that moves viewers to actions or retains a long-lasting message requires elaborative directing skills 3 . Based on a narrative analysis used in the educational domain and observed rules and conventions in the production of this media, the authors in [21] propose a hierarchy of narrative structures at the shot level as shown in Figure-4. In this paper, we select the five most meaningful structures from this hierarchy for experimentation. This set includes: direct-narration (DN), assisted-narration (AN), voice-over (VO), expressive-linkage (EL), and functional-linkage (FL). We shall now briefly describe these narratives. Direct-narration (DN) and assisted-narration (AN) are re-ferred to jointly as on-screen narration, which refer to the segments with the appearance of the narrator. The purpose of these sections is to speak to the viewers with the voice of authority, and is commonly used to demarcate a new topic or subtopic, to clarify a concept or to lead the viewers through a procedure with examples. DN is a more strict form of on-screen narration. It involves eye-to-eye contact where the narrator speaks to the viewers directly. An analogy from news video is the anchor-shot. AN refers to parts of the video when a narrator appears in a more diverse style, and the attention of the viewers is not necessarily focused on him or her. Here, the purpose is not only to talk to the viewers, but also to emphasize a message by means of text captions and/or to convey an experience via background scenes. A similar structure from news for AN is the reporting shot. Assisted narration can be used both in the introduction of a topic or in the main body, and thus this structure should be shared 4 by both higher semantics `introduction' and `main body'. As we see later, this knowledge is explicitly modeled and incorporated in the design of the topology for the S-HSMM. An important feature is that although the semantics of AN is shared, the typical durations are different when it is used in the introduction or the main body respectively. An AN section used to demarcate a new topic usually contains only one, and sometimes two shots, while an AN section used in the main body is typically long, spanning a number of shots. Conditioning on the parent (i.e., introduction or main body), the typical duration distribution of the AN section is learned automatically for each case by our model. The voice-over (VO) structure is identified as sections where the audiotrack is dominated by the voice of the narrator , but without his or her appearance. The purpose of these segments is to communicate with the viewers via the narrator's voice. Additional pictorial illustration is usually further shown in the visual channel. Expressive linkage (EL) and Functional linkage (FL) belong to the same broader linkage group in the hierarchy in Figure-4. The purpose of the linkage structure is to maintain the continuity of a story line but there is neither on-screen nor voice-over narration involved. Functional linkage contains transition shots encountered in switching from one subject to the next. Usually, large superimposed text captions are used and the voice narration is completely stopped 3 We note that the two closest video genre to educational videos is news and documentaries. In the description of what follows on educational genre, we can spot several similarities across these genre. 4 In terms of parameterization, it is a form of parameter tying . 16 Linkage Narration On-screen Narration S u p p o r t i v e N a r r a t i o n N a r r a t i o n V o i c e O v e r vo educational videos on lk f u lk e x lk Ex pre ssi ve Lin kag e Fu nct ion al Lin kag e sn Supportive Narration Di r ect i on Na r rat i on an w t dn a n w s vo w. Texts/Scenes vo w ith Sc enes v o t s vo w s vo w t Ass Na rrw .S cen es As sN arr w. Te xts vo with Texts Figure 4: The hierarchy of narrative structures in educational videos proposed in [21]. with possibly music played in the background. Expressive linkage, on the other hand, is used to create `mood' for the subject being presented. For example, in the video presenting the fire safety topic, there is a segment in which the narration is completely stopped and then a sequence of pictures of the house on fire is shown. These scenes obviously do not give any direct instruction, rather they create a sense of `mood' that helps the video to be more appealing and interesting . 4.2 Feature extraction and shot classification The feature set and method for shot classification described in [21] is employed in this paper. The feature set is extracted from both visual and audio streams at the shot-based level. From the image sequence, we choose to detect the frontal faces to reflect the appearance of the narrator using the CMU face detection algoritm [25]; and captioned texts as one of the common means of conveying information in educational videos using the algorithm described in [27]. In order to classify a shot into direct-narration, voice-over, linkage, etc., further information is sought from the audio stream. Audio features are computed as the percentage of the following audio classes within a shot: vocal speech, music , silence, and non-literal sound. A shot is then classified into one of the elements of = {DN, AN, V O, EL, F L} using the classification framework reported in [21]. Since we claim no contribution at this stage, we shall refer readers to [21] for full details on this classification scheme. EXPERIMENTAL RESULTS Our dataset D consists of 12 educational and training videos containing different types of subjects and presentational styles, and thus this constitutes a relatively noisy set of data. We manually provide groundtruth for these videos with topic transitions. In some cases, the groundtruth for topic transitions comes directly from the hardcopy guidelines supplied by the producer. At the pre-processing stage, Webflix [15] is used to perform shot transition detection and all detection errors are corrected manually. Since our contribution from this paper is at the semantic level, the latter step is to ensure an error at the shot detection does not influence the performance of the system at higher levels. Since educational videos mainly contain cut and dissolve transitions, the shot detection accuracy is found to be very high with rare cases being erroneous. Given shot indices, each video is processed as described in Section 4, and then each shot S is labeled as one of the elements of = {DN, AN, V O, EL, F L}. 5.2 Model topology and parameterization We will use four models in this experiments: the flat HMM and HSMM (as the baseline cases), the HHMM and the S-HSMM . For the flat HMM and HSMM, we range the number of states from 2 to 5 with the observation space , where 2 is intended to be the minimum number of states required (like `intro' and `main body') and 5 is the number of alphabets (i.e., in the relaxed way that the number of states equates to the number of alphabets). The semi-Markov version HSMM is further parameterized by 3-phase Coxian distributions as the duration distributions of the states. The choice of M = 3 phases is hinted by the results reported in [5] where M = 3 has resulted in best performances. For the HHMM and the S-HSMM, the topology shown in the top of Figure-1 is used to construct the S-HSMM in this experiment. This topology specifies Q = 2 states at the top level where state 1 and 2 correspond to the introduction and the main body of the topic respectively. The Markov chain at this level is similar to the flat HMM used in [4] for news story segmentation 5 reviewed in Section 2. We incorporate the assumed prior knowledge that a topic usually starts with either direct-narration, assisted-narration or functional linkage , thus state 1 has {1, 2, 5} as its children set. Similarly, the main body can contain assisted-narration, voice-over or expressive linkage, hence its children set is {2, 3, 4}. Here state 2 (assisted narration) has been shared by both parent state 1 (`intro') and 2 (`main body'). The bottom level has 5 states corresponding to 5 shot labels. To map the labels to the bottom states, we construct a diagonal-like B observation matrix and fix it, i.e., we do not learn B. The diagonal entries of B are set to 0.99 to relax the uncertainty during the classification stage. The duration models in the S-HSMM are used with M = 3 phases Coxian. 5.3 Detection Results Given the dataset D, our evaluation employs a leave-one-out strategy to ensure an objective cross-validation. We sequentially pick out a video V and use the remainder set {D \ V } to train the model, and then use V for testing. In the results that follow, this method is used for all cases including the flat HMM, the flat HSMM, hierarchical HMM, and the S-HSMM. A topic transition is detected when the introduction state at the top level is reached during the Viterbi decoding. Examples of Viterbi decoding with the S-HSMM and HHMM are shown in Figure-5. To measure the performance, in addition to the well-known 5 They called `transition' and `internal' states instead of `introduction' and `main body'. 17 recall (recall) and precision (prec) metrics, we include the F-score (f-score) metric defined as: f-score = 2 recall prec recall + prec = 2 1 recall + 1 prec -1 While the recall rate measures how well the system can recover the true topic transitions, and high precision ensures that it does not over-segment the video, the F-score shows the overall performance of the system. In the ideal case when recall=prec=100%, clearly f-score = 1, i.e., the highest performance the system can achieve. The baseline cases: flat HMM and HSMM Since initialization is crucial during EM learning, we apply multiple random restart points when conducting the experiments , including the uniform initialization. Although several restarts were used, the flat HMM is found to yield extremely poor results in all cases. Even when we train and test on the same dataset, the flat HMM still produces poor detection results, proving to be unsuitable in our topical transition detection settings. The flat HSMM produces slightly better results than the flat HMM, but still in all ten runs, the performance is still very low (recall= 7.74% and prec= 48% in the best case). The poor performance of the HMM and HSMM is of no surprise, since their forms are too strict to model a rather high concept - the `topic'. Furthermore, with the flat structures , they offer no mechanism to incorporate prior domain knowledge such as those that we use in the topology of the S-HSMM and HHMM. This clearly shows that hierarchical models are much more suitable for video analysis than the flat ones. Given the poor results in the flat structure cases, we will omit the HMM and HSMM in the discussion of what follows below. Detection with the S-HSMM and HHMM The recall rate, precision and F-score for representative runs are reported in Table 1, in which the best performance are highlighted in bold. The detection results for each individual video for the best cases are shown in Table 2. With different random restarting points, including the uniform initialization , the performance of the HHMM ranges from poor to very good (41.29% 83.23% for recall and 80.00% 84.47% for precision), whereas the S-HSMM consistently yields good results (83.87% 84.52% for recall and 87.92% 88.51% for precision). Since during training there is nothing exposed to the testing examples, we also report (in the second part of Table 1) the performances of the HHMM and S-HSMM in a likelihood-based `best model selection' scheme. This scheme works as follows. As in the leave-one-out strategy, let V be a video selected from D, and N is the number of times we train the model using the dataset {D \ V } (i.e., without V ). Let i (V ) and L i (V ) (i = 1 . . . N ) respectively be the learned model and the likelihood (at convergence) obtained for i-th run. We then use the model i to test on the unseen video V where i = argmax i=1...N L i (V ). Simply speaking , we sequentially `throw away' a video V , then select the best model (i.e., highest likelihood) among all runs to test on V . For the HHMM, the result stays the same as when we choose the best performance based on the F-score. For the S-HSMM, the recall stays the same, while the precision slightly decreases from 88.51% to 87.92%. Nevertheless, the S-HSMM is still superior to the HHMM. recall (%) prec (%) f-score results for best performance selection Uniform 42.58 81.48 0.559 Rand. 1 83.23 84.47 0.840 HHMM Rand. 2 83.23 84.87 0.840 Rand. 3 83.23 84.87 0.840 Rand. 3 41.29 80.00 0.545 Rand. 4 83.87 83.87 0.839 Uniform 84.52 87.92 0.862 Rand. 1 84.52 88.51 0.865 S-HSMM Rand. 2 83.87 87.25 0.855 Rand. 3 84.52 88.51 0.865 Rand. 4 83.87 87.25 0.855 Rand. 5 84.52 88.51 0.865 results for best model selection HHMM 83.23 84.87 0.840 S-HSMM 84.52 87.92 0.862 Table 1: Detection Performances for the S-HSMM and the HHMM. Best performance for each case is highlighted in bold (we note that best performances are attained in multiple cases and we select one of them to highlight). Table 1 and 2 show that modeling with the S-HSMM results in better performances than the HHMM in both recall and precision rates. And as a result, the F-score improves from 0.840 to 0.865. While the recall rate improves only slightly, the 4% improvement in the precision indicates that the HHMM tends to over-segment the video more frequently than the S-HSMM. This has confirmed our belief that duration information is an important factor in our topic transition detection settings. The semi-Markov modeling has effectively overcome the limitation of the strict Markov assumption of {future past | present} 6 in the flat HMM, allowing longer temporal dependency to be captured via the duration of the state. Nevertheless, given a somewhat more contained set of data used in this experiment, the results from both the S-HSMM and HHMM are better than the previous detection results of news story reported in [4] (which came first in TRECVIC2003 testbed) and the heuristics and Bayesian approaches on topic detection in [23, 21]. These results do not only imply the advantages of the S-HSMM over the HHMM, but also show the contribution of the HHMM in its own right. CONCLUSION In this paper we explore the difficult problem of detecting topic transitions through the use of two probabilistic models , the HHMM and the S-HSMM. Both allow the modeling of hierarchy and the sharing of substructures within the hierarchy , whilst the S-HSMM additionally allows the explicit modeling of durative properties. Coupled with the use of the Coxian model, we show how this unified framework performs better than the baseline cases (the flat HMM and HSMM) and previous results reported. In particular the use of the S-HSMM demonstrates that the modeling of duration is a 6 i.e., the future is conditionally independent of the past given the present. 18 Video TP FP Miss GT 1 - "EESafety" 10 8 1 3 3 5 13 2 - "SSFall" 4 4 1 1 2 2 6 3 - "ElectS" 6 6 2 1 2 2 8 4 - "TrainHaz" 18 20 2 2 3 1 21 5 - "EyeS" 10 10 0 1 0 0 10 6 - "FootS" 10 10 1 1 1 1 11 7 - "HKeeping" 11 11 3 3 1 1 12 8 - "Maintn" 9 8 1 3 4 5 13 9 - "HandS" 9 9 1 1 1 1 10 10 - "SBurning" 19 19 1 1 2 2 21 11 - "HeadProt" 6 5 1 3 1 2 7 12 - "WeldingS" 19 19 3 3 4 4 23 Sum 131 129 17 23 24 26 155 Table 2: Detection results for each video in the best performance cases of the S-HSMM and the HHMM (TP: True Positive, FP: False Positive, GT: Ground-Truth). powerful tool in the extraction of higher level semantics. The results demonstrate the promise of the approach and although the results are demonstrated with the educational and training film genre, the method can easily be applied to other genres. We believe that the promise of the approach lies in its unified probabilistic handling of durative properties and shared hierarchical structure, allowing it to handle long video sequences with inherent variability and complicated semantics. Acknowledgement Hung Bui is supported by the Defense Advanced Research Projects Agency (DARPA), through the Department of the Interior, NBC, Acquisition Services Division, under Contract No. NBCHD030010. REFERENCES [1] B. Adams, C. Dorai, and S. Venkatesh. Automated film rhythm extraction for scene analysis. In IEEE International Conference on Multimedia and Expo, pages 10561059, Tokyo, Japan, August 2001. [2] P. Aigrain, P. Jolly, and V. Longueville. Medium knowledge-based macro-segmentation of video into sequences. In M. Maybury, editor, Intelligent Multimedia Information Retrieval, pages 159174. AAAI Press/MIT Press, 1998. [3] H. H. Bui, D. Q. Phung, and S. Venkatesh. Hierarchical hidden markov models with general state hierarchy. In D. L. McGuinness and G. Ferguson, editors, Proceedings of the Nineteenth National Conference on Artificial Intelligence, pages 324329, San Jose, California, USA, 2004. AAAI Press / The MIT Press. [4] L. Chaisorn, T.-S. Chua, C.-H. Lee, and Q. Tian. A hierarchical approach to story segmentation of large broadcast news video corpus. In IEEE International Conference on Multimedia and Expo, Taipei, Taiwan, June 2004. [5] T. V. Duong, H. H. Bui, D. Q. Phung, and S. Venkatesh. Activity recognition and abnormality detection with the Switching Hidden Semi-Markov Model. In IEEE Int. Conf. on Computer Vision and Pattern Recognition, volume 1, pages 838845, San Diego, 20-26 June 2005. IEEE Computer Society. [6] S. Fine, Y. Singer, and N. Tishby. The hierarchical hidden markov model: Analysis and applications. Machine Learning, 32(1):4162, 1998. [7] A. Hanjalic. Shot-boundary detection: Unraveled and resolved? IEEE Transaction in Circuits and Systems for Video Technology, 12(2):90105, 2002. [8] A. Hanjalic, R. L. Lagendijk, and J. Biemond. Automated high-level movie segmentation for advanced video retrieval systems. IEEE Transactions in Circuits and Systems for Video Technology, 9(4):580588, 1999. [9] I. Ide, K. Yamamoto, and H. Tanaka. Automatic video indexing based on shot classification. In First International Conference on Advanced Multimedia Content Processing, pages 99114, Osaka, Japan, November 1998. [10] U. Iurgel, R. Meermeier, S. Eickeler, and G. Rigoll. New approaches to audio-visual segmentation of TV news for automatic topic retrieval. In IEEE Int. Conf. on Acoustics, Speech, and Signal Processing, volume 3, pages 13971400, Salt Lake City, Utah, 2001. [11] E. Kijak, L. Oisel, and P. Gros. Hierarchical structure analysis of sport videos using HMMs. In Int. Conf. on Image Processing, volume 2, pages II10258 vol.3, 2003. [12] S. E. Levinson. Continuously variable duration hidden markov models for automatic speech recognition. Computer Speech and Language, 1(1):2945, March 1986. [13] T. Lin and H. J. Zhang. Automatic video scene extraction by shot grouping. Pattern Recognition, 4:3942, 2000. [14] Z. Liu and Q. Huang. Detecting news reporting using audio/visual information. In International Conference on Image Processing, pages 2428, Kobe, Japan, October 1999. [15] Mediaware-Company. Mediaware solution webflix professional V1.5.3, 1999. http://www.mediaware.com.au/webflix.html. [16] C. D. Mitchell and L. H. Jamieson. Modeling duration in a hidden markov model with the exponential family. In Proc. of IEEE Int. Conf. on Acoustics, Speech, and Signal Processing, pages II.331II.334, Minneapolis, Minnesota, April 1993. [17] K. Murphy and M. Paskin. Linear-time inference in hierarchical HMMs. In T. G. Dietterich, S. Becker, and Z. Ghahramani, editors, Advances in Neural Information Processing Systems, Cambridge, MA, 2001. MIT Press. [18] M. R. Naphade and T. S. Huang. Discovering recurrent events in video using unsupervised methods. In Int. Conf. om Image Processing, volume 2, pages 1316, Rochester, NY, USA, 2002. [19] D. Q. Phung. Probabilistic and Film Grammar Based Methods for Video Content Analysis. PhD thesis, Curtin University of Technology, Australia, 2005. [20] D. Q. Phung, H. H. Bui, and S. Venkatesh. Content structure discovery in educational videos with shared 19 2 2 2 2 1 1 2 2 2 2 2 2 1 1 1 2 2 2 2 2 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 1 1 1 1 2 2 2 4 4 4 4 5 2 3 3 3 3 3 2 5 2 1 3 3 3 3 3 5 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 5 2 5 1 3 3 3 1 1 2 3 3 3 2 2 2 2 3 3 3 3 3 2 2 2 2 3 3 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 2 2 2 2 2 2 1 2 1 2 2 2 2 2 1 2 1 1 2 2 2 2 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 1 2 2 2 1 2 2 2 2 2 2 1 1 1 2 2 2 2 1 1 1 1 1 2 2 2 2 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 1 1 1 1 1 2 2 2 Detected Topic Transitions z t x t m t t e t main body intro 13 2 2 2 2 1 1 2 2 2 2 2 1 1 1 1 2 2 2 2 2 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 1 1 1 1 2 2 2 4 4 4 4 5 2 3 3 3 3 3 2 5 2 1 3 3 3 3 3 5 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 5 2 5 1 3 3 3 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 shot number z t x t t e t S H S M M H H M M 39 Detected Topic Transitions 5 21 Ground-Truth Figure 5: Example of Viterbi decoding for the S-HSMM and the HHMM for the first 45 shots of video `EESafety'. These results should be read together with Figure-3 to see the semantics of the DBN structure. structures in the hierarchical HMMs. In Joint Int. Workshop on Syntactic and Structural Pattern Recognition, pages 11551163, Lisbon, Portugal, August 1820 2004. [21] D. Q. Phung and S. Venkatesh. Structural unit identification and segmentation of topical content in educational videos. Technical report, Department of Computing, Curtin University of Technology, 2005. TR-May-2005. [22] D. Q. Phung, S. Venkatesh, and H. H. Bui. Automatically learning structural units in educational videos using the hierarchical HMMs. In International Conference on Image Processing, Singapore, 2004. [23] D. Q. Phung, S. Venkatesh, and C. Dorai. High level segmentation of instructional videos based on the content density function. In ACM International Conference on Multimedia, pages 295298, Juan Les Pins, France, 1-6 December 2002. [24] L. R. Rabiner. A tutorial on hidden markov models and selected applications in speech recognition. In Procs. IEEE, volume 77, pages 257286, February 1989. [25] H. A. Rowley, S. Baluja, and T. Kanade. Neutral network-based face detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(1):2338, January 1998. [26] K. Shearer, C. Dorai, and S. Venkatesh. Incorporating domain knowlege with video and voice data analysis. In Workshop on Multimedia Data Minning, Boston, USA, August 2000. [27] J.-C. Shim, C. Dorai, and R. Bolle. Automatic text extraction from video for content-based annotation and retrieval. In International Conference on Pattern Recognition, volume 1, pages 618620, Brisbane, Australia, August 1998. [28] C. G. Snoek and M. Worring. Multimodal video indexing: A review of the state-of-the-art. Multimedia Tools and Applications, 2004. In Press. [29] H. Sundaram. Segmentation, Structure Detection and Summarization of Multimedia Sequences. PhD thesis, Columbia University, 2002. [30] H. Sundaram and S.-F. Chang. Computable scenes and structures in films. IEEE Transactions in Multimedia, 4(4):482491, 2002. [31] B. T. Truong. An Investigation into Structural and Expressive Elements in Film. PhD thesis, Curtin University of Technology, 2004. [32] J. Vendrig and M. Worring. Systematic evaluation of logical story unit segmentation. IEEE Transactions on Multimedia, 4(4):492499, 2002. [33] C. Wang, Y. Wang, H. Liu, and Y. He. Automatic story segmentation of news video based on audio-visual features and text information. In Int. Conf. on Machine Learning and Cybernetics, volume 5, pages 30083011, 2003. [34] J. Wang, T.-S. Chua, and L. Chen. Cinematic-based model for scene boundary detection. In The Eight Conference on Multimedia Modeling, Amsterdam, Netherland, 5-7 November 2001. [35] L. Xie and S.-F. Chang. Unsupervised mining of statistical temporal structures in video. In A. Rosenfield, D. Doreman, and D. Dementhons, editors, Video Mining. Kluwer Academic Publishers, June 2003. [36] L. Xie, S.-F. Chang, A. Divakaran, and H. Sun. Learning hierarhical hidden markov models for unsupervised structure discovery from video. Technical report, Columbia University, 2002. [37] X. Zhu, L. Wu, X. Xue, X. Lu, and J. Fan. Automatic scene detection in news program by integrating visual feature and rules. In IEEE Pacific-Rim Conference on Multimedia, pages 837842, Beijing, China, 2001. 20

domain knowledge;Topic Transition Detection;A variety in directional styles;semantic relationship of neighborhood scenes;coxian switching hidden semi-markov model;natural hierarchical organization of videos;model educational video content;extended Hierarchical Hidden Markov Model;unified and coherent probabilistic framework;Educational Videos;shot-based semantic classification;their semantically shared substructures;topic transition detection;probabilistic framework;Coxian Switching Hidden semi-Markov Model;Coxian;Hierarchical Markov (Semi-Markov) Models;typical durations of important semantics;modeling temporal correlation;hierarchical hidden markov model

197

Towards Content-Based Relevance Ranking for Video Search

Most existing web video search engines index videos by file names, URLs, and surrounding texts. These types of video metadata roughly describe the whole video in an abstract level without taking the rich content, such as semantic content descriptions and speech within the video, into consideration. Therefore the relevance ranking of the video search results is not satisfactory as the details of video contents are ignored. In this paper we propose a novel relevance ranking approach for Web-based video search using both video metadata and the rich content contained in the videos. To leverage real content into ranking, the videos are segmented into shots, which are smaller and more semantic-meaningful retrievable units, and then more detailed information of video content such as semantic descriptions and speech of each shots are used to improve the retrieval and ranking performance. With video metadata and content information of shots, we developed an integrated ranking approach, which achieves improved ranking performance. We also introduce machine learning into the ranking system, and compare them with IRmodel (information retrieval model) based method. The evaluation results demonstrate the effectiveness of the proposed ranking methods.

INTRODUCTION Multimedia search has become an active research field due to the rapid increase of online-available content and new practical applications. Search technology is considered the key to navigating the Internet's growing media (video, audio and image) collections. Google Yahoo, Blinkx and other search companies have provided elementary video search engines. However, existing video search engines are all based on the text information related to the video which can be retrieved from web pages, such as file names, URLs, and surrounding texts. These types of textual information can be considered as "metadata" of the video since they only roughly describe the video. There is no doubt that text searching is the most efficient way to retrieve information (even when searching for videos), because it well matches the manner of human thinking. However, only using metadata is far form people's expectation in video searching, because even the best case scenario, the metadata is only the highly concentrated overview of a video, with many losses on details. In general, a video consists of many shots and sub-events with a temporal main thread. The video should be segmented into smaller retrievable units that are directly related to what users perceive as meaningful. Much research has concentrated on segmenting video streams into "shots" using low level visual features [1]. Each segment has its own scenes and meanings. In many cases, when users query a video, they intend to find some desired clips in the video instead of viewing it thoroughly. However, this can seldom be achieved by searching the surrounding text which is related to the whole video. Much content information can be used to search videos and shots. In content-based video retrieval systems, video shots can be classified into or annotated by several semantic concepts. The most substantial works in this field are presented in the TREC Video Retrieval Evaluation (TRECVID) community [2]. In addition, speech is also significant information which has close connection to video contents. Some videos are associated with transcripts/closed captions which are provided by content provider. Using ASR (automatically speech recognition) to generate speech text is another practical solution. In this paper, with the video metadata and content information of video shots, we index and rank the videos in a way similar to general text-based web page search engines. The IR-model, which is widely employed in text information retrieval and web page search, will be applied to rank the search results by examining relevance between query and indexed information (including both metadata and content information). To fully utilize the content information and get a better ranking performance, we integrate the "shot relevance" into "video relevance". That is, the ranking is decided not only by the relevance of video (metadata of the entire video), but also by all the relevant shots within the video. We also apply learning based method to rank the search results based on a set of features extracted from the corresponding query, video metadata, and content information. The rest of this paper is organized as follows. Section 2 introduces the IR-model based ranking, including extraction of video metadata and content information, and a ranking method integrating these two types of information. In section 3, a learning based ranking approach is presented. Section 4 compares the ranking performance evolution results, and Section 5 concludes the paper. IR-MODEL BASED RANKING In the traditional text retrieval and web page search, IR (Information Retrieval) models are usually used to evaluate the relevance between a query and a document. BM25 [3] is one of the frequently used evaluation methods. Given a query Q and a document D, the relevance between Q and D can be modeled as the summation of the relevance between each query term (word) t in Q and D: and b are parameters. tf(t,D) is term frequency, means the frequency of term t appears in document D. df(t) is document frequency, means the frequency of document which contains term t within all documents. |D| stands for the length of document D. The basic idea of this IR model can be explained as, if the query term appears in document more frequently (higher tf), and the query term is more unique that less documents contain it (lower df), the query will be more relevant to the document. 2.2 Index and Rank the Video Information The video data used in our experimental system are from MSN Video (http://video.msn.com/), which contains 7230 videos. 2.2.1 Metadata of the video Because the videos in our data set are made by professional content provider, there is rich meta information that describes each entire video with brief text. Each video has the following metadata fields: headline, caption, keywords, source, video URL, thumbnail image URL, related web page anchor text, page URL, etc. Besides these types of textual information, some format information of the video, such as video length, frame size, bit rate, and creation date are also extracted. Some selected information fields of video metadata are listed in Table 1. Table 1. Video metadata Field Example Value Headline Discovery launches Caption July 26: Watch the entire launch of space shuttle D... Source MSNBC Keywords Technology, science, Space, Partner Codes ... Video URL http://www.msnbc.msn.com/default.cdnx/id/871313... Link anchor MSNBC.com's Technology and Science front Link URL http://www.msnbc.msn.com/id/3032118 date 7/26/2005 4:40:48 PM video length 609.72 seconds Frame size 320 x 240 Bit rate 180 Kbps For the videos contained in general web pages, some attributes mentioned above may not be obtained directly, but the surrounding texts, URL, filename can be extracted as the metadata fields of the video. These information fields correspond to document D in Section 2.1. Different fields can be represented by different type of D (D i in Equation 4). The overall relevance can be calculated by the weighted summation of the relevance of all fields. The weight of the fields (DW i in Equation 4) can be determined by their importance, significance, and representativeness to the video. = ) , ( ) , ( Q D R DW Q Video R i i (4) In our system, four major information fields from video metadata are selected to be indexed: headline, caption, keywords, and source. Headline is a highly representative description of the video content. Keywords are also good recapitulative terms. For these two fields, higher weights are set. Caption is a more general and detail depicts for the video; Source provided a higher level and less relevant information, they will be set lower weights for ranking. Table 2 gives out the weights of fields in our experimental system. Table 2. Weights for relevance evaluation Fields weight Headline 10 Keywords 10 Caption 5 source 1 2.2.2 Content information of the video shots There is plenty of information in the visual/audio content of the video sequence, which can not be sufficiently presented by the aforementioned textual video metadata. We can build a set of models that can be applied to automatically detect a corresponding set of concepts such that each video shot can be annotated with a detection confidence score for each concept. Successful concept modeling and detection approaches have been introduced in TRECVID, relying predominantly on visual/aural analysis and statistical machine learning methods [4]. The LSCOM-lite Lexicon [5] designed for the TRECVID 2005 Benchmark consists of more than 40 concepts spread across multiple concept-types such as object, events, site etc. Though the size of the lexicon is still far from practical application for general Web-based video search, this semantic information is promising to enable real content-based video search, and therefore it is applied in our ranking system. Besides visual contents, information from audio channel, especially the speech, is also very useful for searching videos.. In our experimental system, we use Microsoft speech recognition engine (with a large vocabulary). This engine gives recognized words with a start timestamp, length, and a recognition confidence value, which are very useful for later indexing and ranking. The speech texts are allotted and assigned into video shots, according to the timestamp of words and video shots. The content information is associated with individual video shot, which consist of semantic keywords (with corresponding detection confidences), and speech words (with recognition confidences). The confidences of words will act as weights of term frequencies tf to calculate the relevance in Equation (2). 628 2.3 Integrated Ranking with Metadata and Content Information To combine metadata and content to rank the videos, we index the videos by metadata and index the video shots by content information separately, and then integrated these two rank lists, named video list and shot list, to form a final ranking. The integrated ranking returns search result by video, but taking all the relevance shots within this video into consideration. For video list, each item is a video. Let item iv .vid denotes the video ID of the i th item, item iv .score denotes the ranking score of the i th item. For shot list, each item is a shot from a video. Let item is .vid, item is .sid, item is .score donote the video ID which the shot belong to, the shot ID within the video, and the ranking score of the i th item respectively. The integrating process is presented in Algorithm 1. The basic idea is that, all the ranking score of the relevance shots within the video are accumulated to the ranking score of the video, with corresponding weights. The relevant shots in the video will be highlight when displaying the video as search result. new a integrated result list (item denotes as item iI ) for each item iv in video list{ new item ic ; item iI .vid = item is .vid; item iI score = item iv . score * Weight_v; for each item item is in shot list{ if(item iv .vid == item is .vid) { item iI addshot(item is .sid); item iI score += item is . score * Weight_s; remove item is from shot list } } remove item iv from video list add item iI to integrated list } add the remaining video list and shot list into the integrated list sort the integrated list by item iI .score. // Weight_v and Weight_s are weights for score accumulating Algorithm 1. Generate integrated rank list. LEARNING BASED RANKING IR-model based ranking just consider some basic features such as term frequency tf, document frequency df, and document length, etc. In the learning based approach, more features are extracted from the query, metadata, and content information. To be clear, suppose the query contains three terms "a b c", we compute the following features from each document field: Ordered match: the frequency that both "a" and "b" appeared in the indexed text, and "b" appears after "a". Partly exact match: the frequency that "a b" or "b c" appeared in the indexed text. Exact match: the frequency that "a b c" appeared in the indexed text. Query length: number of query terms For the content information, each word has a confidence value, we also consider: Weighted tf: Term frequency with confidence weighted, High confident match: query term match with words with high confidence. High confident words: words with high confidence in the indexed text. Some non-textual, query-independent features, such as shot length, video length, frame size, bit rate, etc, are also taken into account. By counting in the combinations of several document fields and query terms (or part of query), we have about 50 dimensional features in total for a query and search result to form a sample.. The GroundTruth of sample is the relevance judgments between a query and a result, which is collected by a user labeling system introduced in the next section. 3.2 Neutral Network based Ranking Traditionally the learning algorithms are used for classification problems. For ranking problems, one possible way is organize the samples by pairs. Pair sample (x 1 , x 2 ) are considered as a positive sample, if x 1 are ranked ahead of x 2 , and vice versa. The loss function is also formulated in pair wise. RankNET [6] is used in our implementation to train the ranking model and to validate the performance. About half of the labeled data are used in training, and the second half are used for validation. EVALUATIONS To evaluate the ranking performance of our proposed methods, we developed a user labeling tool to collect some query-result relevance judgments. The video data set we used in our experiment includes news video, TV programs, movie trailers, advertisements, etc. According to the characteristics of the content of these videos, we selected some news related queries, such as hot events, hot place, and hot person names, to evaluate the ranking performance. For each query, we use the IR-model based ranking describe in Section 2 to generate a result list, and randomly select some results form the list to label. Considering the labeling workload, for each labeler and each query, 9 results are select from the list. To make the selected query-result samples have a good uniformity on distribution, 3 results are randomly selected from the first 1/3 part of the list, 3 are from the second 1/3 part, and the other 3 are from the last 1/3 part. The order of these 9 selected results is shuffled and then provided to users to do relevance labeling. In the labeling tool, for a query and a result, user can see all the information of the result, including the file format information (frame size, bit rate, video length, etc), description (headline, caption, keywords), video thumbnail, video (in a video player), thumbnails of video shots, the speech text of the relevant shots. The words matched with query terms are highlighted. See Figure 1. If there are relevant shots in the result, the thumbnails of them are displayed with doubled size. The shot number, time 629 information, and the speech are also shown in the interface. Users are asked to read the displayed information, browse the thumbnails, and play the video and shots (a tool button is provided to play from one shot) to give a relevance judgment from 1, 2, 3, 4, and 5, which represent bad, fair, good, excellent, and perfect, respectively. Figure 1. Relevance labeling tool In our experiment, ten users are invited to do labeling, and about 2,000 relevance judgments of query-result samples are collected. 4.2 Precision Performance of Ranking We have conducted a comparison between the 4 approaches listed below: MR: Ranking only based on video metadata (Section 2.2.1). CR: Ranking only by content information (Section 2.2.2). RI: Integrated Ranking described in Section 2.3 RN: RankNET based ranking described in Section 3.2. The precision in top N of the rank lists of all the labeled queries is used to evaluate the performance of ranking method. N top in results labeled total N top in results labeled relevant N Precision = @ (5) In our implementation, the judgment Perfect or Excellent are considered as relevant results, while other judgments are treated as irrelevant results. The Presicion@N (N=1 to 5) of the 4 ranking methods are shown in Table 3. From the results, we can see that: 1) Precisions of MR are very low. Only using video metadata will result in a poor performance, since details of the video content are ignored. The content information is more effective to search and rank video than metadata, as precisions of CR are higher than that of MR. 2) Precisions of RI are much higher than that of MR and CR. By combining video metadata and content information, the performance is significantly improved and reaches an acceptable level, which shows that content-based relevance ranking is a promising approach. 3) RN has a good performance, even better than RI. Comparing to IR-model based ranking, more features are included to learning the relevance. The result implies that the learning method can organize the information for ranking in a more effective way. Table 3. Precision of the ranking approaches Precision@? 1 2 3 4 5 MR 0.305 0.326 0.299 0.259 0.228 CR 0.544 0.526 0.571 0.550 0.522 RI 0.796 0.727 0.684 0.634 0.606 RN 0.805 0.746 0.763 0.717 0.669 DISCUSSIONS AND CONCLUSION We have presented a novel content-based approach to rank video search results . In addition to the video metadata, more detailed content information in the video is used to improve the relevance ranking of video search results. The videos are segmented into shots, which can carry rich content information such as semantic concept keywords and speech. With the video metadata and content information, we proposed an IR-model based ranking method and a learning-based ranking method. Evaluation of the top ranked results shows that the proposed ranking methods have significantly improved performance comparing to the approach use video metadata only, which is frequently used in existing web video search engines. In future work, more types of content information can be integrated into our ranking scheme, such as content-based quality metric, user comments and rating for videos shared in web communities. Moreover, how to define effective semantic concepts, i.e., video semantic ontology, that facilitate video searching and ranking is also a challenging problem., which is also one of our future works. REFERENCES [1] Hong-Jiang Zhang, A. Kankanhalli, and S. Smoliar, "Automatic Partitioning of Full-motion Video," A Guided Tour of Multimedia Systems and Applications, IEEE Computer Society Press, 1995. [2] http://www-nlpir.nist.gov/projects/trecvid [3] S. E. Robertson, S. Walker, and M. Beaulieu. Okapi at TREC7: automatic ad hoc, filtering, VLC and filtering tracks. In Proceedings of TREC'99. [4] M. Naphade, J.R. Smith, F. Souvannavong, "On the Detection of Semantic Concepts at TRECVID," ACM Multimedia, ACM Press, New York, NY, pp. 660-667, Oct. 10-16, 2004 [5] M. Naphade, L. Kennedy, J.R. Kender, S.F. Chang, J.R. Smith, P. Over, A. Hauptmann, "LSCOM-lite: A Light Scale Concept Ontology for Multimedia Understanding for TRECVID 2005," IBM Research Tech. Report, RC23612 (W0505-104), May, 2005. [6] Chris Burges, et.al, "Learning to Rank using Gradient Descent", ICML 2005, Bonn, Germany, pp.89-96, August 7-11 , 2005. 630

video index, relevance ranking;content-based relevance ranking;video retrieval;metadata;learning based ranking;neutral network based ranking;Relevance ranking;content information;content-based approach;ranking method;integrated ranking;video metadata;IR-model;segmented;Content-based ranking;machine learning model;video segmentation;IR model based ranking;Video search;video search

198

Towards Reasonability Properties for Access-Control Policy Languages

The growing importance of access control has led to the definition of numerous languages for specifying policies. Since these languages are based on different foundations, language users and designers would benefit from formal means to compare them. We present a set of properties that examine the behavior of policies under enlarged requests, policy growth, and policy decomposition. They therefore suggest whether policies written in these languages are easier or harder to reason about under various circumstances. We then evaluate multiple policy languages, including XACML and Lithium, using these properties.

INTRODUCTION Access-control policies should not be write-only. Because they govern both the containment and availability of critical information, they must be highly amenable to analysis by both humans and by reasoning software such as verifiers. An access-control policy dictates a function from requests for access to decisions about whether or not to grant access . The competing requirements of expressive power and computational speed makes the design of policy languages a delicate balancing act. Contemporary policy languages have largely followed one of two routes. Some are based on logics, restricting first-order logic (e.g., Lithium [9]) or augmenting Datalog (e.g., Cassandra [2]). Others are custom languages such as XACML [12] and EPAL [13], which behave roughly by rule-evaluation and do not depend on theorem-proving capabilities to determine a response to a query. The custom language approach often produces fairly limited languages. For example, to express hierarchical role-based access-control (RBAC) [14] in XACML requires a fairly cumbersome encoding [1]. On the other hand, its more direct request evaluation strategy suggests that policies written in XACML are more transparent than policies written in languages based on first-order logic (as we motivate in Section 2). How, then, do we distinguish different policy languages? Studies of complexity and expressive power may ensure tractable verification and the ability to capture certain policies, but do not directly classify the ease of reasoning about policies in a language. In this paper we take a step towards formalizing reasonability properties that make languages more amenable to reasoning. We then apply these properties to actual policy languages. Such properties are useful even when verification is computationally tractable because they provide a guide to where and how to edit a policy for a desired effect. Concretely, our properties study three main questions: how decisions change as requests include more information, how decisions change as policies grow, and how amenable policies are to compositional reasoning. The last of these is especially important for two reasons. First, organizations in-creasingly have different divisions creating policy fragments that must be combined into a whole while preserving the intent of each division; second, to mirror these use cases, and to scale better as policies grow in size, it becomes important for analysis and verification tools to function modularly. These properties codify our observations made while writing and studying policies for non-trivial systems. (We do not, however, presume to make broad statements about the impact of these properties for manual reasoning.) They are meant to be descriptive rather than prescriptive: which ones a language should satisfy depends on the context of its use. We do expect these properties to help both language designers and policy authors, the former to set goals and the latter to evaluate languages. We first motivate the work with an example. Section 3 presents background on policy languages. Section 4 presents the heart of our formalism. Section 5 applies this framework to XACML, and Section 6 to logical approaches such as 160 Lithium. The remainder discusses related work and offers concluding remarks. MOTIVATING EXAMPLE Consider the following natural-language policy: 1 1. If the subject is a faculty member, then permit that subject to assign grades. 2. If the subject is a student, then do not permit that subject to assign grades. 3. If the subject is not a faculty member, then permit that subject to enroll in courses. We might represent this policy as follows: faculty( s) = Permit(s, grades, assign) ( p 1 ) student( s) = Permit(s, grades, assign) ( p 2 ) faculty(s) = Permit(s, courses, enroll) ( p 3 ) Let the above formalization be p and the first line of the policy be sub-policy p 1 , the second p 2 , and the third p 3 . Consider the following natural-language request: A student requests to enroll in courses. Assume that requests list the subject, resource, and action by name if possible and by variable if the name is unknown, along with any other known facts. In this representation, the request becomes: (s, courses, enroll) with student(s) ( q 1 ) Should the policy grant access? At least three interpretations of the policy are possible: 1. p grants access due to p 3 . The request does not show the subject being a faculty member; thus, p 3 applies and p produces the decision to permit access. This relies on the assumption that since the request does not show the subject being faculty, that the subject is in fact not faculty. One could drop this assumption. 2. The policy does not apply to the request. One would reason that p 1 and p 2 do not apply since they are dealing with assigning grades and not enrolling in courses. Furthermore, one could conclude that p 3 does not apply since the request does not prove that the subject is not faculty. To do so, the request would have been (s, courses, enroll) with student(s) faculty(s) Since the policy does not apply to the request, the system should have and enact some default behavior. 3. By reasoning different than that used in the first interpretation , p could still grant the request. As in the second interpretation, one could conclude that the request alone fails to establish that the subject is not a faculty member. However, if the subject were a faculty member, then the first two lines together would yield a contradiction: p 1 would imply that the subject could enroll in courses and p 2 would imply that the subject could not. Thus, student-faculty members do not exist. Since the subject of the request is clearly a student, he must not be faculty member. Thus, p 3 applies to grant access. 1 This example is adapted from Halpern and Weissman [9]. In the first two interpretations the user may limit his reasoning to each sub-policy independent of one another. However , under the third interpretation (which, in fact, is the one chosen by Halpern and Weissman), the user must reason about all three sub-policies at once. Furthermore, under Interpretation 2, the user must reason about both positive and negative attributes, unlike under Interpretation 1. These semantic differences drastically affect a reader's ability to comprehend policies. For example, Interpretation 3 requires both global analysis and demands rich reasoning power to deduce the contradiction. This paper formalizes these differences and their burdens. BACKGROUND Despite the differences between access-control policy languages , we can still identify many common elements. First we describe features common to most languages, and then we discuss in detail two areas in which many languages differ : the available decisions and policy combinators. 3.1 Common Features A policy language must provide a way of describing the different forms of access and the environment in which they could occur. This information forms a request. Many languages break requests into four different parts: Subject the person or process making the request, Resource the object, subsystem, person, or process that the subject would like to affect (e.g., a file name or a process id), Action the command or change that the subject would like to execute on the resource, and Environment describes any other relevant information including the time of day, location, or the previous actions of the subject. The first three make up the form of access requested while the last gives the context in which this access would occur. Each of these parts lists attributes associated with its respective topic. In some languages, the absence or negation of an attribute might also be explicitly listed (see Section 4.2). Languages must also provide a set of decisions. Such a set must include some decisions that grant access and some that prohibit access. A policy will associate with each request a decision (or in the case of nondeterministic policies, a policy will relate each request with some number of decisions). Definition 3.1. An access-control policy language is a tuple L = (P, Q, G, N, ) with P a set of policies, Q a family of sets of requests indexed by policies, G the set of decisions that stipulate that the system should grant access (granting decisions), N the set of decisions that stipulate that the system should not grant access (non-granting decisions), a function taking a policy p P to a relation between Q p and G N, where G N = . 161 When clear from context, the above symbols will be ref-erenced without explicitly relating them to L, and D will represent G N. The function gives the meaning of policy p, and we write q p d for p P , q Q p , and d D, when p assigns a decision of d to the request q. If for a language Q p = Q p for all p and p in P , then we drop the subscript on Q and treat it as a set of requests common to all policies. Given L define the partial order on D to be such that d d if either d, d N, d, d G, or d N and d G. 3.2 Decisions Policy languages must provide decisions to indicate a pol-icy's intent to grant or not to grant a request. Some languages might just provide two decisions: permit for granting access and deny for not granting access. A policy in such a language associates various subsets of requests with one of these two decisions. For example, p 1 explicitly identifies a subset of permitted requests and p 2 gives denied requests. However, a policy might assign some requests to neither permit nor deny (e.g., q 1 under Interpretation 2 of p). To err on the side of safety, the policy language should provide for such requests a default decision that does not imply a grant of access creating a closed policy [10]. However, assigning them the decision of deny may limit the ability to compose policies. For example, while combining the policies of two departments, one would like to distinguish between those requests that each department really would like to prohibit and those about which they do not care [3]. The decision of not applicable serves this purpose. With a decision of not applicable sufficing to prevent access , some languages elect not to include statements associating requests with deny. This leaves only statements permitting some set of requests. The uniformity of statements in such languages might make the policy easier to read and compose (see Section 3.3). However, allowing for the explicate denial of requests can quickly rule out exceptional cases and provides a means to determine when a policy does not grant access by desire rather than by default. Some requests might not have a logical interpretation under a given policy. For example, a request of (s, grades, assign) with faculty(s) student(s) (q 2 ) under Interpretation 3 of p contradicts the policy itself. A request might even contain illogical values or require undefined computation (such as division by zero). For generality, a system might like to assign a decision to such inputs rather than excluding them from the set of requests and leaving the policy undefined on them. In such cases, a decision of error or some refinement of it might be appropriate. One may view the fact that an error state is reached given a request to be a weakness in the policy. However, one may also take it to be a statement about the world in which the policy is to function: that no such requests may logically exist. Error decisions can enforce these preconditions or assumptions that the policy has made. 3.3 Policy Combinators The policy of an organization often consists of the composition of policy fragments, or sub-policies, from a variety of internal units (e.g., legal, accounting, and execute departments of a corporation). Thus, policy languages provide combinators to create a single policy from these fragments. Under p, the request given above (q 2 ) is permitted by p 1 but denied p 2 . The method used to combine the three sub-policies of p into one policy determines how to resolve this conflict. Some languages, like the hypothetical language in which p is written, might have only one policy combinator that is implicitly applied. Other languages provide multiple combinators. If a policy has sub-policies nested inside of sub-policies, the different layers may be resolved differently. Some of the possible policy combinators are: Permit Overrides If any of the sub-policies produces a permit, return only permit. Otherwise, if any produces a deny, return only deny. Else, return not applicable. Deny Overrides If any of the sub-policies produces a deny, return only deny. Otherwise, if any produces a permit, return only permit. Else, return not applicable. First Applicable Return the decision reached by the first sub-policy to reach one other than not applicable. All Seen Return a set containing the decisions reached by all the sub-policies. Either Permit or Deny Nondeterministically return one of the produced decisions. Error Return a error if the sub-policies produces both permit and deny. Otherwise return permit if produced, or deny if produced. Else, return not applicable. And Conjoin the sub-policies together by logical And and return the implied decision(s). De Capitani di Vimercati et al. [4] list additional combinators . The nature of the combinators available in a language can greatly impact the clarity of policies written in it. Notice that many of the above combinators behave the same in the absence of the decision of deny. One might conclude from this observation that allowing the explicit denial of a request is an undesirable complication in a language. To formalize the role of policy combinators, let policies be either an atomic policy or a set of sub-policies combined by some policy combinator. Let p be a policy that consists of sub-policies p i with 1 i n. Then p = (p 1 , p 2 , . . . , p n ) represents the composition of the sub-policies using . 2 Since sub-policies are themselves policies, one may apply to them. 3 The relationship between (p 1 , p 2 , . . . , p n ) and meaning of each sub-policy p i affects the clarity of the policy and is studied in the next section. POLICY LANGUAGE PROPERTIES Having formalized policy languages, we are now ready to describe properties of them. 2 We assume that the set of combinators in a given language L = (Q, P, G, N, ) is clear from the structure of P and . If this is not the case for a language, one could explicitly add it to the definition of an policy language. 3 Some languages may permit contextual information from enclosing policies to affect the meaning of the sub-policies. For example, a language might have a notation of variable binding. For such a language, might be extended to take a second argument that carries such contextual information . All the following definitions could be extended, e.g., monadically, to deal with such an extended . 162 4.1 Determinism and Totality Definition 4.1. A language L = (P, Q, G, N, ) is deterministic if p P, q Q p , d, d D, q p d q p d = d = d For a deterministic language, we can define a function which takes a policy p P and returns a function from Q p to D as p . q . d D s.t. q p d. For a deterministic language, may be given instead of to define the language. (We only even mention nondeterministic languages due to the existence of one: XACML with obligations.) Definition 4.2. A language L = (P, Q, G, N, ) is total if p P, q Q p , d D s.t. q p d The policies of total languages will always make a decision. 4.2 Safety Under Interpretation 2, the request contained too little information to determine which of the sub-policies of p applied . Interpretation 1 avoids such indecision by having requests implicitly refute the presence of any attribute not listed. These two interpretations produce different meanings for statements like faculty(s) found in p 3 . Under Interpretation 1, faculty(s) holds if faculty(s) is not in the request, while under the second, the request must explicitly list faculty(s) for it to hold. We call the former implicit and the latter explicit. The explicit approach permits distinguishing between unknown information and attributes known to be absent. The explicit interpretation, however, incurs the cost of listing a possibly large set of absent attributes and can lead to indecision as shown above. Such indecision, however, allows the system to recognize when the policy requires more information to yield a decision . In contrast, the implicit interpretation can grant undue access. If, for example, a request does not list faculty(s) simply because the system did not determine whether s was a faculty member or not, then the system might erroneously allow s to enroll in courses. Thus, the sub-system producing requests must be sure to include all the relevant facts in each request. For large scale systems, collecting or even determining the germane information might consume large amounts of time. For such systems, the explicate approach might prove better since requests may leave out information safely and be refined until the policy yields a decision. Furthermore, overzealous optimizations and other coding errors might result in the system producing requests that do not contain all the relevant facts. Having a policy drive which information requests include allows for the system to collect only the information really needed to reach a decision from the policy. Under this approach , the sub-system evaluating the policy starts with a request that contains only the readily available information. If this sub-system needs additional information to reach a decision from the policy, it requests the necessary additional information. Thus, the system does not need to know what information the policy requires at the time of generating the initial request. This approach may allow for more efficient implementations. Once a datum has been published, it cannot easily be retracted . Therefore, preventing unwanted access is usually preferable to granting it. As a result, such incomplete requests should only result in a grant of access if the complete one would have. We can formally state this safety concern: Definition 4.3. Let be a family of partial ordering on requests of a language L = (P, Q, G, N, ) indexed by the policies of L. L is safe with respect to iff p P, q, q Q p , q p q = p (q) p (q ). Due to differences in the contents of a requests, for each language a different family of partial orderings will interest users. The relation should be such that if q p q , then q contains all the information contained in q and possibly more. Often one partial ordering may serve every policy. For example, consider a language in which requests are sets of non-contradictory facts and the set of decisions is {permit, deny}. Then using the subset partial ordering for p (for every policy p) will make sense since it matches the intuition of information content. If the language is safe with respect to such a defined , then one may omit facts from a request without causing undue access. Informally, in a safe language, undue access is impossible provided that requests tell no lies; whereas, in an unsafe language, the requests must additionally tell the whole truth. The choice between these is a function of trust in the program generating requests, comprehensiveness of analysis to generate requests, efficiency, and so on. Nevertheless, the ability to conclude, given a request that will yield access, that all requests with more information will also yield access , can potentially be a great boon to policy reasoning. Some languages might choose to avoid the complications introduced by a policy testing for the absence of an attribute all together. In some contexts, such as certificate passing systems in which a certificate may be withheld, negated attributes may not make sense. 4 In such a context, requests would not list negated attributes and the policy would not test for the absence of an attributes at all. 4.3 Independent Composition Consider the third interpretation of p. Under this interpretation , the meaning of p can only be determined by looking at the interactions of the different sub-policies as a whole. Notice that any one of these sub-policies would produce a decision of not applicable in isolation, and yet together they interact to produce a permit decision. The third interpretation thus inhibits the easy use of local reasoning to reach conclusions about the policy as a whole. This increases the possibility of unintended results from combining sub-policies into a policy. The alternative, as found in the first two interpretations, is for the sub-policies to be combined in such a way that only the result of each in isolation matters. This property is formalized as follows: Definition 4.4. A policy combinator of a language L = (P, Q, G, N, ) independently composes its sub-4 One may argue that certificate passing systems may use negative certificates to achieve the checking of attribute absence . Whether this captures the notion of the absence of an attribute or just the presence of another related attribute is unclear. For example, one could conceivably hold both a positive and a negative certificate for an attribute. 163 policies iff p 1 , p 2 , . . . , p n P, i, 1 i n = Q (p 1 ,p 2 ,...,p n ) Q p i (1) and there exists a function : Q D D such that p 1 , p 2 , . . . , p n P, q Q (p 1 ,p 2 ,...,p n ) , (p 1 , p 2 , . . . , p n ) (q) = (q)( p 1 (q), p 2 (q), . . . , p n (q)) (2) If all the combinators of L independently compose, then L has the independent composition property. The first requirement forces a request defined for a policy to also be defined on each of its sub-policies. This is necessary for the second requirement to be well defined. The second requirement ensures that one can determine the decision of the whole policy from the request and decisions of its sub-policies on that request; no other properties of the sub-policies matter. One might alternatively be tempted to define independent composition thus: Definition 4.5. A policy combinator of a language L = (P, Q, G, N, ) semantically composes its sub-policies iff : (Q D) (Q D), p 1 , p 2 , . . . , p n P, (p 1 , p 2 , . . . , p n ) = ( p 1 , p 2 , . . . , p n ) (3) If all the combinators of L semantically compose, then L has the semantic composition property. Semantic composition ensures that all sub-policies with the same meaning in isolation will behave the same under the combinator. A language with the semantic composition property is arguably more clear than one without it, since only the isolated meaning of the sub-policy must known to reason about its use under the combinator. Theorem 4.6. If a policy combinator of an policy language L has independent composition, then it has semantic composition. Proof. To prove that has semantic composition, : (Q D) (Q D) required for Equation 3 will be constructed from the : Q D D known to exist since independently composes. Let (f 1 , f 2 , . . . , f n ) = q . (q)(f 1 (q), f 2 (q), . . . , f n (q)) Then ( p 1 , p 2 , . . . , p n ) = q . (q)( p 1 (q), p 2 (q), . . . , p 3 (q)) = q . (p 1 , p 2 , . . . , p n ) (q) = (p 1 , p 2 , . . . , p n ) Theorem 4.7. The semantic composition of a policy combinator does not imply that it independently composes. Proof. Consider a rather odd language that has only one unary policy combinator, , atomic policies that are sets of values, G = {permit}, N = {deny}, and requests that are sets of values. Let the semantics be (p 1 ) = q . (permit p 1 ( {v }) = permit deny p 1 ( {v }) = deny for some distinguished value v , and for atomic policies p, p (q) equals permit iff p q = and equals deny otherwise. The language has semantic composition: for such that (f 1 ) = (permit f 1 ( {v }) = permit deny f 1 ( {v }) = deny clearly, (p 1 ) = ( p 1 ). To show that the language does not have independent composition, assume that it does. Then there exists such a : Q D D to satisfy Equation 2. Let p 1 = {v} and p 2 = {v, v } for some value v such that v = v . Then, deny = ({v}) ({v}) = ( {v})( {v} ({v})) = ({v})(permit) = ( {v})( {v, v } ({v})) = ({v, v }) ({v}) = permit A contradiction is reached since permit = deny. Only with independent composition can a policy reader with a specific request in mind know the decision of the whole policy from each of the component policies. This enables a reader to ask what-if questions like "What if Bob requests to write the log?" and determine the answer from recursively asking that question of the sub-policies. Such an ability is particularly helpful to readers interested in only a subset of the possible requests or already familiar with some of the sub-policies. 4.4 Monotonicity As noted at the end of Section 3.3, the decision of deny complicates the policy combinators. One of the reasons for this is that, under combinators like Deny Overrides, a back-and -forth pattern can arise when considering the decision of the whole policy from the sub-policies. Consider each sub-policy in p with the request q 2 . Under a reasonable interpretation p 1 yields a decision of permit, p 2 a decision of deny, and p 3 not applicable. Thus, if the order of p was changed to p 3 , p 1 , p 2 and we assume a Deny Overrides policy combinator , the apparent decision would go from non-granting to granting to non-granting. Note that Permit Overrides does not exhibit this pattern since it is impossible to go from a granting decision to a non-granting one under it. Thus, the formalization of this pattern focuses on the transition from a granting to a non-granting decision. Definition 4.8. A policy combinator of a language L = (P, Q, G, N, ) is monotonic iff p 1 , . . . , p n , p P, q Q , (p 1 , . . . , p n ) (q) (p 1 , . . . , p i , p , p i+1 , . . . , p n ) (q) where Q = Q (p 1 ,...,p n ) Q (p 1 ,...,p i ,p ,p i+1 ,...,p n ) . We say L is monotonic if every combinator is monotonic. Adding another sub-policy to a monotonic combinator cannot change the decision from granting to non-granting. Having motivated and established these criteria, we now apply them to concrete access control languages. 164 CORE XACML In its entirety, XACML [12] exceeds the bounds of the definitions given in Section 3. Full XACML includes obligations , which act as annotations on the decisions of permit and deny. These annotations specify actions that the system enforcing the access controls must preform before granting access or upon prohibiting access. Thus, an XACML policy may have effects beyond just granting or prohibiting access that the model presented fails to address. Handling all of XACML is beyond the scope of this paper . For illustrative purposes, we employ a formalized subset of XACML, which we will call Core XACML (CXACML), which corresponds to the input of the tool Margrave [7, 8]. This subset is expressive enough to capture RBAC 0 [14]. 5.1 Syntax CXACML has two syntaxes: one for policies and one for requests. We present the policy syntax first, with the start non-terminal P. For syntactic brevity, we use a Lisp-like parenthetical syntax in place of XML notation. P ::= (Policy C T P ) | (Rule T F) C ::= FirstApp | DenyOver | PermitOver T ::= ( (L ) (L ) (L ) (L ) ) L ::= (A + ) A ::= (id val) F ::= Permit | Deny Those policies formed by using solely the right choice of the production rule for P are called rules. XACML does not consider rules to be policies. However, since the semantics assigned to rules allows them to behave as policies, we will consider them policies. The elements of the syntax category T are called targets. The four parts of the target are the requirements placed on the subject, resource, action, and environment, respectively, for the policy to apply to a request . The non-terminals id and val are strings representing the attribute IDs and values. Example 5.1. The following is a CXACML policy: (Policy FirstApp ((()) (((name log))) (()) (())) (Rule (((role dr)) (()) (()) (())) Deny) (Rule ((()) (()) (()) (())) Permit)) This policy permits all requests for access to a log except those made by doctors, which it denies. In detail, the policy is composed of two sub-policies using the combinator First Applicable and applies only to requests where the resource has the name of log. The first sub-policy denies requests where the subject has the role of dr regardless of the resource, action, or environment. The second permits all requests. The syntax for requests is (with start non-terminal Q): Q ::= ( (A ) (A ) (A ) (A ) ) They simply list the attributes possessed by the subject, resource, action, and environment in turn. 5.2 Semantics In the following natural semantics, we will use the convention that a lower-case letter represents an element of the set or syntactic category represented by the upper-case equivalent . For example, P is the set of all policies and p is a policy. Let D be the set of all decisions (D = {permit, deny, na}). The core of the semantics of CX ACML compares requests to targets. We will denote this relation by qt for request q and target t. The natural semantics of Table 1 defines . Next we define . Table 2 gives the result of evaluating rules. The following tables defines over policies. Table 3 deals with two cases where a policy does not apply to a request. Finally, we must define the policy combinators: Permit Overrides in Table 4, Deny Overrides in Table 5, and First Applicable in Table 6. 5.3 Analysis The syntax and semantics of CXACML defines L CXACML = (P, Q, G, N, ). The syntax determines P and Q where the same set of requests is used for every policy (and thus, we treat Q as a set of requests). From the semantics, G = {permit}, N = {na, deny}. CXACML allows for explicit denials and the checking of the implicit absence of attributes. Theorem 5.2. L CXACML is deterministic. Proof. Inspection of the inference rules for atomic policies (Table 2) shows that only one of them can hold at a time. Thus, atomic policies are deterministic. Table 4 combined with Table 3 gives the semantics of the policy combinator Permit Overrides. The antecedents of all these inference rules are disjoint, that is, at most one them can hold for any policy and request. Thus, Permit Overrides is deterministic. The same argument holds for Deny Overrides and First Applicable using Tables 5 and 6. Thus, all the combinators are deterministic. Thus, one may view a CXACML policy as a function from requests to decisions with in place of . Further inspection establishes that is a total function. For two requests q = (s r a e) and q = (s r a e ), let q p q (for every policy p) if s s , r r , a a , and e e where is defined in Table 1. Theorem 5.3. CXACML is not safe with respect to . Proof. Consider the policy p shown in Example 5.1 and the requests q = (() ((name log)) () ()) and q = (((role dr)) ((name log)) () ()). Clearly q p q . Yet p (q) = permit deny = p (q ) Theorem 5.4. L CXACML has independent composition. Proof. A combination algorithm c and target t together determine a policy combinator. For each pair of values for c and t, the needed function t c : Q D D exists to provide the meaning of the policy (Policy c t p p ) and satisfy Equation 2. For Permit Overrides (when c = PermitOver, or PO for short), the function t PO (q)(d d ) is equal to na if q / t permit else if d = permit t PO (q)(d ) = permit deny else if d = deny t PO (q)(d ) = deny na otherwise The function t DO (q)(d d ) for Deny Overrides is equal to na if q / t deny else if d = deny t DO (q)(d ) = deny permit else if d = permit t DO (q)(d ) = permit na otherwise For First Applicable, t FA (q)(d d ) is 165 ( a 1 ) (l 1 ) ( a 2 ) (l 2 ) ( a 3 ) (l 3 ) ( a 4 ) (l 4 ) (( a 1 ) ( a 2 ) ( a 3 ) ( a 4 )) ((l 1 ) ( l 2 ) ( l 3 ) ( l 4 )) i s.t. l i ( a ) ( a ) (l 1 l 2 . . . l n ) i j s.t. a i = a j ( a 1 a 2 . . . a n ) ( a 1 a 2 . . . a m ) Table 1: The Match Relationship q / t q (Rule t f) na qt q (Rule t Permit) permit qt q (Rule t Deny) deny Table 2: on Rules q (Policy c t) na q / t q (Policy c t p ) na Table 3: Default na Inference Rules qt i s.t. q p i permit q (Policy PermitOver t p 1 p 2 . . . p n ) permit qt i s.t. q p i deny j, (q p j permit) q (Policy PermitOver t p 1 p 2 . . . p n ) deny qt i, q p i na q (Policy PermitOver t p 1 p 2 . . . p n ) na Table 4: Inference Rules for Permit Overrides qt i s.t. q p i deny q (Policy DenyOver t p 1 p 2 . . . p n ) deny qt i s.t. q p i permit j, (q p j deny) q (Policy DenyOver t p 1 p 2 . . . p n ) permit qt i, q p i na q (Policy DenyOver t p 1 p 2 . . . p n ) na Table 5: Inference Rules for Deny Overrides qt q p 1 permit q (Policy FirstApp t p 1 p 2 . . . p n ) permit qt q p 1 deny q (Policy FirstApp t p 1 p 2 . . . p n ) deny qt q p 1 na q (Policy FirstApp t p 2 , . . . , p n ) d q (Policy FirstApp t p 1 p 2 . . . p n ) d Table 6: Inference Rules for First Applicable na if q / t d else if d = permit d = deny t FA (q)(d ) otherwise where t PO ( ) = t DO ( ) = t FA ( ) = na for the empty sequence . Theorem 5.5. L CXACML is not monotonic. Proof. Consider the policy p in Example 5.1 and the policy p that would be p without the first rule. Let the request q be (((role dr)) ((name log)) () ()). p (q) = permit, but p (q) = deny. Thus, adding a rule to p results in a request going from being granted to not being granted. ADAPTATIONS OF FIRST-ORDER LOGIC Whereas XACML is an attempt to create a policy language from whole cloth, other languages are adaptations of first-order logic. Halpern and Weissman present several schemata for such languages [9]. Here we present and analyze the languages produced by two of their schemata, Lithium and L 5 . For the ease presentation, first, we define the language schemata FOL, a more readily identifiable restriction of first-order logic. To ensure efficiency (and decidability!), the languages of L 5 and Lithium use additional context-sensitive constraints to restrict FOL. We discuss these restrictions after giving a semantics to FOL. (The semantics of L 5 and Lithium will be the same as that of FOL, restricted to subsets of the language.) The schemata FOL is a restriction of many-sorted first-order logic. Each language of FOL corresponds to giving the logic a different vocabulary (the parameters including quantifier symbols, predicate symbols, constant symbols, and function symbols). We assume that includes the sorts S for subjects, R for resources, A for actions, and the predicate symbol Permit of the sort S R A {T, F}. 5 may also include sorts to represent environmental data such as the current time and location. 6.1 Syntax of FOL A standard policy under the vocabulary is an expression with one of the following forms: ( y 1 x 1 , . . ., y m x m ( 1 . . . n Permit(s, r, a))) ( y 1 x 1 , . . ., y m x m ( 1 . . . n Permit(s, r, a))) where each x i names a variable over the sort identified by y i , s is a term over the sort S, r is a term over the sort R, a is a term over the sort A, and each j is a literal over that may include the variables x 1 , . . . , x m . The policies of the language FOL() are the standard policies under and conjunctions of policies: P ::= StandardPolicy | (and P ) Example 6.1. Let the vocabulary contain 1. the sorts S = {amy, bob, joe}, R = {grades, courses}, and A = {assign, enroll}; 5 Halpern and Weissman treat the Permit predicate as taking two arguments, a subject and a resource-action, instead of three. 166 2. the predicates Permit : S R A {T, F}, faculty : S {T, F}, and student : S {T, F}. FOL( ) includes the following policy: (and ( S x (faculty(x) Permit(x, grades, assign))) ( S x (student(x) Permit(x, grades, assign))) ( S x ( faculty(x) Permit(x, courses, enroll)))) where S identifies the sort S. As the semantics will soon show, this policy has the same meaning as policy p from Section 2 does under Interpretation 3. The requests of FOL() have the form (s, r, a, e) where s S is the subject making the request; r R is the requested resource; a A is the action the subject would like to preform on the resource; and e is a conjunction of ground literals and universal formulas of the form y 1 x 1 , . . . , y m x m ( 1 . . . n n+1 ) where each x i names a variable over the sort identified by y i and each i is a literal over that may include the variables x l , . . . , x m . The expression e provides information about s, r, a, and the environment. Example 6.2. The four-tuple (bob, courses, enroll, student(bob) faculty(amy) student(amy)) is a request of FOL( ) where is defined in Example 6.1. 6.2 Semantics of FOL The semantics of a policy follows from interpreting it as a formula in many-sorted first-order logic. The policy combinator and becomes conjunction. The standard policies and e are interpreted as the corresponding logic formulas. A policy p defines a relation p between requests and {permit, deny} as follows: ( s, r, a, e) p permit iff p e Permit( s, r, a) ( s, r, a, e) p deny iff p e Permit(s, r, a) where is interpreted as the standard "proves" relation for many-sorted first-order logic over . To define a deterministic and total version of FOL, we expand the set of decisions to D = {na, permit, deny, error} and define p ((s, r, a, e)) to be error if pe Permit( s, r, a) and pe Permit(s, r, a) permit if pe Permit( s, r, a) and pe Permit(s, r, a) deny if pe Permit( s, r, a) and pe Permit(s, r, a) na if pe Permit( s, r, a) and pe Permit(s, r, a) Since a policy composed of sub-policies, each composed of standard policies, is semantically equivalent to a policy composed of all the standard policies without the intermediate sub-policies, we will henceforth treat all policies as either a standard policy or a conjunction of standard policies. 6.3 Analysis of FOL The language FOL() defines the deterministic and total policy language (P, Q, G, N, ). The syntax determines P and Q where Q may be treated as a set of requests since for all policies p and p , Q p = Q p . The semantics requires that G = {permit} and N = {na, deny, error}. The languages of FOL has the policy combinator and. FOL allows for explicit denials and checking for the explicit absence of attributes. Given two requests q = (s, r, a, e) and q = (s , r , a , e ), if s = s , r = r , or a = a , we consider the two requests incomparable . If s = s , r = r , and a = a , then we would like to order requests according to their information content. One might conclude that q p q if e = e. However, suppose e = , where is logical contradiction. Then e contains no information and yet it implies e. Similarly, if p e = , then e contains no information with respect to p. Thus, we define p as follows: Let (s, r, a, e) p ( s , r , a , e ) iff 1. s = s , r = r , and a = a ; and 2. p e implies p e but not , or p e implies . Theorem 6.3. FOL() is safe with respect to for any vocabulary . Proof. Assume FOL() is not safe. Then there must exist p P and q, q Q such that q p q and p (q) p (q ). Let q = (s, r, a, e) and q = (s, r, a, e ). Since permit is the only granting decision, p (q) = permit and thus p e Permit( s, r, a). Since N = {na, deny, error}, p (q ) must be either na, deny, or error. Since q p q , two cases arise: 1. pe implies pe but not : Since pe = pe and pe Permit( s, r, a), pe Permit( s, r, a). Thus, p (q ) is either permit or error. However, if p (q ) = error, then pe = , a contradiction. Furthermore, p (q ) = permit / N is also a contradiction. 2. p e implies : In this case, p (q) = error = permit, a contradiction. We can thus conclude that FOL() must be safe. Theorem 6.4. FOL() does not have independent composition for some . Proof. Consider the policy p a : (and ( S x, student(x) Permit(x, log, read)) ( S x, student(x) Permit(x, log, read))) the policy p b : (and ( S x, student(x) Permit(x, log, edit)) ( S x, student(x) Permit(x, log, edit))) and request q = (bob, log, read, T). On q, p a produces the decision of permit while its sub-policies yield na. However, p b produces the decision of na while its sub-policies also yield na on q. Thus, the required function and would have to satisfy permit = and (q)(na na) = na A contradiction, and hence and cannot exist. Theorem 6.5. FOL() is not monotonic for some . Proof. Consider the policy p c : (and ( S x, student(x) Permit(x, log, read)) ( S x, student(x) Permit(x, log, read))) with and without the second sub-policy, and the request (bob, log, read, student(bob)). In the absence of the second sub-policy, the decision is permit, whereas p c produces error. 167 6.4 Analysis of Lithium Halpern and Weissman restrict FOL to create the language they dub Lithium. 6 A slightly modified form follows. Lithium relies heavily on the notion of "bipolarity". A literal of f is labeled bipolar in f relative to the equality statements in e if the following holds: there exists a term in f and variable substitutions and such that it follows from e that = . Lithium also makes use of the notion of equality-safety. (p, e 0 , e 1 ) is equality-safe if 1. e 1 p when written in CNF (i.e., of the form c 1 . . . c n where each c j has the form y 1 x 1 , . . . , y m x m ( ) where is a qualifier-free disjunction of literals) has no clause with a disjunct of the form t = t , and 2. it is not the case that f 0 t = t where f 0 is the conjunction of the equality statements in e 0 , where t and t are closed terms such that (1) they both appear in e 0 ; and (2) either t is a sub-term of t , or both t and t mention function symbols. Like FOL, Lithium is a set of languages each with a different vocabulary. Let Li() be the instance of Lithium using the vocabulary . Li() has the same set of policies as FOL(). However, each policy p of Li() has a different set of requests for which it is defined (a different value for Q p ). A request (s, r, a, e 0 e 1 ) of FOL() is in Q p iff: 1. e 0 is a basic environment (a conjunction of ground terms), 2. e 1 is a conjunction of universally quantified formulas, 3. (p, e 0 , e 1 ) is equality-safe, and 4. every conjunct of e 1 p has at most one literal that is bipolar in e 1 p relative to the equality statements in e 0 . Lithium is safe since its requests are a subset of those of FOL. Theorem 6.6. Lithium does not have independent composition for some . Proof. Consider the policies p a and p b and the request given with them in the proof of Theorem 6.4. The request is in Q p a . To show this, we check that the request satisfies all four of the requirements for a request to be in Q p a given above. Since e 0 e 1 = T, the first three requirements hold. The last requirement holds since student(x) and student(x) are the only bipolars and they are each in a different conjunct. By similar reasoning, the request is also in Q p b . Thus, the proof follows as before. 6.5 Analysis of L 5 In their work, Halpern and Weissman define a further restriction of FOL, which they call L 5 . Like Lithium, L 5 () includes all the policies of FOL() with each policy having a different set of requests for which it is defined. For a policy p of L 5 (), Q p consists of all the requests (s, r, a, e) of FOL() such that: 6 The name Lithium only appears in the 2006 version of their work [9]. 1. e is a basic environment, 2. equality is not used in e or p, 3. for every atomic policy p in p, all variables appearing in p appears as an argument to Permit in p , and 4. there are no bipolars in p relative to the empty set of equality statements. As with Lithium, L 5 () is safe since it is a subset of a safe language. Halpern and Weissman have proven the following theorem (Proposition 4.2 in their updated document [9]): Theorem 6.7. Let p be a compound policy and (s, r, a, e) be a request of L 5 . Then e p Permit( s, r, a) iffthere is a sub-policy p of p such that e p Permit( s, r, a). Using the same approach as given in their proof, one can generalize this proof to include statements of the form ep Permit(s, r, a) also. Theorem 6.8. L 5 () has independent composition for all . Proof. Allowing p i to range over the sub-policies of p, the above result yields: p (q) = 8 &gt; &gt; &gt; &lt; &gt; &gt; &gt; : error i, j, p i (q) = permit p j (q) = deny permit else if i, p i (q) = permit deny else if i, p i (q) = deny deny otherwise From this, it is easy to construct an appropriate value for and (q)(d 1 , d 2 , . . . , d n ): error if i, j, d i = permit d j = deny permit else if i, d i = permit deny else if i, d i = deny deny otherwise Notice that and does not use the value of q: it merely composes the results from its sub-policies. A policy author concerned solely with expressive power would select Lithium over L 5 . However, the choice becomes more complicated when concerned about the ability to reason about policies, because only L 5 features independent composition. We hope that elucidating this trade-off with a combination of proof and illustrative examples, as we have done above, will help authors choose better between the policy languages they use, even when the languages are within the same family. RELATED WORK De Capitani di Vimercati et al. discuss explicit denial and how it introduces the need for policy combinators that reduce the clarity of the language [4]. The authors list various policy combinators that are possible, many of which are more complex than those we present. The paper includes discussion of a few policy languages, including XACML and a language grounded in first-order logic. The paper does not, however, attempt to systemically compare them. The work of Mark Evered and Serge B ogeholz concerns the quality of a policy language [5]. After conducting a case study of the access-control requirements of a health 168 information system, they proposed a list of criteria for policy languages. They state that languages should be concise, clear, aspect-oriented (i.e., separate from the application code), fundamental (i.e., integrated with the middleware, not an ad hoc addition), positive (i.e., lists what is allowed, not what is prohibited), supportive of needs-to-know, and efficient. Although they compare four languages based on these criteria, they do not formalize the criteria. Some authors have considered formal treatments of programming language expressiveness [6, 11]. Felleisen's is the closest in spirit to ours. His framework examines the ability to translate the features of one language in the other with only local transformations. That work does not, however, directly address reasoning. DISCUSSION This paper presents our analysis framework and its findings . Some differences between languages lie in the realms of decision sets, policy combinators, and checking for the absence of attributes, but these are clear from the language definitions. Our framework highlights the following more subtle, semantic differences: Independence Core XACML and L 5 feature independent composition of polices into compound policies, and thus allow for reasoning about a policy by reasoning about the sub-policies separately. Lithium, in contrast , does not exhibit this property, and therefore potentially requires reasoning about a policy all at once. Safety L 5 and Lithium provide safety for the most natural definition of the "contains more information" ordering. Core XACML, in contrast, does not, which implies that information missing from a Core XACML request could result in unintended access being granted. These differences are not orthogonal. Clearly, the combinators selected determine whether the language will have independent composition. Furthermore, implicit checking of attributes will result in the loss of safety. These properties may guide policy language designers. For example, suppose a designer wishes to create a safe variant of XACML. One way to achieve this would be to eliminate rules that deny access and thus the decision of deny. (We provide further details in an extended version of this work [15].) As noted in Section 5, the comparison framework must be generalized to treat language with more exotic constructs like obligations. More importantly, we need to perform user studies to determine whether, and how well, our properties correlate with policy comprehension by humans. Lastly, this framework should be coupled with one for measuring the expressive power of a policy language before fair judgment may be passed on languages. ACKNOWLEDGMENTS We thank Joe Halpern and Vicky Weissman for useful conversations and for sharing their ongoing work. We also thank Konstantin Beznosov, Kathi Fisler, and Steve Reiss. This work was partially supported by NSF grant CPA-0429492 to Brown University and by the Army Research Office through grant number DAAD19-02-1-0389 ("Perpetually Available and Secure Information Systems") to Carnegie Mellon Uni-versity's CyLab. REFERENCES [1] A. Anderson. Core and hierarchical role based access control (RBAC) profile of XACML, version 2.0. Technical report, OASIS, Sept. 2004. [2] M. Y. Becker and P. Sewell. Cassandra: Flexible trust management, applied to electronic health records. In IEEE Computer Security Foundations Workshop, pages 139154, 2004. [3] E. Bertino, P. Samarati, and S. Jajodia. Authorizations in relational database management systems. In ACM Conference on Computer and Communications Security, pages 130139, 1993. [4] S. De Capitani di Vimercati, P. Samarati, and S. Jajodia. Policies, models, and languages for access control. In Databases in Networked Information Systems: 4th International Workshop, volume 3433 of Lecture Notes in Computer Science. Springer-Verlag, Mar. 2005. [5] M. Evered and S. B ogeholz. A case study in access control requirements for a health information system. In Workshop on Australasian Information Security, Data Mining and Web Intelligence, and Software Internationalisation, pages 5361, 2004. [6] M. Felleisen. On the expressive power of programming languages. Science of Computer Programming, 17:3575, 1991. [7] K. Fisler, S. Krishnamurthi, L. A. Meyerovich, and M. C. Tschantz. Verification and change-impact analysis of access-control policies. In International Conference on Software Engineering, pages 196205, May 2005. [8] M. M. Greenberg, C. Marks, L. A. Meyerovich, and M. C. Tschantz. The soundness and completeness of Margrave with respect to a subset of XACML. Technical Report CS-05-05, Department of Computer Science, Brown University, Apr. 2005. [9] J. Halpern and V. Weissman. Using first-order logic to reason about policies. In IEEE Computer Security Foundations Workshop, pages 187201, 2003. Updated 2006 version available at http://www.citebase.org/ cgi-bin/citations?id=oai:arXiv.org:cs/0601034. [10] S. Jajodia, P. Samarati, V. S. Subrahmanian, and E. Bertino. A unified framework for enforcing multiple access control policies. In ACM SIGMOD International Conference on Management of Data, pages 474485, 1997. [11] J. C. Mitchell. On abstraction and the expressive power of programming languages. Science of Computer Programming, 212:141163, 1993. [12] OASIS. eXtensible Access Control Markup Language (XACML) version 2.0. OASIS Standard, Feb. 2006. [13] C. Powers and M. Schunter. Enterprise privacy authorization language (EPAL 1.2). W3C Member Submission, Nov. 2003. [14] R. S. Sandhu, E. J. Coyne, H. L. Feinstein, and C. E. Youman. Role-based access control models. IEEE Computer, 29(2):3847, 1996. [15] M. C. Tschantz and S. Krishnamurthi. Towards reasonability properties for access-control policy languages with extended XACML analysis. Technical Report CS-06-04, Department of Computer Science, Brown University, Apr. 2006. 169

common features;Access control;lithium;modularity;reasonability property;policy decomposition;properties;access control;policy combinator;XACML;comtemporary policy;access-control policy;policy language property;first order logic;xacml;multiple policy language;policy language;policy;security;formalize;policy languague

199

Tracking Dynamics of Topic Trends Using a Finite Mixture Model

In a wide range of business areas dealing with text data streams, including CRM, knowledge management, and Web monitoring services, it is an important issue to discover topic trends and analyze their dynamics in real-time.Specifically we consider the following three tasks in topic trend analysis: 1)Topic Structure Identification; identifying what kinds of main topics exist and how important they are, 2)Topic Emergence Detection; detecting the emergence of a new topic and recognizing how it grows, 3)Topic Characterization ; identifying the characteristics for each of main topics. For real topic analysis systems, we may require that these three tasks be performed in an on-line fashion rather than in a retrospective way, and be dealt with in a single framework. This paper proposes a new topic analysis framework which satisfies this requirement from a unifying viewpoint that a topic structure is modeled using a finite mixture model and that any change of a topic trend is tracked by learning the finite mixture model dynamically.In this framework we propose the usage of a time-stamp based discounting learning algorithm in order to realize real-time topic structure identification .This enables tracking the topic structure adaptively by forgetting out-of-date statistics.Further we apply the theory of dynamic model selection to detecting changes of main components in the finite mixture model in order to realize topic emergence detection.We demonstrate the effectiveness of our framework using real data collected at a help desk to show that we are able to track dynamics of topic trends in a timely fashion.

INTRODUCTION In a wide range of business areas dealing with text streams, including CRM, knowledge management, and Web monitoring services, it is an important issue to discover topic trends and analyze their dynamics in real-time.For example, it is desired in the CRM area to grasp a new trend of topics in customers' claims every day and to track a new topic as soon as it emerges.A topic is here defined as a seminal event or activity.Specifically we consider the following three tasks in topic analysis: 1) Topic Structure Identification; learning a topic structure in a text stream, in other words, identifying what kinds of main topics exist and how important they are. 2) Topic Emergence Detection; detecting the emergence of a new topic and recognizing how rapidly it grows, similarly, detecting the disappearance of an existing topic. 3) Topic Characterization; identifying the characteristics for each of main topics. For real topic analysis systems, we may require that these three tasks be performed in an on-line fashion rather than in a retrospective way, and be dealt with in a single framework. The main purpose of this paper is to propose a new topic analysis framework that satisfies the requirement as above, and to demonstrate its effectiveness through its experimental evaluations for real data sets. Our framework is designed from a unifying viewpoint that a topic structure in a text stream is modeled using a finite mixture model (a model of the form of a weighted average of a number of probabilistic models) and that any change of a topic trend is tracked by learning the finite mixture model dynamically.Here each topic corresponds to a single mixture component in the model. All of the tasks 1)-3) are formalized in terms of a finite mixture model as follows: As for the task 1), the topic structure is identified by statistical parameters of a finite mixture model.They are learned using our original time-stamp based discounting learning algorithm, which incrementally and adaptively estimates statistical parameters of the model by gradually forgetting out-of-date statistics, making use of time-stamps of data.This makes the learning procedure adaptive to changes of the nature of text streams. As for the task 2), any change of a topic structure is rec-ognized by tracking the change of main components in a mixture model.We apply the theory of dynamic model selection [7] to detecting changes of the optimal number of main components and their organization in the finite mixture model.We may recognize that a new topic has emerged if a new mixture component is detected in the model and remains for a while.Unlike conventional approaches to statistical model selection under the stationary environment, dynamic model selection is performed under the non-stationary one in which the optimal model may change over time.Further note that we deal with a complicated situation where the dimension of input data, i.e., the number of features of a text vector, may increase as time goes by. As for the task 3), we classify every text into the cluster for which the posterior probability is largest, and then we characterize each topic using feature terms characterizing texts classified into its corresponding cluster.These feature terms are extracted as those of highest information gain, which are computed in real-time. We demonstrate the validity of the topic trend analysis framework, by showing experimental results on its applications to real domains.Specifically we emphasize that it is really effective for discovering trends in questions at a help desk. 1.2 Related Work The technologies similar to 1)-3) have extensively been ex-plored in the area of topic detection and tracking (TDT) (see [1]).Actually 1) and 2) are closely related to the subprob-lems in TDT called topic tracking and new event detection, respectively.Here topic tracking is to classify texts into one of topics specified by a user, while new event detection, formerly called first story detection, is to identify texts that discuss a topic that has not already been reported in earlier texts.The latter problem is also related to work on topic-conditioned novelty detection by Yang et.al.[16]. In most of related TDT works, however, topic tracking or new event detection is conducted without identifying main topics or a topic structure, hence the tasks 1)-3) cannot be unified within a conventional TDT framework.Further topic timeline analysis has not been addressed in it. Swan and Allen [12] addressed the issue of how to auto-matically overview timelines of a set of news stories.They used the -method to identify at each time a burst of feature terms that more frequently appear than at other times. Similar issues are addressed in the visualization community [3].However, all of the methods proposed there are not designed to perform in an on-line fashion. Kleinberg [4] proposed a formal model of "bursts of ac-tivity" using an infinite-state automaton.This is closely related to topic emergence detection in our framework.A burst has a somewhat different meaning from a topic in the sense that the former is a series of texts including a specific feature, while the latter is a cluster of categorized texts.Hence topic structure identification and characterization cannot be dealt with in his model.Further note that Kleinberg's model is not designed for real-time analysis but for retrospective one. Related to our statistical modeling of a topic structure, Liu et.al. [2] and Li and Yamanishi [6] also proposed methods for topic analysis using a finite mixture model.Specifically , Liu et.al. considered the problem of selecting the optimal number of mixture components in the context of text clustering.In their approach a single model is selected as an optimal model under the assumption that the optimal model does not change over time.Meanwhile, in our approach, a sequence of optimal models is selected dynamically under the assumption that the optimal model may change over time. Related to topic emergence detection, Matsunaga and Yamanishi [7] proposed a basic method of dynamic model selection , by which one can dynamically track the change of number of components in the mixture model.However, any of all of these technologies cannot straightforwardly be applied to real-time topic analysis in which the dimension of data may increase as time goes by. Related to topic structure identification, an on-line discounting learning algorithm for estimating parameters in a finite mixture model has been proposed by Yamanishi et. al.[14]. The main difference between our algorithm and theirs is that the former makes use of time-stamps in order to make the topic structure affected by a timeline of topics while the latter considers only the time-order of data ignoring their time-stamps. The rest of this paper is organized as follows: Section 2 describes a basic model of topic structure.Section 3 gives a method for topic structure identification.Section 4 gives a method for topic emergence detection.Section 5 gives a method for topic characterization.Section 6 gives experimental results.Section 7 gives concluding remarks. MODEL We employ a probabilistic model called a finite mixture model for the representation of topic generation in a text stream.Let W = {w 1 , , w d } be the complete vocabulary set of the document corpus after the stop-words removal and words stemming operations.For a given document x, let tf(w i ) be the term frequency of word w i in x.Let idf(w i ) be the idf value of w i , i. e. , idf(w i ) = log(N/df(w i )) where N is the total number of texts for reference and df(w i ) is the frequency of texts in which w i appears.Let tf-idf(w i ) be the tf-idf value of w i in x, i. e. , tf-idf(w i ) = tf(w i ) log(N/df(w i )). We may represent a text x of the form: x = (tf(w 1 ), ..., tf (w d )) or x = (tf-idf(w 1 ), ..., tf -idf(w d )). We may use either type of the representation forms. Let K be a given positive integer representing the number of different topics.We suppose that a text has only one topic and a text having the i-th topic is distributed according to the probability distribution with a density: p i (x| i ) (i = 1, 2, , K), where i is a real-valued parameter vector.We suppose here that x is distributed according to a finite mixture distribution (see e.g., [8]) with K components given by p(x| : K) = K X i=1 i p(x| i ), (1) where i &gt; 0 (i = 1, , K) and P K i=1 i = 1. We set = ( 1 , , K-1 , 1 , , K ).Here i denotes the degree to what the i-th topic is likely to appear in a text stream. Note that each component in the mixture defines a single cluster in the sense of soft-clustering. Throughout this paper we suppose that each p i (x| i ) takes a form of a Gaussian density: Letting d be the dimension of 812 Industry/Government Track Poster Text Data Stream Discount Learning Finite Mixture Model 1 Dynamic Model Selection Topic Emergence Detection Topic Characteriztion Timeline of Topics Discount Learning Discount Learning Finite Mixture Model 2 Finite Mixture Model K Figure 1: Topic Trend Analysis System each datum, p i (x| i ) = i (x| i , i ) (2) = 1 (2) d/2 | i | 1/2 exp ,, - 12(x - i ) T -1 i (x - i ) , where i is a d-dimensional real-valued vector, i is a d d-dimensional matrix, and we set i = ( i , i ).In this case (1) is so-called a Gaussian mixture.Note that a Gaussian density may be replaced with any other form of probability distributions, such as a multinomial distribution. In terms of a finite mixture model, a topic structure is identified by A) the number of components K (how many topics exist), B) the weight vector ( 1 , , K ) indicating how likely each topic appears, and C) the parameter values i (i = 1, , K) indicating how each topic is distributed.A topic structure in a text stream must be learned in an on-line fashion.Topic emergence detection is conducted by tracking the change of main components in the mixture model. Topic characterization is conducted by classifying each text into the component for which the posterior is largest and then by extracting feature terms characterizing the classified texts.Topic drift may be detected by tracking changes of a parameter value i for each topic i.These tasks will be described in details in the sessions to follow. The overall flow of the tasks is illustrated in Figure 1. A text is sequentially input to the system.We prepare a number of finite mixture models, for each of which we learn statistical parameters using the time-stamp based learning algorithm to perform topic identification.These tasks are performed in parallel.On the basis of the input data and learned models, we conduct dynamic model selection for choosing the optimal finite mixture model.We then compare the new optimal model with the last one to conduct topic emergence detection.Finally for each component of the optimal model, we conduct topic characterization. TOPIC STRUCTURE IDENTIFICATION WITH DISCOUNTING LEARNING In this section we propose an algorithm for learning a topic structure, which we call an time-stamp based discounting topic learning algorithm. The algorithm is basically designed as a variant of the incremental EM algorithm for learning a finite mixture model (see, e.g., Neal and Hinton [9]). Our proposed one is distinguished from existing ones with regards to the following three main features: 1) Adaptive to the change of the topic structure. The parameters are updated by forgetting out-of-date statistics as time goes on.This is realized by putting a larger weight to the statistics for a more recent data. 2) Making use of time stamps for texts. Not only the time order of texts but also their time stamps are utilized to make the topic structure depend on the timeline.For example, for two text data x t 1 , x t 2 (t 1 &lt; t 2 ), if the length t 2 -t 1 is larger, the topic structure learned at time t 2 will be less affected by that at time t 1 . 3) Normalizing data of different dimensions. We consider the on-line situation where the dimension of a datum may increase as time goes by.This situation actually occurs because new words may possibly be added to the list of words every time a new text is input.Hence it is needed for normalizing data of different dimensions. We suppose that text data x 1 , x 2 , ... are given in this order, and each has a time-stamp indicating when it appeared .Here is a description of the algorithm, in which is a discounting parameter, i denotes the posterior density of the ith component, and m is introduced for calculation of weights for old statistics. Time-stamp Based Discounting Learning Algorithm Initialization: Set initial values of (0) i , (0) i , (0) i , m (0) (i = 1, , k).Let &gt; 0, 0 &lt; &lt; 1 be given. Iteration: For t = 1, 2, .. do the following procedure. For the t-th data be x t and its time stamp be t new .Let the time stamp of the (t - 1)-th data be t old . For i = 1, , k, update the parameters according to the 813 Industry/Government Track Poster following rules: p(i|x t ) := (t-1) i p i (x t | (t-1) i , (t-1) i ) P k l=1 (t-1) l p l (x t | (t-1) l , (t-1) l ) (t) i := WA(p(i|x t ), 1/k|1, ) (t) i := WA( (t) i , (t) i |m (t-1) , -(t new -t old ) ) (t) i := WA( (t-1) i , x t | i m (t-1) , -(t new -t old ) (t) i ) (t) i := WA( (t-1) i , x i x T i | i m (t-1) , -(t new -t old ) (t) i ) (t) i := (t) i - i T i m (t) := (t new -t old ) m (t-1) + 1, where WA denotes the operation such that WA(X, Y |A, B) = A A + B X + B A + B Y. Generally, we set the initial value (0) i = 1/K, m (0) = 0, a small value to (0) i , and set (0) i the first x t s that are different each other.This algorithm updates i , i , and i as the weighted average of the latest parameter value and the new statistics.The weight ratio is m (t-1) : -(t new -t old ) for i , and i m (t-1) : -(t new -t old ) (t) i for i and i , respectively. Note that Yamanishi et.al.'s sequentially discounting learning algorithm [14] can be thought of as a special case of this algorithm in which the time interval t l+1 - t l is independent of l.In that case if we further let = 1, the algorithm becomes an ordinary incremental EM algorithm. In real implementation, we supposed that i is a diagonal matrix for the sake of computational complexity issues.The scalability issue for dealing with a general matrix i remains for future study. TOPIC EMERGENCE DETECTION WITH DYNAMIC MODEL SELECTION In this section we are concerned with the issue of topic emergence detection, i.e., tracking the emergence of a new topic.We reduce here this issue to that of selecting the optimal components in the mixture model dynamically.We call this statistical issue dynamic model selection (see also [7]). The key idea of dynamic model selection is to first learn a finite mixture model with a relatively large number of components , then to select main components dynamically from among them on the basis of Rissanen's predictive stochastic complexity [10]. The procedure of dynamic model selection is described as follows: Initialization: Let K max (maximum number of mixture components) and W (window size) be given positive integers. Set initial values of (0) i , (0) i = ( (0) i , (0) i ) (i = 1, , K max ). Iteration: For t = 1, 2, , do the following procedure 1 to 4: 1. Model Class Construction: Let G t i be the window average of the posterior probability ( t-W i + + t i )/W .For k = 1, , K max , do the following procedure: Let 1 , , k be the indices of k highest scores such that G (t-1) 1 G (t-1) k .Construct the following mixture model with k components: For s = t - W, , t, p (t-1) (x| 1 , , k ) = k-1 X j=1 (t-1) j p j (x| (t-1) j ) + 1 k -1 X j=1 (t-1) j ! U. where U is a uniform distribution over the domain. 2. Predictive Stochastic Complexity Calculation: When the t-th input data x t with dimension d t is given, compute S (t) (k) = t X s=t-W "-logp (s) (x s | 1 , , k )/d s " . (3) 3. Model Selection: Select k t minimizing S (t) (k).Let p j (x| (t-1) j ) (j = 1, , k t ) be main components at time t, which we write as {C (t) 1 , C (t) k t }. 4. Estimation of Parameters: Learn a finite mixture model with K max components using the time-stamp based discounting learning algorithm.Let the estimated parameter be ( (t) 1 , , (t) K max , (t) 1 , , (t) K max ). Note that the S (t) (k) can be thought of as a variant of Rissanen's predictive stochastic complexity [10] normalized by the dimension for each datum, which can be interpreted as the total code length required for encoding a data stream x t-W , ..., x t into a binary string sequentially. Once main components C (t) 1 , , C (t) k t are obtained, we compare them with C (t-1) 1 , , C (t-1) k t-1 to check the emergence of a new topic or the disappearance of an existing topic in the following way.If a new component is selected at some point and remains for a longer time than a specified threshold, we may determine that a new topic has emerged. Specifically, if the optimal number k t of components becomes larger than k t-1 , we can recognize that a new topic has emerged.Similarly, if an existing component is not selected at some time and does not appear any longer, then we may determine that the topic has disappeared. TOPIC CHARACTERIZATION WITH INFORMATION GAIN Once the optimal finite mixture model is obtained, we are concerned with the issue of how to characterize each topic.We address this issue by extracting terms characterizing each topic and by observing the growth or decay of each topic component.Details are shown below. A) Extracting terms characterizing each topic. We attempt to characterize each topic by extracting characteristic words for it.We perform this task by computing the information gain of possible words. In the time-stamp based discounting topic learning algorithm , the posterior probability distribution over the set of clusters is estimated every time a text data is input.According to that posterior distribution an input text will be 814 Industry/Government Track Poster categorized into the component for which the posterior probability is largest.This clustering task can be performed in an on-line fashion. After observing the t-th datum x t , for i = 1, , k, let S t (i) be the set of texts in x t = x 1 , .., x t classified into the i-th component and let t i be the size of S t (i).Let S t = k i=1 S t (i). Below we show the method for computing the information gain of a term w for each topic component.For any term w, let S(w) be a set of vectors in S t such that the frequency of w is larger than a given threshold, and let m w be the size of S(w).Let S( w) be a set of vectors in S t such that the frequency of w is not larger than the threshold, and m w be the size of S( w). For a specified topic component, say, the i-th component, let m + w be the number of vectors in S(w) that are also included in S t (i).Let m + w be the number of vectors in S( w) that are also included in S t (i). Then we define the information gain of w for the i-th topic component as follows: IG(w|i) = I(t, t i ) - `I(m w , m + w ) + I(m w , m + w ) , where I(x, y) is the information measure such as stochastic complexity [10], extended stochastic complexity [13][5].The stochastic complexity [10] is given as follows: I(x, y) = xH " y x " + 1 2 log " x 2 " , where H(x) = -x log x - (1 - x) log(1 - x) log(1 - x) is the binary entropy function, and log's base is 2.A special case of extended stochastic complexity is given as follows [13][5]: I(x, y) = min{y, x - y} + cpx log x, where c is a constant. We select a specified number of terms ws of largest information gains.We can think of them the set of terms characterizing the i-th topic component. The statistics: m w , m + w , m w , m + w needed for computing information gain can be calculated in an on-line fashion. Hence topic characterization task is conducted in real-time. B) Observing the growth or decay of each cluster. Let G (t) i be the window average of the posterior probability of the ith topic component, that is, G t i = ( t-W i +, , + t i )/W . G (t) i increases when texts corresponding to the i-th topic is input, and decreases when the other is input.We can see how rapidly this topic grows by observing G (t) i as t goes by. EXPERIMENTAL RESULTS We conducted an experiment on real data: contact data of a help desk for an internal e-mail service.An example of the data record is presented in Table 1.It has the field of contact date/time, question/request, answered date/time, answer, and so on.The number of the records is 1202.The date of the first/last contact is Feb 21 2004/May 20 respectively. We input contact dates as the time-stamps, and ques-tions/requests as the text data to our system.We set K max to 50 and to 0.99.Our system ran on an NEC Express5800 with 1GHz Pentium III and a 1GB memory.The system was implemented using C, and OS was Windows 2000 Server. Processing 1202 records of data took about five minute. Figure 2 shows the number of components k t selected by our system as main topics .The number increases at the Date k *t Figure 2: Number of main topics beginning of March, and has a peak in the middle of April. Since a fiscal year begins at April in Japan, we can suppose that the number of topics at the help desk is increasing around the first day of April. Let us look into a few of the components, because we do not have enough space for all of the components.Here, we observe Component 27 and 42 in detail.Figure 3 shows the window averages G 27 , G 42 of the posterior probabilities and the periods where the components are selected as main topics . G 27 increases in the beginning of April and has the first G 27 G 42 C27 is main. C42 is main. Figure 3: G i and Period of Component 27, 42 peak at April 12.Then it repeats increase and decrease until the middle of May.The corresponding component is selected as main from the first week of April, and remains as main until the middle of May (with short discontinuances). G 42 is positive during April, and also the corresponding topic is main during April.The lines of G i s indicate how important the corresponding topics are in each time.Moreover, we can observe how the emerged topics grows and disappears from the figure.The topic corresponding to Component 42 emerges at the beginning of April, grows for two weeks, is attenuated, then drops out from the main topics at the end of April. Term "transfer" was extracted as a characteristic word 815 Industry/Government Track Poster Table 1: Examples of help desk data records Contact date/time Question/Request Answered date/time Answer ,,, Feb 26 2004 14:05 I forgot my password. How can I ... Feb 26 2004 14:48 You can get a new ... Feb 26 2004 14:08 Until what time is an account for a ... Feb 26 2004 14:25 It is valid for 14 days after retirement. Feb 26 2004 14:09 Is it possible to forward mails from ... Feb 26 2004 15:09 Yes. You can set up by .... .... .... .... .... for Component 27.Texts classified into this component are questions like "Is it possible to use Service XXX after I am transfered to YYY?".That kind of questions may increase around the beginning of a fiscal year."Service ZZZ" and "failure" were extracted as chracteristic words for Component 42.Actually, Service ZZZ failed in the beginning of April, then, the topic consists of related complaints and questions. In this way we can recognize the emergence, growth, and decay of each topic from the system.Through this example it has turned out that our framework for topic trend analysis are very effective for tracking dynamics of topic trends in contact data at a help desk. CONCLUSION AND FUTURE STUDY In this paper we have proposed a framework for tracking dynamics of topic trends using a finite mixture model.In this framework the three main tasks: topic structure identification , topic emergence detection, and topic characterization are unified within a single framework.Topic structure identification has been realized by our unique time-stamp based learning algorithm.It enables tracking topic structures adaptively by forgetting out-of-date statistics.Topic emergence detection has been realized on the basis of the theory of dynamic model selection.It enables detecting changes of the optimal number of components in the finite mixture model to check whether a new topic has appeared or not.Topic characterization has been realized by on-line text clustering and feature extraction based on information gain.Through the experiments using real data collected at a help desk, it is demonstrated that our framework works well in the sense that dynamics of topic trends can be tracked in a timely fashion. The following issues remain open for future study: Context-based topic trend analysis: In this paper we have proposed an approach to word-based topic trend analysis. However, we need to further analyze contexts, i.e., relations among words, in order to more deeply analyze the semantics of topics. Multi-topics analysis: We supposed that one text comes from a single mixture component corresponding to a single topic.It is our future study how to deal with texts having multi topics. REFERENCES [1] J.Allen, R.Papka, and V.Lavrenko: On-line new event detection and tracking, in Proceedings of SIGIR International Conference on Information Retrieval, pp:37-45, 1998. [2] X.Liu, Y.Gong, W.Xu, and S.Zhu: Document clustering with cluster refinement and model selection capabilities, in Proceedings of SIGIR International Conference on Information Retrieval, pp:191-198, 2002. [3] S.Harve, B.Hetzler, and L.Norwell: ThemeRiver: Visualizing theme changes over time, in Proceesings of IEEE Symposium on Information Visualization, pp:115-123, 2000. [4] J.Kleiberg: Bursty and hierarchical structure in streams, in Proceedings of KDD2002, pp:91-101, ACM Press, 2003. [5] H.Li and K.Yamanishi: Text classification using ESC-based decision lists, Information Processing and Management, vol.38/3, pp:343-361, 2002. [6] H.Li and K.Yamanishi: Topic analysis using a finite mixture model, Information Processing and Management, Vol.39/4, pp 521-541, 2003. [7] Y.Matsunaga and K.Yamanishi: An information-theoretic approach to detecting anomalous behaviors, in Information Technology Letters vol.2 (Proc. of the 2nd Forum on Information Technologies), pp:123-124, (in Japanese) 2003. [8] G.McLahlan and D.Peel: Finite Mixture Models, Wiley Series in Probability and Statistics, John Wiley and Sons, 2000. [9] R.M.Neal and G.E.Hinton: A view of the EM algorithm that justifies incremental sparse, and other variants, Learning in Graphical Models, M. Jordan (editor), MIT Press, Cambridge MA, USA. [10] J.Rissanen: Universal coding, information, and estimation, IEEE Trans. on Inform. Theory, 30:629-636, 1984. [11] R.Swan and J.Allen: Extracting significant time-varying features from text, in Proceedings of 8th International Conference on Information Knowledge Management, pp:38-45, 1999. [12] R.Swan and J.Allen: Automatic generation of overview timelines, in Proceedings of SIGIR International Conference on Information Retrieval, pp:49-56, 2000. [13] K.Yamanishi: A Decision-theoretic Extension of Stochastic Complexity and Its Applications to Learning, IEEE Trans. on Inform. Theory, vol.44/4, pp:1424-1439, 1998. [14] K.Yamanishi, J.Takeuchi, G.Williams, and P.Milne: On-line unsupervised outlier detection using finite mixtures with discounting learning algorithms," in Proceedings of KDD2000, ACM Press, pp:320324 2000. [15] Y.Yang, T.Pierce, J.G.Carbonell: A study on retrospective and on-line event detection, in Proceedings of SIGIR International Conference on Information Retrieval, pp:28-30, 1998. [16] Y.Yang, J.Zang, J.Carbonell, and C.Jin: Topic-conditioned novelty detection, in Proceedings of KDD 2002, pp:688-693, 2002. 816 Industry/Government Track Poster

finite mixture model;CRM;time-stamp based discounting learning algorithm;topic structure identification;topic characterization;topic detection and tracking;time-stamp based learning algorithm;Topic Structure Identification;topic emergence detection;text mining;Topic Emergence Detection;tracking dynamics;dynamic model selection;Data Mining;information gain;topic trends;Topic Characterization;text data streams;model selection;topic trend;topic analysis

2

A Case Study on How to Manage the Theft of Information

This paper shows the importance that management plays in the protection of information and in the planning to handle a security breach when a theft of information happens. Recent thefts of information that have hit major companies have caused concern. These thefts were caused by companies' inability to determine risks associated with the protection of their data and these companies lack of planning to properly manage a security breach when it occurs. It is becoming necessary, if not mandatory, for organizations to perform ongoing risk analysis to protect their systems. Organizations need to realize that the theft of information is a management issue as well as a technology one, and that these recent security breaches were mainly caused by business decisions by management and not a lack of technology.

INTRODUCTION After counter-terrorism and counter-intelligence, cyber crime is the third highest priority for the U.S. Federal Bureau [4]. With the rise of the theft of information and the lure of big profits for this stolen information, it is necessary for information systems to have the ability to protect this valuable asset. It is estimated that a credit card number unsupported by any other documentation is worth $10, and a credit history report retails for $60 [2]. Recent breaches of information systems that have lead to thefts of information have shown that management practices and not technology was part of the issue and in some cases the primary cause of the theft of the information. With each of these thefts, there is a third party committing a crime, but in each case, risk analysis could have been used to avoid or to help mitigate the theft. It is becoming a necessity that companies examine their business practices and company policies to avoid risks associated with information stealing. The solution to information stealing does not reside in technology alone but also requires an understanding by management of the business and the risks associated with it. This paper examines the theft of information from different companies in order to explain the short coming of management practices that lead to the theft. . CASE STUDIES In May of 2005, Citigroup lost computer tapes that were being sent to the credit bureau via UPS that included Social Security numbers and payment history information for 3.9 million customers. After this event, this New York based company has decided that it will start sending its data to the credit bureau electronically using encryption [8]. Citigroup should have learned a lesson from Time Warner who lost a shipment of backup tapes that contained personal information of 600,000 employees that was being sent to an offsite data storage company in March of 2005 [9]. But the question remains, why was Citigroup sending sensitive information unsecured? Why did they not encrypt the data in the first place, and why did they realize that these tapes could get lost or stolen as evident to what happened with Time Warner? The answer is because they did not correctly identify the risk. Citigroup believed that UPS was a secure method for sending this information and that the data would be difficult to retrieve off the tapes because of the hardware needed to read the tapes. Citigroup needed to evaluate this risk of properly protecting confidential information while in transmission. Now, Citigroup has the issue of dealing with the negative public associated with this event, and the loss of any potential customers/revenue it will lose because of 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. Information Security Curriculum Development (InfoSecCD) Conference '05, September 23-24, 2005, Kennesaw, GA, USA. Copyright 2005 ACM 1-59593-261-5/05/0009...$5.00. 135 it. This issue would have been avoided if Citigroup would have properly identified this risk and taken the steps to protect this information. If the tapes were lost and the data was encrypted, then this story would have never happened. 2.2 Case II: ChoicePoint Choicepoint has made more than 50 acquisitions since 1997 to make it one of the largest collections of personal data in the United States. Choicepoint sells data "to clients doing background checks on job and loan applicants and conducting criminal investigations" [10]. On February 16, 2005, ChoicePoint went public to tell 145,000 people that identity thieves may have gained access to their personal information including their Social Security numbers and credit reports. "Authorities believe it was the work of a group of people who used IDs stolen from legitimate business people to set up phony businesses that contracted with ChoicePoint for ID checks, Bernknopf (ChoicePoint's spoke person) said" [5]. With ChoicePoint's security incident, there was no firewall hacked, or an IDS fooled. This was a deceptive scheme that took advantage of security holes in the business process. ChoicePoint's CISO, Rich Baich, stated "The mislabeling of this event as a hack is killing ChoicePoint. It's such a negative impression that suggests we failed to provide adequate protection. Fraud happens everyday. Hacks don't" [10]. ChoicePoint seemed to push that they were the victims of fraud, and not at fault. The bottom line is that confidential information was stolen, and the individuals who had their information stolen do not care if it was hacker or if the company was a victim of fraud. ChoicePoint failed to identify holes in the business process to allow this event to happen. Which if someone hacked into their system, it would have lead to the same result, the theft of information. ChoicePoint needs to recognize that identifying risks with their business process is just as important as securing their information system from an external hacker. 2.3 Case III: Egghead.com Egghead Software was a company that opened in 1984 to sell computer hardware and software that grew to have more than 205 stores worldwide. Then in 1998 the company moved its business to the internet as Egghead.com. In December of 2000, Egghead.com stated that "a hacker has breached its computer system and may have gained access to its customer database" [6]. Jerry Kaplan, Egghead.com's co-chairman , stated that there was "no evidence" to support that the database with the credit card numbers for its customer was stolen but, he also could not give confirmation that they were not stolen. "Egghead's inability to determine how many of it's customers credit cards had been compromised may mean that the company does not have a real-time auditing system in place, said Paul Robertson, senior developer for security service firm TruSecure Corp. `If you don't know how many credit-card numbers you lost, you are giving a quick, blanket, worst-case answer--and then finding out what happened afterwards,' he said." [1]. The way that Egghead.com handled its security incident showed that they did not have a good plan to manage the theft of information, and it appeared as if they made the plan to handle this situation as it happened. This lack of planning and risk analysis by management caused Egghead.com's business to suffer tremendously. Shortly thereafter this event, Egghead.com went into bankruptcy, and on November 26, 2001, Amazon.com acquired Egghead.com's assets in the Bankruptcy Court [6]. It appears the inability for Egghead.com to successful determine with certainty the extent of information stolen caused more damage to the company's reputation then the actual event itself. If Egghead.com had a well developed incident response plan in place to handle this security breach and a way to handle the media that followed, Egghead.com may have been able to weather the storm and stay in business. But all customer confidence was lost and Egghead.com was not able to recover. 2.4 Case IV: New Jersey Crime Ring Bank employees for Wachovia Corporation, Bank of America Corporation, Commerce Bancorp Inc., and PNC Bank stole information on 676,000 customer accounts that are all New Jersey residents. It is considered the largest banking security breach in history by the U.S. Department of the Treasury. "The suspects pulled up the account data while working inside their banks, then printed out screen captures of the information or wrote it out by hand, Lomia (a New Jersey Police Detective) said. The data was then provided to a company called DRL Associates Inc., which had been set up as a front for the operation. DRL advertised itself as a deadbeat-locator service and as a collection agency, but was not properly licensed for those activities by the state, police said" [13]. With this security breach, there was no technology involved. No hackers breached the information system. This was completely an inside job. The question becomes of how this could have been prevented? The answer is that in some cases the theft of information can not be prevented. The only the thing that management can do is be prepared for when it does happen. Because of incidents like this, it is becoming a duty of management to have an incident response plan in place long before a security breach happens. From a risk analysis viewpoint, an incident like this is difficult to detect and almost impossible to stop before it happens. But when it does happen and the criminals are caught, it becomes a necessity to punish the ones responsible to the full extent of the law to deter others from following suit. 2.5 Case V: LexisNexis LexisNexis is provider of legal and business data. In March of 2005, LexisNexis announced that the information on 32,000 people was stolen. These breaches occurred at one of the subsidiary companies, Seisint Inc. Seisnt Inc. was the company who was the provider of data to the Multistate Anti-Terrorism Information Exchange (MATRIX) system. "LexisNexis, which acquired Seisint of Boca Raton, Florida, in September for $775 million, expressed regret over the incident and said that it is notifying the individuals whose information may have been accessed and will provide them with credit-monitoring services" [12]. In this incident, hackers stole username and passwords of legitimate users to access the confidential information. In a statement, "Kurt Sanford, president and CEO of LexisNexis Corporate and Federal Markets, said that the company will improve the user ID and password administration procedures that its customers use and will devote more resources to protecting user's privacy and reinforcing the importance of privacy" [12]. This security breach is very similar to the incident that happened at ChoicePoint who is one of LexisNexis's competitors. 136 There are several policies that should have been implemented that could have reduced the risk of this security breach. Since LexisNexis gives third parties access to its confidential information, there becomes a need to educate these organizations on certain practices to protect the data. Where was this education, and was there a lack of education due to the possible effect that it could have on business? Also, what was the password policy for its customers? LexisNexis has not elaborated on the details of the security breach, but considering the statement of the CEO of LexisNexis after the incident, there clearly seems that there was a failure to detect the risk associated with their customer's password policy that could result in a theft of information. LexisNexis inability to properly assess this risk caused the security breach. Through education and a secure password administration policy, this event could have been avoided. RESULTS AND DISCUSSION When analyzing these case studies, an important thing to ponder is that for every security breach reported, how many go unreported? These security breaches could have been avoided with proper risk assessment and risk analysis, or at least the probability of a security breach could have been reduced greatly. For all security breaches, the prevention or at least the reduction of the probability of the security breach begins and ends with decisions that management makes. In an organization, when a security breach occurs it causes a company to re-evaluate their policies that guide their information security. With this rash of security incidents that have recently taken place, companies do not need to wait until a security breach happens to evaluate their security policies and analyze their risks. Companies need to have an ongoing risk analysis that is continually developed and re-developed. They need policies that are ever changing to meet new threats and new security weaknesses from a both business practices and technology viewpoints. Looking at the incidents that happened at ChoicePoint, LexisNexis, and Citigroup, these companies have technological solutions to protect their data from being stolen, but they failed at weighing equal importance the security of the data from a business issue perspective. This showed in their inability to properly evaluate the risk in the business practices. In several of the cases, the theft of information occurred because of the business practices of the company, and technology was not even involved. Also, companies need to learn from the mistakes of others because history will repeat itself if the lesson is not learned. There is an age old saying that is a wise person learns from their mistakes, but an even wiser person learns from the mistakes of others. Citigroup needed this advice. With Citigroup's loss of their backup tapes, they should have learned from the mistake that Time Warner made just months earlier, but they did not. Security policies and practices need the flexibility to change, and management has a responsibility to make these changes when new threats or new weakness surface so that they can protect their data. Companies and organizations need to realize the importance of making information security a business issue as well as a technological one. With the issue that happened with Egghead.com, they did have security systems in place to protect their data from being stolen, "but it lacked the kind of coordinated organizational response necessary to convince customers and shareholders that their sensitive data were actually secure." Egghead.com lost 25% of its stock value when their customer data was stolen [7]. Egghead.com was not ready for the media storm that followed the security breach which ultimately caused their collapse. By making information security a business issue, as well as a technological one, companies can add strategic, operational, and organizational defenses to protect their data. CONCULSIONS As more identity thefts occur, companies that make their money from storing this information are going to become liable. " `The ChoicePoint scandal has been a wake up call for how vulnerable consumers are to identity theft because of the lack of security standards for the largely unregulated information broker industry,' said Gail Hillebrand, Senior Attorney for Consumers Union's West Coast Office. `This bill will ensure that information brokers are held accountable for enforcing tough security practices to prevent thieves from gaining access to sensitive consumer data. And it gives consumers important new rights to examine the information maintained about them and to correct any errors they may find' [3]. Companies need to find the importance of protecting their data from both technology and business practices weaknesses. Companies view the protection of their data from a technology issue, but fail to realize the importance that management plays in protecting their systems with the creation of policies and understanding the risks that face their information systems. From a consumer standpoint, if a company is making profit from someone's personal information and they fail to protect this data, should they not give some sort of reputation? Companies own and manage consumer information, and individuals have little power over their information that is controlled by these organizations. As identity theft continues and companies fail at protecting their data, legislation will be passed that will force companies to comply with regulator standards that may force companies to give this reputation to individuals who have their identity stolen. Today, there are only laws to protect data in certain industries. This includes the Health Insurance Portability and Accountability Act for healthcare and the Gramm-Leach-Bliley Act for financial services. With consumer groups voicing their opinions regarding the theft of information from companies, the US Congress and other state legislators are getting prepared to pass broader data privacy protection to protect consumers [11]. There are steps that companies and organizations need to take to protect themselves from the theft of information. First, companies need to be prepared when a security breach occurs because a risk to an asset is never zero percent. Organizations need to establish policies and risk assessments that protect their data from both technology risks and business practices well before a security breach occurs. This is achieved by companies having the organizational structure that allows management to fully understand the business processes and technology that expose their information systems to threats. Also, companies need the ability to change and adapt to new threats that oppose their information. It is not possible to prevent all security breaches that lead to a theft of information, but companies will need to have policies and practices in place to better protect the 137 data. Companies will need not only to weigh technology risk to their information, but also understand business issues that expose their information to theft. It no longer matters how the information stolen, whether it was a hacker or a social engineer that committed the crime; companies need to protect their information from all threats and minimize their risks from all aspects. REFERENCES [1] Charny, Ben and Lemos, Robert. December 22, 2000. Egghead Scrambles to Guage Damage. Retrieved 06/19/2005 from http://seclists.org/lists/isn/2000/Dec/0134.html [2] Crawford, Michael. June 16, 2005. Criminals Grasp the Metrics of Information Value. Retrieved 06/20/2005 from http://www.computerworld.com.au/index.php?id=550545875 &eid=-255 [3] ConsumersUnion.org. Consumers Union applauds Nelson (FL) bill to extend federal oversight to information brokers like ChoicePoint. Retrieved 06/28/2005 from http://www.consumersunion.org/pub/core_financial_services /002027.html [4] Easen, Nick. April 21, 2004. Cyber Crime is Right Under Your Nose. Retrieved 06/25/2005 from http://www.cnn.com/2004/BUSINESS/04/20/go.cyber.securi ty/index.html [5] Gross, Grant. February 23, 2005. ChoicePoint's Error Sparks Talk of ID Theft Law. Retrieved 06/22/2005 from http://pcworld.com/news/article/0,aid,119790,00.asp [6] Liu, Bob. December 3, 2001. Eggheacd.com Becomes Amazon.com Property. Retrieved 06/22/2005 from http://www.internetnews.com/ec-news/article.php/932871 [7] McKinsey & Company, Inc. June 6, 2002. Managing Information Security. Retrieved 06/22/2005 from http://news.com.com/2009-1017-933185.html [8] McMillian, Robert. June 7, 2005. Citigroup to Encrypt Data Sent to Credit Bureaus. Retrieved 06/20/2005 from http://www.computerworld.com/hardwaretopics/hardware/st ory/0,10801,102315,00.html [9] Mearian, Lucas. May 2, 2005. Time Warner Says Data of 600,000 Workers Lost. Retrieved 06/21/2005 from http://www.computerworld.com/databasetopics/data/story/0, 10801,101500,00.html [10] Mimoso, Michael. April 2005. Damage Control. Retrieved 06/21/2005 from http://informationsecurity.techtarget.com/magItem/1,291266, sid42_gci1073914,00.html [11] Rasmussen, Michael. March 3, 2005. ChoicePoint Security Breach Will Lead to Increased Regulation. Retrieved 06/25/2005 from http://www.csoonline.com/analyst/report3416.html [12] Robert, Paul. March 9, 2005. Hackers Grab LexisNexis Info on 32,000 People. Retrieved 06/24/2005 from http://www.pcworld.com/resource/article/0,aid,119953,pg,1, RSS,RSS,00.asp [13] Weiss, Todd. May 20, 2005. Scope of Bank Data Theft Grows to 676,000 Customers. Retrieved 06/24/2005 from http://www.computerworld.com/securitytopics/security/cybe rcrime/story/0,10801,101903,00.html 138

security breach;risk analysis;Information Security;business practises and policy;information system;cases of information theft;privacy;management issue;Information Security Management;theft of information;human factor;data protection procedure;Security Management;information security;cyber crime;confidential information;incident response plan;encryption;data protection;personal information

20

A Survey of Collaborative Information Seeking Practices of Academic Researchers

Information seeking and management practices are an integral aspect of people's daily work. However, we still have little understanding of collaboration in the information seeking process. Through a survey of collaborative information seeking practices of academic researchers, we found that researchers reported that (1) the lack of expertise is the primary reason that they collaborate when seeking information; (2) traditional methods, including face-to-face, phone, and email are the preferred communication mediums for collaboration; and (3) collaborative information seeking activities are usually successful and more useful than individually seeking information. These results begin to highlight the important role that collaborative information seeking plays in daily work.

INTRODUCTION Information seeking and management practices are an integral aspect of people's daily work. In organizational work, information is vital for making decisions and coordinating activities. Therefore, organizations have developed a wide variety of processes and technologies to support their workers' information seeking activities. Much of this support has been for the individual information seeker; in most organizations, information seeking has been traditionally viewed as an individual activity [1, 2]. Yet, collaboration is becoming an increasingly important component of work in organizations. Multidisciplinary teams are a common feature of modern organizations [3, 4]. To successfully accomplish their work, team members must collaborate with each other efficiently and effectively. One important aspect of the team's work is seeking information [5]. Yet, we have little understanding of collaborative information seeking practices [6, 7]. Therefore, to help team members work together effectively and to design information systems that support their work, we must understand the collaborative information seeking practices of team members. To examine collaborative information seeking (CIS) practices, we conducted a survey of academic researchers in a small technology-focused research university. Researchers have traditionally collaborated with each other on research projects because of the often cross-disciplinary nature of the work. This collaboration has increased in recent years as information and communication technologies have improved. Although the survey asked a variety of questions, in this paper, we focus on three particular areas of interest: What triggers are most likely to lead to CIS activities? When engaging in CIS, what media or channel of communication is most likely used to collaborate? How successful are these CIS activities? In a previous study, we identified three triggers that cause team members to collaborate when seeking information. These triggers are (1) lack of expertise (2) complex information need and (3) information not easily accessible [8]. In this study, we were interested in identifying which of these triggers researchers reported to be the most important reason for them to collaborate when seeking information. We also wanted to identify what were the primary mechanisms of collaboration (e.g., e-mail, face-to-face , etc.). We were also interested in determining the 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. GROUP'05, November 69, 2005, Sanibel Island, Florida, USA. Copyright 2005 ACM 1-59593-223-2/05/0011...$5.00. 85 degree to which researchers found collaborative information seeking to be successful, particularly in comparison to individual information seeking. COLLABORATIVE INFORMATION SEEKING Although there is limited research on collaborative information seeking, researchers are beginning to explore this phenomena in various domains [9]. In a study of two design teams, Poltrock et al [10] found that each team had different communication and information seeking practices. Interestingly, they did not examine an individual's role in the information seeking process but rather how team members actively worked together to identify information needs. They argue that an understanding of collaborative information retrieval will allow for the informed design of technologies meant to support such work, and will also allow teams to work more effectively with these sources of information. In a study of information behavior in a hierarchical work environment a military command and control environment Sonnenwald and Pierce [11] described information seeking as a dynamic activity in which "individuals must work together to seek, synthesize and disseminate information." They examined how team members maintained awareness of each other's information activities and how this awareness influenced their information sharing with each other. Finally, in a study of collaborative information seeking in the medical domain, Reddy and Dourish [12] argue that work rhythms play a role in healthcare providers' collaborative information seeking practices. Although a few studies have examined collaborative information seeking in small group settings through ethnographic field studies, there have been, to the best of our knowledge, no studies that have used surveys to gather data on CIS from a larger population sample. METHODS Seventy researchers at a small a small technology-focused research university participated in this study. The majority were faculty researchers and a small percentage were graduate research assistants. Most participants were from science and technology disciplines. 3.2 Materials and Procedures One-hundred and fifty potential participants were emailed a request to participate, which included an email link to an online survey. The response rate was 47%. The survey included the following items: 1. What causes you to work together when looking for information? (a) The information needed is complex. (b) The information needed requires a different expertise. (c) The information is not immediately accessible. 2. What medium are you most likely to use when collaborating with your teammates to look for information? (a) Electronic forum; (b) Email; (c) Face-to-face; (d) Fax; (e) Instant message; (f) Telephone; (f) Web conferencing 3. When collaborating with teammates to look for information, we usually find the information for which the team is searching. 4. Participating in collaborative information seeking is easier than individual information seeking. 5. Participating in collaborative information seeking leads to more relevant information being found than when individually seeking information. 6. Participating in collaborative information seeking leads to information being found more quickly than when individually seeking information. Participants responded to each phrase under item 1 and to items 3 6 on a scale ranging from 1 (strongly disagree) to 10 (strongly agree) and to item 2 on a scale ranging from 1 (not at all likely) to 10 (very likely). The survey also included free-text opportunities for the respondents to provide more information about their answers, if they chose to do so. RESULTS & DISCUSSION In order to determine the triggers that are most likely to lead to collaborative information seeking, the responses to questionnaire items 1a, 1b, and 1c where considered. A one-way within-subjects analyses of variance (ANOVA) was computed with trigger serving as the independent variable with three levels (complexity, expertise, and accessibility) and rating as the dependent variable. This ANOVA was statistically significant F(2, 132) = 16.878 and p &lt; .001. Bonferroni's post hoc tests indicated that expertise was rated significantly higher (M = 8.17) than both complexity (M = 6.80) and accessibility (M = 6.73), while complexity did not significantly differ from accessibility. The findings indicate that academic researchers will most often collaborate because they find the information requires a different expertise than their own. Many academic research projects are multidisciplinary in nature and require particular knowledge that a researcher may not have. As one researcher stated, "The basic reason is that frequently a wide range of expertise is needed and no one person can possibly have all the skills needed to be successful." 86 Although the complexity of the information need and accessibility of information could lead to collaboration, they are not viewed as strongly as expertise. In regards to information accessibility, one researcher points out, "information is usually accessible; however, someone else will likely understand it better." For this researcher the difficulty was not in accessing the information but rather in understanding its relevance which may require different expertise. During the CIS process, different researchers bring their own particular expertise and perspective to the team. When a researcher seeks information outside her domain of expertise, she will often turn to another researcher for help. These different expertises play an important role in the collaborative information seeking activities of the research team. 4.2 Communication Mediums for CIS Activities In order to examine the relationship between communication mediums, and to reduce the number of variables for subsequent analysis, a principal component factor analyses with a Varimax rotation was computed on the responses to questionnaire item 2. A four factor solution was selected because all Eigen values were above 1, and a logical grouping of sources emerged. We labeled the first factor "traditional", and it included: email, face-to-face, and telephone. We labeled the second factor "web" and it included: instant messenger, web conferencing, and web sites. The third and fourth factor each included one item, "electronic forum" and "fax", and were, thus labeled accordingly. Factor scores were created by using the mean of all the items that loaded on a given factor, and these factor scores were used in subsequent analyses. In order to identify the media that are most likely used for collaborate information seeking, a one-way within-subjects analyses of variance (ANOVA) was computed with medium factor scores serving as the independent variable with four levels (traditional, web, electronic forum and fax) and rating as the dependent variable. This ANOVA was statistically significant as F(3, 195) = 84.709 and p &lt; .001. Bonferroni's post hoc tests indicated that traditional media (M= 8.10) significantly outscored all other types; both web media (M=3.64) and electronic forum (M=4.58) significantly outscored fax (M=2.70) but did not significantly differ one from the other. Researchers preferred traditional media for their communication. Within this category, we included e-mail. Although e-mail may not seem to fit in the same category as face-to-face and telephone, it has become such a ubiquitous communication medium that respondents viewed it as being similar to face-to-face and the telephone. Furthermore, email has been in existence much longer than other types of electronic mediums such as web conferences. People are more comfortable and experienced with email and personal conversations, whether these conversations are in person or on the phone. The other media were not as strongly embraced. For instance, we had anticipated that web-based media such as web-conferencing and instant messaging would have higher rating than it did. One possible explanation is "newness" of the technology. For instance, instant messenger tools are still relatively new and have not permeated to all groups and ages. Furthermore, some of the web-based media take time to set-up. Web conferences and web sites require time and effort unlike picking up the phone to talk to someone. Interestingly, although not included as a medium to rate, some participants added campus mail and "snail mail" as a medium for communication. Whether collaborators are physically co-located or geographically dispersed, communication is an essential component of collaborative information seeking. The researchers orient towards the mediums that are familiar to them. 4.3 Success of Collaborative Information Seeking Activities In order to address the question of whether collaborations are successful when engaging in CIS, a dichotomous variable was created for each success item (3 6), whereby a rating of 0 to 5 was considered "disagree", and a rating of 6 to 10 was considered "agree". We initially used a 10- point scale in order to be consistent with the rest of the survey questions. We then made the decision to reduce the scale to a dichotomous variable in order to evaluate this question with a test of statistical significance. Using this dichotomous variable, a chi-square analysis was performed on the frequencies for each success item. The results of these analyses as well as the mean rating for each item, with mean representing degree of agreement from 1 to 10 (10 representing "strongly agree"), is displayed in Table 1. Table 1. Means and Chi-Square for Success Factors Agreement Success Factor Mean Agree Disagree Chi-Square Usually find info 8.0152 64 2 X 2 = 58.242, p &lt; .001 Easier than individual info seeking 7.1061 50 16 X 2 = 17.515, p &lt; .001 Find more relevant info than individual info seeking 7.3788 55 11 X 2 = 29.333, p &lt; .001 Quicker than individual info seeking 6.9394 48 10 X 2 = 24.897, p &lt; .001 87 Success is often subjective and difficult to define, particularly with ill-defined tasks such as information seeking. Therefore, we asked four questions related to success to gain a better understanding of this important area. Most researchers agreed that when collaborating with colleagues to look for information, they usually found the needed information. They also thought that collaboratively seeking information was easier and lead to more relevant information than individually seeking information. Collaborative information seeking allows researchers to rely on other colleagues for help and guidance; therefore, allowing them to focus on their own area of expertise. This could be one possible reason why researchers strongly believe that CIS allows them to quickly find more relevant information when compared to individual information seeking. At the same time, one researcher provided a note of caution stating that the success "depends on your team of seekers." As in many collaborative activities, the success depends on how well the team of information seekers can work together when looking for information. CONCLUSIONS Collaborative information seeking is an important aspect of the work done by teams. The findings presented here raise issues that are important to consider when conceptualizing collaborative information seeking and how to best support this activity. One important issue is how to support information seeking in geographically dispersed teams. Physical co-located team members can have face-to-face interaction. However, for "virtual" teams technical support becomes even more important because they do not have the advantages of the face-to-face interaction. This technical support could include features that allow individuals to exchange ideas, or share searches while collaboratively searching for information [9]. For the next stages of this study, we plan on conducting a field study of academic research teams to better understand the actual interaction of team members during the collaborative information seeking process. ACKNOWLEDGMENTS We would like to thank the anonymous participants who answered the survey. This research was supported in part by Missouri Research Board grant 1734. REFERENCES 1. Ellis, D. (1989). A behavioral model for information retrieval system design. Journal of Information Science, 15: p. 237-247. 2. Ellis, D. and M. Haugan. (1997) Modeling the Information Seeking Patterns of Engineers and Research Scientists in an Industrial Environment. The Journal of Documentation. 53(4): p. 384-403. 3. Hackman, R. ed. (1990) Groups that Work (and Those That Don't): Creating Conditions for Effective Teamwork. Jossey-Bass Publications: San Francisco. 4. Mankin, D., S. Cohen, and T. Bikson. (1996). Teams and Technology. Boston, MA: Harvard Business School Press. 5. Bruce, H., et al. (2002). A comparison of the collaborative information retrieval (CIR) behaviors of two design teams. in Information Seeking In Context: The Fourth International Conference on Information Needs, Seeking and Use in Different Contexts. Lisbon, Portugal. 6. Sonnenwald, D.H. and L.A. Lievrouw. (1996). Collaboration during the Design Process: A Case Study of Communication, Information Behavior, and Project Performance. in Proc Int Conf on Research in Information Needs, Seeking, and Use in Different Contexts. Tampere, Finland: London: Taylor Graham. 7. Haythornthwaite, C., B. Wellman, and M. Mantei. (1995). Work Relationships and Media Use: A Social Network Analysis. Group Decision and Negotiation. 4(3): p. 193-211. 8. Reddy, M. (In submission) Collaborative Information Seeking: Supporting the work of multi-disciplinary patient care teams. Journal of American Medical Informatics Association (JAMIA). 9. Twidale, M. and D.M. Nichols. (1998). Designing Interfaces to Support Collaboration in Information Retrieval. Interacting with Computers. 10(2): p. 177-193. 10. Poltrock, S., et al. (2003). Information Seeking and Sharing in Design Teams. in Proceedings of the 2003 International ACM SIGGROUP Conference on Supporting Group Work. 11. Sonnenwald, D.H. and L.G. Pierce. (2000). Information behavior in dynamic group work contexts: interwoven situational awareness, dense social networks and contested collaboration in command and control. Information Processing and Management.36: p. 461-479. 12. Reddy, M. and P. Dourish. (2002). A Finger on the Pulse: Temporal Rhythms and Information Seeking in Medical Care. In Proc. of ACM Conf. on Computer Supported Cooperative Work (CSCW'02). New Orleans, LA: New York: ACM. p. 344-353. 88

Academic Researchers;communication media;information seeking;Group Work;Survey;collaboration;Collaborative Information Seeking

200

Transactional Agent Model for Fault-Tolerant Object Systems

A transactional agent is a mobile agent which manipulates objects in multiple computers by autonomously finding a way to visit the computers. The transactional agent commits only if its commitment condition like atomicity is satisfied in presence of faults of computers. On leaving a computer , an agent creates a surrogate agent which holds objects manipulated. A surrogate can recreate a new incarnation of the agent if the agent itself is faulty. If a destination computer is faulty, the transactional agent finds another operational computer to visit. After visiting computers, a transactional agent makes a destination on commitment according to its commitment condition. We discuss design and implementation of the transactional agent which is tolerant of computer faults.

INTRODUCTION A transaction manipulates multiple objects distributed in computers through methods. Objects are encapsulations of data and methods for manipulating the data. A transaction is modeled to be a sequence of methods which satisfies the ACID (atomicity, consistency, isolation, and dura-bility ) properties [8, 9]. Huge number and various types of peer computers are interconnected in peer-to-peer (P2P) networks [3]. Personal computers easily get faulty not only by crash but also by hackers and intrusions. A mobile agent can autonomously escape from faulty computers by moving to another operational computer. Mobile agents [5, 19] are programs which move to remote computers and then locally manipulate objects on the computers. An ACID transaction initiates a subtransaction on each database server, which is realized in mobile agents [16, 9, 13]. In this paper, a transactional agent is a mobile agent which autonomously decides in which order the agent visits computers in presence of computer faults, and locally manipulates objects in a current computer with not only atomicity but also other types of commitment conditions like at-least-one condition [6]. After manipulating all or some objects in computers, an agent makes a decision on commit or abort. For example, an agent atomically commits only if all objects in the computers are successfully manipulated [4]. An agent commits if objects in at least one of the computers are successfully manipulated. In addition, an agent negotiates with another agent which would like to manipulate a same object in a conflicting manner. Through the negotiation, each agent autonomously makes a decision on whether the agent holds or releases the objects [6, 14]. If an agent leaves a computer, objects locked by the agent are automatically released. Hence, once leaving a computer, an agent cannot abort. An agent creates a surrogate agent on leaving a computer. A surrogate agent still holds locks on objects in a computer on behalf of the agent after the agent leaves. A transactional agent autonomously finds another destination computer if a destination computer is faulty. An agent and surrogate are faulty if the current computer is faulty. Some surrogate of the agent which exists on another computer recreates a new incarnation of the agent. Simi-larly , if a surrogate may be faulty, another surrogate detects the fault and takes a way to recover from the fault. For example, if an agent takes an at least one commitment condition , a fault of the surrogate can be neglected as long as at-least-one surrogate is operational. In section 2, we present a system model. In section 3, we discuss transactional agents. In section 4, we discuss fault-tolerant mechanism. In sections 5 and 6, we discuss implementation and evaluation of transactional agents. SYSTEM MODEL A system is composed of computers interconnected in reliable networks. Each computer is equipped with a class base (CB) where classes are stored and an object base (OB) which is a collection of persistent objects. A class is composed of attributes and methods. An object is an instantia-tion of a class which is an encapsulation of data and meth-1133 2005 ACM Symposium on Applied Computing ods. If result obtained by performing a pair of methods op 1 and op 2 on an object depends on the computation order, op 1 and op 2 conflict with one another. For example, a pair of methods increment and reset conflict on a counter object . On the other hand, increment and decrement do not conflict, i.e. are compatible. A transaction is modeled to be a sequence of methods, which satisfies the ACID properties [4]. Especially, a transaction can commit only if all the objects are successfully manipulated. If a method op 1 from a transaction T 1 is performed before a method op 2 from another transaction T 2 which conflicts with op 1 , every method op 3 from T 1 has to be performed before every method op 4 from T 2 conflicting with the method op 3 . This is the serializability property [2, 4]. Locking protocols [2, 4, 7] are used to realize the serializability of transactions. Here, a transaction locks an object before manipulating the object. A mobile agent is a program which moves around computers and locally manipulates objects in each computer [5, 18, 19]. A mobile agent is composed of classes. A home computer home(c) of a class c is a computer where the class c is stored. For example, each class c is identified by a pair of IP address of a home computer home (c) and a local path to the directory where the class c is stored. A home computer home (A) of a mobile agent A is a home computer of the class of the agent A. TRANSACTIONAL AGENTS A transactional agent is a mobile agent which satisfies the following properties: 1. autonomously decides on which computer to visit. 2. manipulates objects on multiple computer. 3. commits only if some commitment condition of the agent is satisfied, otherwise aborts. For simplicity, a term agent means a transactional agent in this paper. Target objects are objects to be manipulated by an agent. T arget computers have the target objects. An agent A is composed of routing RC(A), commitment CC(A), and manipulation agents M C(A, D 1 ), ..., M C(A, D n ), where D i stands for a target computer of the agent A. Here, let Dom(A) be a set of target computers D 1 , ..., D n of an agent A. First, an agent A on a current computer has to move to a computer in Dom(A). A computer D j to which an agent A on D i moves is a destination computer. An agent A has to autonomously make a decision on which computer to visit. In the routing agent RC(A), a destination computer is selected. Then, the agent A moves to the destination computer. Here, an agent first finds a candidate set of possible destination computers. Then, the agent selects one target computer in the candidate computers and moves to the computer. Secondly, a transactional agent A manipulates objects in a current computer D. The agent A initiates a manipulation agent M C(A, D) for manipulating objects in the current computer D from the home computer. If an object base is realized in a relational database system [11], objects are manipulated by issuing SQL commands in M C(A, D). Lastly, a transactional agent makes a decision on whether the agent can commit or abort after visiting target computers . A traditional transaction [2] atomically commits only if objects in all the target computers are successfully manipulated . In this paper, we consider other types of commitment conditions [6]. For example, in the at-least-one commitment, a transaction can commit only if objects in at least one target computer are successfully manipulated. 3.2 Routing agent A transactional agent A locally manipulates objects in a computer D i through the manipulation agent M C(A, D i ) and then outputs intermediate objects OU T (A, D i ). In the meanwhile, the agent A visits another computer D j . Here, objects in D j are manipulated through the manipulation agent M C(A, D j ) by using the intermediate objects In(A, D j ) (=OU T (A, D i )). Thus, the manipulation classes are related with input-output relation. Here, D i x D j shows that the manipulation agent M C(A, D i ) outputs an intermediate object x which is used by M C(A, D j ). If D i x D j , the agent A has to visit D i before D j and the intermediate object x has to be delivered to D j . The input-output relation is shown in an input-output graph as shown in Figure 1. D 5 x y z w D 1 D 2 D 4 D 3 :computer :temporary object Figure 1: Input-output graph There are computer and object nodes. Directed edges D i x and x D i show that the manipulation agent M C(A, D i ) outputs and inputs an object x, respectively. In Figure 1, the agent A outputs an intermediate object w in D i . The agent A uses x in D 3 , D 4 , and D 5 . This means the agent A is required to visit D 3 , D 4 , and D 5 after D 1 . From the input-output graph, a transactional agent A decides in which order the agent visits. A directed acyclic graph (DAG) M ap(A) named a map is created from the input-output graph [Figure 2]. Here, a node D shows a computer D with a manipulation agent M C(A, D). A directed edge D 1 D 2 a computer D 2 is required to be manipulated after D 1 . D 1 D 2 if and only if (if f ) D 1 D 2 or D 1 D 3 D 2 for some computer D 3 . D 1 and D 2 are independent (D 1 D 2 ) if neither D 1 D 2 nor D 2 D 1 . Here, a transactional agent A can visit the computers D 1 and D 2 in any order and can in parallel visit the computers D 1 and D 2 . Figure 2 shows an example of a map M ap(A) obtained from the input-output graph of Figure 1. Here, an agent A is required to visit a computer D 3 after D 1 , D 4 after D 2 and D 3 , and D 5 after D 4 . On the other hand, an agent A can visit D 1 and D 2 in any order, even in parallel. In Figure 1, the intermediate object w has to be delivered to D 3 , D 4 , and D 5 . There are following ways to bring an intermediate object x obtained in D i to D j : 1. A transactional agent A carries the intermediate object x to D j . 2. x is transfered from D i to D j before A arrives at D j . 3. x is transfered from D i to D j after A arrives at D j . A routing agent RC(A) of a transactional agent A with a map M ap(A) is moving around computers [Figure 3]. First, 1134 D 1 D 2 D 3 D 4 D 5 Figure 2: Map. a collection I of computers which do not have any in-coming edge are found in M ap(A). For example, I = {D 1 , D 2 } in Figure 2. One computer D i is selected in I so as to satisfy some condition, e.g. D i nearest to the current computer is selected. For example, an agent takes a computer D 1 in Figure 2. The agent A moves to D i . Here, a manipulation agent M C(A, D i ) is loaded to D i from the home computer. After manipulating objects in D i , D i is removed from M ap(A). Another destination D j is selected and A moves to D j . Initially, a routing agent RC(A) of the agent A is loaded and started on a computer. The computer is a base computer base (A) of the agent A. An agent A leaves the base computer for a computer D i . Here, D i is a current computer current (A) of A. If the agent A invokes a method t of a class c on D i , the class c is searched: 1. The cache of the current computer D i is first searched for the class c. If c is found in the cache, the method t in the cache is invoked. 2. If not, the class base (CB i ) of D i is locally searched. If found, the class c in CB i is taken to invoke t. 3. Otherwise, the class c is transferred from the home computer home (c) into D i . A history H(A) shows a sequence of computers which an agent A has visited. CB D1 D2 :routing agent :manipulation agent :map Figure 3: Mobile agent. 3.3 Manipulation agent A manipulation agent is composed of not only application-specific classes but also library classes like JDBC [17] and JAVA classes [18]. Each computer is assumed to support a platform to perform a mobile agent on an object base (OB). A platform includes cache and class base (CB). The routing, manipulation, and commitment agents of a transactional agent A are stored in the class base (CB) of the home computer home (A). If an agent A invokes a method t of a class c in a computer D i , the class c is loaded from the home computer home (c) to the cache in D i . Then, the method t of the class c is performed in D i . If a method u of another class d is invoked in the method t, the class d is loaded from the home computer home (d) as well as the class c. Meanwhile, if another agent B invokes a method t of the class c in D i , the class c in the cache is used to invoke the method t without loading the class c. Thus, if classes are cashed in a computer D i , methods in the classes are locally invoked in D i without any communication. Otherwise, it takes a longer time to invoke methods since classes with the methods are transferred from the home computers in networks . Here, the class c is loaded i.e. cached to D i . The method t of the class c is performed on D i . If another agent B comes to D i after A has left D i , B can take usage of the class c in the cache. 3.4 Commitment agent If a transactional agent A finishes manipulating objects in each computer, the following commitment condition is checked by the commitment agent CC(A): 1. Atomic commitment : an agent is successfully performed on all the computers in the domain Dom(A), i.e. all-or -nothing principle used in the traditional two-phase commitment protocol [4, 15]. 2. Majority commitment : an agent is successfully performed on more than half of the computers in Dom(A). 3. At-least-one commitment : an agent is successfully performed on at least one computer in Dom(A). 4. n r commitment : an agent is successfully performed on more than r out of n computers (r n) in Dom(A). 5. Application specific commitment : condition specified by application is satisfied. 3.5 Resolution of confliction Suppose an agent A moves to a computer D j from another computer D i . The agnet A cannot be performed on D j if there is an agent or surrogate B conflicting with A. Here, the agent A can take one of the following ways: 1. W ait: The agent A in the computer D i waits until the agent A can land at a computer D j . 2. Escape: The agent A f inds another computer D k which has objects to be manipulated before D j . 3. N egotiate: The agent A negotiates with the agent B in D j . After the negotiation, B releases the objects or aborts. 4. Abort: The agent A aborts. Deadlock among agents may occur. If the timer expires, the agent A takes a following way: 1. The agent A retreats to a computer D j in the history H(A). All surrogates preceding D j are aborted. 2. Then, the surrogate agent A j on D j recreates a new incarnation of the agent A. The agent A finds another destination computer D h . The surrogate A j to which the agent A retreats plays a role of checkpoint [12]. Suppose a surrogate agent B holds an object in a computer D j . An agent A would like to manipulate the object but conflicts with B in D j . The surrogate B makes a following decision: 1. Atomic commitment : The agent A waits until the surrogate B finishes. 2. At-least-one commitment : If the surrogate B knows at least one sibling surrogate of B is committable, B releases the object and aborts after informing the other sibling surrogates of this abort. 3. Majority commitment : If the surrogate B knows more than half of the sibling surrogates are committable, B releases the object and aborts after informing the other surrogates. 4. n r commitment : If the surrogate B knows more than 1135 or equal to r sibling surrogate agents are committable, the surrogate B releases the object and aborts. FAULT-TOLERANT AGENT We assume computers may stop by fault and networks are reliable. A transactional agent is faulty only if a current computer of the agent is faulty. Suppose an agent A finishes manipulating objects on a computer D i . The agent A selects one computer D j from the map M ap(A, D i ). The agent A detects by timeout mechanism that D j is faulty. The agent A tries to find another destination computer D k [Figure 4]. If found, A moves to D k as presented here. If A cannot find another destination computer in M ap(A, D i ), the agent A backs to the preceding computer D k [Figure 5]. D i is removed from M ap(A, D k ). Then, the agent in D k tries to find another destination computer in M ap(A, D k ). Di Dj Dk Figure 4: Forwarding recovery. Dk Di Dj Figure 5: Backwarding recovery. An agent A leaves its surrogate agent A i on a computer D i . The surrogate A i holds objects even after the agent A leaves D i . An agent A and surrogate agent A i stop if the current computers are faulty. First, suppose an agent A stops on the current computer D j . Suppose that the agent A comes from D i to D j . The surrogate A i on D i detects that the agent A stops on D j . Here, A i takes one of the following actions: 1. Find a succeeding surrogate A k of A i and skips A j . 2. Recreate a new incarnation of the agent A. If the commitment condition is not atomic, the surrogate A j takes the first one, i.e. skips the fault of A j . For the atomic condition, A i recreates a new incarnation of the agent A. The agent A takes another destination computer D k in M ap(A, D i ). If found, the agent A moves to D k . Otherwise, A waits until the computer D i is recovered or backs to the precedent computer from D j . A surrogate A i on a computer D i may be faulty as well. A preceding surrogate A j on D j detects the fault of A i . Suppose a surrogate agent A i of A exists on D i . A i+1 and A i-1 show the succeeding and precedeing surrogate agents of A i , respectively [Figure 6]. A i periodically sends an enquiry message AYL (are you alive) to A i+1 and A i-1 to check if A i+1 and A i are alive. On receipt of the AYL message, a surrogate sends back a response message IAL (I am alive). Thus, a faulty surrogate is detected by the succeeding and preceding a surrogate with timeout mechanism. If A i detects the stop of A i+1 , A i does the follwings: 1. A new incarnation of the agent A is recreated on D i . 2. From the map M ap(A, D i ), a new destination D different from D i-1 is detected. A i-1 A i A i+1 AYL AYL IAL IAL Figure 6: Fault detection. 3. If detected, the agent A moves to D. Otherwise, A i informs A i-1 of abort and then aborts. A i-1 does the procedure from step 1. If the surrogate A i detects the stop of the preceding surrogate A i-1 or receives an abort message for A i-1 , A i informs the succeeding surrogate A i+1 of abort. On receipt of the abort message from A i , A i+1 forwards the abort message to A i+2 and then aborts. Thus, abort messages are eventually forwarded up to the agent A. In Figure 7, suppose A 2 stops. A pair of surrogates A 1 and A 3 detect the stop of A 2 . A 1 creates a new incarnation A of the agent A. The obsolete incarnation A still is moving to D 6 . The succeeding surrogate A 3 of A 2 sends an abort message to A 4 . If the abort message catches up the agent A, A can be aborted. Otherwise , the obsolete incarnation A cannot stop. Thus, there might exist multiple incarnations of an agent. A0 A1 A2 A3 A4 A D1 D2 D3 D4 D5 D6 A' Ai : surrogate A : agent Figure 7: Incarnations of an agent. On receipt of an AY L message from the preceding surrogate A i-1 , A i sends an IAL message with the address information which A i knows of surrogates are backwarding to preceding surrogates. If the surrogate A i finds A i-1 to be faulty, A i sends an abort message to not only A i+1 but also a surrogate whose address A i knows and which is nearest to the current computer of A. By this method, an abort message can more easily catch up with the agent mapped the agent can be aborted. IMPLEMENTATION We discuss how to implement transactional agents in Aglets. A transactional agent A is composed of routing, manipulation , and commitment subagents. An routing agent RC(A) with a map M ap(A) is transfered from one computer to another . When an agent A, i.e. routing agent RC(A) arrives at a computer D i , a manipulation agent M C(A, D i ) is created by loading the manipulation class. An object base (OB) is realized in a relational database system, Oracle [11]. A transactional agent manipulates table objects by issuing SQL commands, i.e. select, insert, delete, and update in a current computer D i . The computation of each agent A on a computer D i is realized as a local transaction on a database system. If the agent A leaves D i , the transaction for A commits or aborts. That is, objects manipulated by A are released. Even if the agent A leaves D i , the objects manipulated by A are required to be still held because A may abort after leaving D i . If the 1136 objects are released, the agent is unrecoverable. Therefore, a surrogate agent is created on D i . The surrogate agent is composed of a manipulation agent M C(A, D i ) and an object agent OBA i . OBA i behaves as follows: 1. On arrival at a computer D i , the routing agent RC(A) of an agent A initiates a manipulation agent M C(A, D i ) and an object agent OBA i on D i , i.e. M C(A, D i ) and OBA classes are loaded. OBA i initiates a transaction on an object base OB i . 2. If M C(A,D i ) issues a method for manipulating objects , OBA i issues SQL commands to the database system in D i . 3. If the agent A finishes, A leaves D i . However, OBA i is still operational and holding the objects in D i . 4. OBA i commits and aborts if the agent A sends commit and abort requests to the surrogate A i , respectively. An object agent OBA i stays on a computer D i while holding objects even if the agent A leaves D i . OBA i is a local transaction on an object base OB i . On completion of the agent A, OBA i and M C(A, D i ) are terminated. OB i D i OBA OBA i SQL XA MC(A, D i ) RC(A) Figure 8: Object agent (OBA). An OBA class can be loaded to a computer with any type of database system. If a transactional agent comes to D i from another home computer, an OBA class is loaded to D i from the home computer. Thus, OBA instances are accumulated in the cache. In order to resolve this problem, an OBA class is loaded as follows: 1. If the OBA class is not cached in the current computer, the OBA class is loaded from home (OBA). 2. If the OBA class could not be loaded from home (OBA), an OBA class in the home computer of the agent is loaded to a computer. The routing agent RC(A) leaves a computer D i if the manipulation agent M C(A, D i ) finishes manipulating objects. M C(A, D i ) recreates a new incarnation of a routing agent RC(A) if the agent A stops due to the computer fault. A transactional agent A can commit if all or some of the surrogates commit depending on the commitment condition. For example, a transactional agent commits if all the surrogate agents successfully exist. Communication among an agent and its surrogate agents is realized by using the XA interface [20] which supports the two-phase commitment protocol [15] [Figure 8]. Each surrogate agent issues a prepare request to a computer on receipt of a prepare message from A. If prepare is successfully performed, the surrogate agent sends a prepared message to A. Here, the surrogate agent is committable. Otherwise, the surrogate agent aborts after sending aborted to A. The agent A receives responses from the surrogate agents after sending prepare to the surrogates. On receipt of the responses from surrogate agents, the agent A makes a decision on commit or abort based on the commitment condition. For example, if the atomic condition holds, A sends commit only if prepared is received from every surrogate . The agent A sends abort to all committable agents if an aborted message is received from at least one surrogate. On receipt of abort, a committable surrogate aborts. In the at-least-one commitment condition, A sends commit to all committable surrogates only if prepared is received from at least one surrogate. Next, we discuss how to support robustness against faults of computers. Suppose a surrogate agent A i of a transactional agent A stops after sending prepared. Here, A i is committable. On recovery of the committable surrogate A i , A i unilaterly commits if the surrogate agent is committable in the at-least-one commitment condition. In the atomic condition, A i asks the other surrogates if they had commit-ted . Suppose A i is abortable, i.e. faulty before receiving prepared. On recovery, A i unilaterly aborts. EVALUATION Client create D 1 move result Routing Agent Database Server Manipulation Agent Object Agent M 1 M 2 M 1 D 2 result M 2 Home Computer move move D 3 manipulate result M 3 M 3 result result result manipulate manipulate Figure 9: Evaluation model We evaluate the transactional agent which is implemented in Aglets. In the evaluation, There are three server computers D 1 , D 2 , and D 3 . A transactonal agent is created in a computer C by loading classes from the home computer h. The servers D 1 , D 2 , and D 3 are realized in personal computers (Pentium 3) with Oracle database systems, which are interconnected in the 1Gbps Ethernet. First, a transactional agent A is initiated in a base computer C. The agent A finds in which order D 1 , D 2 , and D 3 to be visited. Here, the agent A visits D 1 , D 2 , and D 3 in this order as shown in Figure 9. On arrival of the agent A on D i , the manipulation agent M C(A, D i ) and object agent OBA i are loaded to D i [Figure 9]. We consider that following types of transactional agents: A. The manipulation agents M C(A, D 1 ) derives intermediate object I from the object base. The object bases in D 2 and D 3 are updated by using the object I, i.e. objects in I are added to the object base. B. M C(A, D 1 ) and M C(A, D 2 ) derive objects to intermediate objects I 1 and I 2 , respectively. Then, the object base in D 3 is manipulated by using I 1 and I 2 . There are three ways to deliver intermediate objects de-rived to another computer: 1. The transactional agent A carries intermediate objects to a destination computer D j from D i . 1137 2. After the agent A arrives at a computer D j , the agent A requests D i to send the intermediate objects. 3. The agent A transfers the intermediate object I to a computer D j before leaving D i . The total response time of a transactional agent is measured for number of intermediate objects, i.e. number of tuples deriverd in computeres. Figures 10 and 11 show the response time for the types of transactional agents A and B, respectively. The second and third ways to deliver intermediate objects to destination computers imply shorter responce time than the first way. Figure 10: Response A Figure 11: Response B CONCLUDING REMARKS The authors discussed a transactional agent model to manipulate objects in multiple computers with types of commitment constraints in presence of computer faults. A transactional agent autonomausly finds a distination computer, moves to a computer, and then locally manipulates objects. We discussed how to implement transactional agents in Aglets and Oracle. We evaluated the transactional agent in terms of response time. REFERENCES [1] American National Standards Institute. The Database Language SQL, 1986. [2] P. A. Bernstein, V. Hadzilacos, and N. Goodman. Concurrency Control and Recovery in Database Systems. Addison-Wesley, 1987. [3] L. Gong. JXTA: A Network Programming Environment, pages 8895. IEEE Internet Computing,, 2001. [4] J. Gray and A. Reuter. Transaction Processing : Concepts and Techniques. Morgan Kaufmann, 1993. [5] IBM Corporation. Aglets Software Development Kit Home. http://www.trl.ibm.com/aglets/. [6] T. Komiya, T. Enokido, and M. Takizawa. Mobile agent model for transaction processing on distributed objects. Information Sciences, 154:2338, 2003. [7] F. H. Korth. Locking primitives in a database system. Journal of ACM, 30(1):5579, 1989. [8] N. A. Lynch, M. Merritt, A. F. W. Weihl, and R. R. Yager. Atomic Transactions. Morgan Kaufmann, 1994. [9] K. Nagi. Transactional Agents : Towards a Robust Multi-Agent System. Springer-Verlag, 2001. [10] A. Omicini, F. Zambonelli, M. Klusch, and R. Tolksdorf. Coordination of Internet Agents. Springer-Verlag, 2001. [11] Oracle Corporation. Oracle8i Concepts Vol. 1 Release 8.1.5, 1999. [12] R. S. Pamula and P. K. Srimani. Checkpointing strategies for database systems. Proc. of the 15th Annual Conf. on Computer Science, IEEE Computer Society, pages 8897, 1987. [13] S. Pleisch. State of the Art of Mobile Agent Computing - Security, Fault Tolerance, and Transaction Support. IBM Corporation, 1999. [14] M. Shiraishi, T. Enokido, and M. Takizawa. Fault-tolerant mobile agent in distributed objects systems. Proc. of the 9th IEEE International Workshop on Future Trends of Distributed Computing Systems (FTDCS 2003), pages 145151, 2003. [15] D. Skeen. Nonblocking commitment protocols. Proc. of ACM SIGMOD, pages 133147, 1982. [16] A. D. Stefano, L. L. Bello, and C. Santoro. A distributed heterogeneous database system based on mobile agents. Proc. of the 7th Workshop on Enabling Technologies (WETICE'98), IEEE Computer Society, pages 223229, 1998. [17] Sun Microsystems Inc. JDBC Data Access API. http://java.sun.com/products/jdbc/. [18] Sun Microsystems Inc. The Source for Java (TM) Technology. http://java.sun.com/. [19] J. E. White. Telescript Technology : The Foundation for the Electronic Marketplace. General Magic Inc., 1994. [20] X/Open Company Ltd. X/Open CAE Specification Distributed Transaction Processing: The XA Specification, 1991. 1138

fault-tolerant agent;transactional agent;Transaction;ACID transaction;surrogate agent;Mobile agent;Fault-Tolerant;fault-tolerant;computer fault;mobile agent;transaction processing

201

Translating Unknown Cross-Lingual Queries in Digital Libraries Using a Web-based Approach

Users' cross-lingual queries to a digital library system might be short and not included in a common translation dictionary (unknown terms). In this paper, we investigate the feasibility of exploiting the Web as the corpus source to translate unknown query terms for cross-language information retrieval (CLIR) in digital libraries. We propose a Web-based term translation approach to determine effective translations for unknown query terms by mining bilingual search-result pages obtained from a real Web search engine. This approach can enhance the construction of a domain-specific bilingual lexicon and benefit CLIR services in a digital library that only has monolingual document collections. Very promising results have been obtained in generating effective translation equivalents for many unknown terms, including proper nouns, technical terms and Web query terms.

INTRODUCTION With the development of digital library technologies, large amounts of library content and cultural heritage material are being digitized all over the world. As digital library systems become commonly constructed and digitized content becomes widely accessible on the Web, digital libraries that cross language and regional boundaries will be in increasingly high demand globally. Unfortunately, most of existing digital library systems only provide monolingual content and search support in certain target languages. To facilitate a cross-language information retrieval (CLIR) service in digital library systems, it is important to develop a powerful query translation engine. This must be able to automatically translate users' queries from multiple source languages to the target languages that the systems accept. Conventional approaches to CLIR incorporate parallel texts [16] as the corpus. These texts contain bilingual sentences, from which word or phrase translations can be extracted with appropriate sentence alignment methods [7]. The basic assumption of such an approach is that queries may be long so query expansion methods can be used to enrich query terms not covered in parallel texts. However, this approach presents some fundamental difficulties for digital libraries that wish to support practical CLIR services. First, since most existing digital libraries contain only monolingual text collections, there is no bilingual corpus for cross-lingual training. Second, real queries are often short, diverse and dynamic so that only a subset of translations can be extracted through the corpora in limited domains. How to efficiently construct a domain-specific translation dictionary for each text collection has become a major challenge for practical CLIR services in digital libraries. In this paper, we propose a Web-based approach to deal with this problem. We intend to exploit the Web as the corpus to find effective translations automatically for query terms not included in a dictionary (unknown terms). Besides, to speedup online translation process of unknown terms, we extract possible key terms from the document set in digital libraries and try to obtain their translations in advance. For some language pairs, such as Chinese and English, as well as Japanese and English, the Web offers rich texts in a mixture of languages. Many of them contain bilingual translations of proper nouns, such as company names and personal names. We want to realize if this positive characteristic makes it possible to automatically extract bilingual translations of a large number of query terms. Real search engines, such as Google and AltaVista , allow us to search English terms for pages in a certain language, e.g. Chinese or Japanese. This has motivated us to develop the proposed approach for mining bilingual search-result pages, which are normally returned in a long, ordered list of snippets of summaries to help users locate interesting documents. The proposed approach uses the bilingual search-result pages of unknown queries as the corpus for extracting translations by utilizing the following useful techniques: (1) Term extraction methods that extract translation candidates with correct lexical boundaries. (2) Term translation methods that determine correct translations based on co-occurrence and context similarity analysis. Several preliminary experiments have been conducted to test the performance of the proposed approach. For example, very promising translation accuracy has been obtained in generating effective translation equivalents for many unknown terms, including proper nouns, technical terms and Web query terms. Also, it has been shown that the approach can enhance bilingual lexicon construction in a very efficient manner and thereby benefit CLIR services in digital libraries that only have monolingual document collections. In Section 2 of this paper, we examine the possibility of using search-result pages for term translation. The technical details of the proposed approach, including the term extraction and term translation methods, are presented with some experiments in Sections 3 and 4 respectively. An application of the proposed approach to bilingual lexicon construction is described in Section 5. Finally, in Section 6 we list our conclusions. OBSERVATIONS AND THE PROPOSED APPROACH A large number of Web pages contain a mixture of multiple languages. For example, Chinese pages on the Web consist of rich texts in a mixture of Chinese (main language) and English (auxiliary language), many of which contain translations of proper nouns and foreign terms. In fact, in the Chinese writing style, the first time a foreign term appears in the text, we might also write its original word, e.g., "" (Yahoo). In our research, we are seeking to determine if the percentage of correct translations for real queries is high enough in the top search-result pages. If this is the case, search-result-based methods can be useful in alleviating the difficulty of term translation. According to our observations, many query terms are very likely to appear simultaneously with their translations in search-result pages. Figure 1 illustrates the search-result page of the English query "National Palace Museum", which was submitted to Google to search Chinese pages. Many relevant results were obtained, including both the query itself and its Chinese aliases, such as "" (National Palace Museum), "" (an abbreviation of National Palace Museum) and "" (Palace Museum), which might not be covered in general-purpose translation dictionaries. Figure 1. An illustration showing translation equivalents, such as National Palace Museum/"" (""), which co-occur in search results returned from Google. Although search-result pages might contain translations, the difficulties in developing a high-performance search-result-based term translation approach still remain. For example, it is not straightforward to extract translation candidates with correct lexical boundaries and minimum noisy terms from a text. It is also challenging to find correct translations for each unknown term within an acceptable number of search-result pages and an acceptable amount of network access time. To deal with these problems, the proposed approach contains three major modules: search-result collection, term extraction and term translation, as shown in Figure 2 (a). In the search-result collection module, a given source query (unknown term) is submitted to a real-world search engine to collect top search-result pages. In the term extraction module, translation candidates are extracted from the collected search-result pages using the term extraction method. Finally, the term translation module is used to determine the most promising translations based on the similarity estimation between source queries and target translations. In fact there are two scenarios to which the proposed approach can be applied. Except online translation of unknown queries, another application is offline translation of key terms as in Figure 2 (b). To reduce unnecessary online translation processes, the proposed approach can be used to augment the bilingual lexicon via translating key terms extracted from the document set in a digital library. These extracted key terms are likely to be similar to terms that users may use in real user queries. The proposed approach can be applied to those unknown key terms to obtain their translations with an offline batch process (the extracted translations might be edited by indexers). Furthermore, the constructed bilingual lexicon can be incrementally updated with the input of unknown queries from users and the performing of online translation processes. To facilitate the above scenarios the proposed term extraction and term translation techniques are required, which will be further described in the following sections. Figure 2. (a) An abstract diagram showing the concept of the proposed approach for translating an unknown query. (b) Two application scenarios of the proposed Web-based term translation approach: online translation of unknown queries and offline translation of key terms extracted from the document set. Search-Result Collection Search-Result Pages Translation Candidates Term Extraction Term Translation Search Engine Source Query Target Translations (a) (b) Proposed Approach Q Doc Online Translation of Unknown Queries Offline Translation of Key Terms 109 TERM EXTRACTION The first challenge of the proposed approach is: how to efficiently and effectively extract translation candidates for an unknown source term from a set of search-result pages. Other challenging issues include: whether all possible translations can be extracted and whether their lexical boundaries can be correctly segmented. Conventionally, there are two types of term extraction methods that can be employed. The first is the language-dependent linguistics-based method that relies on lexical analysis, word segmentation and syntactic analysis to extract named entities from documents. The second type is the language-independent statistics-based method that extracts significant lexical patterns without length limitation, such as the local maxima method [19] and the PAT-tree-based method [3]. Considering the diverse applications in digital library and Web environments, we have adopted the second approach. Our proposed term extraction method, i.e., the PAT-tree-based local maxima method, is a hybrid of the local maxima method [19] and the PAT-tree-based method [3], which has been found more efficient and effective. First, we construct a PAT tree data structure for the corpus, in this case, a set of search-result pages retrieved using the source term as query. (The same term extraction method will be applied to extract key terms from digital libraries in Section 5 where the corpus is the documents in digital libraries). By utilizing the PAT tree, we can efficiently calculate the association measurement of every character or word n-gram in the corpus and apply the local maxima algorithm to extract the terms. The association measurement is determined not only by the symmetric conditional probability [19] but also by the context independency ratio [3] of the n-gram. We detail the proposed method in the following subsections. 3.1 Association Measurement The proposed association measurement, called SCPCD, combines the symmetric conditional probability (SCP) [19] with the concept of context dependency (CD) [3]. SCP is the association estimation of the correlation between its composed sub n-grams, which is as defined below: = + = + = = 1 1 1 1 2 1 1 1 1 1 2 1 1 ) ... ( ) ... ( 1 1 ) ... ( ) ( ) ( 1 1 ) ( ) ( n i n i i n n i n i i n n w w freq w w freq n w w freq w w p w w p n w w p w w SCP K K K K (1) where w 1 ... w n is the n-gram to be estimated, p(w 1 ... w n ) is the probability of the occurrence of the n-gram w 1 ... w n , and freq(w 1 ...w n ) is the frequency of the n-gram. To a certain degree, SCP can measure the cohesion holding the words together within a word n-gram, but it cannot determine the lexical boundaries of the n-gram. An n-gram with complete lexical boundaries implies that it tends to have free association with other n-grams appearing in the same context. Therefore, to further ensure that an n-gram has complete lexical boundaries, the concept of context dependency is introduced. Moreover, we consolidate the concept with SCP to form one association measurement. In order to achieve this goal, a refined measure, the context independency ratio - which is a ratio value between 0 and 1 - is extended from [3]. It is defined as follows: 2 1 1 1 1 ) ( ) ( ) ( ) ( n n n n w w freq w w RC w w LC w w CD K K K K = (2) where LC(w 1 ... w n ) is the number of unique left adjacent words in western languages, or characters in oriental languages, for the n-gram in the corpus, or is equal to the frequency of the n-gram if there is no left adjacent word/character. Similarly, RC(w 1 ... w n ) is the number of unique right adjacent words/characters for the n-gram, or is equal to the frequency of the n-gram if there is no right adjacent word/character. Using this ratio we are able to judge whether the appearance of an n-gram is dependent on a certain string containing it. For example, if w 1 ... w n is always a substring of string xw 1 ... w n y in the corpus, then CD(w 1 ... w n ) is close to 0. Combining formulae (1) and (2), the proposed association measure SCPCD is as follows = + = = 1 1 1 1 1 1 1 1 1 ) ( ) ( 1 1 ) ( ) ( ) ( ) ( ) ( n i n i i n n n n n w w freq w w freq n w w RC w w LC w w CD w w SCP w w SCPCD K K K K K K K (3) Note that the difference between the formulae of SCPCD and SCP is in their numerator items. For SCP, those n-grams with low frequency tend to be discarded, which is prevented in the case of SCPCD. The proposed new measure determines a highly cohesive term because of the frequencies of its substrings and the number of its unique left and right adjacent words/characters. 3.2 Local Maxima Algorithm The local maxima algorithm, called LocalMaxs in [18], is based on the idea that each n-gram has a kind of cohesion that holds the words together within the n-gram. This is a heuristic algorithm used to combine with the previous association measurements to extract n-grams, which are supposed to be key terms from the text. We know different n-grams usually have different cohesion values. Given that: An antecedent (in size) of the n-gram w 1 w 2 ... w n , ant(w 1 ... w n ), is a sub-n-gram of the n-gram w 1 ... w n , having size n - 1. i.e., the (n-1)-gram w 1 ... w n-1 or w 2 ... w n . A successor (in size) of the n-gram w 1 w 2 ... w n , succ(w 1 ... w n ), is a (n+1)-gram N such that the n-gram w 1 ... w n is an ant(N). i.e., succ(w 1 ... w n ) contains the n-gram w 1 ... w n and an additional word before (on the left) or after (on the right) it. The local maxima algorithm extracts each term whose cohesion, i.e. association measure, is local maxima. That is, the term whose association measure is greater than, or equal to, the association measures of its antecedents and is greater than the association measures of its successors. 3.3 The PAT-Tree Based Local Maxima Algorithm Despite the usefulness of the local maxima algorithm, without a suitable data structure the time complexity of the algorithm is high. The main time complexity problems occur in two areas. One is calculating the context independency ratio (CD) for each unique n-gram in the corpus and the other is to find the successor of an n-gram. The two problems can be treated as one, i.e. finding the successors of an n-gram. An intuitive way to do this is to find out all (n+1)-grams and then compare the n-gram with them sequentially to see if they are the successors of it. As this is time-consuming, we introduce PAT tree as the data structure. 110 The above method is time consuming, however, so we use the PAT tree, which is a more efficient data structure. It was developed by Gonnet [8] from Morrison's PATRICIA algorithm (Practical Algorithm to Retrieve Information Coded in Alphanumeric) [15] for indexing a continuous data stream and locating every possible position of a prefix in the stream. The PAT tree structure is conceptually equivalent to a compressed digital search tree, but smaller. The superior feature of this structure mostly resulted from its use of semi-infinite strings [14] to store the substream values in the nodes of the tree. This also makes it easier and more efficient to find the successors of an n-gram. More details on the PAT tree can be found in [3]. By utilizing the constructed PAT tree as the corpus, we can efficiently retrieve all n-grams from the corpus, obtain their frequencies and context dependency values, and then calculate the association measures, SCPCD, of all of them. 3.4 Experiments on Term Extraction To determine the effectiveness of the proposed association measure SCPCD and the efficiency of the PAT-tree data structure, we conducted several experiments on Web search-result pages using the proposed PAT-tree-based local maxima algorithm. First, to test whether SCPCD can perform better than SCP and CD, we randomly selected 50 real queries in English from a Chinese search engine called Openfind 3 . We then submitted each of them to Google to search Chinese result pages. Most of these query terms such as proper nouns and technical terms were not covered in the common translation dictionary. After using the term extraction method, the top 30 extracted Chinese translation candidates were examined and the extraction accuracy of each candidate to the source query was manually determined. We applied this test mainly to determine whether the SCPCD measurement can extract more relevant translation candidates and segment them with correct lexical boundaries. A translation candidate was taken as correctly extracted only if it was correctly segmented and contained meanings relevant to the source term. A relevant translation candidate was not necessarily a correct translation. The whole relevant set was determined by examining the terms extracted by all of the test methods, e.g., CD, SCP, and SCPCD. Table 1 clearly shows that the method based on the SCPCD measurement achieves the best performance. Table 1. The obtained extraction accuracy including precision, recall, and average recall-precision of auto-extracted translation candidates using different methods. Association Measure Precision Recall Avg. R-P CD 68.1 % 5.9 % 37.0 % SCP 62.6 % 63.3 % 63.0 % SCPCD 79.3 % 78.2 % 78.7 % In order to determine the efficiency of the PAT-tree data structure, we compared the speed performance of the local maxima method and the PAT-tree-based local maxima method. As Table 2 shows, the PAT-tree data structure is more efficient in term extraction. Although the PAT-tree construction phase took a little more time 3 http://www.openfind.com/ in a small corpus, in a real-world case for a large corpus - where 1,367 and 5,357 scientific documents were tested (refer to Section 5.2 for the details) - the PAT-tree-based local maxima method performed much better than the local maxima method. Table 2. The obtained average speed performance of different term extraction methods. Term Extraction Method Time for Preprocessing Time for Extraction LocalMaxs (Web Queries) 0.87 s 0.99 s PATtree+LocalMaxs (Web Queries) 2.30 s 0.61 s LocalMaxs (1,367 docs) 63.47 s 4,851.67 s PATtree+LocalMaxs (1,367 docs) 840.90 s 71.24 s LocalMaxs (5,357 docs) 47,247.55 s 350,495.65 s PATtree+LocalMaxs (5,357 docs) 11,086.67 s 759.32 s TERM TRANSLATION In the term translation module, we utilize the co-occurrence relation and the context information between source queries and target translations to estimate their semantic similarity and determine the most promising translations. Several similarity estimation methods were investigated based on co-occurrence analysis. These included mutual information, DICE coefficient, and statistical tests including the chi-square test and the log-likelihood ratio test [17, 20], where the chi-square test and the context vector analysis achieved the best performance. These will be introduced below. 4.1 The Chi-Square Test The chi-square test ( 2 ) was adopted as the major method of co-occurrence analysis in our study. One major reason is that the required parameters for the chi-square test can be effectively computed using the search-result pages, which alleviates the data sparseness problem. It also makes good use of all relations of co-occurrence between the source and target terms, especially the information that they do not co-occur. For source term s and target term t, the conventional chi-square test can be transformed as the similarity measure defined below [6]: ) ( ) ( ) ( ) ( ) ( ) , ( 2 2 d c d b c a b a c b d a N t s S + + + + = (4) w here a: the number of pages containing both terms s and t; b: the number of pages containing term s but not t; c: the number of pages containing term t but not s; d: the number of pages containing neither term s nor t; N: the total number of pages, i.e., N= a+b+c+d. Since most search engines accept Boolean queries and can report the number of pages matched, the required parameters for the chi-square test can be obtained by submitting Boolean queries such as `st', `~st', `s~t' to search engines and utilizing the returned page counts. On the other hand, it is easy to get number N using some search engines (e.g., Google), which indicates the 111 total number of their collected Web pages. The number d may not be directly available from the search engine, but it can be calculated using the formula N= a+b+c+d, i.e., d = N-a-b-c. 4.2 Context Vector Analysis Co-occurrence analysis is applicable to higher frequency terms since they are more likely to appear with their translation candidates. On the other hand, lower frequency terms have little chance of appearing with candidates on the same pages. The context vector method (CV) is therefore adopted to deal with this problem. As translation equivalents may share similar terms, for each query term, we take the co-occurring feature terms as the feature vector. The similarity between query terms and translation candidates can be computed based on their feature vectors. Thus, lower frequency query terms still have a chance to extract correct translations. The context vector-based method has been used to extract translations from comparable corpora, such as the use of Fung et al.'s seed word [5]. In our method, real users' popular query terms are used as the feature set, which should help to avoid many inappropriate feature terms. Like Fung et al.'s vector space model, we also use the TF-IDF weighting scheme to estimate the significance of context features. This is defined as follows: ) n log( ) , ( max ) , ( N d t f d t f w j j i t i = (5) where f(t i ,d) is the frequency of term t i in search-result page d, N is the total number of Web pages in the collection of search engines, and n is the number of pages containing t i . Given the context vectors of a source query term and each target translation candidate, their similarity is estimated with cosine measure as follows: ) ( ) ( ) , ( 1 2 1 2 1 = = = = m i t m i s t s m i cv i i i i w w w w t s S (6) It is not difficult to construct context vectors for source query terms and their translation candidates. For a source query term, we can use a fixed number of the top search results to extract translation candidates. The co-occurring feature terms of each query can also be extracted, and their weights calculated, which together form the context vector of the query. The same procedure is used to construct a context vector for each translation candidate. 4.3 The Combined Method Benefiting from real-world search engines, the search-result-based method using the chi-square test can reduce the work of corpus collection, but has difficulty in dealing with low-frequency query terms. Although context vector analysis can deal with difficulties encountered by the chi-square test, it is not difficult to see that the feature selection issue needs to be carefully handled. Intuitively, a more complete solution is to integrate the above two methods. Considering the various ranges of similarity values in the two methods, we use a linear combination weighting scheme to compute the similarity measure as follows: = m m m t s R t s S all ) , ( ) , ( (7) where m is an assigned weight for each similarity measure S m , and R m (s,t) - which represents the similarity ranking of each target candidate t with respect to source term s - is assigned to be from 1 to k (the number of candidates) in decreasing order of similarity measure S m (s,t). 4.4 Experiments on Term Translation 4.4.1 The Test Bed To determine the effectiveness of the proposed approach, we conducted several experiments to extract translation pairs for Chinese and English terms in different domains. Web Queries: We collected query terms and the logs from two real-world Chinese search engines in Taiwan, i.e., Dreamer and GAIS. The Dreamer log contained 228,566 unique query terms for a period of over 3 months in 1998, while the GAIS log contained 114,182 unique query terms for a period of two weeks in 1999. We prepared two different test query sets based on these logs. The first, called the popular-query set, contained a set of 430 frequent Chinese queries in the logs. These queries were obtained from the Chinese translations of 1,230 English terms out of the most popular 9,709 query terms (with frequencies above 10 in both logs), which co-occurred with their English counterparts in the logs. The popular-query set was further divided into two types: type Dic (the terms covered in the dictionary), consisting of about 36% (156/430) of the test queries and type OOV (out of vocabulary; the terms not in the dictionary), consisting of about 64% (274/430) of the test queries. The second set, called the random-query set, contained 200 Chinese query terms, which were randomly selected from the top 20,000 queries in the Dreamer log, where 165 (about 82.5%) were not included in general-purpose translation dictionaries. Proper Names and Technical Terms: To further investigate the translation effectiveness for proper names and technical terms, we prepared two other query sets containing 50 scientists' names and 50 disease names in English. These were randomly selected from the 256 scientists (Science/People) and 664 diseases (Health/Diseases and Conditions) in the Yahoo! Directory. It should be noted that 76% (38/50) of the scientists' names and 72% (36/50) of the disease names were not included in the general-purpose translation dictionary, which contained 202,974 entries collected from the Internet. To evaluate the search-result-based methods, we obtained search-result pages of the source query terms by submitting them to real-world Chinese search engines, such as Google Chinese and Openfind. Basically, we used only the first 100 retrieved results (snippets) to extract translation candidates. The context vector of each source query and the required parameters (page counts) for the chi-square test were also extracted from the retrieved search-result pages. To evaluate the performance of translation extraction, we used the average top-n inclusion rate as a metric. For a set of test queries, the top-n inclusion rate was defined as the percentage of queries whose translations could be found in the first n extracted translations. Also, we wished to know if the coverage rate of translations, i.e. the percentage of queries whose translations could be found in the whole extracted candidate set, was high enough in the top search-result pages for real queries. 112 4.4.2 Performance Web Queries We carried out experiments to determine the performance of the proposed approach by extracting translations for the popular-query set. Tables 3 and 4 show the results in terms of top 1-5 inclusion rates and coverage rates for Chinese and English queries respectively. In this table, "CV", " 2 " and "Combined" represent the context-vector analysis, the chi-square test, and the combined method, respectively. In addition, "Dic", "OOV" and "All" represent the terms covered in a dictionary, the terms not in a dictionary, and the total test query set, respectively. The coverage rates we obtained were promising, which shows that the Web contains rich mixed texts in both languages. The performance of the English query set was not as good as the Chinese query set. The reason for this was that the English queries suffered from more noise in Chinese translation candidates since the search-result pages in the Chinese Web generally contain much more Chinese than English content. We also conducted an experiment for random queries. As Table 5 shows, the coverage rates were encouraging. Proper Names, Technical Terms and Common Terms To further determine the effectiveness of the proposed approach in dealing with the translation of proper names and technical terms, we conducted an experiment on the test sets of scientists' names and medical terms using the combined method. As the results in Table 6 show, the top-1 inclusion rates for the scientists' and disease names were 40% and 44% respectively. Some examples of the extracted correct translations are shown in Table 7. Although the achieved performance for real queries looked promising, we wished to know if it was equally effective for common terms. We randomly selected 100 common nouns and 100 common verbs from a general-purpose Chinese dictionary. Table 8 shows the results obtained using the combined method. It is easy to see that the proposed approach is less reliable in Table 3. Coverage and inclusion rates for popular Chinese queries using different methods. Method Query Type Top-1 Top-3 Top-5 Coverage Dic 56.4% 70.5% 74.4% 80.1% OOV 56.2% 66.1% 69.3% 85.0% CV All 56.3% 67.7% 71.2% 83.3% Dic 40.4% 61.5% 67.9% 80.1% OOV 54.7% 65.0% 68.2% 85.0% 2 All 49.5% 63.7% 68.1% 83.3% Dic 57.7% 71.2% 75.0% 80.1% OOV 56.6% 67.9% 70.9% 85.0% Combined All 57.2% 68.6% 72.8% 83.3% Table 4. Coverage and inclusion rates for popular English queries using different methods. Method Top-1 Top-3 Top-5 Coverage CV 50.9% 60.1% 60.8% 80.9% 2 44.6% 56.1% 59.2% 80.9% Combined 51.8 % 60.7% 62.2% 80.9% Table 5. Coverage and inclusion rates for random queries using the different methods. Method Top-1 Top-3 Top-5 Coverage CV 25.5% 45.5% 50.5% 60.5% 2 26.0% 44.5% 50.5% 60.5% Combined 29.5% 49.5% 56.5% 60.5% Table 6. Inclusion rates for proper names and technical terms using the combined method. Query Type Top-1 Top-3 Top-5 Scientist Name 40.0% 52.0% 60.0% Disease Name 44.0% 60.0% 70.0% 113 extracting translations of such common terms. One possible reason is that the usages of common terms are diverse on the Web and the retrieved search results are not highly relevant. It is fortunate that many of these common words can be found in general-purpose translation dictionaries. Table 8. Top 1, 3, 5 inclusion rates obtained using the combined method for extracting translations of common nouns and verbs. Query Type Top-1 Top-3 Top-5 100 Common Nouns 23.0% 33.0% 43.0% 100 Common Verbs 6.0% 8.0% 10.0% BILINGUAL LEXICON CONSTRUCTION To enhance CLIR services in a digital library that only has monolingual document collections, the proposed approach can be used to construct a domain-specific bilingual lexicon. We take the document set in digital libraries into consideration. The document set in the target language is first analyzed and possible key terms that are representative of the document set are extracted, using the proposed term extraction method. These extracted key terms are likely to be similar to terms that users may use in real user queries, since they are relatively more significant than other terms in the documents. The proposed term translation method can then be applied to those key terms not included in common translation dictionaries to obtain the translation of key terms in the source language. Therefore, a bilingual lexicon can then be constructed where the mappings between key terms and relevant terms in the source and target languages are maintained. As we have already indicated, the constructed bilingual lexicon can benefit CLIR services. For a given source query, the similarity with candidate source relevant terms can be calculated using the context vector method presented in Section 4. Also, and the top-ranked relevant terms can be extracted using the constructed bilingual lexicon. After the corresponding translations of relevant terms are obtained, relevant documents in the target language can be retrieved, using these relevant translations. The source query can then be expanded with the relevant translations and conventional CLIR methods can be used to retrieve documents in the target language. 5.2. An Application We tested the STICNET Database 4 , which is a government-supported Web-accessible digital library system providing a search service for scientific documents collected in Taiwan. The system contained documents in either English or Chinese, but no cross-language search was provided. To test the performance of bilingual lexicon construction, we selected 1,367 Information Engineering documents and 5,357 Medical documents respectively from the STICNET Database for the period 1983 to 1997 as the test bed. Using the PAT-tree-based term extraction method, key terms were automatically extracted from each document collection and their relevant translations were extracted by the proposed term translation approach. In the collection of Information Engineering documents, 1,330 key terms (with a threshold of 2 to 6-gram character strings, a term frequency&gt;10, and an association value&gt;0.1) were automatically extracted. Meanwhile, 5,708 key terms (with a threshold of 2 to 6-gram character strings and a term frequency&gt;40) were automatically extracted from the Medical document collection. Among the 1,330 auto-extracted key terms from the Information Engineering documents, 32% were not included in KUH Chinese Dictionary 5 (unknown terms) - one of the largest Chinese dictionaries with 158,239 term entries - where 75% of these unknown terms were found useful. In the case of Medical documents, 71% of the 5,708 auto-extracted key terms were not included in KUH Chinese Dictionary where 36.6% of these unknown terms were found useful. Table 9 shows the accuracy of the extracted translations for these useful unknown terms. The promising result shows the potential of the proposed approach to assist bilingual lexicon construction. 4 http://sticnet.stic.gov.tw/ 5 http://www.edu.tw/mandr/clc/dict/ Table 7. Some examples of the test English proper names and technical terms, and their extracted Chinese translations. Query Type English Query Extracted Translations (in Traditional Chinese) Scientist Name Galilei, Galileo (Astronomer) Crick, Francis (Biologists) Kepler, Johannes (Mathematician) Dalton, John (Physicist) Feynman, Richard (Physicist) // / // // Disease Name Hypoplastic Left Heart Syndrome Legionnaires' Disease Shingles Stockholm Syndrome Sudden Infant Death Syndrome (SIDS) / 114 Table 9. The top-n inclusion rates of translations for auto-extracted useful unknown terms. Query Type Top-1 Top-3 Top-5 Auto-extracted useful terms in Information Engineering 33.3% 37.5% 50.0% Auto-extracted useful terms in Medicine 34.6% 46.2% 50.0% RELATED WORK Many effective retrieval models have been developed for CLIR. For example, the Latent Semantic Indexing (LSI) method [4] has been utilized to model inter-term relationships, instead of exact term matching. Other methods include the cross-lingual relevance model [11], which integrates popular techniques of disambiguation and query expansion. However, translation of queries not covered in a bilingual dictionary remains one of the major challenges in practical CLIR services [9]. To deal with the translation of out-of-dictionary terms, conventional research on machine translation has generally used statistical techniques to automatically extract translations from domain-specific, sentence-aligned parallel bilingual corpora [20]. However, a large parallel corpus is difficult to obtain. Some work has been done on term translation extraction from comparable texts, such as bilingual newspapers [5], which are easier to obtain. Using a non-parallel corpus is more difficult than a parallel one, due to the lack of alignment correspondence for sentence pairs. On the other hand, research on digital libraries has made the same endeavor. Larson et al. [10] proposed a method for translingual vocabulary mapping using multilingual subject headings of book titles in online library catalogs - a kind of parallel corpus. However, book titles are still limited in coverage, compared to the rich resources on the Web. A new potential research direction is to perform query translation directly, through mining the Web's multilingual and wide-range resources [16]. Web mining is a new research area that focuses on finding useful information from large amounts of semi-structured hypertexts and unstructured texts [1]. Chen et al. [2] proposed a dictionary-based approach in which the search results returned from Yahoo China search engine were utilized to extract translations for terms not covered in the dictionary. In their work only an English term appearing (maybe in parenthesis) immediately or closely after a Chinese term was considered a possible translation. In our previous research, we proposed an approach for extracting translations of Web queries through the mining of anchor texts and link structures and obtained very promising results [12, 13]. Previous experiments showed that the anchor-text-based approach can achieve a good precision rate for popular queries. Its major drawback is the very high cost of the hardware and software required to collect sufficient anchor texts from Web pages. Collecting anchor texts requires a powerful Web spider and takes cost of network bandwidth and storage. Because of the practical needs of digital libraries, search-result pages, which are easier to obtain are, therefore, investigated in this paper. CONCLUSION In this paper, we have introduced a Web-based approach for dealing with the translation of unknown query terms for cross-language information retrieval in digital libraries. With the proposed term extraction and translation methods, it is feasible to translate unknown terms and construct a bilingual lexicon for key terms extracted from documents in a digital library. With the help of such bilingual lexicons, it would be convenient for users to formulate cross-lingual queries. The simplicity of the approach not only makes it very suitable for digital library systems, but would also facilitate the implementation of CLIR services. REFERENCES [1] Chakrabarti, S. Mining the Web: Analysis of Hypertext and Semi Structured Data, Morgan Kaufmann, 2002. [2] Chen, A., Jiang, H., and Gey, F. Combining Multiple Sources for Short Query Translation in Chinese-English Cross-Language Information Retrieval. In Proceedings of the 5th International Workshop on Information Retrieval with Asian Languages (IRAL 2000), 2000, 17-23. [3] Chien, L.F. PAT-Tree-based Keyword Extraction for Chinese Information Retrieval. In Proceedings of the 20 th Annual International ACM Conference on Research and Development in Information Retrieval (SIGIR 1997), 1997, 50-58. [4] Dumais, S. T., Landauer, T. K., and Littman, M. L. Automatic Cross-Linguistic Information Retrieval Using Latent Semantic Indexing. In Proceedings of ACM-SIGIR Workshop on Cross-Linguistic Information Retrieval (SIGIR 1996), 1996, 16-24. [5] Fung, P. and Yee, L. Y. An IR Approach for Translating New Words from Nonparallel, Comparable Texts. In Proceedings of the 36th Annual Conference of the Association for Computational Linguistics (ACL 1998), 1998, 414-420. [6] Gale, W. A. and Church, K. W. Identifying Word Correspondences in Parallel Texts. In Proceedings of DARPA Speech and Natural Language Workshop, 1991, 152-157. [7] Gale, W.A. and Church, K.W. A Program for Aligning Sentences in Bilingual Corpora. Computational Linguistics, 19, 1 (1993), 75-102. [8] Gonnet, G.H., Baeza-yates, R.A. and Snider, T. New Indices for Text: Pat Trees and Pat Arrays. Information Retrieval Data Structures & Algorithms, Prentice Hall, 1992, 66-82. [9] Kwok, K. L. NTCIR-2 Chinese, Cross Language Retrieval Experiments Using PIRCS. In Proceedings of NTCIR workshop meeting, 2001, 111-118. [10] Larson, R. R., Gey, F., and Chen, A. Harvesting Translingual Vocabulary Mappings for Multilingual Digital Libraries. In Proceedings of ACM/IEEE Joint Conference on Digital Libraries (JCDL 2002), 2002, 185-190. [11] Lavrenko, V., Choquette, M., and Croft, W. B. Cross-Lingual Relevance Models. In Proceedings of ACM Conference on Research and Development in Information Retrieval (SIGIR 2002), 2002, 175-182. [12] Lu, W. H., Chien, L. F., and Lee, H. J. Translation of Web Queries using Anchor Text Mining. ACM Transactions on Asian Language Information Processing, 1 (2002), 159-172. [13] Lu, W. H., Chien, L. F., and Lee, H. J. Anchor Text Mining for Translation of Web Queries: A Transitive Translation Approach. ACM Transactions on Information Systems, 22 (2004), 128. [14] Manber, U. and Baeza-yates, R. An Algorithm for String Matching with a Sequence of Don't Cares. Information Processing Letters, 37 (1991), 133-136. [15] Morrison, D. PATRICIA: Practical Algorithm to Retrieve Information Coded in Alphanumeric. JACM, 1968, 514-534. 115 [16] Nie, J. Y., Isabelle, P., Simard, M., and Durand, R. Cross-language Information Retrieval Based on Parallel Texts and Automatic Mining of Parallel Texts from the Web. In Proceedings of ACM Conference on Research and Development in Information Retrieval (SIGIR 1999), 1999, 74-81. [17] Rapp, R. Automatic Identification of Word Translations from Unrelated English and German Corpora, In Proceedings of the 37th Annual Conference of the Association for Computational Linguistics (ACL 1999), 1999, 519-526. [18] Silva, J. F., Dias, G., Guillore, S., and Lopes, G. P. Using LocalMaxs Algorithm for the Extraction of Contiguous and Non-contiguous Multiword Lexical Units. Lecture Notes in Artificial Intelligence, 1695, Springer-Verlag, 1999, 113-132. [19] Silva, J. F. and Lopes, G. P. A Local Maxima Method and a Fair Dispersion Normalization for Extracting Multiword Units. In Proceedings of the 6 th Meeting on the Mathematics of Language, 1999, 369-381. [20] Smadja, F., McKeown, K., and Hatzivassiloglou, V. Translating Collocations for Bilingual Lexicons: A Statistical Approach, Computational Linguistics, 22, 1 (1996), 1-38. 116

Information Search and Retrieval;Web Mining;Term Translation;translation dictionary;Context Vector Analysis;Unknown Cross-Lingual Queries;Web-based term translation approach;Cross-Language Information Retrieval;BILINGUAL LEXICON CONSTRUCTION;Digital Library;PAT-Tree Based Local Maxima Algorithm;CLIR services;Term Extraction;Digital Libraries

202

TypeCase: A Design Pattern for Type-Indexed Functions

A type-indexed function is a function that is defined for each member of some family of types. Haskell's type class mechanism provides collections of open type-indexed functions, in which the indexing family can be extended by defining a new type class instance but the collection of functions is fixed. The purpose of this paper is to present TypeCase: a design pattern that allows the definition of closed type-indexed functions, in which the index family is fixed but the collection of functions is extensible. It is inspired by Cheney and Hinze's work on lightweight approaches to generic programming. We generalise their techniques as a design pattern . Furthermore, we show that type-indexed functions with type-indexed types, and consequently generic functions with generic types, can also be encoded in a lightweight manner, thereby overcoming one of the main limitations of the lightweight approaches.

Introduction A type-indexed function is a function that is defined for each member of a family of types. One of the most popular mechanisms implementing this notion is the Haskell [31] type class system. A type class consists of a collection of related type-indexed functions; the family of index types is the set of instances of the type class. Type classes provide just one possible interpretation of the notion of type-indexed functions. In particular, they assume an open-world perspective: the family of index types is extensible, by defining a new type class instance for that type, but the collection of type-indexed functions is fixed in the type class interface so needs to be known in advance. For some applications -- particularly when providing a framework for generic programming -- the family of index types is fixed (albeit large) and the collection of type-indexed functions is not known in advance, so a closed-world perspective would make more sense. The original concept of a design pattern has its origins in Christopher Alexander's work in architecture, but it has been picked up with enthusiasm by the object-oriented programming community. The idea of design patterns is to capture, abstract and record beneficial recurring patterns in software design. Sometimes those patterns can be captured formally, as programming language constructs or software library fragments. Often, however, the appropriate abstraction cannot be directly stated, either because of a lack of expressiveness in the language, or because there is inherent ambiguity in the pattern -- Alexander describes a pattern as a solution `you can use [. . . ] a million times over, without ever doing it the same way twice' [1]. In this case, one must resort to an informal description. Even if the abstraction itself can be captured formally, one might argue that a complete description of the pattern includes necessarily informal information: a name, motivation, examples, consequences, implementation trade-offs, and so on. In this paper, we present a technique that allows the definition of closed type-indexed functions, as opposed to the open type-indexed functions provided by type classes; we do so in the format of a design pattern. Our inspiration comes from previous research on lightweight approaches to generic programming (LAGP). In particular , Hinze's two papers "A Lightweight Implementation of Generics and Dynamics" [4] (LIGD, with James Cheney) and "Generics for the Masses" [19] (GM) provide our motivation and basis. Those two papers focus on the particular context of generic programming, and provide a number of techniques that can be used to encode first-class generic functions in Haskell. However, those techniques have a wider applicability, not addressed by Hinze. We propose a generalisation of the technique, and demonstrate its use in a variety of applications. Our specific contributions are: Generalisation of the lightweight approaches. We provide templates for designing closed type-indexed functions, abstracting away from generic programming. The techniques in LIGD and GM are instances of these templates. A design pattern for type-indexed functions. We document this generalisation as a design pattern. Type-indexed functions with type-indexed types. We show that with our more general interpretation of the design pattern, type-indexed functions with type-indexed types are also instances of the design pattern. As a consequence, generic functions with generic types can also be encoded in a lightweight manner. Thus, we remove one of the main limitations of the lightweight approaches. Other applications. We present two other interesting applications of the pattern: PolyP in Haskell 98, and a very flexible printf function. The remainder of this paper is structured as follows. In Section 2 we review the lightweight approaches to generic programming. In Section 3 we abstract the essence of the technique as a design pattern . Section 4 presents two other small applications of the design pattern, and Section 5 uses it to model type-indexed functions with type-indexed types. Section 6 concludes. Lightweight generic programming We start by summarising the earlier work on lightweight approaches to generic programming underlying our generalisation. 2.1 "A Lightweight Implementation of Generics and Dynamics" Cheney and Hinze [4] show how to do a kind of generic programming , using only the standard Hindley-Milner type system extended with existential types. The index family consists of hierarchical sums and products of integers and characters. This family is enough to represent a large subset of Haskell 98 datatypes (including mutually recursive and nested datatypes). data Sum a b = Inl a | Inr b data Prod a b = Prod a b data Unit = Unit This style of generic programming requires a representation of types as values in order to support typecase analysis. The key idea of the LIGD paper is to use a parametrised type as the type representation , ensuring that the type parameter reflects the type being represented. Some Haskell implementations have recently been extended with generalised algebraic datatypes (GADTs) [32], which can be used for this purpose; but LIGD predates that extension, and depends only on existential quantification. data Rep t = RUnit (t Unit) | RInt (t Int) | RChar (t Char) | a b. RSum (Rep a) (Rep b) (t (Sum a b)) | a b. RProd (Rep a) (Rep b) (t (Prod a b)) data a b = EP{from :: a b,to :: b a} (Note that the universal quantifications are in contravariant positions , so act existentially.) The intention is that the equivalence type a b represents embedding/projection pairs witnessing to an isomorphism between types a and b, thereby enforcing a correspondence between types t and Rep t. Of course, within Haskell, it is not possible to automatically verify the isomorphisms (from to = id and to from = id), so these laws should be externally checked. Furthermore, we follow the convention of ignoring the `ugly fact' of bottom values destroying the `beautiful theory' of many such isomorphisms [8]. A common case is with the trivial embedding/projections. self :: a a self = EP{from = id,to = id} Using self , we can provide a set of smart constructors for the Rep type, yielding representations of types by themselves. rUnit :: Rep Unit rUnit = RUnit self rInt :: Rep Int rInt = RInt self rChar :: Rep Char rChar = RChar self rSum :: Rep a Rep b Rep (Sum a b) rSum ra rb = RSum ra rb self rProd :: Rep a Rep b Rep (Prod a b) rProd ra rb = RProd ra rb self Using these smart constructors, we can build representations for recursive datatypes, by making explicit the structure isomorphism of the datatype. For instance, the isomorphism defining lists is [a] = 1 + a [a], and so the corresponding type representation is as follows. rList :: a. Rep a Rep [a] rList ra = RSum rUnit (rProd ra (rList ra)) (EP from to) where from [ ] = Inl Unit from (x : xs) = Inr (Prod x xs) to (Inl Unit) = [ ] to (Inr (Prod x xs)) = x : xs Note that the representation of a recursive datatype is an infinite value; but, because of laziness, this poses no problem. Having constructed representation values for arbitrary types, the final step is to define generic functions. Using the representation as a basis for structural case analysis, it is possible to simulate a typecase [16]. For example, here is a definition of generic equality: eq :: t. Rep t t t Bool eq (RInt ep) t 1 t 2 = from ep t 1 from ep t 2 eq (RChar ep) t 1 t 2 = from ep t 1 from ep t 2 eq (RUnit ep) = True eq (RSum ra rb ep) t 1 t 2 = case (from ep t 1 ,from ep t 2 ) of (Inl x,Inl y) eq ra x y (Inr x,Inr y) eq rb x y False eq (RProd ra rb ep) t 1 t 2 = case (from ep t 1 ,from ep t 2 ) of (Prod x y,Prod x y ) eq ra x x eq rb y y Using Haskell type classes, it is possible to make the use of generic functions even more convenient: the class TypeRep can be used to build values of type Rep t implicitly. class TypeRep t where rep :: Rep t instance TypeRep Unit where rep = rUnit instance TypeRep Int where rep = rInt instance TypeRep Char where rep = rChar instance (TypeRep a,TypeRep b) TypeRep (Sum a b) where rep = rSum rep rep instance (TypeRep a,TypeRep b) TypeRep (Prod a b) where rep = rProd rep rep instance TypeRep a TypeRep [a] where rep = rList rep For example, we can now express generic equality with an implicit rather than explicit dependence on the representation. ceq :: t. TypeRep t t t Bool ceq t 1 t 2 = eq rep t 1 t 2 2.2 "Generics for the Masses" Hinze's later GM approach [19] has a very similar flavour to LIGD; however, somewhat surprisingly, Hinze shows how to do generic programming strictly within Haskell 98, which does not support rank-n types or even existential types. Nevertheless, there is a close relationship between type classes and polymorphic records (for example, one possible translation of type classes into System F uses polymorphic records), and these require something like existential types for their encoding. Thus, type class instances can be seen as implicitly-passed records. Hinze uses this observation to deliver two implementations of generics. 2.2.1 Generic functions on types The first implementation of generics in GM ("GM1", from now on) can be seen as a direct descendent of LIGD. Instead of using a datatype with an existential quantification, Hinze uses a type class Generic. 99 class Generic g where unit :: g Unit sum :: (TypeRep a,TypeRep b) g (Sum a b) prod :: (TypeRep a,TypeRep b) g (Prod a b) datatype :: TypeRep a (b a) g b char :: g Char int :: g Int The parameter g of the type class represents the generic function, and each of the member functions of the type class encodes the behaviour of that generic function for one structural case. Generic functions over user-defined types can also be defined using the datatype type case. In this case, the isomorphism between the datatype and its structural representation must be provided. The type class TypeRep is used to select the appropriate behaviour of the generic function, based on the type structure of its argument . The role of this type class is somewhat analogous to the synonymous one in Section 2.1. One contrast with LIGD is that TypeRep for GM1 is not optional, because the type representations are always implicitly passed. class TypeRep a where typeRep :: Generic g g a instance TypeRep Unit where typeRep = unit instance (TypeRep a,TypeRep b) TypeRep (Sum a b) where typeRep = sum instance (TypeRep a,TypeRep b) TypeRep (Prod a b) where typeRep = prod instance TypeRep Char where typeRep = char instance TypeRep Int where typeRep = int For GM, the type class TypeRep directly selects the appropriate behaviour for a particular structural case from the generic function. In contrast, for LIGD, the corresponding type class TypeRep builds a value as a type representation for a particular structural case, and this representation is then used by a generic function to select the appropriate behaviour. The effect is the same, but GM is more direct. A new generic function is defined via an instance of Generic, providing an implementation for each structural case. For instance, the generic function gSize that counts all the elements of type Int and Char in some structure could be encoded as follows. newtype GSize a = GSize{appGSize :: a Int} instance Generic GSize where unit = GSize ( 0) sum = GSize (t case t of Inl x gSize x Inr y gSize y) prod = GSize (t case t of Prod x y gSize x + gSize y) datatype iso = GSize (t gSize (from iso t)) char = GSize ( 1) int = GSize ( 1) gSize :: TypeRep a a Int gSize = appGSize typeRep A record of type GSize a contains a single function appGSize of type a Int, which can be used to compute the number of elements in some structure of type a. The function gSize, which is the actual generic function, simply extracts the sole appGSize field from a record of the appropriate type, built automatically by typeRep. 2.2.2 Generic functions on type constructors The second implementation of generics in GM ("GM2") permits parametrisation by type constructors rather than by types. For example , whereas the generic function gSize of the previous section has type a Int for all first-order types a in the type class TypeRep, in this section we show a generic function gSize with type f a Int for all type constructors f in the constructor class FunctorRep. Lifting in this fashion introduces the possibility of ambiguity: a type g (f a) may be considered a type constructor g applied to a type f a, or the composition of constructors g and f applied to type a. Therefore we must explicitly pass type representations, increasing flexibility but decreasing brevity. This is reflected in the analogous type class Generic, where the implicitly-passed TypeRep contexts are now changed to explicitly-passed functions. class Generic g where unit :: g Unit sum :: g a g b g (Sum a b) prod :: g a g b g (Prod a b) datatype :: (b a) g a g b char :: g Char int :: g Int However, this modification of the type class restricts expressivity, since the only generic function we can call is the one being defined, recursively. Consequently, generic functions that perform calls to other generic functions (as when defining generic membership in terms of generic equality) become harder to define. With the new Generic class it is also possible to build the values for type representations automatically, using another type class TypeRep. Just as with LIGD, this class now becomes optional. Alternatively, we can use a type class FunctorRep to capture the notion of unary type constructor or functor. class FunctorRep f where functorRep :: Generic g g a g (f a) We have to define similar classes for each arity of type constructor. Generic functions are defined in a very similar fashion to GM1. For instance, the type Count a below represents a generic function that counts zero for each occurrence of a value of type Int or Char in some structure of type a. newtype Count a = Count{applyCount :: a Int} instance Generic Count where unit = Count ( 0) sum a b = Count (x case x of Inl l applyCount a l Inr r applyCount b r) prod a b = Count ((Prod x y) applyCount a x + applyCount b y) datatype iso a = Count (x applyCount a (from iso x)) char = Count ( 0) int = Count ( 0) While this function by itself approximates const 0, it is the basis for other more useful functions that really count the number of elements in some structure in some way, by overriding the behaviour of the basic generic function for occurrences of the type parameter: gSize :: FunctorRep f f a Int gSize = applyCount (functorRep (Count ( 1))) The payback of using FunctorRep is that we can define the behaviour of the generic function for its parameters. For instance, we could sum all the integers in some integer-parametrised datatype by using the identity function to define the behaviour of the generic function for the type parameter. gSum :: FunctorRep f f Int Int gSum = applyCount (functorRep (Count id)) 100 Closed type-indexed functions In LIGD and GM, we are shown three methods for implementing closed type-indexed functions. Those three variations give us different expressive power, and impose different constraints on the type system. A choice of implementation techniques, together with technical trade-offs making no one method superior in all circumstances , is characteristic of design patterns. In this section, we introduce the TypeCase design pattern, capturing the different techniques for implementing closed type-indexed functions. The TypeCase design pattern Intent: Allowing the definition of closed type-indexed functions. Motivation: The typecase design pattern captures a closed-world view of ad-hoc polymorphism. In Haskell, the type class system is a mechanism that supports ad-hoc polymorphism, but from an open-world point of view: they can be extended with cases for new datatypes, at the cost of a non-extensible set of functions. Under the closed-world assumption, there is a fixed set of type-structural cases but arbitrarily many type-indexed functions ranging over those cases. An example where the closed-world perpective works better than the open-world one is generic programming, in which we take a structural perspective on types as opposed to the more traditional nominal one. Using just a few operations on types, it is possible to represent the whole family of structural definitions of interest. For instance, here is a possible definition for a generic function that counts all the elements of some structure t: gsize t :: :: t Int gsize Unit = 0 gsize Int = 1 gsize Sum (Inl x) = gsize x gsize Sum (Inr y) = gsize y gsize Prod (Prod x y) = gsize x + gsize y With an open-world perspective, we can present a fixed number of type-indexed definitions that range over those few cases; but we cannot easily introduce new definitions. This is clearly not appropriate for generic programming. In fact, what we expect from a generic programming facility is the ability to a introduce new generic definition without affecting the surrounding context. This is precisely what the closed-world perspective provides us. Applicability: Use this pattern: to encode collections of definitions that are indexed by some fixed family of types, while allowing new definitions to be added to the collection without affecting modularity; when a definition is variadic, that is, it has a variable number of arguments (see Section 4.2 for an example); to try to avoid type-class trickery, such as multiple-parameter type classes, functional dependencies, overlapping instances or even duplicate instances (just consider a direct encoding of the examples presented in the paper into type classes [30]); to capture some shape invariants, like the ones captured by some nested types or phantom types [29, 18]. Structure: See Figure 1. Participants: Structural Cases: a set of datatypes which represent the possible structural cases for the type-indexed function; Typecase: representing the structure of a type-indexed function; Dispatcher: a type class, containing a single function, that is responsible for dispatching a value of one of the structural cases into the corresponding branch of the typecase, based on the type of the value; Type-indexed function: defining the type-indexed function using an instance of the typecase. Collaborations: The typecase uses the structural cases in order to create a corresponding number of cases that can be used to define the type-indexed function. The dispatcher uses the structural cases in order to create a corresponding number of instances that will forward some value of that family of structural cases into the corresponding case in the typecase component. The type-indexed function (TIF) uses an instance of the typecase in order to implement the desired functionality for the type-indexed function. Implementation: Typically, a typecase component is created using the structural cases. There are three main variations for the implementation of a typecase: two of them are based on type classes and the other one on a smart datatype. A smart datatype is a parametrised type where the type parameters are dependent on the constructors . The idea of a smart datatype can be represented in various forms: existential datatypes with an equivalence type ( la LIGD), GADTs, phantom types, among others. The goal of this design pattern is to simulate a closed type-indexed function. In general, a type-indexed function f has the following structure. f t :: | d 1 ... d k :: f t 1 a 1 ... a i = x 11 ... x 1n e 1 . . . f t m z 1 ... z j = x m1 ... x mn e m The type signature tells us that f has one type parameter t and optional type parameters d 1 ... d k with the same structure and kind as t. The type of the TIF may depend on t and d 1 ... d k . We should note that this is not the same as having a TIF with multiple type arguments. There is no problem, in principle, in having multiple-parameter type arguments, but it would lead to an explosion in the number of typecases. This would be a generalisation of this design pattern. For simplicity, we will only consider type parameters with the same structure. The usefulness of this simpler case is reflected in applications such as generic map where the input and output structures of the generic map function are the same. The body of f contains (at least) m branches, providing the behaviour of the TIF for each member of the family of types t (that is, t 1 a 1 ... a i ,...,t m z 1 ... z j ). This family of types corresponds to the structural cases participant of the design pattern . For each branch of the definition, we bind possible variables x 11 ... x 1n ,...,x m1 ... x mn and define each typecase of f with e 1 ,...,e m . We now discuss the three main variations of the design pattern. 1. Smart datatypes: This variation is inspired by the LIGD approach . Hindley-Milner typing extended with existential datatypes (supported in most Haskell compilers) is enough to encode it. However, with extensions such as GADTs (supported by GHC 6.4) the encoding becomes much more direct. Unfortu-nately , neither of those extensions conforms to Haskell 98. We will present this version of the design pattern using a GADT syntax for simplicity. Using the structural cases given by t 1 a 1 ... a i ,...,t m z 1 ... z j , we can derive the typecase and dispatcher seen in Figure 1. Since there are m structural cases in a standard instance of the design pattern, one would create m constructors c t 1 ,...,c t m and also m instances for Rep . TIFs can now be defined using those components, by creating some function f that takes a first argument of type Rep and returns a value of type . 101 Smart Datatype Implicit/Explicit Representations Typecase data t d 1 ... d k where c t 1 :: (a 1 ... a i ) (t 1 a 1 ... a i ) d 11 ... d 1k . . . c t m :: (z 1 ... z j ) (t m z 1 ... z j ) d m1 ... d mk class (g :: k +1 ) where case t 1 :: (a 1 ... a i ) g (t 1 a 1 ... a i ) d 11 ... d 1k . . . case t m :: (z 1 ... z j ) g (t m z 1 ... z j ) d m1 ... d mk Dispatcher class Rep t d 1 ... d k where rep :: Rep t d 1 ... d k instance (a 1 ... a i ) Rep (t 1 a 1 ... a i ) d 11 ... d 1k where rep = c t 1 rep i . . . instance (z 1 ... z j ) Rep (t m z 1 ... z j ) d m1 ... d mk where rep = c t m rep j class Rep t d 1 ... d k where rep :: g g t d 1 ... d k instance (a 1 ... a i ) Rep (t 1 a 1 ... a i ) d 11 ... d 1k where rep = case t 1 {rep i } . . . instance (z 1 ... z j ) Rep (t m z 1 ... z j ) d m1 ... d mk where rep = case t m {rep j } Type-indexed function f :: t d 1 ... d k f (c t 1 r a 1 ... r a i ) = x 11 ... x 1n [[e 1 ]] . . . f (c t m r z 1 ... r z j ) = x m1 ... x mn [[e m ]] f :: Rep t d 1 ... d k f = f rep newtype F t d 1 ... d k = F{f :: } f :: Rep t d 1 ... d k f = f rep instance Rep F where case t 1 {r a 1 ... r a i } = x 11 ... x 1n [[e 1 ]] . . . case t m {r z 1 ... r z j } = x m1 ... x mn [[e m ]] Figure 1. The structure of the TypeCase design pattern. The dispatcher component is optional in this variation. The TIFs created with this variation are fully closed to extension; no customisation is possible. This means that if we want to add extra functionality we need to modify the smart datatype (and the dispatcher if we have one). However, TIFs that call other TIFs are trivial to achieve; there is no need for tupling. 2. Implicit representations: The implicit representation version of the design pattern is inspired by GM1. Perhaps surprisingly, some implementations of this instance require only Haskell 98. However, if we need to have structurally-dependent variables, then we also require multiple-parameter type classes. Proceeding in a similar fashion to the smart datatype approach , we use the structural cases to derive the typecase and dispatcher seen in Figure 1. Again, because we have m structural cases, we create m functions case t 1 ,...,case t m and m instances of Rep . The dispatcher is not an optional component: it always needs to be defined in this variation. As with the smart datatype variation, TIFs defined in this way are fully closed to extension, and calls to other TIFs are trivial. 3. Explicit representations: The explicit representation variation of the design pattern is inspired by GM2. Like the implicit approach, Haskell 98 is enough to handle the simpler forms (one type parameter). However, if we discard the optional dispatcher , then Haskell 98 can handle all forms. Using the structural cases to derive the typecase and dispatcher seen in Figure 1, we would obtain a very similar structure to the implicit representation version. The most noticeable difference is that, with the explicit representation, the definition of rep needs to provide the corresponding case function with the representations for each of its type parameters. The second difference is that , which corresponds to the representations of the type parameters, reflects the fact that we are providing explicit representations. Thus, corresponds in this instance to explicit arguments of the function, while with the implicit representation it corresponds to (implicitly passed) type class constraints. The dispatcher is an optional component. Variations of this instance of the design pattern can also be found in the literature [10, 37], as described in Section 4.2. TIFs defined in this fashion are not fully closed to extension: it is possible to override default behaviour. However, the extra flexibility comes at a cost: recursive calls to other TIFs are not possible. One common solution for this problem is to tuple together into a record the mutually-dependent functions. Another possibility would be to have a notion of dependencies: if a TIF f requires calls to another TIF g, then the record that defines f has a field that is an instance of g. Although this work is quite tedious, Lh [26] shows how a type system can lighten the burden. An associated problem for TIFs in this setting is the issue of composability. If two TIFs are defined using different instances (this is, they are not tupled together), then we cannot, in a straightforward manner, use the same representation to compose them. To illustrate the problem, consider: newtype F v 1 ... v n = F{f :: } newtype G v 1 ... v n = G{g :: } instance Generic F where ... instance Generic G where ... Now let us suppose that we define a type-indexed abstraction (that is, a function that uses one or more TIFs and is not defined over the structure of types): h rep = ... f rep ... g rep ... The interpretation of this definition as a type-indexed function could be thought of as: h a = ... f a ... g a .... While this is a perfectly reasonable interpretation, in practice f requires inconsistent types F v 1 ... v n and G v 1 ... v n for rep: F and G are two different type constructors, so in a Hindley-Milner type system, unification obviously fails. However, F and G do have something in common. In particular, they are both 102 instances of Generic. So, in Haskell extended with higher-order polymorphism, we can capture this relation with a rank-2 type, thus providing a possible solution for the problem of composability. h :: ( g. Generic g g v 1 ... v n ) h rep = ... f rep ... g rep ... We should note that even though we have presented three main variations of the design pattern, the concept of a design pattern is, by itself, quite informal and thus prone to different interpretations. For instance, as we will see later, applications of the pattern (such as GM) can have more type cases than there are datatype variants, because some cases overlap. It is important to note that, depending on the context of a problem, a design pattern can be adapted to better fit that problem. Applications We present two applications of the design pattern. In Section 4.1, still within the context of generic programming, we show how one can build a library inspired by PolyP [21, 22] but working in Haskell 98. In Section 4.2, we present a very flexible version of a C-style printf function. 4.1 Light PolyP It probably comes as no surprise to the reader that the technique introduced in GM and LIGD can be applied to other generic programming approaches as well. PolyP was one of the first attempts to produce a generic programming language. It is a simpler language than Generic Haskell, working in a much more restricted family of datatypes, namely one-parameter regular types. But this restriction allows stronger properties to be stated: its simplicity and strong theoretical background make it an appropriate language for teaching both the theory [3] and practice of generic programming. Our proposal Light PolyP encourages this, because no external PolyP compiler is required (although one might still be desirable, for a more convenient syntax). Norell [30] shows how to use the Haskell type class system (extended with multiple-parameter type classes and functional dependencies ) to obtain first-class PolyP generic functions in Haskell. In this section, we will present a "lighter" version of PolyP, requiring only Haskell 98 (without extensions such as multiple-parameter type classes and functional dependencies) but with the same expressive power. Instead of using sums of products like LAGP or Generic Haskell, PolyP uses lifted pattern functors as structural cases. The pattern functors Empty, Plus and Prod have counterparts in LAGP. The pattern functors Rep and Par correspond respectively to the recursive argument and the parameter of the unary regular datatype. The pattern functor Const t for some type t represents the constant functor, and Comp handles the composition of functors required for regular types. data Empty p r = Empty data Plus g h p r = Inl (g p r) | Inr (h p r) data Prod g h p r = Prod (g p r) (h p r) newtype Par p r = Par{unPar ::p} newtype Rec p r = Rec{unRec :: r} newtype Comp d h p r = Comp{unComp :: d (h p r)} newtype Const t p r = Const{unConst :: t} The equivalence type is used to establish the isomorphism between a regular datatype and its top-level structure. The embedding/projection functions are traditionally called inn and out. data Iso a b = Iso{inn :: a b,out :: b a} listIso = Iso inL outL where inL (Inl Empty) = [ ] inL (Inr (Prod (Par x) (Rec xs))) = x : xs outL [ ] = Inl Empty outL (x : xs) = Inr (Prod (Par x) (Rec xs)) In PolyP no generic customisation is allowed, thus we can use an implicit representation version of the design pattern and consequently , it is possible for one generic function to use other generic functions in its definition. The typecase component corresponds to: class Generic f where empty :: f Empty plus :: (Rep g,Rep h) f (Plus g h) prod :: (Rep g,Rep h) f (Prod g h) par :: f Par rec :: f Rec comp :: (Functor d,Rep h) f (Comp d h) constant :: f (Const t) The dispatcher simply selects the corresponding case based on the type of the argument of the generic function g. class Rep g where rep :: Generic f f g instance Rep Empty where rep = empty instance (Rep g,Rep h) Rep (Plus g h) where rep = plus instance (Rep g,Rep h) Rep (Prod g h) where rep = prod instance Rep Par where rep = par instance Rep Rec where rep = rec instance (Functor d,Rep h) Rep (Comp d h) where rep = comp instance Rep (Const t) where rep = constant Like GM, defining a generic function is a matter of declaring a record with a single field, a function of the appropriate type. As an example, we could define fmap2, the map operation for binary functors, as follows. newtype FMap2 a b c d f = FMap2{ appFMap2 :: (a c) (b d) f a b f c d} instance Generic (FMap2 a b c d) where empty = FMap2 ( Empty) plus = FMap2 (f g t case t of Inl x Inl (fmap2 f g x) Inr y Inr (fmap2 f g y)) prod = FMap2 (f g t case t of Prod x y Prod (fmap2 f g x) (fmap2 f g y)) par = FMap2 (f g (Par t) Par (f t)) rec = FMap2 (f g (Rec t) Rec (g t)) comp = FMap2 (f g (Comp t) Comp (fmap (fmap2 f g) t)) constant = FMap2 ( (Const t) (Const t)) fmap2 :: Rep f (a c) (b d) f a b f c d fmap2 = appFMap2 rep With fmap2 it is now possible to define several widely-applicable recursion operators [28, 14] using PolyP. For example, the cata-morphism operator could be defined as: cata iso f = f fmap2 id (cata iso f ) out iso Note that one must give explicitly the isomorphism that converts between the datatype and its representation. This contrasts with the original PolyP approach, in which that translation is inferred . This is the common trade-off of brevity for flexibility; being forced to state the isomorphism allows the programmer to choose a different one, giving something analogous to Wadler's ideas about 103 views [34]. We might say that this style of generic programming is isomorphism-parametrised instead of datatype-parametrised. In the original PolyP, the polytypic construct provides a convenient syntax for encoding generic functions. Furthermore, combinators for pointfree programming may be provided, making generic definitions even more compact. These combinators are just normal Haskell functions, and so there is no problem in implementing them in pure Haskell; but to keep the example short, we have stuck with pointwise definitions. The advantages of this translation when compared with the one proposed in [30] are that it requires only Haskell 98, and that the types of the generic functions are much closer to what one would expect. In Norell's translation, the type class constraints posed some problems because both the two-parameter class FunctorOf and the classes for the generic functions propagated throughout the code. With the Light PolyP approach, only instances of Rep propagate, leading usually to just one type class constraint. 4.2 Printf The C-style printf function, which takes a variable number of parameters, has always been a challenge for programmers using strongly and statically typed languages. The problem with printf is that, in its true essence, it requires dependent types. This happens because the value of the format string determines the type of the function. However, it has been shown by Danvy [10] that by changing the representation of the control string it is possible to encode printf in any language supporting a standard Hindley-Milner type system. 4.2.1 A solution using explicit representations In this section, we will demonstrate that Danvy's solution is another instance of the TypeCase design pattern, using an explicit representation . Furthermore, we will show a new use of the printf function by making use of the fact that we can (in some cases) infer the format string. Danvy's original solution had the following combinators: lit :: String (String a) String a lit x k s = k (s ++ x) eol :: (String a) String a eol k s = k (s ++ &quot;\n&quot;) int :: (String a) String Int a int k s x = k (s ++ show x) str :: (String a) String String a str k s x = k (s ++ x) eod :: String String eod = id If we capture all the occurrences of the form String t with a newtype Printf , and modify the definitions in order to reflect this newtype, we obtain the following code. newtype Printf t = Printf {printfApp :: String t} lit :: String Printf a Printf a lit x k = Printf (s printfApp k (s ++ x)) eol :: Printf a Printf a eol k = Printf (s printfApp k (s ++ &quot;\n&quot;)) int :: Printf a Printf (Int a) int k = Printf (s x printfApp k (s ++ show x)) str :: Printf a Printf (String a) str k = Printf (s x printfApp k (s ++ x)) eod :: Printf String eod = Printf id Taking one step further, we can now abstract over Printf and create a type class that replaces it with some functor f . class Format f where lit :: String f r f r eol :: f r f r int :: f r f (Int r) str :: f r f (String r) eod :: f String With this last transformation, we can start seeing an instance of the TypeCase design pattern. The structural cases participant consists of functions of the form Int r or String r, or a String -- lit and eol are overlapping cases. The class Format constitutes the typecase participant. Because the dispatcher is optional in explicit versions of the design pattern, there is no obligation to define it. Now, using the newtype Printf , we can define an instance of Format that implements the functionality of printf . instance Format Printf where lit x k = Printf (s printfApp k (s ++ x)) eol k = Printf (s printfApp k (s ++ &quot;\n&quot;)) int k = Printf (s x printfApp k (s ++ show x)) str k = Printf (s x printfApp k (s ++ x)) eod = Printf id The final touch is provided by the definition of printf in terms of printfApp. The printf function is expected to receive the formatting argument of type Printf t as its first parameter. The parameter t defines the type of printf , which can involve a variable number of arguments. Analysing the type of printfApp, we see that the first parameter is the formatting argument, the resulting type is the type that we expect for printf , and there is a second argument which is a String. Now, what does that String represent? Danvy's solution uses a continuation-passing style and the second argument of printfApp corresponds to the value fed to the initial continuation. Thus using the string &quot;&quot; for that argument does the trick. printf :: Printf t t printf p = printfApp p &quot;&quot; We have shown, informally, that Danvy's solution is indeed an instance of the TypeCase design pattern. However, some questions might be asked at this point. Do we really need to create a class in order to implement printf ? What other instances of the class would we be able to provide? In fact there are not many other uses for the type class; printf seems to be the only natural instance. Perhaps we could consider scanf , another C function that uses the same format string; but the derived type for scanf would be different, and so it is not possible to reuse the same type class. Another possibility would be considering other versions of printf , such as one for the IO monad. However, if we think that printf is really the only useful instance of the type class, why not get rid of the type class all together? A design pattern is a flexible design, and depending on the context of the problem, it can be adapted to fit the problem. If a type-indexed function is used at just one type index, it is reasonable to simplify the pattern and eliminate the type class. The result would be the specialised solution using the newtype Printf t presented before . We could go even further and argue that Danvy's original solution is already an instance of the design pattern, corresponding to one further simplification of the design pattern, namely getting rid of the newtype. 4.2.2 An alternative solution using smart datatypes In the previous section, we have argued that Danvy's version of printf is an instance of the TypeCase design pattern. However, Danvy's solution and explanation for printf is not, perhaps, very intuitive to understand. In this section, we take a different perpective and will look at the formatting parameter of printf as a special kind of list. This perpective corresponds to an instance of the design pattern using a smart datatype. The datatype (the typecase participant) encodes a list, which has an empty case that corres-104 ponds to the combinator eod, and a number of recursive cases that correspond to lit, eol, int and str. data Printf t where Lit :: String Printf t Printf t Eol :: Printf t Printf t Int :: Printf t Printf (Int t) Str :: Printf t Printf (String t) Eod :: Printf String Informally speaking, we have reused the types from the newtype solution and lifted the functions to constructors. However, using a datatype instead of a number of functions makes it easier to view the format parameter of printf as a list. For instance, the Lit constructor takes the literal string that we wish to print and also the list corresponding to the rest of the format parameter of printf . The printfApp from the previous section would, in this setting, correspond to a dependently-typed function (in the sense that the types of its branches are determined by the constructors used to perform pattern matching). printfApp :: Printf t String t printfApp (Lit x k) s = printfApp k (s ++ x) printfApp (Eol k) s = printfApp k (s ++ &quot;\n&quot;) printfApp (Int k) s = (x printfApp k (s ++ show x)) printfApp (Str k) s = (x printfApp k (s ++ x)) printfApp Eod s = s The final step is to define printf . Little effort is required; we just need to copy the definition of printf from the previous section. The only apparent difference between the two versions is that, where the first version uses functions like lit and int, this version uses constructors like Lit and Int. However, despite the similarity of the two solutions, their expressive power is not the same. The smart datatype solution in this section is fully closed to extension. That is, in order to add another case in the formatting list, such as a constructor Chr that handles characters, we would need to modify the GADT itself. On the other hand, the solution in the previous section using the explicit version of the design pattern allows some form of extensibility. Adding a new case for printf that handles characters corresponds to adding a new function, which could even be in a different module. 4.2.3 Making use of a dispatcher The two solutions that we presented did not make any use of a dispatcher. In this section we will show how the dispatcher can be useful. The version of the dispatcher presented here is for the explicit representation solution in Section 4.2.1, but could be easily adapted to the smart datatype solution in Section 4.2.2. Suppose that we want to define a function that prints a pair of integers. Equipped with printf , we could try to encode that with either one of the following two functions. printPair x y = printf fmt &quot;(&quot; x &quot;, &quot; y &quot;)&quot; where fmt = str $ int $ str $ int $ str $ eod printPair2 x y = printf fmt x y where fmt = lit &quot;(&quot; $ int $ lit &quot;, &quot; $ int $ lit &quot;)&quot; $ eod The function printPair tackles the problem using a printf that takes a format argument expecting five arguments: three strings and two integers. The function printPair2, on the other hand, makes use of the fact that the string arguments are constants, and uses lit instead. Thus, in this case, printf takes the format argument and two integer arguments. Although relatively compact, the format argument is not as convenient to use as it would be in C, where one would write something like &quot;(%d, %d)&quot;. The role of the dispatcher is to infer automatically the corresponding type representation for some type t. In the case of printf , it is not possible to infer all possible representations. Consider, for instance, the end of line case eol ::f r f r, which takes an existing format with some type r, adds a newline and returns a format of the same type. Clearly, there is no way to deduce that there is an occurrence of eol based on the type alone. Similarly, the lit case has no effect on the type. Nevertheless, the other, more type-informative, cases of printf can be inferred. class Rep t where rep :: Format f f t instance Rep String where rep = eod instance Rep r Rep (Int r) where rep = int rep instance Rep r Rep (String r) where rep = str rep We should note that these instance declarations are outside the scope of Haskell 98 -- types are used where type variables should occur. However, this is a quite mild extension, and is supported by most Haskell compilers. Making use of the fact that now we can infer some cases of the string format, we could define: printPair :: Int Int String printPair x y = printf rep &quot;(&quot; x &quot;, &quot; y &quot;)&quot; printTrio :: Int Int Int String printTrio x y z = printf rep &quot;(&quot; x &quot;, &quot; y &quot;, &quot; z &quot;)&quot; The function printPair does the same as before. However, with this new definition, the format directive is automatically inferred. The function printTrio is doing the same as printPair, except that it does it for triples. We should emphasise that the occurrences of printf in those two functions use different numbers of arguments. We should also mention that, in some situations, we will need to provide explicit types, otherwise the type checker would not be able to infer the correct instances of the type class Rep. This use of printf seems to be practical, and for this simple version of it we might even argue that everything that we could do with a manually-provided parameter could be done with an automatically-inferred one. We simply do not need lit and eol, because those can be simulated using str (with, of course, extra String arguments). Nevertheless, if we decided to go for a more powerful version of printf , this might not be the case. Consider, for instance, the formatting directive &quot;%2d&quot;. In this case the number 2 is specifying the minimum width of the string that represents that number. If we wanted to allow this kind of behaviour, we could add an extra parameter of type Int to the int case. However, the problem now is to choose a value for that parameter when we automatically build the format directive. In this case we need to use some default value (for instance 1). However we are no longer able, for all possible cases, to simulate the functionality of printf with manual format strings using only automatically-built ones. Type-indexed types Until now we have been discussing type-indexed functions, that is, families of functions indexed by types. We turn now to type-indexed types, that is, families of types indexed by types. In the context of generic programming, we call these generic types. Generic functions with generic types are functions that have different result types for each structural case. In this section, we will show how to implement type-indexed types as another variation of the TypeCase design pattern. We do this by translating a standard example of Generic Haskell [20], namely generic tries [17], into our approach. 105 5.1 Encoding type-indexed types Section 3 presents templates for encoding type-indexed functions. In this section, we show how to translate a type-indexed type into an instance of the TypeCase design pattern. In general, a type-indexed type has the form t :: :: t 1 a 1 ... a i = d 11 ... d 1n 1 . . . t m z 1 ... z j = d m1 ... d mn m where is the type-level function that defines the type-indexed type; t is the family of types (or type constructors) (t 1 a 1 ... a i ),..., (t m z 1 ... z j ) of kind that corresponds to the structural cases of the design pattern; and, finally, is the kind of t :: . For each type that is member of that family, we have a corresponding branch for . The type-level lambda abstraction on the right side of each branch is optional, and corresponds to possible parametrically polymorphic variables d 1 ... d n that the type-indexed type might depend on. Finally, 1 ... m corresponds to the family of types (or type constructors) that defines the type-indexed type. 5.1.1 Type class translation We can now derive an instance of the TypeCase design pattern to capture type-indexed functions with type-indexed types. The typecase participant, for instances of the design pattern using either implicit or explicit representations, could be defined as follows. class (g :: ) where case t 1 :: (a 1 ... a i ) g (t 1 a 1 ... a i ) d 11 ... d 1n 1 . . . case t m :: (z 1 ... z j ) g (t m z 1 ... z j ) d m1 ... d mn m We reuse the name for the name of the type class that encodes the typecase component. The parameter g is a type constructor with kind , where is the literal occurrence of (if we were to use instead of its literal occurrence , we would obtain the wrong kind). There are m functions case t 1 ,...,case t m that correspond to the typecases for each type (t 1 a 1 ... a i ),...,(t m z 1 ... z j ). Each case of the typecase function is defined by providing the type constructor g with the corre-ponding types. Finally, (a 1 ... a i ) ,..., (z 1 ... z j ) corresponds to the representations for the types (a 1 ... a i ),...,(z 1 ... z j ). The only difference between explicit and implicit versions of the design pattern for the typecase component is that in the explicit version the occurrences of are expanded into explicitly-passed representations of the form g a ..., whereas with the implicit representations those occurrences are replaced by type class constraints of the form Rep a .... The dispatcher can also be derived; but to do so requires extensions to Haskell 98 -- specifically, multiple-parameter type classes with functional dependencies. The problem is that, even in its simplest form, a type-indexed type requires at least two type arguments : the first one corresponding to the index type, and the second one that is the resulting type-indexed type for that index, and thus depending on the index. This problem is not too serious if we use the explicit representations variant of the pattern, since the dispatcher is optional, but using implicit representations forces us outside Haskell 98. class Rep t d 1 ... d n | t d 1 ... d n where rep :: g g t d 1 ... d n instance (a 1 ... a i ) Rep (t 1 a 1 ... a i ) d 11 ... d 1n 1 where rep = case t 1 {rep i } . . . instance (z 1 ... z j ) Rep (t m z 1 ... z j ) d m1 ... d mn m where rep = case t m {rep j } The type class Rep has at least two type arguments: t and . If there are parametric types that depends on, then the type class also needs to account for those types (d 1 ... d n ). The class contains just one member function, rep, used to build representations for . The function rep has a type class constraint ensuring that g is an instance of . There are, at least, m instances of Rep , and those instances define rep with the corresponding case t function. If we are implementing an implicit version of the design pattern, then the definition of rep is complete; otherwise, for an explicit version, we need to apply case t to a number i of rep functions (where i is the number of type parameters of t ). The constraints (a 1 ... a i ) ,..., (z 1 ... z j ) are very similar to the constraints , and in fact for implicit representations they coincide: they correspond to representations for the types a 1 ... a i ,...,z 1 ... z n . 5.1.2 Smart datatype translations Encoding type-indexed functions with smart datatypes proceeds in a similar fashion to the encoding with type classes. We will demonstrate how to do this translation using a GADT syntax (as found in the new GHC 6.4 Haskell compiler). A type-indexed type generates a smart datatype of the following form. data t d 1 ... d n where c t 1 :: (a 1 ... a i ) (t 1 a 1 ... a i ) d 11 ... d 1n 1 . . . c t m :: (z 1 ... z j ) (t m z 1 ... z j ) d m1 ... d mn m Instead of being parametrised by a "function" (like the type class approach), a smart datatype is parametrised by all the types on which it depends. Another difference from the type class approach is that the functions that represent each case are now replaced by constructors c t 1 ,...,c t m that can just be pattern matched (in a dependent manner) by functions defined over those datatypes. A final difference is that (a 1 ... a i ) ,..., (z 1 ... z j ) need to reflect the fact that we are now using a smart datatype. The changes to Rep are minimal; the only change to the type class version is that in the definition of rep we now use the constructors c t 1 ,...,c t m instead of the functions case t 1 ,...,case t m . class Rep t d 1 ... d n | t d 1 ... d n where rep :: t d 1 ... d n instance (a 1 ... a i ) Rep (t 1 a 1 ... a i ) d 11 ... d 1n 1 where rep = c t 1 rep i . . . instance (z 1 ... z j ) Rep (t m z 1 ... z j ) d m1 ... d mn m where rep = c t m rep j 5.2 Tries Tries or digital search trees are a traditional example of a generic type. Tries make use of the structure of search keys in order to organise information, which can then be efficiently queried. In this section we will show how to implement generic tries using a variation of the LAGP type representations. For a more theoretical presentation of tries, see [20, 17]; the implementation of tries presented here follows closely the implementations found in those papers. In [20], the generic type for tries is given as follows. 106 FMap t :: :: FMap Unit v = Maybe v FMap Int v = MapInt v FMap Plus t 1 t 2 v = OptPair (FMap t 1 v ) (FMap t 2 v ) FMap Prod t 1 t 2 v = FMap t 1 (FMap t 2 v ) It is clear that the type-indexed function FMap takes a type parameter t :: and another type of kind and returns another type of kind . Only the shape of parameter t is analysed; the other parameter v needs to be used in the definition because the resulting type is parametrically polymorphic in relation to v. We encode this characterisation of FMap as follows. class FMap g where unit :: g Unit v Maybe plus :: g a v c g b v d g (Plus a b) v (PlusCase c d) prod :: g a (d v) c g b v d g (Prod a b) v (ProdCase c d) data :: g a v c Iso b a Iso (d v) (c v) g b v d int :: g Int v MapInt This class forms the typecase participant of an explicit representation variant of the TypeCase pattern. The class FMap is a variation of the Generic class from Section 2.2.2. The functor g :: ( ) takes the necessary information to rebuild the type-indexed type. The three parameters of the functor correspond, respectively , to the type parameter t, the second parameter and the resulting type of FMap. (The kind of the resulting type is now . We could have used kind as in FMap, but we believe this version is slightly more readable.) The function unit just reflects the change of the functor g and adds the information for the parametric type v and the functor Maybe that is used to define the trie for the Unit case. The cases for plus and prod have explicit arguments that correspond to the recursive calls of the function; and the functors PlusCase c d and ProdCase c d correspond to the respective cases of the type-indexed type. The data function handles user-defined datatypes, having a recursive case and two isomorphisms: the first between the structural cases and a second between the tries corresponding to those cases. Finally, we could also define some extra base cases to handle primitive types such as Int and Char. The auxiliary definitions for the newtypes PlusCase a b v and ProdCase a b v are defined as follows. data OptPair a b = Null | Pair a b newtype PlusCase a b v = PlusCase {unPlus :: OptPair (a v) (b v)} newtype ProdCase a b v = ProdCase {unProd :: a (b v)} The introduction of OptPair a b is for efficiency reasons [20]. In order to use a user-defined type (or a built-in type that does not have a special case for it), we need to do much the same work as for GM2 in Section 2.2.2. As an example, we show what to do for Haskell's built-in lists. list :: FMap g g a (FList c v) c g [a] v (FList c) list ra = data (plus unit (prod ra (list ra))) listEP (Iso unFList FList) listEP :: Iso [a] (Plus Unit (Prod a [a])) listEP = Iso fromList toList where fromList [ ] = Inl Unit fromList (x : xs) = Inr (Prod x xs) toList (Inl Unit) = [ ] toList (Inr (Prod x xs)) = x : xs newtype FList c v = FList{ unFList :: (PlusCase Maybe (ProdCase c (FList c))) v} The function list defines the encoding for the representation of lists. Because lists are a parametrised datatype with one type parameter, list is a function that takes one argument; this argument corresponds to the representation of the list type argument, and list returns the representation for lists. The definition is nearly the same as the equivalent for GM, but it takes an extra isomorphism describing the mapping between the structural representation of a list trie and a newtype FList c v that is introduced to represent the resulting list trie. The function listEP is just the isomorphism [a] = 1 + a [a]. This means that listEP can be shared with other versions of generics that use the same structural cases. However, list and FList c v still have to be introduced for each type-indexed datatype. Nevertheless, that is boilerplate code, and, with compiler support, it is should be possible to avoid writing it. Having set up the main components of the design pattern, we can now move on to define our first function over tries. The function empty creates a new empty trie and can be defined as follows. newtype EmptyTrie a v t = EmptyTrie{empty :: t v} instance FMap EmptyTrie where unit = EmptyTrie Nothing int = EmptyTrie (MapInt [ ]) plus ra rb = EmptyTrie (PlusCase Null) prod ra rb = EmptyTrie (ProdCase (empty ra)) data ra iso iso2 = EmptyTrie (to iso2 (empty ra)) This function is very simple but, nonetheless, it has a type-indexed type: the unit case returns Nothing; the int case returns a value of a user-defined type for integer tries; the cases for prod and plus return, respectively, values for the previously defined ProdCase and PlusCase types; finally, the data returns a value of the newtype used to represent the trie of some user-defined datatype. Another function that we will probably want to have in a library for tries is the lookUp function which, given a key, returns the corresponding value stored in the trie. newtype LUp a v t = LUp{lookUp :: a t v Maybe v} instance FMap LUp where unit = LUp ( fm fm) int = LUp (i fm lookUpInt i fm) plus ra rb = LUp (t fm case (unPlus fm) of Null Nothing (Pair fma fmb) case t of (Inl l) lookUp ra l fma (Inr r) lookUp rb r fmb) prod ra rb = LUp (t (ProdCase fma) case t of (Prod x y) (lookUp ra x lookUp rb y) fma) data ra iso iso2 = LUp (t r lookUp ra (from iso t) (from iso2 r)) (The operator represents monadic composition.) The functions empty and lookUp have definitions that only have generic function calls to themselves. However, that is not the case for all generic functions. One such function is the generic function that creates a trie containing a single element; a possible definition makes use of the generic function empty. We discussed in Section 3 that, using an explicit version of the design pattern, there are some issues with generic functions calling generic functions other than themselves. One solution for this problem is using tupling. Just as one does with a type class, we would choose a fixed set of functions and group them together in a record. For instance, in the case of tries, we could have the following. data Tries a v t = Tries{ empty :: t v , isempty :: t v Bool, single :: a v t v, lookup :: a t v Maybe v, insert :: (v v v) a v t v t v, merge :: (v v v) t v t v t v, delete :: a t v t v} 107 With our definition we could, for any function in the record, make mutual generic calls. Whilst we could have used a multiple-parameter type class with functional dependencies in order to implement this library of functions over tries, there would be one important disadvantage in doing so (apart from the fact that we need to leave Haskell 98): we can only have functions on types of kind . With type classes, contexts are implicitly passed, and there is no way to redefine those implicit behaviours. In other words, type classes have the same limitation as implicit representations as a version of the TypeCase design pattern, in that they can only work on types. On the other hand, derived from the fact that we use external representations, with this implementation we can define generic functions over type constructors. Tupling is not the only option to solve the problem of generic function calls. Another possibility is to have the notion of dependencies : instead of tupling all functions together, we can, for each generic function that we need to use, include one instance of that function. Here is a possible definition of single using this strategy. data Single a v t = Single{ emptyT :: EmptyTrie a v t , single :: a v t v} instance FMap Single where unit = Single unit ( v Just v) int = Single int (i v MapInt [(i,v)]) plus ra rb = Single (plus (emptyT ra) (emptyT rb)) (i v case i of Inl l PlusCase (Pair (single ra l v) (empty (emptyT rb))) Inr r PlusCase (Pair (empty (emptyT ra)) (single rb r v))) prod ra rb = Single (prod (emptyT ra) (emptyT rb)) (i v case i of Prod x y ProdCase (single ra x (single rb y v))) data ra iso iso2 = Single (data (emptyT ra) iso iso2) (i v to iso2 (single ra (from iso i) v)) The idea of dependencies is motivated by Dependency-Style Generic Haskell [26, 27]. In this version of Generic Haskell, the type system reflects the uses of generic functions in the definitions by keeping track of constraints that identify such uses. With this definition , we have to manually introduce those dependencies by adding extra fields to the record that keep track of all the functions on which the definition depends. That change is also reflected in the instance that defines the generic function, where we need to provide values for the extra fields; the values for those fields just reconstruct the dependent functions with their values for those fields. Discussion and conclusions The goal of design patterns is not to come up with a miraculous solution for a problem. Instead, design patterns capture good techniques that appear in the literature or in practice, in a variety of contexts, and document them to make them easier to identify and implement. In this paper we have generalised the technique found in LIGD and GM to a design pattern, and presented a number of applications of the pattern. Furthermore, we have identified other occurrences of the design pattern in the literature. 6.1 Related work The technique used by Danvy [10] and generalised by Yang [37] allows us to encode type-indexed values in a Hindley-Milner type system. This encoding is directly related to the explicit representation version of the TypeCase pattern. This technique influenced many other works, ranging from type-directed partial evaluation [37, 9, 12], through embedded interpreters [2], to a generalisation of families of functions like zipWith [13] -- these are all possible applications of the TypeCase design pattern. Our paper revises that technique and shows how slightly richer type systems can be used to improve it. In particular, the use of a dispatcher makes it possible to automatically built the values encoding types. Moreover, the issue of composability (identified by Yang), while still a problem , can benefit from stronger type systems: the use of rank-two types combined with type classes provides a good solution. The work on extensional polymorphism [11] presents an approach that allows functions to implicitly bind the types of their arguments in a modified version of ML. Furthermore, using a typecase construct it is possible to support generic programming. Harper and Morrisett's work on intensional type analysis [16] presents an intermediate language where run-time type analysis is permitted, using typecase and Typecase constructs to define type-indexed functions and type-indexed types, respectively. However, approaches based on run-time type analysis have important drawbacks ; for instance, they cannot support abstract datatypes, and they do not respect the parametricity theorem [35, 33]. Subsequent approaches to intensional type analysis by Crary and others [7, 6] use a type-erasure semantics that does not suffer from those problems . Still, those approaches were limited to first-order type analysis . More recently, Weirich [36] proposed a version of intensional type analysis covering higher-order types with a type-erasure semantics . Furthermore, she presented an implementation in Haskell (augmented with rank-two types). This work inspired Hinze's implementation of GM, which shows, in essence, how to avoid rank-two types by using Haskell's class system. Our work makes use of those results and explains how to simulate typecase constructs. Furthermore, we show that the limitation of GM that generic functions with generic types cannot be defined can be lifted with our more general interpretation. Generic programming (or perhaps datatype-generic programming [15]) is about defining functions and types that depend on the structure of other types. One of the first attempts to produce a generic programming language was PolyP [21]. This language allowed the definition of generic functions over regular datatypes with one type parameter. In Section 4.1 we show that, using our design pattern, it is possible to define PolyP-like generic functions just using Haskell 98. A previous attempt [30] to define first-class PolyP functions in Haskell required extensions to the language. The Generic Haskell [26, 5] project is more ambitious than PolyP, and aims at defining generic functions for nearly all types defin-able in Haskell 98. Furthermore, Generic Haskell features generic types and generic function customisation (which were not present in PolyP). Dependency-Style Generic Haskell [26, 27] introduces a rather complex type system that keeps track of dependencies on generic function calls. The need for this sophisticated type system is a consequence of a model for generic programming that allows generic function customisation. The approach presented in [24] is another kind of lightweight approach to generic programming, relying on a run-time type-safe cast operator. With that operator it is possible to define a number of traversals that allow a very interesting model of generic programming based on nominal typing. Our design pattern can be used to encode many of the generic definitions that these generic programming techniques allow. However, it can be less practical than approaches providing a special-purpose compiler. Nevertheless, the advantage of our technique is that we do not need to commit in advance to a model of generic programming : we have the freedom to choose our own model of generic programming. Design patterns in the object-oriented programming community have been given a great deal of attention. Whilst amongst the 108 functional programming community there has been some work on -- or, at least, involving the concept of -- design patterns [24, 23, 25], the concept is still much less popular than in the object-oriented community. Moreover, most of this work presents patterns that are really more like algorithmic patterns rather than design patterns. Perhaps the reason why this happens is that functional languages are very expressive, and often natural features of those languages, such as laziness or higher-order functions, can be used to remove the need for complex designs. Nevertheless, we believe that our design pattern is more related to the OO concept of a design pattern with type classes/datatypes taking the role of OO interfaces and class instances taking the role of OO concrete classes. One difficulty found in this work had to do with the fact that, unlike OO design patterns which are documented using informal notations such as UML, we do not have a notation to "talk" about the design of Haskell programs. The notation that we used is quite ad-hoc and it can be difficult to read. 6.2 Future work We mentioned that this design pattern seems to be very similar to OO design patterns. It would be interesting to explore the applicability of this design pattern in an OO environment. Design patterns are useful to overcome the lack of certain features in programming languages. In our case, we overcome the lack of a typecase construct. The work on intensional type analysis investigates the possibility of languages supporting typecase constructs directly in the language. Combining these results in order to extend Haskell with a more natural support for typecase programming is something we would like to try in the future. Problems that use multiple instances of the design pattern are not composable. For instance, in a generic programming context, we could have a class Generic that allowed us to define generic functions with one type parameter; and we could also have a class FMap for working with tries. Although, those classes are structured in a similar way, they require two distinct representations of types, one for each of the classes; we hope to address this impracticality. Acknowledgements We would like to thank Ralf Hinze for the discussion that inspired this paper. Stefan Holdermans, the anonymous referees and the members of the Algebra of Programming group at Oxford and the EPSRC-funded Datatype-Generic Programming project made a number of helpful suggestions. References [1] C. Alexander. A Pattern Language. Oxford University Press, 1977. [2] N. Benton. Embedded interpreters. Microsoft Research, Cambridge, Jan. 2005. [3] R. Bird and O. de Moor. Algebra of Programming. International Series in Computer Science. Prentice Hall, 1997. [4] J. Cheney and R. Hinze. A lightweight implementation of generics and dynamics. In Haskell Workshop, pages 90104, 2002. [5] D. Clarke and A. Lh. Generic Haskell, specifically. In Generic Programming, pages 2147. Kluwer, B.V., 2003. [6] K. Crary and S. Weirich. Flexible type analysis. In International Conference on Functional Programming, pages 233248, 1999. [7] K. Crary, S. Weirich, and J. G. Morrisett. Intensional polymorphism in type-erasure semantics. In International Conference on Functional Programming, pages 301312, 1998. [8] N. A. Danielsson and P. Jansson. Chasing bottoms: A case study in program verification in the presence of partial and infinite values. In D. Kozen, editor, LNCS 3125: Mathematics of Program Construction, pages 85109. Springer-Verlag, 2004. [9] O. Danvy. Type-directed partial evaluation. In Principles of Programming Languages, 1996. [10] O. Danvy. Functional unparsing. Journal of Functional Programming , 8(6):621625, 1998. [11] C. Dubois, F. Rouaix, and P. Weis. Extensional polymorphism. In Principles of Programming Languages, pages 118129, 1995. [12] P. Dybjer and A. Filinski. Normalization and partial evaluation. In LNCS 2395: Applied Semantics, pages 137192. Springer, 2002. [13] D. Fridlender and M. Indrika. Do we need dependent types? Journal of Functional Programming, 10(4):409415, 2000. [14] J. Gibbons. Calculating functional programs. In Algebraic and Coalgebraic Methods in the Mathematics of Program Construction, pages 149202, 2000. [15] J. Gibbons. Patterns in datatype-generic programming. In Declarative Programming in the Context of Object-Oriented Languages, 2003. [16] R. Harper and G. Morrisett. Compiling polymorphism using intensional type analysis. In Principles of Programming Languages, pages 130141, San Francisco, California, 1995. [17] R. Hinze. Generalizing generalized tries. Journal of Functional Programming, 10(4):327351, 2000. [18] R. Hinze. Fun with phantom types. In J. Gibbons and O. de Moor, editors, The Fun of Programming, pages 245262. Palgrave, 2003. [19] R. Hinze. Generics for the masses. In International Conference on Functional Programming, pages 236243. ACM Press, 2004. [20] R. Hinze, J. Jeuring, and A. Lh. Type-indexed data types. Science of Computer Programming, 51(1-2):117151, 2004. [21] P. Jansson. Functional Polytypic Programming. PhD thesis, Chalmers University of Technology, May 2000. [22] J. Jeuring and P. Jansson. Polytypic programming. In J. Launchbury, E. Meijer, and T. Sheard, editors, LNCS 1129: Advanced Functional Programming, pages 68114. Springer-Verlag, 1996. [23] T. Khne. A Functional Pattern System for Object-Oriented Design. Verlag Dr. Kovac, ISBN 3-86064-770-9, Hamburg, Germany, 1999. [24] R. Lmmel and S. Peyton Jones. Scrap your boilerplate: a practical design pattern for generic programming. In Types in Language Design and Implementation, 2003. [25] R. Lmmel and J. Visser. Design patterns for functional strategic programming. In Workshop on Rule-Based Programming, 2002. [26] A. Lh. Exploring Generic Haskell. PhD thesis, Utrecht University, 2004. [27] A. Lh, D. Clarke, and J. Jeuring. Dependency-style Generic Haskell. In International Conference on Functional Programming, pages 141 152, 2003. [28] E. Meijer, M. Fokkinga, and R. Paterson. Functional programming with bananas, lenses, envelopes and barbed wire. In LNCS 523: Functional Programming Languages and Computer Architecture, pages 124144. Springer-Verlag, 1991. [29] D. Menendez. Fixed-length vectors in Haskell. http://www. haskell.org/pipermail/haskell/2005-May/015815.html . [30] U. Norell and P. Jansson. Polytypic programming in Haskell. In Implementing Functional Languages, 2003. [31] S. Peyton Jones, editor. Haskell 98 Language and Libraries: The Revised Report. Cambridge University Press, 2003. [32] S. Peyton Jones, G. Washburn, and S. Weirich. Wobbly types: Type inference for generalised algebraic data types. Microsoft Research, Cambridge, 2004. [33] J. C. Reynolds. Types, abstraction and parametric polymorphism. In Information Processing 83, pages 513523. Elsevier, 1983. [34] P. Wadler. Views: a way for pattern matching to cohabit with data abstraction. In Principles of Programming Languages, pages 307 313. ACM Press, 1987. [35] P. Wadler. Theorems for free! In Functional Programming and Computer Architecture, 1989. [36] S. Weirich. Higher-order intensional type analysis in type-erasure semantics. http://www.cis.upenn.edu/~sweirich/papers/ erasure/erasure-paper-july03.pdf , 2003. [37] Z. Yang. Encoding types in ML-like languages. In International Conference on Functional Programming, pages 289300, 1998. 109

Generic programming;type-indexed functions;type classes

203

UML-Based Service Robot Software Development: A Case Study

The research field of Intelligent Service Robots, which has become more and more popular over the last years, covers a wide range of applications from climbing machines for cleaning large storefronts to robotic assistance for disabled or elderly people. When developing service robot software, it is a challenging problem to design the robot architecture by carefully considering user needs and requirements, implement robot application components based on the architecture, and integrate these components in a systematic and comprehensive way for maintainability and reusability. Furthermore, it becomes more difficult to communicate among development teams and with others when many engineers from different teams participate in developing the service robot. To solve these problems, we applied the COMET design method, which uses the industry-standard UML notation, to developing the software of an intelligent service robot for the elderly, called T-Rot, under development at Center for Intelligent Robotics (CIR). In this paper, we discuss our experiences with the project in which we successfully addressed these problems and developed the autonomous navigation system of the robot with the COMET/UML method.

INTRODUCTION Robots have been used in several new applications. In recent years, both academic and commercial research has been focusing on the development of a new generation of robots in the emerging field of service robots. Service robots are individually designed to perform tasks in a specific environment for working with or assisting humans and must be able to perform services semi- or fully automatically [1]. Examples of service robots are those used for inspection, maintenance, housekeeping, office automation and aiding senior citizens or physically challenged individuals [2]. A number of commercialized service robots have recently been introduced such as vacuum cleaning robots, home security robots, robots for lawn mowing, entertainment robots, and guide robots [3, 4]. In this context, Public Service Robot (PSR) systems have been developed for indoor service tasks at Korea Institute of Science and Technology (KIST) [5, 6]. The PSR is an intelligent service robot, which has various capabilities such as navigation, manipulation, etc. Up to now, three versions of the PSR systems, that is, PSR-1, PSR-2, and a guide robot Jinny have been built. The worldwide aging population and health care costs of aged people are rapidly growing and are set to become a major problem in the coming decades. This phenomenon could lead to a huge market for service robots assisting with the care and support of the disabled and elderly in the future [8]. As a result, a new project is under development at Center for Intelligent Robotics (CIR) at KIST, i.e. the intelligent service robot for the elderly, called T-Rot. In our service robot applications, it is essential to not only consider and develop a well-defined robot software architecture, but also to develop and integrate robot application components in a systematic and comprehensive manner. There are several reasons for this: First, service robots interact closely with humans in a wide range of situations for providing services through robot application components such as vision recognition, speech recognition, navigation, etc. Thus, a well-defined robot control architecture is required for coherently and systematically combining these services into an integrated system. Second, in robot systems, there are many-to-many relations among software components as well as hardware components. For instance, a local map module requires range data from a laser scanner, ultrasonic sensors, and infrared sensors, as well as prior geometrical descriptions of the environment. On the other hand, the laser scanner should provide its data to a path planner, localizer, and a local map building module. These relationships, as well as interactions among software or hardware modules, must be carefully analyzed and systematically managed from an early stage of development in order to understand the big picture. Third, the functional performance of each software and hardware module becomes highly dependent on the architecture, as the number of robot platforms increases [6], and new services are added, or existing services are removed or updated to address changes in user needs. Fourth, previously developed software modules like maps, localization, and path planners can be directly reused for new tasks or services by service robot developers. Thus, a robot architecture, as well as systematic processes or methods, are required to support the implementation of the system, to ensure modularity and reusability. As a consequence, in the previous work [5,6], the Tripodal schematic control architecture was proposed to tackle the problems. Many related research activities have been done. However, it is still a challenging problem to develop the robot architecture by carefully taking into account user needs and requirements, implement robot application components based on the architecture, and integrate these components in a systematic and comprehensive way. The reason is that the developers of service robots generally tend to be immersed in technology specific components, e.g. vision recognizer, localizer and path planner, at an early stage of product development without carefully considering architecture to integrate those components for various services [9]. Moreover, engineers and developers are often grouped into separate teams in accordance with the specific technologies (e.g., speech processing, vision processing), which makes integration of these components more difficult [7, 9]. In such a project like T-Rot, particularly, several engineers and developers (i.e., approximately, more than 150 engineers) from different organizations and teams participate in the implementation of the service robot. Each separate team tends to address the specific technologies such as object recognition, manipulation, and navigation and so on. Engineers who come from different teams are concerned with different characteristics of the system. Thus, a common medium is required to create mutual understanding, form consensus, and communicate with each other for successfully constructing the service robot. Without such a medium or language, it is difficult to sufficiently understand the service robot system and interact between teams to integrate components for services. Within the domain of software engineering, many approaches have been suggested for a systematic and complete system analysis and design, and for the capture of specifications. The object-oriented paradigm [10,11] is a widely-accepted approach to not only cover the external and declarative view of a system, but also at the same time bridge seamlessly with the internal implementation view of a system [13]. Object-oriented concepts are crucial in software analysis and design because they focus on fundamental issues of adaptation and evolution [14]. Therefore, compared with the traditional structured software development methods, object-oriented methods are a more modular approach for analysis, design, and implementation of complex software systems, which leads to more self-contained and hence modifiable and maintainable systems. More recently, the Unified Modeling Language (UML) [15,16] has captured industry-wide attention for its role as a general-purpose language for modeling software systems, especially for describing object-oriented models. The UML notation is useful to specify the requirements, document the structure, decompose into objects, and define relationships between objects in a software system. Certain notations in the UML have particular importance for modeling embedded systems [17,18], like robot systems. By adopting the UML notation, development teams thus can communicate among themselves and with others using a defined standard [14,17,18]. More importantly, it is essential for the UML notation to be used with a systematic object-oriented analysis and design method in order to be effectively applied [14]. As a result, our aim is to develop the intelligent service robot based on the systematic software engineering method, especially for real-time, embedded and distributed systems with UML. To do so, we applied the COMET method, which is a UML based method for the development of concurrent applications, specifically distributed and real-time applications [14]. By using the COMET method, it is possible to reconcile specific engineering techniques with the industry-standard UML and furthermore to fit such techniques into a fully defined development process towards developing the service robot systems. In this paper, we describe our experience of applying the COMET /UML method into developing the intelligent service robot for the elderly, called T-Rot, developed at CIR. In particular, we focused on designing an autonomous navigation system for the service robot, which is one of the most challenging issues for the development of service robots. Section 2 describes the hardware configuration and services of the T-Rot, and discusses the related work. Section 3 illustrates how to apply the COMET method into designing and developing the autonomous navigation system for the service robot, and discusses the results of experiments. The lessons learned from the project are summarized in section 4, and section 5 concludes the paper with some words on further work. BACKGROUD ON T-Rot Fig. 1. KIST service robots At KIST, intelligent service robots have been developed in large-scale indoor environments since 1998. So far, PSR-1 and PSR-2, which performs delivery, patrol, and floor cleaning jobs, and a guide robot Jinny, which provides services like exhibition guide and guidance of the road at a museum, have been built [5,6] (see Fig. 1). The service robot T-Rot is the next model of the PSR system under development for assisting aged persons. Development of T-Rot, in which our role is developing and integrating robot software, started in 2003 by mainly CIR with 535 more than 10 groups consisting of more than 150 researchers and engineers from academia and industry. This project is based on the needs and requirements of elderly people through the studies and analysis of the commercial health-care market for providing useful services to them. Thus, the aim of this project is to develop the intelligent service robot for the elderly by cooperating and integrating the results of different research groups. This project that is divided into three stages will continue until 2013 and we are now in the first stage for developing the service robot incrementally to provide various services. 2.2 Hardware of T-Rot The initial version of T-Rot, as shown in Fig. 2, has three single board computer (SBC), that is, mobile Pentium 4 (2.2GHz) and 1GB SDRAM on each SBC. In terms of software environment, Linux Red hat 9.0 and RTAI (Real-Time Application Interface) [12] are used as operating system. Fig. 3 shows hardware configuration as a whole. As mentioned earlier, development of T-Rot is conducted incrementally for various services and thus the platform will be extended with manipulators and robot hands later. In our project, we developed the robot software based on the initial version of the platform. The details of the hardware platform are described in Table 1. Fig. 2. T-Rot robot hardware platform Fig. 3. T-Rot robot hardware platform configuration Table 1. T-Rot hardware platform devices Intel Mobile Pentium 4 (2.2 GHz) 1GB SDRAM SBC 30GB Hard Disk 16 microphones for speaker localization 1 microphone for speech recognition Voice 1 speaker for speech generation Vision 2 stereo vision cameras for recognizing users and object s (1288 H x 1032 V maximum resolution and 7Hz fram e rates) Pan/Tilt for controlling the vision part 2 laser scanners (front and back) 2 IR scanners (front and back) 12 Ultrasonic sensors Sensor 1 Gyroscope sensor for measuring balance 2 actuators for two drive wheels (right and left) 2 free wheels (the support wheels) 2 Servo Motors (100 [w]) 2 encoders (2048 ppr) Actuator 2 bumpers 1 TFT LCD & Touch (10.4" 1024x768, 26000 colors) KVM (Keyboard/Mouse) Interface Wireless LAN for communications 2.3 Robot Services Some of the primary services under-developed that the initial version for T-Rot provides for the elderly are described as below. Voice-based Information Services: The robot T-Rot can recognize voice commands from a user (i.e., an aged person) via microphones equipped with the robot and can synthesize voices for services. While a user is watching TV, the user can ask some questions about the specific TV program or request a task to open an Internet homepage by speaking the TV program name. Sound Localization and Voice Recognition: A user can call a robot's predefined name, to let the robot recognize the call while the robot knows the direction to move to the user. This service analyzes audio data from 3 microphones on the shoulder for sound localization and 16 mic array on the head for speech recognition to recognize the command from the user. Autonomous navigation: A user can command the robot to move to a specific position in the map to perform some task. For instance, the robot can navigate to its destination in the home environment via its sensors, which include laser scanners and ultrasonic sensors. The robot plans a path to the specified position, executes this plan, and modifies it as necessary for avoiding unexpected obstacles. While the robot is moving, it constantly checks sensor data from its sensors every 200 ms. An errand service: The robot can carry objects that a user (i.e., an aged person) usually uses, like a plate, books, a cane a cup of tea, beverages, etc according to the user's instructions. For instance, the user can order the robot to bring a cup of tea or beverage by speaking the name of the drink. Of these T-Rot services, our emphasis was on the autonomous navigation service, which is one of the most challenging issues and is essential in developing service robots, particularly mobile service robots to assist elderly people. It includes hardware integration for various sensors and actuators, and the development of crucial navigation algorithms like maps, path planners, and 536 localizers as well as software integration of software modules like a path planner, a localizer, and a map building module. 2.4 Control Architecture of PSR Up to now, there have been many related research activities to develop efficient and well-defined control architectures and system integration strategies for constructing service robots. A recent trend is that many control architectures are converging to a similar structure based on a hybrid approach that integrates reactive control and deliberation [6]. At KIST, for developing service robots, that is PSR-1, PSR-2, and Jinny in the previous work [5,6], the Tripodal schematic control architecture was proposed as the solution to the problem. One important point of Tripodal schematic design is to integrate robot systems by using a layered functionality diagram. The layered functionality diagram is a conceptual diagram of three layers for arrangement of various hardware and software modules and functions. It also shows the connectivity and the information flow between components. Those layers are composed of deliberate, sequencing, and reactive layers based on the hybrid approach. The purposes of the deliberate layer are to interface with a user and to execute a planning process. The sequencing layer is classified into two groups, that is, the controlling part that executes the process by managing the components in the reactive layer and the information part that extracts highly advanced information from sensor data. The reactive layer controls the real-time command and hardware-related modules for sensors and actuators. The detailed description of whole control architecture of the PSR is introduced in [5]. However, as described earlier, in order to effectively apply this approach and the UML notation to developing service robots, it is essential to use a systematic software engineering process or methods like object-oriented analysis and design methods, especially for real-time and embedded systems. We believe that only a systematic and comprehensive software development process and method will be able to resolve the issues discussed before and will be vital for success in developing service robots. 2.5 The COMET method COMET [14] is a method for designing real-time and distributed applications, which integrates object-oriented and concurrent processing concepts and uses the UML notation [15,16]. The COMET object- oriented software life cycle model is a highly iterative software development process based around the use case concept. Therefore, in this project, the COMET method with UML was used to develop a system for autonomous navigation by the intelligent service robot, T-Rot. The method separates requirements activities, analysis activities and design activities, and these activities are briefly described as below. The details are described in section 3 with the case study. Requirements modeling - A use case model is developed in which the functional requirements of the system are defined in terms of actors and use cases. Analysis modeling - Static and dynamic models of the system are developed. The static model defines the structural relationships among problem domain classes. A dynamic model is then developed in which the use cases from the requirements model are refined to show the objects that participate in each use case and how they interact with each other. Design modeling The software architecture of the system is designed, in which the analysis model is mapped to an operational environment. For distributed applications, a component based development approach is taken, in which each subsystem is designed as a distributed self-contained component. APPLYING THE COMET/UML METHOD TO T-ROT In this section, we explain how to develop robot software for the autonomous navigation system with the COMET/UML method. In our project, the UML notation conforms to UML 1.3 and the Rational Rose tool is used. 3.1 Requirements Modeling Capturing the functional requirements of the system is the first phase in software development, which defines what the system should do or provide for the user. In our approach, developers can catch the functional requirements or services by using the use case model in terms of use cases and actors (see Fig. 4). To identify and define the requirements of the system more clearly, the system has to be considered like a black box. In the service robot, the actor can be usually a human user as well as external I/O devices and external timer. Navigation Commander (from 1.0 Actors) Clock (from 1.0 Actors) Obstacle Avoidance &lt;&lt;extend&gt;&gt; Fig. 4. Use case diagram for Navigation Table 2 shows a specification for Navigation use case. In our navigation system, we identified a Commander and a Clock as an actor. While the robot is moving, if the robot recognizes obstacles, it should avoid them for continuing to go to the destination. Even when humans or objects suddenly appear, the robot must be able to stop to avoid crashing into those. However, in order to do this, the robot has to check that there are obstacles by using sensor data more often (e.g., every 50 ms) than the normal navigation system does (e.g., every 200 ms). As a result, the Obstacle Avoidance use case is extended from the Navigation use case. During the Navigation use case is executing, if the obstacles are recognized, then the Obstacle Avoidance use case is triggered to perform the emergency stop of the robot. If the obstacles disappear, the robot moves again to the destination. Table 2. Navigation use case Summary The Commander enters a destination and the robot system moves to the destination. Actor Commander Precondition The robot system has the grid map and the current position is known Description 1. The use case begins when the commander enters a destination. 2. The system calculates an optimal path to the destination. 3. The system commands the wheel actuator to start 537 moving to the destination. 4. The wheel actuator notifies the system that it has started moving 5. The system periodically reads sensor data and calculates the current position. 6. The system determines that it arrives at the destination and commands the wheel actuator to stop. 7. The wheel actuator notifies the system that it has stopped moving and the use case is finished. Alternative 6.1. If the system doesn't arrive at the destination, it keeps moving. Postcondition The robot system is at the destination and waiting for the next destination. 3.2 Analysis Modeling 3.2.1 Static Modeling The objective of static modeling is to understand the interface between the system and the external environment and to describe the static structure of the system under development by developing a system context class diagram. It is specifically important for real-time and embedded systems like robot systems [14]. The system context class diagram can be determined by static modeling of the external classes that connect to the system. Commander (from 1.0 Actors) Sensor &lt;&lt;external input device&gt;&gt; WheelActuator &lt;&lt;external output device&gt;&gt; CommandLine &lt;&lt;external user&gt;&gt; 1 1 1 1 Robot Navigation System &lt;&lt;System&gt;&gt; 1..* 1 1..* 1 Inputs To 1 1 1 1 Outputs to 1 1 1 1 interacts with Clock (from 1.0 Actors) Clock &lt;&lt;external timer&gt;&gt; 1 1 1 1 Awakens Fig. 5. Robot Navigation System context class diagram The system context class diagram of the Robot Navigation System is shown in Fig. 5, which illustrates the external classes to which the system has to interface. In our navigation system, a commander enters a destination via a command line, to which the robot should move. The system uses sensor data via various sensors such as laser scanners, IR scanners, ultrasonic sensors, etc and it controls the wheels of the robot via the wheel actuator. Therefore, the external classes correspond to the users (i.e., a Commander who interacts with the system via a Command Line), and I/O devices (i.e., a Sensor and Wheel Actuator). A Clock actor needs an external timer class called Clock to provide timer events to the system. This external timer class is needed to periodically check sensor data via those sensors for avoiding obstacles (i.e., doing the emergency stop) while the robot is moving. Next, to structure the Robot Navigation System into objects, object structuring needs to be considered in preparation for dynamic modeling. The objective of the object structuring is to decompose the problem into objects within the system. We identified the internal objects according to the object structuring criteria in COMET (see Fig. 6). In our system, interface objects, i.e. a Command Line Interface, Sensor Interface and Wheel Actuator Interface are identified by identifying the external classes that interface to the system, i.e. the Command Line, Sensor, and Wheel Actuator, respectively. There are four entity objects identified, that is, a Current Position, Destination, Navigation Path and Navigation Map, which are usually long-living object that stores information. In addition to those objects, there is a need for control objects, which provide the overall coordination for objects in a use case and may be coordinator, state-dependent control, or timer objects. The Navigation System has a state-dependent control object called Navigation Control that controls the wheel actuator and sensors. The states of the Navigation Control object are shown on a Navigation Control statechart (this will be discussed in the dynamic modeling). There are two timer objects, i.e. a Navigation Timer and an Obstacle Avoidance Timer. The Obstacle Avoidance Timer is activated by a timer event from an external timer to periodically check that there is any obstacle around the robot. On the other hand, the Navigation Timer is started by the Navigation Control object and generates a timer event for navigation. Also, a Localizer algorithm object and Path Planner algorithm object are identified, which encapsulate an algorithm used in the problem domain, namely the autonomous navigation. &lt;&lt; Robot Navigation System &gt;&gt; Commander (from 1.0 Actors) CommandLineInterface &lt;&lt;user interface&gt;&gt; CommandLine &lt;&lt;external user&gt;&gt; 1 1 1 1 1 1 1 1 SensorInterface &lt;&lt;input device interface&gt;&gt; Sensor &lt;&lt;external input device&gt;&gt; 1 1..* 1 1..* WheelActuator &lt;&lt;external output device&gt;&gt; WheelActuatorInterface &lt;&lt;output device interface&gt;&gt; 1 1 1 1 Destination &lt;&lt;entity&gt;&gt; Navigation Path &lt;&lt;entity&gt;&gt; Navigation Map &lt;&lt;entity&gt;&gt; Current Position &lt;&lt;entity&gt;&gt; Navigation Control &lt;&lt;state dependent&gt;&gt; Navigation Timer &lt;&lt;timer&gt;&gt; ObstacleAvoidanceTimer &lt;&lt;timer&gt;&gt; Clock &lt;&lt;external timer&gt;&gt; 1 1 1 1 Localizer &lt;&lt;algorithm&gt;&gt; PathPlanner &lt;&lt;algorithm&gt;&gt; Fig. 6. Object structuring class diagram for Navigation System 3.2.2 Dynamic Modeling Dynamic modeling emphasizes the dynamic behavior of the system and plays an important role for distributed, concurrent and real-time system analysis. The dynamic model defines the object interactions that correspond to each use case and thus is based on the use cases and the objects identified during object structuring. In our case, collaboration diagrams are developed to show the sequence of object interactions for each use case. Additionally, if the collaboration involves the state-dependent object, which executes a statechart, the event sequence is shown on a statechart. : Navigation Control : CommandLine : Sensor : WheelActuator : WheelActuatorInterface : SensorInterface : Destination : Navigation Path : Navigation Map : Current Position : CommandLineInterface : Navigation Timer Path Planner Localizer Sequencing Layer &lt;&lt;external user&gt;&gt; &lt;&lt;user interface&gt;&gt; &lt;&lt;state dependent control&gt;&gt; &lt;&lt;timer&gt;&gt; &lt;&lt;entity&gt;&gt; &lt;&lt;algorithm&gt;&gt; &lt;&lt;entity&gt;&gt; &lt;&lt;entity&gt;&gt; &lt;&lt;entity&gt;&gt; &lt;&lt;algorithm&gt;&gt; &lt;&lt;external input device&gt;&gt; &lt;&lt;input device interface&gt;&gt; &lt;&lt;output device interface&gt;&gt; &lt;&lt;external output device&gt;&gt; Deliberate Layer Reactive Layer 1.2a: Store Destination 2.11, 3.11 : Check Destination 2.12 : No , 3.12: Yes 1.13, 2.18: Planned Path 1.10, 2.15: Read a Path 1.14: Start 2.19: Move 3.13: Stop 1.17: Started 3.16: Stopped 1.4, 2.7, 3.7: Read Current Position 1.7, 2.10, 3.10: Current Position 1.2, 2.5, 3.5: Read Map 1.8, 2.13: Update Map 1.3, 2.6, 3.6 : Map 1.9, 2.14: Updated Map 1: Enter Destination 1.1: Destination Entered 2.1, 3.1: Read Sensors 2.4, 3.4: Sensor Data 2.2, 3.2: Read 2.3, 3.3: Data 1.15: Start WheelActuator Output 2.20:Move WheelActuator Output 3.14: Stop WheelActuator Output 1.16, 5.8: Started Ack 3.15: Stopped Ack 1.5, 2.8, 3.8: Localize 1.6, 2.9, 3.9: Current Position 2, 3: After(Elapsed Time) 1.18, 5.10: Start Timer 3.17, 4.10: Stop Timer 1.12, 2.17: Path 1.11, 2.16: Plan a path Fig. 7. Collaboration diagram for Navigation use case 538 In the navigation system, the Localizer has the algorithm which can calculate the current position based on sensor data via sensors. So, the role of the Localizer is to update the current position of the service robot. In the Path Planner object, there is a method for calculating a path to arrive at the destination based on both sensor information and the current position that is calculated at the Localizer. The Navigation Timer is an internal timer that is controlled by the Navigation Control. After the destination is entered from the external user, the Navigation Control starts the Navigation Timer, then the timer generates a timer event periodically (i.e., every 200ms) until the Navigation Control stops the timer. The Navigation use case starts with the commander entering the destination into the navigation system. The message sequence number starts at 1, which is the first external event initiated by the actor. Subsequence numbering in sequence is from 1.1 to 1.18 as shown in Fig. 7. The next message sequence activated by the Navigation Timer is numbered 2, followed by the events 2.1, 2.2, and so forth. The following message sequences are illustrated in the collaboration diagram (see Fig. 7). The collaboration diagram for the Obstacle Avoidance use case is shown in Fig. 8. When activated by the Obstacle Avoidance Timer every 50 ms, the Sensor Interface object reads sensor data via various sensors (Events 4.1, 5.1, 6.1). If an obstacle is recognized, the Obstacle Avoidance Timer sends the emergency stop message to the Wheel Actuator Interface (Event 4.5). Afterwards, the timer also sends a suspend event to the Navigation Control. If the obstacle disappears, the timer sends a restart event to the Navigation Control for the robot to move again. : Sensor : WheelActuator : WheelActuatorInterface : SensorInterface : Clock : ObstacleAvoidanceTimer &lt;&lt;state dependent control&gt;&gt; &lt;&lt;external timer&gt;&gt; &lt;&lt;external input device&gt;&gt; &lt;&lt;timer&gt;&gt; &lt;&lt;input device interface&gt;&gt; &lt;&lt;output device interface&gt;&gt; &lt;&lt;external output device&gt;&gt; : Navigation Control Sequencing Layer Reactive Layer 5.6: Start 5.9: Started 4, 5, 6: Timer Event 4.2, 5.2, 6.2: Read 4.3, 5.3, 6.3: Data 4.1, 5.1, 6.1: Read Sensors 4.4, 5.4, 6.4: Sensor Data 4.5: Stop 4.8: Stopped 4.9: Suspend 5.5: Restart 6.5: Time Expired 5.7: Start WheelActuator Output 4.6 : Stop WheelActuator Output 5.8: Started Ack 4.7: Stopped Ack Fig. 8. Collaboration diagram for Obstacle Avoidance use case With COMET, the software architecture can be based on a software architectural style (pattern) such as client/server or layers of abstraction. In our project, the layered strategy of the Tripodal schematic design described in section 2 is applied for design and modeling of the robot system, which provides a conceptual diagram of three layers (i.e., deliberate, sequencing, and reactive layers) for arrangement of various hardware and software modules and functions. Therefore, in the collaboration diagrams (see Fig. 7 and 8), the Command Line Interface is located in the deliberate layer and the Sensor Interface, Wheel Actuator Interface, and Obstacle Avoidance Timer are in the reactive layer. The others are positioned in the sequencing layer. In our navigation system, after drawing the collaboration diagrams for the Navigation and Obstacle Avoidance use cases which include the Navigation Control state-dependent object, we develop a Navigation Control statechart, which is executed by the Navigation Control object. The statechart needs to be considered in connection with the collaboration diagram. Specifically, it is required to take into account the messages that are received and sent by the control object, which executes the statechart [14]. An input event (e.g., 1.1: destination entered) into the Navigation Control object on the collaboration diagram should be consistent with the same event shown on the statechart. The output event, which causes an action, enable or disable activity, like 1.2: Read Map (which cases an action) on the statechart must be consistent with the output event depicted on the collaboration diagram. Because the statechart modeling involves two state-dependent use cases in the navigation system, it is also required to consolidate the two partial statecharts to create a complete statechart. The complete statechart for both the Navigation and Obstacle Avoidance use cases is shown in Fig. 9. Idle Starting Planning a Path Checking Destination Stopping 3.16: Stopped / 3.17: Stop Timer Reading Sensors Localizing Moving 1.17, 5.9: Started / 1.18, 5.10: Start Timer 1.13: Planned Path[ Start ] / 1.14: Start 2.18: Planned Path[ Move ] / 2.19: Move Reading Map 2.4, 3.4: Sensor Data / 2.5, 3.5: Read Map 1.1: Destination Entered / 1.2 : Read Map, 1.2a: Store Destination 1.3, 2.6, 3.6: Map / 1.4, 2.7, 3.7: Read Current Position Updating Map 2.10, 3.10:Current Position[ Move ] / 2.11, 3.11: Check Destination 1.7: Current Position[ Start ] / 1.8: Update Map 1.9, 2.14: Updated Map / 1.10, 2.15: Read a Path 2.12: No / 2.13: Update Map 3.12 : Yes / 3.13: Stop Suspending 2, 3: After( Elapsed Time ) / 2.1, 3.1: Read Sensors 4.9: Suspend / 4.10: Stop Timer 5.5: Restart / 5.6: Start 6.1: Time Expired Fig. 9. Statechart for Navigation Control 3.3 Design Modeling 3.3.1 Software Architecture In this phase, all collaboration diagrams developed for use cases in the analysis model are merged into the consolidated collaboration diagram. The consolidated collaboration diagram is thus intended to be a complete description of all objects and their interactions. The consolidation of the two collaboration diagrams respectively supporting the two use cases is shown in Fig. 10. Some objects and message interactions appear on more than one collaboration diagram. For instance, the Navigation Control, Navigation Timer, Sensor Interface and Wheel Actuator Interface objects participate in both the Navigation and Obstacle Avoidance use cases. For those objects, their message interactions are only shown once in the consolidated collaboration diagram. 3.3.2 Architectural Design of Distributed Real-time Systems The robot system is a distributed embedded system and executes on distributed nodes by the communication methods like TCP/IP, CAN (Controller Area Network), and Wire/Wireless LAN. With COMET, a distributed real-time system is structured into distributed subsystems. Tasks in different subsystems may communicate with each other via several types of message communication, such as asynchronous, synchronous with reply, synchronous without reply, and client/server communications, etc. 539 Hence, we should define distributed nodes and their messages to each node. The overall distributed software architecture for the robot navigation system is depicted in Fig. 11. In the robot system, objects that are part of the navigation are located in the robot navigation system. The robot navigation system communicates with the external I/O devices via synchronous message without reply communication and with the external timer via asynchronous message communication. : Navigation Control : CommandLine : Sensor : WheelActuator : WheelActuatorInterface : SensorInterface : Destination : Navigation Path : Navigation Map : Current Position : CommandLineInterface : Navigation Timer Path Planner Localizer &lt;&lt;external user&gt;&gt; &lt;&lt;user interface&gt;&gt; &lt;&lt;state dependent control&gt;&gt; &lt;&lt;timer&gt;&gt; &lt;&lt;entity&gt;&gt; &lt;&lt;algorithm&gt;&gt; &lt;&lt;entity&gt;&gt; &lt;&lt;entity&gt;&gt; &lt;&lt;entity&gt;&gt; &lt;&lt;algorithm&gt;&gt; &lt;&lt;external input device&gt;&gt; &lt;&lt;input device interface&gt;&gt; &lt;&lt;output device interface&gt;&gt; &lt;&lt;external output device&gt;&gt; : Clock : ObstacleAvoidanceTimer &lt;&lt;external timer&gt;&gt; &lt;&lt;timer&gt;&gt; Deliberate Layer Sequencing Layer Reactive Layer Store Destination Check Destination Yes/No Planned Path Read a Path Start Move Stop Started Stopped Read Current Position Current Position Read Map Update Map Map Enter Destination Start WheelActuator Output Move WheelActuator Output Stop WheelActuator Output Started Ack Stopped Ack Read Sensors Sensor Data Read Data Read Sensors Sensor Data Localize Current Position Destination Entered After(Elapsed Time) Start Timer Stop Timer Path Plan a path Timer Event Suspend Restart Time Expired Stop Stopped Fig. 10. Consolidated collaboration diagram for Navigation System : CommandLine : Sensor : WheelActuator : Robot Navigation System &lt;&lt; synchronous message without reply&gt;&gt; &lt;&lt; synchronous message without reply&gt;&gt; &lt;&lt; synchronous message without reply&gt;&gt; : Clock &lt;&lt;asynchronous message&gt;&gt; Enter Destination Start WheelActuator Output Stop WheelActuator Output Move WheelActuator Output Read Timer Event Fig. 11. Distributed software architecture for Navigation System 3.3.3 Task Structuring During the task structuring phase, a task architecture can be developed in which the system is structured into concurrent tasks, and the task interfaces and interconnections are defined. A task is an active object and has its own thread of control. In this sense, the term "object" will be used to refer to a passive object in this paper. In COMET, task structuring criteria are provided to help in mapping an object-oriented analysis model of the system to a concurrent tasking architecture. At the end of this phase, a task behavior specification (TBS) is developed. The task architecture for the Navigation System is shown in Fig. 12. In order to determine the tasks in the system, it is necessary to understand how the objects in the application interact with each other based on the collaboration diagrams. In the collaboration diagram of Fig. 7, the Localizer object reads sensor data and the map from the Current Position object, calculates a new current position, and sends the current position to the Current Position object for updating it. Thus, the Localizer object is structured as an asynchronous algorithm task called Localizer. There are two asynchronous algorithms, i.e. Localizer and Path Planner, which are internal asynchronous tasks. There are four passive entity objects, i.e. Destination, Current Position, Navigation Map, and Navigation Path, which do not need a separate thread of control and are further all categorized as data abstraction objects. The Sensor and Wheel Actuator are a passive input device and a passive output device, respectively because they do not generate an interrupt on completion of the input or output operation. : CommandLine : Sensor : WheelActuator &lt;&lt; external user &gt;&gt; &lt;&lt; passive input device &gt;&gt; &lt;&lt; control clustering &gt;&gt; &lt;&lt; passive output device &gt;&gt; : Destination : Navigation Path : Navigation Map : Current Position &lt;&lt; data abstraction &gt;&gt; : Localizer : Path Planner &lt;&lt; asynchronous algotithm &gt;&gt; &lt;&lt; asynchronous algotithm &gt;&gt; : Navigation Controller &lt;&lt; data abstraction &gt;&gt; &lt;&lt; data abstraction &gt;&gt; &lt;&lt; data abstraction &gt;&gt; : Clock : Navigation Controller &lt;&lt;sequential clustering&gt;&gt; &lt;&lt;external timer&gt;&gt; Reactive Layer Deliberate Layer Sequencing Layer enter (in destination) read(out sensorData) read(out sensorData) read(out sensorData, out map) update(in currentPosition) read(out destination,out currentPosition,out map) update(in path) store(destination) check(currentPosition, yes/no) read(in sensorData,in map,out currentPosition) Start WheelActuator Output Move WheelActuator Output Stop WheelActuator Output read(in destination,in currentPosition,in map, out path) read(out map) update(in sensorData, in currentPosition, out map) suspend() restart() timerEvent stopWheelActuatorOutput Fig. 12. Task architecture for Navigation System The Navigation Control is a state-dependent control object that executes the Navigation Control statechart and is structured as a control task because it needs to have a separate thread of control. The Navigation Control object can be combined with the Command Line Interface, Navigation Timer, Sensor Interface, and Wheel Actuator Interface objects into one task, Navigation Controller, based on the control clustering task structuring criterion because it is not possible for them to execute concurrently (see the middle of Fig. 12). The Obstacle Avoidance Timer object is structured as a periodic task, activated periodically to read sensor data. It can be grouped with the Sensor Interface and Wheel Actuator Interface into one sequentially clustered task, Obstacle Avoidance Controller based on sequential clustering since those are carried out in a sequential order. The design of those composite tasks, the Navigation Controller and Obstacle Avoidance Controller are considered in the next section (i.e., detailed software design). After developing the task architecture, a task behavior is described for specifying the characteristics of each task based on COMET. During the task structuring, the TBS focuses on the task inputs and outputs. One part of the TBS, i.e. the task's event sequencing logic is defined in the detailed software design phase. 3.3.4 Detailed Software Design The internals of composite tasks which have passive objects nested inside them are designed, detailed task synchronization issues are addressed, and each task's internal event sequencing logic is defined in this phase. Before this is done, the information hiding classes (from which the passive objects are instantiated) are designed. In particular, the operations of each class and the design of the class interfaces are determined and specified in a class interface specification (because of space limitation, the detailed TBS and the class interface specification have not been included). 540 Let us consider the internal design of the Navigation Controller, which is a composite task designed as a control clustering task, to show the nested information hiding objects (see Fig. 13). The information hiding objects are the Navigation Control state-dependent control object, the Sensor Interface and Wheel Actuator Interface objects, the Navigation Timer object and the user interface object, the Command Line Interface. In addition, the Navigation Controller contains one coordinator object called Navigation Coordinator, which receives incoming messages and coordinates the execution of the other objects. That is, the Navigation Coordinator extracts the event from the request and calls Navigation Control.processEvent (in event, out action) (see Fig. 13). The Navigation Control returns the action to be performed like store, check, start, etc according to the state transition table. Afterwards, the Navigation Coordinator initiates the action. &lt;&lt;control clustering&gt;&gt; :NavigationController : Navigation Control : Navigation Timer : CommandLineInterface : Current Position : Navigation Map : Navigation Path : WheelActuatorInterface : WheelActuator : SensorInterface : Sensor : Destination : Navigation Coordinator : CommandLine &lt;&lt; external user &gt;&gt; &lt;&lt;user interface&gt;&gt; &lt;&lt;timer&gt;&gt; &lt;&lt;coordinator&gt;&gt; &lt;&lt;data abstraction&gt;&gt; &lt;&lt;data abstraction&gt;&gt; &lt;&lt;data abstraction&gt;&gt; &lt;&lt;input device interface&gt;&gt; &lt;&lt;state dependent control&gt;&gt; &lt;&lt;output device interface&gt;&gt; &lt;&lt;data abstraction&gt;&gt; Start WheelActuator Output Move WheelActuator Output Stop WheelActuator Output read(out sensorData) store(in destination) check(in currentPosition,out yes/no) read(out sensorData) startTimer( ) stopTimer( ) activate( ) read(in sensorData, in map, out CurrentPosition) read(out map) update(sensorData, currentPosition, map) read(destination, currentPosition, map) start(in path,out started) move(in path) stop(out stopped) processEvent(in event,out action) startRobot(in destination) enter(in destination) Fig. 13. Detailed software design for Navigation Controller In our system, communication between tasks such as the Navigation Controller, Localizer, and Path Planner is through data abstraction classes like the Current Position and Navigation Path. As a result, connector objects [14] are not used for the message communication interface between tasks. : CommandLineInterface : Navigation Control : Navigation Coordinator : Destination : Navigation Map : Current Position : Navigation Path : WheelActuatorInterface : Navigation Timer : SensorInterface 1. startRobot(destination) 1.2. store(destination) 1.4. read(map) 1.6. read(sensorData, map, currentPosition) 1.1. processEvent(event, action) 1.3. processEvent(event, action) 1.5. processEvent(event, action) 1.7. processEvent(event, action) 1.8. update(sensorData, currentPosition, map) 1.9. processEvent(event, action) 1.10. read(destination, currentPosition, map, path) 1.11. processEvent(event, action) 1.12. start(path, started) 1.13. processEvent(event, action) 1.14. startTimer( ) 2. activate( ) 2.1. processEvent(event, action) 2.2. read(sensorData) 2.3. processEvent(event, action) 2.4. read(map) 2.5. processEvent(event, action) 2.6. read(sensorData, map, currentPosition) 2.7. processEvent(event, action) 2.8. check(currentPosition, yes/no) 2.10. update(sensorData, currentPosition, map) 2.9. processEvent(event, action) 2.12. read(destination, currentPosition, map, path) 2.11. processEvent(event, action) 2.13. processEvent(event, action) 2.14. move(path) 3. stop(stopped) 4. processEvent(event, action) 5. stopTimer( ) if not desitniation if destination Fig. 14. The task event diagram for Navigation Controller Lastly, the task's event sequencing logic is specified, which describes how the task responds to each of its message or event inputs. However, instead of using informally Pseudo code in COMET, in this project, task event diagrams are developed for tasks by using the UML sequence diagrams for understanding and readability, which turned out to be very useful when to implement the tasks. Fig. 14 illustrates the task event diagram for the Navigation Controller. LESSONS LEARNED This section summarizes the lessons learned from the project where we successfully applied the object-oriented method with UML to developing the service robot. 4.1 UML for Service Robot Domain Through the case study, we found that the UML standard was very useful as a notation for specifying the requirements, documenting the structure, decomposing into objects, and defining relationships between objects especially in a service robot system. Certain diagrams and notations were particularly importance for analyzing, designing, and modeling service robot systems as follows. Use case diagrams: With the use case model, services or functions (i.e., functional requirements), which a service robot performs or provides, can be defined in terms of actors who are users of the robot system and use cases. Each use case defines the behavior of some aspect of the robot system without revealing its internal structure. Class diagrams: The class diagram notation is used to depict the static model, which focuses on the static structure of the robot system. The class diagram shows the classes of objects in the system, their internal structure including attributes, their operations, and their relationships to other classes (such as associations and generalization/inheritance). Collaboration diagrams: This diagram shows how objects that participate in each use case interact with each other by sending and receiving messages in the dynamic model. It defines a specific way to use objects in the robot system by showing the possible interactions between them, especially to satisfy the needs described in the use case, namely provide the services. Compared to a sequence diagram, the diagram in particular is useful for synthesizing the collaboration diagrams to create the software architecture of the system as discussed in section 3.3. Sequence diagrams: This diagram show objects interaction arranged in time sequence and in particular could be used to describe the task event sequencing logic, which describes how the task responds to each of its message or event inputs. In COMET, the event sequencing logic is usually described informally in Pseudo code. We found that the sequence diagram can help the engineers describe the task event sequencing logic and implement the tasks by showing the order in which messages are passed between tasks and objects. State chart diagrams: The service robot system is highly state-dependent like real-time embedded systems. This diagram describes how state-dependent aspects of the system are defined by a finite state machine and can help design and developing highly state-dependent systems. It is 541 also possible for this diagram to model object behavior over several use cases with the collaboration diagrams. In addition, by using the UML notation as a defined standard, different research groups and development teams can communicate among themselves and with others to develop and integrate specific components for providing various services. 4.2 Importance of Systematic Process/Method for Service Robot Domain In order to effectively apply the UML notation and the robot control architecture like the Tripodal schematic control architecture to developing service robots, it is essential to use them with a systematic software engineering process or method, like an object-oriented analysis and design method, especially for real-time and embedded systems. It is not possible to resolve the issues in integrating and developing the service robots discussed before without systematic and comprehensive software development methods, particularly for service robots. In our case study, we applied COMET/UML method to developing the service robot. The COMET object-oriented software life cycle model is a highly iterative software development process based around the use case concept. In the requirements model, the service or functions (i.e., the function requirements) of the robot system are defined in terms of actors and use cases. In the analysis model, the use case is refined to describe the objects that participate in the use case, and their interactions. In the design model, the robot software architecture is developed, emphasizing issues of distribution, concurrency, and information hiding. This project showed that this was a viable approach because applying the COMET method with UML led to developing an effective service robot architecture by carefully taking into account user needs and requirements, implementing technical components based on the architecture, and integrating these components in a systematic and comprehensive fashion. 4.3 Customizing the COMET Method for Service Robot Domain Service robots like PSR-1, PSR-2, and Jinny have been built at KIST based on the Tripodal schematic control architecture. The Tripodal schematic design addressed on developing efficient and well-defined control architecture and a system integration strategy for constructing service robots. T-Rot is the next model of the PSR system under development for assisting aged persons. One of our aims is to develop the intelligent service robot for the elderly by cooperating and integrating the results of different research groups in accordance with the Tripodal schematic control architecture that has already been implemented on the PSR and successfully tested. Thus, the layered strategy of the Tripodal schematic design has been applied for design and modeling of the T-Rot. In the collaboration diagrams of the analysis modeling, and the consolidated collaboration diagram and the task architecture of the design modeling, the Command Line Interface is located in the deliberate layer for interfacing with a user, while the Sensor Interface, Wheel Actuator Interface, and Obstacle Avoidance Timer are in the reactive layer for controlling and managing the components in the reactive layer. The Navigation Control, Navigation Timer, Destination, Current Position, Navigation Path, Navigation Map, Localizer, and Path Planer are positioned in the sequencing layer for controlling the robot motion by executing relatively simple computations in real-time. As a result, the Tripodal schematic control architecture was helpful in arranging various hardware and software modules and functions. Additionally, as stated in section 4.1, in COMET, the event sequencing logic is usually described informally in Pseudo code. We found that the sequence diagram can help the engineers describe the task event sequencing logic and implement the tasks by showing the order in which messages are passed between tasks and objects. Hence, instead of using informal Pseudo code, task event diagrams were developed for tasks by using the UML sequence diagrams to improve understanding and readability. It turned out that these task event diagrams are very useful when implementing these tasks. 4.4 Human Communication Human communication to understand and develop what is desired of the service robot is likely to be more difficult than expected. In our case study, most engineers who are involved in the project come from the mechanical or robotics engineering field. The different research groups and teams tend to focus on their own technology and components and thus it is not easy to realize how much knowledge they have and how much information will need to be made explicit and communicated to integrate those components for the service robot. Several things can be done to improve the situation. One is for engineers from different teams, especially software engineers and mechanical engineers to work together for analyzing, designing, and developing the robot system during the project. It is very important that all engineers and developers from different groups and teams interact directly. Also, in order to develop a common ground for understanding the domain, technology, process and method, a common medium or language such as UML is critical. In addition to the standard notation like UML, guidelines about what notation to use, when to use it, and how to use the notation comprehensively and systematically are required. This is why the method like COMET is needed. Domain knowledge and experiences in each area will make it much easier to communicate what is desired, e.g. service robot domain, the autonomous robot navigation, vision processing, and so on for software engineers, and object-oriented concepts, software development process, and UML, etc for mechanical engineers. If there is relatively little domain knowledge and experience, to have one day or half-day technical workshop is needed. This has proved useful in a variety of settings in the development of the robot system, such as developing and increasing background knowledge of the domain and technology. 4.5 Necessity of Multi-Aspect Integration Method for Service Robot Domain A service robot should be able to perform several tasks autonomously to provide various services for human beings in a dynamic and partially unknown environment by applying both technology and knowledge. In order to be able to achieve complex tasks, perform new tasks, and integrate data learned from experience for the robot services, it is required to consider not only the robot's behavior, but also other robot's characteristics such as learning, planning, decision-making, and knowledge representation. It is necessary to allow existing robot behaviors to be used in new ways, to plan for accomplishing more complex tasks, to reuse the knowledge of one task in other tasks, and to 542 complete tasks more efficiently by learning various action sequences. In the case study, we focused on designing and modeling the robot's behavioral aspect, which is related to the sequencing and reactive layers in the Tripodal layered design, by applying the COMET/UML method. However, it is clear that planning and learning abilities have to also be considered when designing and developing a service robot, which correspond to the deliberate layer that is responsible for interfacing with a user and executing the planning process. As a consequence, a task manager, which is located in the deliberate layer, has been in charge of these robotic abilities in the project. Because the planning process is knowledge based and not reactive, a different analysis and design approach is needed for the task manager. Hence, we are convinced that methods to model the robot's learning, planning and decision making aspects as well as to incorporate, use and maintain task knowledge are necessary. Furthermore, it is essential to integrate these methods with the COMET method into a multi-aspect integration method for developing service robot software. CONCLUSIONS AND FUTHER WORK Service robots have been suggested for a growing number of applications. A service robot is a complex system as it includes various technical components (i.e., hardware and software) to be integrated correctly and many different research groups to develop the components. As a result, it is not only essential to develop complex algorithms or technical components, but also to integrate them adequately and correctly to provide the various robot services. In the paper, we have presented our case study where we developed the autonomous navigation system for the intelligent service robot for the elderly, T-Rot. The object-oriented method for real-time embedded systems, COMET has been applied to the service robot T-Rot with the industry standard UML. It makes it possible to reconcile specific engineering techniques like robot technologies with the UML notation and furthermore to fit such techniques into a fully defined development process towards developing the service robot system. In this way, we contribute to developing service robot software with UML in a systematic manner. The service robot T-Rot is still under development (at this point, we are at the first stage of total three stages). Thus, the current status of our work is to extend applications that include vision processing, speech processing and manipulation for providing various robot services. Also, we work on designing the knowledge-based task manager for improving the robot's ability. ACKNOWLEDGMENTS This research (paper) was performed for the Intelligent Robotics Development Program, one of the 21st Century Frontier R&D Programs funded by the Ministry of Commerce, Industry and Energy of Korea. REFERENCES [1] K. Kawamura and M. Iskarous, Trends in service robots for the disabled and the elderly, Proc. of the 1994 IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, Vol. 3 (1994) 1674. [2] R. D. Schraft, "Mechatronics and robotics for service applications," in IEEE Robotics and Automation Magazine, no. 4, pp. 31 - 35, Dec. 1994. [3] Rofer T., Lankenau A. and Moratz R., Service Robotics-Applications and Safety Issues in an Emerging Market, Workshop W20 proc. ECAI2000, Berlin, 2000. [4] B. You et al., "Development of a Home Service Robot `ISSAC'", Proc. of the 1994 IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, Las Vegas, Nevada, 2003, pp. 2630-2635. [5] G. Kim, W. Chung, M. Kim, and C. Lee, "Tripodal Schematic Design of the Control Architecture for the Service Robot PSR," in Proc. of the IEEE Conf. on Robotics and Automation, Taipei, Taiwan, pp.2792-2797, 2003. [6] G. Kim, W. Chung, M. Kim, and C. Lee, &quot;Implementation of Multi-Functional Service Robots Using Tripodal Schematic Control Architecture&quot;, in Proc. of IEEE Conf. on Robotics and Automation, New Orleans, LA, USA, 2004 [7] A. C. Dominguez-Brito, D.Hernandez-Sosa, J. Isern-Gonzalez , and J. Cabrera-Games. Integrating robotics software. IEEE International Conference on Robotics and Automation, 2004. [8] Q. Meng and M.H. Lee, "Learning and Control in Assistive Robotics for the Elderly", Proc. Of the 2004 IEEE Conf. on Robotics, Automation and Mechartonics, Singapore, Dec., 2004, pp. 71-76. [9] M. Kim, J. Lee, K. Kang, Y. Hong, and S. Bang, "Re-engineering Software Architecture of Home Service Robots: A Case Study", Proc. Of 27th Int. Conf. on Software Engineering (ICSE2005), St. Louis, USA, May, 2005, pp.505-513. [10] G. Booch, Object-Oriented Analysis and Design with Applications, 2nd ed. Redwood City, CA: Benjamin Cummings, 1994. [11] I. Jacobson, Object-Oriented Software Engineering, Addison Wesley, 1992. [12] Real-Time Application Interface, 2004. Available at: http:// www.rtai.org [13] Gjalt de Jong, "A UML-Based Design Methodology for Real-Time and Embedded Systems", DATE 2002, March, 2002. [14] H. Gomaa, Designing Concurrent, Distributed, and Real-Time Application with UML, Addison-Wesley, 2000. [15] OMG Unified Modeling Language, Version 1.5, March 2003. Available at:http://www.uml.org [16] M. Fowler and K. Scott, UML Distilled 2nd Edition, Addison Wesley, 2000. [17] G. Martin, L. Lavagno, and J. Louis-Guerin, "Embedded UML: a merger of real-time UML and codesign", CODES 2001, Copenhagen, April 2001, pp.23-28. [18] G. Martin, &quot;UML for Embedded Systems Specification and Design: Motivation and Overview&quot;, DATE 2002, March, 2002. 543

Software engineering;object-oriented analysis and design methods;service robot development;UML

204

Unified Utility Maximization Framework for Resource Selection

This paper presents a unified utility framework for resource selection of distributed text information retrieval. This new framework shows an efficient and effective way to infer the probabilities of relevance of all the documents across the text databases. With the estimated relevance information, resource selection can be made by explicitly optimizing the goals of different applications. Specifically, when used for database recommendation, the selection is optimized for the goal of high-recall (include as many relevant documents as possible in the selected databases); when used for distributed document retrieval, the selection targets the high-precision goal (high precision in the final merged list of documents). This new model provides a more solid framework for distributed information retrieval. Empirical studies show that it is at least as effective as other state-of-the-art algorithms.

INTRODUCTION Conventional search engines such as Google or AltaVista use ad-hoc information retrieval solution by assuming all the searchable documents can be copied into a single centralized database for the purpose of indexing. Distributed information retrieval, also known as federated search [1,4,7,11,14,22] is different from ad-hoc information retrieval as it addresses the cases when documents cannot be acquired and stored in a single database. For example, "Hidden Web" contents (also called "invisible" or "deep" Web contents) are information on the Web that cannot be accessed by the conventional search engines. Hidden web contents have been estimated to be 2-50 [19] times larger than the contents that can be searched by conventional search engines. Therefore, it is very important to search this type of valuable information. The architecture of distributed search solution is highly influenced by different environmental characteristics. In a small local area network such as small company environments, the information providers may cooperate to provide corpus statistics or use the same type of search engines. Early distributed information retrieval research focused on this type of cooperative environments [1,8]. On the other side, in a wide area network such as very large corporate environments or on the Web there are many types of search engines and it is difficult to assume that all the information providers can cooperate as they are required. Even if they are willing to cooperate in these environments, it may be hard to enforce a single solution for all the information providers or to detect whether information sources provide the correct information as they are required. Many applications fall into the latter type of uncooperative environments such as the Mind project [16] which integrates non-cooperating digital libraries or the QProber system [9] which supports browsing and searching of uncooperative hidden Web databases. In this paper, we focus mainly on uncooperative environments that contain multiple types of independent search engines. There are three important sub-problems in distributed information retrieval. First, information about the contents of each individual database must be acquired (resource representation) [1,8,21]. Second, given a query, a set of resources must be selected to do the search (resource selection) [5,7,21]. Third, the results retrieved from all the selected resources have to be merged into a single final list before it can be presented to the end user (retrieval and results merging) [1,5,20,22]. Many types of solutions exist for distributed information retrieval. Invisible-web.net resource selection components. This solution is useful when the users want to browse the selected databases by themselves instead of asking the system to retrieve relevant documents automatically. Distributed document retrieval is a more sophisticated task. It selects relevant information sources for users' queries as the database recommendation system does. Furthermore, users' queries are forwarded to the corresponding selected databases and the returned individual ranked lists are merged into a single list to present to the users. The goal of a database recommendation system is to select a small set of resources that contain as many relevant documents as possible, which we call a high-recall goal. On the other side, the effectiveness of distributed document retrieval is often measured by the Precision of the final merged document result list, which we call a high-precision goal. Prior research indicated that these two goals are related but not identical [4,21]. However, most previous solutions simply use effective resource selection algorithm of database recommendation system for distributed document retrieval system or solve the inconsistency with heuristic methods [1,4,21]. This paper presents a unified utility maximization framework to integrate the resource selection problem of both database recommendation and distributed document retrieval together by treating them as different optimization goals. First, a centralized sample database is built by randomly sampling a small amount of documents from each database with query-based sampling [1]; database size statistics are also estimated [21]. A logistic transformation model is learned off line with a small amount of training queries to map the centralized document scores in the centralized sample database to the corresponding probabilities of relevance. Second, after a new query is submitted, the query can be used to search the centralized sample database which produces a score for each sampled document. The probability of relevance for each document in the centralized sample database can be estimated by applying the logistic model to each document's score. Then, the probabilities of relevance of all the (mostly unseen) documents among the available databases can be estimated using the probabilities of relevance of the documents in the centralized sample database and the database size estimates. For the task of resource selection for a database recommendation system, the databases can be ranked by the expected number of relevant documents to meet the high-recall goal. For resource selection for a distributed document retrieval system, databases containing a small number of documents with large probabilities of relevance are favored over databases containing many documents with small probabilities of relevance. This selection criterion meets the high-precision goal of distributed document retrieval application. Furthermore, the Semi-supervised learning (SSL) [20,22] algorithm is applied to merge the returned documents into a final ranked list. The unified utility framework makes very few assumptions and works in uncooperative environments. Two key features make it a more solid model for distributed information retrieval: i) It formalizes the resource selection problems of different applications as various utility functions, and optimizes the utility functions to achieve the optimal results accordingly; and ii) It shows an effective and efficient way to estimate the probabilities of relevance of all documents across databases. Specifically, the framework builds logistic models on the centralized sample database to transform centralized retrieval scores to the corresponding probabilities of relevance and uses the centralized sample database as the bridge between individual databases and the logistic model. The human effort (relevance judgment) required to train the single centralized logistic model does not scale with the number of databases. This is a large advantage over previous research, which required the amount of human effort to be linear with the number of databases [7,15]. The unified utility framework is not only more theoretically solid but also very effective. Empirical studies show the new model to be at least as accurate as the state-of-the-art algorithms in a variety of configurations. The next section discusses related work. Section 3 describes the new unified utility maximization model. Section 4 explains our experimental methodology. Sections 5 and 6 present our experimental results for resource selection and document retrieval. Section 7 concludes. PRIOR RESEARCH There has been considerable research on all the sub-problems of distributed information retrieval. We survey the most related work in this section. The first problem of distributed information retrieval is resource representation. The STARTS protocol is one solution for acquiring resource descriptions in cooperative environments [8]. However, in uncooperative environments, even the databases are willing to share their information, it is not easy to judge whether the information they provide is accurate or not. Furthermore, it is not easy to coordinate the databases to provide resource representations that are compatible with each other. Thus, in uncooperative environments, one common choice is query-based sampling, which randomly generates and sends queries to individual search engines and retrieves some documents to build the descriptions. As the sampled documents are selected by random queries, query-based sampling is not easily fooled by any adversarial spammer that is interested to attract more traffic. Experiments have shown that rather accurate resource descriptions can be built by sending about 80 queries and downloading about 300 documents [1]. Many resource selection algorithms such as gGlOSS/vGlOSS [8] and CORI [1] have been proposed in the last decade. The CORI algorithm represents each database by its terms, the document frequencies and a small number of corpus statistics (details in [1]). As prior research on different datasets has shown the CORI algorithm to be the most stable and effective of the three algorithms [1,17,18], we use it as a baseline algorithm in this work. The relevant document distribution estimation (ReDDE [21]) resource selection algorithm is a recent algorithm that tries to estimate the distribution of relevant documents across the available databases and ranks the databases accordingly. Although the ReDDE algorithm has been shown to be effective, it relies on heuristic constants that are set empirically [21]. The last step of the document retrieval sub-problem is results merging, which is the process of transforming database-specific 33 document scores into comparable database-independent document scores. The semi supervised learning (SSL) [20,22] result merging algorithm uses the documents acquired by query-based sampling as training data and linear regression to learn the database-specific, query-specific merging models. These linear models are used to convert the database-specific document scores into the approximated centralized document scores. The SSL algorithm has been shown to be effective [22]. It serves as an important component of our unified utility maximization framework (Section 3). In order to achieve accurate document retrieval results, many previous methods simply use resource selection algorithms that are effective of database recommendation system. But as pointed out above, a good resource selection algorithm optimized for high-recall may not work well for document retrieval, which targets the high-precision goal. This type of inconsistency has been observed in previous research [4,21]. The research in [21] tried to solve the problem with a heuristic method. The research most similar to what we propose here is the decision-theoretic framework (DTF) [7,15]. This framework computes a selection that minimizes the overall costs (e.g., retrieval quality, time) of document retrieval system and several methods [15] have been proposed to estimate the retrieval quality. However, two points distinguish our research from the DTF model. First, the DTF is a framework designed specifically for document retrieval, but our new model integrates two distinct applications with different requirements (database recommendation and distributed document retrieval) into the same unified framework. Second, the DTF builds a model for each database to calculate the probabilities of relevance. This requires human relevance judgments for the results retrieved from each database. In contrast, our approach only builds one logistic model for the centralized sample database. The centralized sample database can serve as a bridge to connect the individual databases with the centralized logistic model, thus the probabilities of relevance of documents in different databases can be estimated. This strategy can save large amount of human judgment effort and is a big advantage of the unified utility maximization framework over the DTF especially when there are a large number of databases. UNIFIED UTILITY MAXIMIZATION FRAMEWORK The Unified Utility Maximization (UUM) framework is based on estimating the probabilities of relevance of the (mostly unseen) documents available in the distributed search environment. In this section we describe how the probabilities of relevance are estimated and how they are used by the Unified Utility Maximization model. We also describe how the model can be optimized for the high-recall goal of a database recommendation system and the high-precision goal of a distributed document retrieval system. 3.1 Estimating Probabilities of Relevance As pointed out above, the purpose of resource selection is high-recall and the purpose of document retrieval is high-precision. In order to meet these diverse goals, the key issue is to estimate the probabilities of relevance of the documents in various databases. This is a difficult problem because we can only observe a sample of the contents of each database using query-based sampling. Our strategy is to make full use of all the available information to calculate the probability estimates. 3.1.1 Learning Probabilities of Relevance In the resource description step, the centralized sample database is built by query-based sampling and the database sizes are estimated using the sample-resample method [21]. At the same time, an effective retrieval algorithm (Inquery [2]) is applied on the centralized sample database with a small number (e.g., 50) of training queries. For each training query, the CORI resource selection algorithm [1] is applied to select some number (e.g., 10) of databases and retrieve 50 document ids from each database. The SSL results merging algorithm [20,22] is used to merge the results. Then, we can download the top 50 documents in the final merged list and calculate their corresponding centralized scores using Inquery and the corpus statistics of the centralized sample database. The centralized scores are further normalized (divided by the maximum centralized score for each query), as this method has been suggested to improve estimation accuracy in previous research [15]. Human judgment is acquired for those documents and a logistic model is built to transform the normalized centralized document scores to probabilities of relevance as follows: ( ) )) ( exp( 1 )) ( exp( | ) ( _ _ d S b a d S b a d rel P d R c c c c c c + + + = = (1) where ) ( _ d S c is the normalized centralized document score and a c and b c are the two parameters of the logistic model. These two parameters are estimated by maximizing the probabilities of relevance of the training queries. The logistic model provides us the tool to calculate the probabilities of relevance from centralized document scores. 3.1.2 Estimating Centralized Document Scores When the user submits a new query, the centralized document scores of the documents in the centralized sample database are calculated. However, in order to calculate the probabilities of relevance, we need to estimate centralized document scores for all documents across the databases instead of only the sampled documents. This goal is accomplished using: the centralized scores of the documents in the centralized sample database, and the database size statistics. We define the database scale factor for the i th database as the ratio of the estimated database size and the number of documents sampled from this database as follows: ^ _ i i i db db db samp N SF N = (2) where ^ i db N is the estimated database size and _ i db samp N is the number of documents from the i th database in the centralized sample database. The intuition behind the database scale factor is that, for a database whose scale factor is 50, if one document from this database in the centralized sample database has a centralized document score of 0.5, we may guess that there are about 50 documents in that database which have scores of about 0.5. Actually, we can apply a finer non-parametric linear interpolation method to estimate the centralized document score curve for each database. Formally, we rank all the sampled documents from the i th database by their centralized document 34 scores to get the sampled centralized document score list {S c (ds i1 ), S c (ds i2 ), S c (ds i3 ),.....} for the i th database; we assume that if we could calculate the centralized document scores for all the documents in this database and get the complete centralized document score list, the top document in the sampled list would have rank SF dbi /2, the second document in the sampled list would rank SF dbi 3/2, and so on. Therefore, the data points of sampled documents in the complete list are: {(SF dbi /2, S c (ds i1 )), (SF dbi 3/2, S c (ds i2 )), (SF dbi 5/2, S c (ds i3 )),.....}. Piecewise linear interpolation is applied to estimate the centralized document score curve, as illustrated in Figure 1. The complete centralized document score list can be estimated by calculating the values of different ranks on the centralized document curve as: ] , 1 [ , ) ( S ^ ^ c i db ij N j d . It can be seen from Figure 1 that more sample data points produce more accurate estimates of the centralized document score curves. However, for databases with large database scale ratios, this kind of linear interpolation may be rather inaccurate, especially for the top ranked (e.g., [1, SF dbi /2]) documents. Therefore, an alternative solution is proposed to estimate the centralized document scores of the top ranked documents for databases with large scale ratios (e.g., larger than 100). Specifically, a logistic model is built for each of these databases. The logistic model is used to estimate the centralized document score of the top 1 document in the corresponding database by using the two sampled documents from that database with highest centralized scores. )) ( ) ( exp( 1 )) ( ) ( exp( ) ( 2 2 1 1 0 2 2 1 1 0 ^ 1 i c i i c i i i c i i c i i i c ds S ds S ds S ds S d S + + + + + = (3) 0 i , 1 i and 2 i are the parameters of the logistic model. For each training query, the top retrieved document of each database is downloaded and the corresponding centralized document score is calculated. Together with the scores of the top two sampled documents, these parameters can be estimated. After the centralized score of the top document is estimated, an exponential function is fitted for the top part ([1, SF dbi /2]) of the centralized document score curve as: ] 2 / , 1 [ ) * exp( ) ( 1 0 ^ i db i i ij c SF j j d S + = (4) ^ 0 1 1 log( ( )) i c i i S d = (5) ) 1 2 / ( )) ( log( ) ( (log( ^ 1 1 1 = i db i c i c i SF d S ds S (6) The two parameters 0 i and 1 i are fitted to make sure the exponential function passes through the two points (1, ^ 1 ) ( i c d S ) and (SF dbi /2, S c (ds i1 )). The exponential function is only used to adjust the top part of the centralized document score curve and the lower part of the curve is still fitted with the linear interpolation method described above. The adjustment by fitting exponential function of the top ranked documents has been shown empirically to produce more accurate results. From the centralized document score curves, we can estimate the complete centralized document score lists accordingly for all the available databases. After the estimated centralized document scores are normalized, the complete lists of probabilities of relevance can be constructed out of the complete centralized document score lists by Equation 1. Formally for the i th database, the complete list of probabilities of relevance is: ] , 1 [ , ) ( R ^ ^ i db ij N j d . 3.2 The Unified Utility Maximization Model In this section, we formally define the new unified utility maximization model, which optimizes the resource selection problems for two goals of high-recall (database recommendation) and high-precision (distributed document retrieval) in the same framework. In the task of database recommendation, the system needs to decide how to rank databases. In the task of document retrieval, the system not only needs to select the databases but also needs to decide how many documents to retrieve from each selected database. We generalize the database recommendation selection process, which implicitly recommends all documents in every selected database, as a special case of the selection decision for the document retrieval task. Formally, we denote d i as the number of documents we would like to retrieve from the i th database and ,.....} , { 2 1 d d d = as a selection action for all the databases. The database selection decision is made based on the complete lists of probabilities of relevance for all the databases. The complete lists of probabilities of relevance are inferred from all the available information specifically s R , which stands for the resource descriptions acquired by query-based sampling and the database size estimates acquired by sample-resample; c S stands for the centralized document scores of the documents in the centralized sample database. If the method of estimating centralized document scores and probabilities of relevance in Section 3.1 is acceptable, then the most probable complete lists of probabilities of relevance can be derived and we denote them as 1 ^ ^ * 1 {(R( ), [1, ]), db j d j N = 2 ^ ^ 2 (R( ), [1, ]),.......} db j d j N . Random vector denotes an arbitrary set of complete lists of probabilities of relevance and ) , | ( c s S R P as the probability of generating this set of lists. Finally, to each selection action d and a set of complete lists of Figure 1. Linear interpolation construction of the complete centralized document score list (database scale factor is 50). 35 probabilities of relevance , we associate a utility function ) , ( d U which indicates the benefit from making the d selection when the true complete lists of probabilities of relevance are . Therefore, the selection decision defined by the Bayesian framework is: d S R P d U d c s d ) . | ( ) , ( max arg * = (7) One common approach to simplify the computation in the Bayesian framework is to only calculate the utility function at the most probable parameter values instead of calculating the whole expectation. In other words, we only need to calculate ) , ( * d U and Equation 7 is simplified as follows: ) , ( max arg * * d U d d = (8) This equation serves as the basic model for both the database recommendation system and the document retrieval system. 3.3 Resource Selection for High-Recall High-recall is the goal of the resource selection algorithm in federated search tasks such as database recommendation. The goal is to select a small set of resources (e.g., less than N sdb databases) that contain as many relevant documents as possible, which can be formally defined as: = = i N j ij i i db d d I d U ^ 1 ^ * ) ( R ) ( ) , ( (9) I(d i ) is the indicator function, which is 1 when the i th database is selected and 0 otherwise. Plug this equation into the basic model in Equation 8 and associate the selected database number constraint to obtain the following: sdb i i i N j ij i d N d I to Subject d d I d i db = = = ) ( : ) ( R ) ( max arg ^ 1 ^ * (10) The solution of this optimization problem is very simple. We can calculate the expected number of relevant documents for each database as follows: = = i db i N j ij Rd d N ^ 1 ^ ^ ) ( R (11) The N sdb databases with the largest expected number of relevant documents can be selected to meet the high-recall goal. We call this the UUM/HR algorithm (Unified Utility Maximization for High-Recall). 3.4 Resource Selection for High-Precision High-Precision is the goal of resource selection algorithm in federated search tasks such as distributed document retrieval. It is measured by the Precision at the top part of the final merged document list. This high-precision criterion is realized by the following utility function, which measures the Precision of retrieved documents from the selected databases. = = i d j ij i i d d I d U 1 ^ * ) ( R ) ( ) , ( (12) Note that the key difference between Equation 12 and Equation 9 is that Equation 9 sums up the probabilities of relevance of all the documents in a database, while Equation 12 only considers a much smaller part of the ranking. Specifically, we can calculate the optimal selection decision by: = = i d j ij i d i d d I d 1 ^ * ) ( R ) ( max arg (13) Different kinds of constraints caused by different characteristics of the document retrieval tasks can be associated with the above optimization problem. The most common one is to select a fixed number (N sdb ) of databases and retrieve a fixed number (N rdoc ) of documents from each selected database, formally defined as: 0 , ) ( : ) ( R ) ( max arg 1 ^ * = = = = i rdoc i sdb i i i d j ij i d d if N d N d I to Subject d d I d i (14) This optimization problem can be solved easily by calculating the number of expected relevant documents in the top part of the each database's complete list of probabilities of relevance: = = rdoc i N j ij Rd Top d N 1 ^ ^ _ ) ( R (15) Then the databases can be ranked by these values and selected. We call this the UUM/HP-FL algorithm (Unified Utility Maximization for High-Precision with Fixed Length document rankings from each selected database). A more complex situation is to vary the number of retrieved documents from each selected database. More specifically, we allow different selected databases to return different numbers of documents. For simplification, the result list lengths are required to be multiples of a baseline number 10. (This value can also be varied, but for simplification it is set to 10 in this paper.) This restriction is set to simulate the behavior of commercial search engines on the Web. (Search engines such as Google and AltaVista return only 10 or 20 document ids for every result page.) This procedure saves the computation time of calculating optimal database selection by allowing the step of dynamic programming to be 10 instead of 1 (more detail is discussed latterly). For further simplification, we restrict to select at most 100 documents from each database (d i &lt;=100) Then, the selection optimization problem is formalized as follows: ] 10 .., , 2 , 1 , 0 [ , * 10 ) ( : ) ( R ) ( max arg _ 1 ^ * = = = = = k k d N d N d I to Subject d d I d i rdoc Total i i sdb i i i d j ij i d i (16) N Total_rdoc is the total number of documents to be retrieved. Unfortunately, there is no simple solution for this optimization problem as there are for Equations 10 and 14. However, a 36 dynamic programming algorithm can be applied to calculate the optimal solution. The basic steps of this dynamic programming method are described in Figure 2. As this algorithm allows retrieving result lists of varying lengths from each selected database, it is called UUM/HP-VL algorithm. After the selection decisions are made, the selected databases are searched and the corresponding document ids are retrieved from each database. The final step of document retrieval is to merge the returned results into a single ranked list with the semi-supervised learning algorithm. It was pointed out before that the SSL algorithm maps the database-specific scores into the centralized document scores and builds the final ranked list accordingly, which is consistent with all our selection procedures where documents with higher probabilities of relevance (thus higher centralized document scores) are selected. EXPERIMENTAL METHODOLOGY It is desirable to evaluate distributed information retrieval algorithms with testbeds that closely simulate the real world applications. The TREC Web collections WT2g or WT10g [4,13] provide a way to partition documents by different Web servers. In this way, a large number (O(1000)) of databases with rather diverse contents could be created, which may make this testbed a good candidate to simulate the operational environments such as open domain hidden Web. However, two weakness of this testbed are: i) Each database contains only a small amount of document (259 documents by average for WT2g) [4]; and ii) The contents of WT2g or WT10g are arbitrarily crawled from the Web. It is not likely for a hidden Web database to provide personal homepages or web pages indicating that the pages are under construction and there is no useful information at all. These types of web pages are contained in the WT2g/WT10g datasets. Therefore, the noisy Web data is not similar with that of high-quality hidden Web database contents, which are usually organized by domain experts. Another choice is the TREC news/government data [1,15,17, 18,21]. TREC news/government data is concentrated on relatively narrow topics. Compared with TREC Web data: i) The news/government documents are much more similar to the contents provided by a topic-oriented database than an arbitrary web page, ii) A database in this testbed is larger than that of TREC Web data. By average a database contains thousands of documents, which is more realistic than a database of TREC Web data with about 250 documents. As the contents and sizes of the databases in the TREC news/government testbed are more similar with that of a topic-oriented database, it is a good candidate to simulate the distributed information retrieval environments of large organizations (companies) or domain-specific hidden Web sites, such as West that provides access to legal, financial and news text databases [3]. As most current distributed information retrieval systems are developed for the environments of large organizations (companies) or domain-specific hidden Web other than open domain hidden Web, TREC news/government testbed was chosen in this work. Trec123-100col-bysource testbed is one of the most used TREC news/government testbed [1,15,17,21]. It was chosen in this work. Three testbeds in [21] with skewed database size distributions and different types of relevant document distributions were also used to give more thorough simulation for real environments. Trec123-100col-bysource: 100 databases were created from TREC CDs 1, 2 and 3. They were organized by source and publication date [1]. The sizes of the databases are not skewed. Details are in Table 1. Three testbeds built in [21] were based on the trec123-100col-bysource testbed. Each testbed contains many "small" databases and two large databases created by merging about 10-20 small databases together. Input: Complete lists of probabilities of relevance for all the |DB| databases. Output: Optimal selection solution for Equation 16. i) Create the three-dimensional array: Sel (1..|DB|, 1..N Total_rdoc/10 , 1..N sdb ) Each Sel (x, y, z) is associated with a selection decision xyz d , which represents the best selection decision in the condition: only databases from number 1 to number x are considered for selection; totally y*10 documents will be retrieved; only z databases are selected out of the x database candidates. And Sel (x, y, z) is the corresponding utility value by choosing the best selection. ii) Initialize Sel (1, 1..N Total_rdoc /10, 1..N sdb ) with only the estimated relevance information of the 1 st database. iii) Iterate the current database candidate i from 2 to |DB| For each entry Sel (i, y, z): Find k such that: ) 10 , min( 1 : ) ) ( ) 1 , , 1 ( ( max arg * 10 ^ * y k to subject d R z k y i Sel k k j ij k + = ) , , 1 ( ) ) ( ) 1 , , 1 ( ( * * 10 ^ * z y i Sel d R z k y i Sel If k j ij &gt ; + This means that we should retrieve * 10 k documents from the i th database, otherwise we should not select this database and the previous best solution Sel (i-1, y, z) should be kept. Then set the value of iyz d and Sel (i, y, z) accordingly. iv) The best selection solution is given by _ /10 | | Toral rdoc sdb DB N N d and the corresponding utility value is Sel (|DB|, N Total_rdoc/10 , N sdb ). Figure 2. The dynamic programming optimization procedure for Equation 16. Table1: Testbed statistics. Number of documents Size (MB) Testbed Size (GB) Min Avg Max Min Avg Max Trec123 3.2 752 10782 39713 28 32 42 Table2: Query set statistics. Name TREC Topic Set TREC Topic Field Average Length (Words) Trec123 51-150 Title 3.1 37 Trec123-2ldb-60col ("representative"): The databases in the trec123-100col-bysource were sorted with alphabetical order. Two large databases were created by merging 20 small databases with the round-robin method. Thus, the two large databases have more relevant documents due to their large sizes, even though the densities of relevant documents are roughly the same as the small databases. Trec123-AP-WSJ-60col ("relevant"): The 24 Associated Press collections and the 16 Wall Street Journal collections in the trec123-100col-bysource testbed were collapsed into two large databases APall and WSJall. The other 60 collections were left unchanged. The APall and WSJall databases have higher densities of documents relevant to TREC queries than the small databases. Thus, the two large databases have many more relevant documents than the small databases. Trec123-FR-DOE-81col ("nonrelevant"): The 13 Federal Register collections and the 6 Department of Energy collections in the trec123-100col-bysource testbed were collapsed into two large databases FRall and DOEall. The other 80 collections were left unchanged. The FRall and DOEall databases have lower densities of documents relevant to TREC queries than the small databases, even though they are much larger. 100 queries were created from the title fields of TREC topics 51-150. The queries 101-150 were used as training queries and the queries 51-100 were used as test queries (details in Table 2). 4.2 Search Engines In the uncooperative distributed information retrieval environments of large organizations (companies) or domain-specific hidden Web, different databases may use different types of search engine. To simulate the multiple type-engine environment, three different types of search engines were used in the experiments: INQUERY [2], a unigram statistical language model with linear smoothing [12,20] and a TFIDF retrieval algorithm with "ltc" weight [12,20]. All these algorithms were implemented with the Lemur toolkit [12]. These three kinds of search engines were assigned to the databases among the four testbeds in a round-robin manner. RESULTS RESOURCE SELECTION OF DATABASE RECOMMENDATION All four testbeds described in Section 4 were used in the experiments to evaluate the resource selection effectiveness of the database recommendation system. The resource descriptions were created using query-based sampling. About 80 queries were sent to each database to download 300 unique documents. The database size statistics were estimated by the sample-resample method [21]. Fifty queries (101-150) were used as training queries to build the relevant logistic model and to fit the exponential functions of the centralized document score curves for large ratio databases (details in Section 3.1). Another 50 queries (51-100) were used as test data. Resource selection algorithms of database recommendation systems are typically compared using the recall metric n R [1,17,18,21]. Let B denote a baseline ranking, which is often the RBR (relevance based ranking), and E as a ranking provided by a resource selection algorithm. And let B i and E i denote the number of relevant documents in the i th ranked database of B or E. Then R n is defined as follows: = = = k i i k i i k B E R 1 1 (17) Usually the goal is to search only a few databases, so our figures only show results for selecting up to 20 databases. The experiments summarized in Figure 3 compared the effectiveness of the three resource selection algorithms, namely the CORI, ReDDE and UUM/HR. The UUM/HR algorithm is described in Section 3.3. It can be seen from Figure 3 that the ReDDE and UUM/HR algorithms are more effective (on the representative, relevant and nonrelevant testbeds) or as good as (on the Trec123-100Col testbed) the CORI resource selection algorithm. The UUM/HR algorithm is more effective than the ReDDE algorithm on the representative and relevant testbeds and is about the same as the ReDDE algorithm on the Trec123-100Col and the nonrelevant testbeds. This suggests that the UUM/HR algorithm is more robust than the ReDDE algorithm. It can be noted that when selecting only a few databases on the Trec123-100Col or the nonrelevant testbeds, the ReDEE algorithm has a small advantage over the UUM/HR algorithm. We attribute this to two causes: i) The ReDDE algorithm was tuned on the Trec123-100Col testbed; and ii) Although the difference is small, this may suggest that our logistic model of estimating probabilities of relevance is not accurate enough. More training data or a more sophisticated model may help to solve this minor puzzle. Collections Selected. Collections Selected. Trec123-100Col Testbed. Representative Testbed. Collection Selected. Collection Selected. Relevant Testbed. Nonrelevant Testbed. Figure 3. Resource selection experiments on the four testbeds. 38 RESULTS DOCUMENT RETRIEVAL EFFECTIVENESS For document retrieval, the selected databases are searched and the returned results are merged into a single final list. In all of the experiments discussed in this section the results retrieved from individual databases were combined by the semi-supervised learning results merging algorithm. This version of the SSL algorithm [22] is allowed to download a small number of returned document texts "on the fly" to create additional training data in the process of learning the linear models which map database-specific document scores into estimated centralized document scores. It has been shown to be very effective in environments where only short result-lists are retrieved from each selected database [22]. This is a common scenario in operational environments and was the case for our experiments. Document retrieval effectiveness was measured by Precision at the top part of the final document list. The experiments in this section were conducted to study the document retrieval effectiveness of five selection algorithms, namely the CORI, ReDDE, UUM/HR, UUM/HP-FL and UUM/HP-VL algorithms. The last three algorithms were proposed in Section 3. All the first four algorithms selected 3 or 5 databases, and 50 documents were retrieved from each selected database. The UUM/HP-FL algorithm also selected 3 or 5 databases, but it was allowed to adjust the number of documents to retrieve from each selected database; the number retrieved was constrained to be from 10 to 100, and a multiple of 10. The Trec123-100Col and representative testbeds were selected for document retrieval as they represent two extreme cases of resource selection effectiveness; in one case the CORI algorithm is as good as the other algorithms and in the other case it is quite Table 5. Precision on the representative testbed when 3 databases were selected. (The first baseline is CORI; the second baseline for UUM/HP methods is UUM/HR.) Precision at Doc Rank CORI ReDDE UUM/HR UUM/HP-FL UUM/HP-VL 5 docs 0.3720 0.4080 (+9.7%) 0.4640 (+24.7%) 0.4600 (+23.7%)(-0.9%) 0.5000 (+34.4%)(+7.8%) 10 docs 0.3400 0.4060 (+19.4%) 0.4600 (+35.3%) 0.4540 (+33.5%)(-1.3%) 0.4640 (+36.5%)(+0.9%) 15 docs 0.3120 0.3880 (+24.4%) 0.4320 (+38.5%) 0.4240 (+35.9%)(-1.9%) 0.4413 (+41.4%)(+2.2) 20 docs 0.3000 0.3750 (+25.0%) 0.4080 (+36.0%) 0.4040 (+34.7%)(-1.0%) 0.4240 (+41.3%)(+4.0%) 30 docs 0.2533 0.3440 (+35.8%) 0.3847 (+51.9%) 0.3747 (+47.9%)(-2.6%) 0.3887 (+53.5%)(+1.0%) Table 6. Precision on the representative testbed when 5 databases were selected. (The first baseline is CORI; the second baseline for UUM/HP methods is UUM/HR.) Precision at Doc Rank CORI ReDDE UUM/HR UUM/HP-FL UUM/HP-VL 5 docs 0.3960 0.4080 (+3.0%) 0.4560 (+15.2%) 0.4280 (+8.1%)(-6.1%) 0.4520 (+14.1%)(-0.9%) 10 docs 0.3880 0.4060 (+4.6%) 0.4280 (+10.3%) 0.4460 (+15.0%)(+4.2%) 0.4560 (+17.5%)(+6.5%) 15 docs 0.3533 0.3987 (+12.9%) 0.4227 (+19.6%) 0.4440 (+25.7%)(+5.0%) 0.4453 (+26.0%)(+5.4%) 20 docs 0.3330 0.3960 (+18.9%) 0.4140 (+24.3%) 0.4290 (+28.8%)(+3.6%) 0.4350 (+30.6%)(+5.1%) 30 docs 0.2967 0.3740 (+26.1%) 0.4013 (+35.3%) 0.3987 (+34.4%)(-0.7%) 0.4060 (+36.8%)(+1.2%) Table 3. Precision on the trec123-100col-bysource testbed when 3 databases were selected. (The first baseline is CORI; the second baseline for UUM/HP methods is UUM/HR.) Precision at Doc Rank CORI ReDDE UUM/HR UUM/HP-FL UUM/HP-VL 5 docs 0.3640 0.3480 (-4.4%) 0.3960 (+8.8%) 0.4680 (+28.6%)(+18.1%) 0.4640 (+27.5%)(+17.2%) 10 docs 0.3360 0.3200 (-4.8%) 0.3520 (+4.8%) 0.4240 (+26.2%)(+20.5%) 0.4220 (+25.6%)(+19.9%) 15 docs 0.3253 0.3187 (-2.0%) 0.3347 (+2.9%) 0.3973 (+22.2%)(+15.7%) 0.3920 (+20.5%)(+17.1%) 20 docs 0.3140 0.2980 (-5.1%) 0.3270 (+4.1%) 0.3720 (+18.5%)(+13.8%) 0.3700 (+17.8%)(+13.2%) 30 docs 0.2780 0.2660 (-4.3%) 0.2973 (+6.9%) 0.3413 (+22.8%)(+14.8%) 0.3400 (+22.3%)(+14.4%) Table 4. Precision on the trec123-100col-bysource testbed when 5 databases were selected. (The first baseline is CORI; the second baseline for UUM/HP methods is UUM/HR.) Precision at Doc Rank CORI ReDDE UUM/HR UUM/HP-FL UUM/HP-VL 5 docs 0.4000 0.3920 (-2.0%) 0.4280 (+7.0%) 0.4680 (+17.0%)(+9.4%) 0.4600 (+15.0%)(+7.5%) 10 docs 0.3800 0.3760 (-1.1%) 0.3800 (+0.0%) 0.4180 (+10.0%)(+10.0%) 0.4320 (+13.7%)(+13.7%) 15 docs 0.3560 0.3560 (+0.0%) 0.3720 (+4.5%) 0.3920 (+10.1%)(+5.4%) 0.4080 (+14.6%)(+9.7%) 20 docs 0.3430 0.3390 (-1.2%) 0.3550 (+3.5%) 0.3710 (+8.2%)(+4.5%) 0.3830 (+11.7%)(+7.9%) 30 docs 0.3240 0.3140 (-3.1%) 0.3313 (+2.3%) 0.3500 (+8.0%)(+5.6%) 0.3487 (+7.6%)(+5.3%) 39 a lot worse than the other algorithms. Tables 3 and 4 show the results on the Trec123-100Col testbed, and Tables 5 and 6 show the results on the representative testbed. On the Trec123-100Col testbed, the document retrieval effectiveness of the CORI selection algorithm is roughly the same or a little bit better than the ReDDE algorithm but both of them are worse than the other three algorithms (Tables 3 and 4). The UUM/HR algorithm has a small advantage over the CORI and ReDDE algorithms. One main difference between the UUM/HR algorithm and the ReDDE algorithm was pointed out before: The UUM/HR uses training data and linear interpolation to estimate the centralized document score curves, while the ReDDE algorithm [21] uses a heuristic method, assumes the centralized document score curves are step functions and makes no distinction among the top part of the curves. This difference makes UUM/HR better than the ReDDE algorithm at distinguishing documents with high probabilities of relevance from low probabilities of relevance. Therefore, the UUM/HR reflects the high-precision retrieval goal better than the ReDDE algorithm and thus is more effective for document retrieval. The UUM/HR algorithm does not explicitly optimize the selection decision with respect to the high-precision goal as the UUM/HP-FL and UUM/HP-VL algorithms are designed to do. It can be seen that on this testbed, the UUM/HP-FL and UUM/HP-VL algorithms are much more effective than all the other algorithms. This indicates that their power comes from explicitly optimizing the high-precision goal of document retrieval in Equations 14 and 16. On the representative testbed, CORI is much less effective than other algorithms for distributed document retrieval (Tables 5 and 6). The document retrieval results of the ReDDE algorithm are better than that of the CORI algorithm but still worse than the results of the UUM/HR algorithm. On this testbed the three UUM algorithms are about equally effective. Detailed analysis shows that the overlap of the selected databases between the UUM/HR, UUM/HP-FL and UUM/HP-VL algorithms is much larger than the experiments on the Trec123-100Col testbed, since all of them tend to select the two large databases. This explains why they are about equally effective for document retrieval. In real operational environments, databases may return no document scores and report only ranked lists of results. As the unified utility maximization model only utilizes retrieval scores of sampled documents with a centralized retrieval algorithm to calculate the probabilities of relevance, it makes database selection decisions without referring to the document scores from individual databases and can be easily generalized to this case of rank lists without document scores. The only adjustment is that the SSL algorithm merges ranked lists without document scores by assigning the documents with pseudo-document scores normalized for their ranks (In a ranked list of 50 documents, the first one has a score of 1, the second has a score of 0.98 etc) ,which has been studied in [22]. The experiment results on trec123-100Col-bysource testbed with 3 selected databases are shown in Table 7. The experiment setting was the same as before except that the document scores were eliminated intentionally and the selected databases only return ranked lists of document ids. It can be seen from the results that the UUM/HP-FL and UUM/HP-VL work well with databases returning no document scores and are still more effective than other alternatives. Other experiments with databases that return no document scores are not reported but they show similar results to prove the effectiveness of UUM/HP-FL and UUM/HP-VL algorithms. The above experiments suggest that it is very important to optimize the high-precision goal explicitly in document retrieval. The new algorithms based on this principle achieve better or at least as good results as the prior state-of-the-art algorithms in several environments. CONCLUSION Distributed information retrieval solves the problem of finding information that is scattered among many text databases on local area networks and Internets. Most previous research use effective resource selection algorithm of database recommendation system for distributed document retrieval application. We argue that the high-recall resource selection goal of database recommendation and high-precision goal of document retrieval are related but not identical. This kind of inconsistency has also been observed in previous work, but the prior solutions either used heuristic methods or assumed cooperation by individual databases (e.g., all the databases used the same kind of search engines), which is frequently not true in the uncooperative environment. In this work we propose a unified utility maximization model to integrate the resource selection of database recommendation and document retrieval tasks into a single unified framework. In this framework, the selection decisions are obtained by optimizing different objective functions. As far as we know, this is the first work that tries to view and theoretically model the distributed information retrieval task in an integrated manner. The new framework continues a recent research trend studying the use of query-based sampling and a centralized sample database. A single logistic model was trained on the centralized Table 7. Precision on the trec123-100col-bysource testbed when 3 databases were selected (The first baseline is CORI; the second baseline for UUM/HP methods is UUM/HR.) (Search engines do not return document scores) Precision at Doc Rank CORI ReDDE UUM/HR UUM/HP-FL UUM/HP-VL 5 docs 0.3520 0.3240 (-8.0%) 0.3680 (+4.6%) 0.4520 (+28.4%)(+22.8%) 0.4520 (+28.4%)(+22.8) 10 docs 0.3320 0.3140 (-5.4%) 0.3340 (+0.6%) 0.4120 (+24.1%)(+23.4%) 0.4020 (+21.1%)(+20.4%) 15 docs 0.3227 0.2987 (-7.4%) 0.3280 (+1.6%) 0.3920 (+21.5%)(+19.5%) 0.3733 (+15.7%)(+13.8%) 20 docs 0.3030 0.2860 (-5.6%) 0.3130 (+3.3%) 0.3670 (+21.2%)(+17.3%) 0.3590 (+18.5%)(+14.7%) 30 docs 0.2727 0.2640 (-3.2%) 0.2900 (+6.3%) 0.3273 (+20.0%)(+12.9%) 0.3273 (+20.0%)(+12.9%) 40 sample database to estimate the probabilities of relevance of documents by their centralized retrieval scores, while the centralized sample database serves as a bridge to connect the individual databases with the centralized logistic model. Therefore, the probabilities of relevance for all the documents across the databases can be estimated with very small amount of human relevance judgment, which is much more efficient than previous methods that build a separate model for each database. This framework is not only more theoretically solid but also very effective. One algorithm for resource selection (UUM/HR) and two algorithms for document retrieval (UUM/HP-FL and UUM/HP-VL) are derived from this framework. Empirical studies have been conducted on testbeds to simulate the distributed search solutions of large organizations (companies) or domain-specific hidden Web. Furthermore, the UUM/HP-FL and UUM/HP-VL resource selection algorithms are extended with a variant of SSL results merging algorithm to address the distributed document retrieval task when selected databases do not return document scores. Experiments have shown that these algorithms achieve results that are at least as good as the prior state-of-the-art, and sometimes considerably better. Detailed analysis indicates that the advantage of these algorithms comes from explicitly optimizing the goals of the specific tasks. The unified utility maximization framework is open for different extensions. When cost is associated with searching the online databases, the utility framework can be adjusted to automatically estimate the best number of databases to search so that a large amount of relevant documents can be retrieved with relatively small costs. Another extension of the framework is to consider the retrieval effectiveness of the online databases, which is an important issue in the operational environments. All of these are the directions of future research. ACKNOWLEDGEMENT This research was supported by NSF grants EIA-9983253 and IIS-0118767. Any opinions, findings, conclusions, or recommendations expressed in this paper are the authors', and do not necessarily reflect those of the sponsor. REFERENCES [1] J. Callan. (2000). Distributed information retrieval. In W.B. Croft, editor, Advances in Information Retrieval. Kluwer Academic Publishers. (pp. 127-150). [2] J. Callan, W.B. Croft, and J. Broglio. (1995). TREC and TIPSTER experiments with INQUERY. Information Processing and Management, 31(3). (pp. 327-343). [3] J. G. Conrad, X. S. Guo, P. Jackson and M. Meziou. (2002). Database selection using actual physical and acquired logical collection resources in a massive domain-specific operational environment. Distributed search over the hidden web: Hierarchical database sampling and selection. In Proceedings of the 28 th International Conference on Very Large Databases (VLDB). [4] N. Craswell. (2000). Methods for distributed information retrieval. Ph. D. thesis, The Australian Nation University. [5] N. Craswell, D. Hawking, and P. Thistlewaite. (1999). Merging results from isolated search engines. In Proceedings of 10th Australasian Database Conference. [6] D. D'Souza, J. Thom, and J. Zobel. (2000). A comparison of techniques for selecting text collections. In Proceedings of the 11th Australasian Database Conference. [7] N. Fuhr. (1999). A Decision-Theoretic approach to database selection in networked IR. ACM Transactions on Information Systems, 17(3). (pp. 229-249). [8] L. Gravano, C. Chang, H. Garcia-Molina, and A. Paepcke. (1997). STARTS: Stanford proposal for internet meta-searching . In Proceedings of the 20th ACM-SIGMOD International Conference on Management of Data. [9] L. Gravano, P. Ipeirotis and M. Sahami. (2003). QProber: A System for Automatic Classification of Hidden-Web Databases. ACM Transactions on Information Systems, 21(1). [10] P. Ipeirotis and L. Gravano. (2002). Distributed search over the hidden web: Hierarchical database sampling and selection. In Proceedings of the 28th International Conference on Very Large Databases (VLDB). [11] InvisibleWeb.com. http://www.invisibleweb.com [12] The lemur toolkit. http://www.cs.cmu.edu/~lemur [13] J. Lu and J. Callan. (2003). Content-based information retrieval in peer-to-peer networks. In Proceedings of the 12th International Conference on Information and Knowledge Management. [14] W. Meng, C.T. Yu and K.L. Liu. (2002) Building efficient and effective metasearch engines. ACM Comput. Surv. 34(1). [15] H. Nottelmann and N. Fuhr. (2003). Evaluating different method of estimating retrieval quality for resource selection. In Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. [16] H., Nottelmann and N., Fuhr. (2003). The MIND architecture for heterogeneous multimedia federated digital libraries. ACM SIGIR 2003 Workshop on Distributed Information Retrieval. [17] A.L. Powell, J.C. French, J. Callan, M. Connell, and C.L. Viles. (2000). The impact of database selection on distributed searching. In Proceedings of the 23rd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. [18] A.L. Powell and J.C. French. (2003). Comparing the performance of database selection algorithms. ACM Transactions on Information Systems, 21(4). (pp. 412-456). [19] C. Sherman (2001). Search for the invisible web. Guardian Unlimited. [20] L. Si and J. Callan. (2002). Using sampled data and regression to merge search engine results. In Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. [21] L. Si and J. Callan. (2003). Relevant document distribution estimation method for resource selection. In Proceedings of the 26th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. [22] L. Si and J. Callan. (2003). A Semi-Supervised learning method to merge search engine results. ACM Transactions on Information Systems, 21(4). (pp. 457-491). 41

resource selection;distributed information retrieval

205

Unwanted Traffic in 3G Networks

The presence of "unwanted" (or background) traffic in the Internet is a well-known fact. In principle any network that has been engineered without taking its presence into account might experience troubles during periods of massive exposure to unwanted traffic, e.g. during large-scale infections. A concrete example was provided by the spreading of Code-Red-II in 2001, which caused several routers crashes worldwide. Similar events might take place in 3G networks as well, with further potential complications arising from their high functional complexity and the scarcity of radio resources. For example, under certain hypothetical network configuration settings unwanted traffic, and specifically scanning traffic from infected Mobile Stations, can cause large-scale wastage of logical resources, and in extreme cases even starvation. Unwanted traffic is present nowdays also in GPRS/UMTS, mainly due to the widespread use of 3G connect cards for laptops. We urge the research community and network operators to consider the issue of 3G robustness to unwanted traffic as a prominent research area.

INTRODUCTION Public wide-area wireless networks are now migrating to third-generation systems (3G), designed to support packet-switched data services and Internet access. Several UMTS networks became operational since 2003 while early GPRS deployments date back to 2000. Since then, the growing popularity of 3G terminals and services has extended the coverage of Internet wireless access to the geographic area, and 3G networks are becoming key components of the global Internet. In a recent CCR contribution Keshav [17] foresees that cell phones will become the dominant component of future Internet population, while Kleinrock expects this role to be played by "small pervasive devices ubiquitously embedded in the physical world" (quoted from [14, p. 112]). Both scenarios underlay that the main access mode in the future Internet will be wide-area wireless. Currently deployed 3G networks, along with their future evolutions, are in pole-position face to concurrent technologies (e.g. WIMAX) to provide such access connectivity in the large-scale. Generally speaking, the 3G network being essentially a mixture of two paradigms, namely mobile cellular and IP, it is exposed to the security and reliability issues affecting each component, plus the new risks emerging from their combination . The 3G environment inherits from the cellular paradigm a number of features like terminal personalization and geolocalization that make privacy and information security particularly critical. When coupled with the IP world, markedly the "openess" of its applications and accessibility, the concerns of privacy and security from the user perspective become even more critical than in legacy 2G networks. Because of that - and of some "lessons learned" from past mistakes in 2G security [5] - privacy and information security aspects have received a thorough treatment in the 3G specifications (see [7] for an exhaustive overview). Nevertheless, the specific topic of 3G network security in relation to the robustness and availability of the network infrastructure itself has not received adequate attention by the research community to date. The problem can be condensed in the following question: What is the level of robustness of a 3G network against deliberate attacks or other unanticipated stimuli? The problem of network security involves issues related to network resilience and stability, and can not be addressed without a deep understanding of the detailed structure and organization of the real network. Considered the relative recent deployment of 3G, and the very limited access that research groups have to these networks, it should be no surprise that the work in this area has been sporadic. Some exploits against 3G network are known and documented in industry reports (e.g. [15] [2]), while the fact that a limited amount of malicious traffic can cause large-case troubles to a wireless cellular network has been "unveiled" in the recent paper [18] with reference to a 2G network supporting open SMS service. But at this stage what is still missing is an exhaustive and systematic recognition of the potential risks, threats and problems to 3G network security, from which a research agenda can be drawn. We provide here a novel contribution towards this goal by introducing an issue that has passed unrecognized so far: the impact onto 3G networks of unwanted traffic, and specifically large-scale worm infections. Remarkably, all the cited previous works consider deliberate DoS attack against the network. Instead here we focus on a slightly more subtle issue, namely the (side-)effects onto the network of (unwanted ) traffic, whose intended target is typically not the network but rather its terminals. Our work was inspired by the consequences of the Code-Red-II infection onto the routers of the wired Internet, reported in [3] and [4]. We claim that under certain conditions and for certain network configuration scenarios large-scale worm infections can cause sensible degradation and risks for the network performances and availability. We urge the research community and network operators to consider the issue of 3G robustness to unwanted traffic as a prominent research area. The goal of this contribution is to trigger interest and at the same time move the first pioneering steps in such direction. The following discussion is based on empirical observations from an operational GPRS/UMTS network collected during an ongoing research project in traffic monitoring and modeling in 3G, the DARWIN project [1], carried out in collaboration with mobilkom austria AG&CoKG (the leading mobile operator in Austria, EU) and Kapsch CarrierCom (provider of equipments and network engineering services). OVERVIEW OF 3G NETWORKS Network structure. A 3G network includes two main sections : a Packet-Switched Core Network (CN), which is based on IP, and one or more Radio Access Network (RAN). Along with the UMTS RAN (UTRAN) based on W-CDMA, several operators maintain a parallel GPRS RAN evolved from the legacy GSM radio. This structure is sketched in Figure 1. It is also possible to connect additional separate RANs to the same CN, typically WLAN [13] and perhaps in the future also WIMAX. Each RAN can evolve independently from the CN: for example in several networks GPRS has been upgraded to EDGE [10, p. 152], while UMTS upgrade towards HSDPA [8, p. 351] is ongoing. Each RAN is connected to the legacy 2G Circuit-Switched Core-Network (not shown in Figure 1) for traditional services like voice calls, and to the Packet-Switched Core-Network (CN for short) for data services. The CN embeds several elements: SGSN, GGSN, and a number of information servers. Some of the latter are shared with the Circuit-Switched Core-Network of the legacy 2G system , e.g. the HLR/AuC. The SGSNs perform functions such as access control, location management , paging, route management [10]. The GGSN is the logical gateway between the CN and external packet networks (Internet and private networks), is endowed with a full IP-stack and handles the IP-level connectivity with the MS. The SGSN and GGSN of the same operator communicate through the Gn interface. The CNs of different opera-1 Notably the close coupling between the circuit- (GSM) and packet-switched (GPRS/UMTS) sections is a source of concern since in principle troubles originated in the latter might cause impairments or side-effect to the former as well. tors are interconnected through the Gp interface for support of roaming. The Gn protocol stack [10, p. 94] shows that a lower UDP/IP layer is used to carry the user data packets across Gn, with an intermediate encapsulation into a 3G-specific protocol (GPRS Tunnelling Protocol, GTP). In fact, the Gn interface is basically a wide-area IP network interconnecting the different SGSN/GGSN sites, and as such it embeds routers, IP subnets etc. Besides that, the CN is rich in IP-based elements, including servers supporting control and management functions (e.g. DNS, DHCP, RADIUS, see [10]) and application elements (e.g. WAP gateway, proxies, internal servers). The latter are always located behind the GGSN, on the Gi side (ref. Figure 1) as they operate directly on the data-plane. Note also that packet filtering and other restiction policies can be located on separate dedicated elements (NAT, IDS, firewalls) at the network boundaries (Gi, Gp) and/or directly configured into the GGSNs. 3G terminals. The population of 3G terminals is highly heterogeneous and includes very different types of device: hand-held phones and PDA, connect-card pluggable into laptops, blackberry, etc. Additionally, a broad range of automatic devices with no human interaction is emerging, taking advantage of the ubiquity of the GPRS/UMTS coverage (e.g. sensors, alarms, presence indicators, remote cameras). Presently the most numerous 3G terminals are hand-held phones. They span a broad range of technological platforms, a major point of difference (for the moment) from the wired Internet that is essentially a monoculture. The last aspect is critical when considering malware infections: such a "bi-ological variety" intrinsically limits the potential infection scope, which in turn reduces somehow the very appeal for programming new pieces of malware. As a result, large-scale infections of cellular phones have not yet been observed, despite a growing number of exploits and pieces of malicious code targeting GPRS/UMTS phones have already appeared in the wild (e.g. Cabir, Mosquito, Comwarrior 2 ). 3G datacards for laptop. Many 3G datacards for laptop were sold starting in 2004, often coupled with flat-rate offers. Most of these laptops are equipped with Microsoft Windows - note that for some datacards drivers are not available for other operating systems. This introduced into the 3G environment a sub-population of homogeneous terminals, i.e. Windows laptops, that are intrinsically exposed to all kinds of exploits and infections that are found in the wired Internet . In case of active infection (e.g. a scanning worm) they introduce into the 3G network the same "unwanted" traffic patterns (e.g. probe SYN packets) that are found in wired LANs and in the Internet. PROBLEM STATEMENT Unwanted traffic. The term "unwanted traffic" has been used in [16] to refer cumulatively to those traffic components originated directly or indirectly by malicious or anyway "non productive" activities. It includes backscatter traffic asso-ciated to remote DoS attacks, scanning probes, spam, exploit attempts etc. Unwanted traffic might have a negative impact onto the underlying network, and in extreme cases drive the network or at least some of its elements to crash. A bright example was provided by the spreading of Code-Red -II in 2001 [3]. Once installed on a victim host, the worm started to scan for new potential victims by sending a high rate of probing TCP SYN packets to random addresses. This caused troubles to the packet forwarding modules of several edge routers all over the Internet, some of which eventually crashed [4]. In simple words, the problem is that route caching mechanisms were designed (and optmized) to operate under "normal" (i.e. expected) traffic conditions, where most of the packets are directed to a relativelly small subset of popular subnets. In such nominal condition, route caching can be very effective. But during the infection probing SYN packet were massively generated and sent to randomly chosen IP addresses, thus driving the cache access mechanisms to explode. In other words, the worm infection built-up a traffic aggregate macroscopically different from the "normal" pattern, and the network proved to be not robust enough to sustain such different conditions. The lesson to be learned is that in terms of the characteristics of the macroscopic traffic aggregate (entropy of the destination IP address distribution, packet size, etc.) large infections or other unwanted traffic components can expose the network to a different "operating point" from what the network was engineered and optmized for, with potentially dramatic effects 3 . Potential impact on 3G. In principle, the 3G network is exposed to the same type of incidents, and perhaps even more given the higher functional complexity inherited by the wireless cellular paradigm. The 3G network is ultimately an IP network, but with important peculiarities. First, the underlying transport stratum, specifically the 3G-specific lower protocols in the RAN, are endowed with very high functional complexity and signaling interactions - mainly for the sake of mobility management and efficient resource management. Second, the population of internal "hosts" is extremely large (from thousands to millions of MSs) and highly dynamic (activity periods can be as short as few seconds). The potential impact of large-scale infections and unwanted traffic in such a system is an intriguing point for research, that has not yet been addressed by the research community. The existence of the problem has been conjectured in a previous work [9, p. 447-448]. In lack of past empirical events, it is not possible to claim that 3G network are exposed to serious damages from large infections. On the other hand, without a systematic risk assessment it is neither possible to provide a priori guarantees about their robustness. Empirical evidence of the very existence of unwanted traffic in a real 3G network has been reported in [6] along with initial but technically-detailed speculations on the potential impact that the observed traffic would have under certain hypothetical conditions and configuration setting. The actual impact, if any, depends on a combination of factors related to the network configuration and equipment features. In the following we illustrate the problem by discussing a few examplary forms of impact that might take place in a real network. Stateful elements. The presence of massive amounts of TCP SYN packets might cause troubles to those stateful elements designed to reserve resources for each TCP con-3 In this regard, this is another example of (lack of) robusteness to unanticipated types of events in HOT systems [11]. nection (e.g. application layer proxies, servers, NATs). Note that some stateful operations might be enabled also on the GGSNs.. In this cases the GGSN logic should be robust to high rates of SYN packets coming from the MSs. Large volumes of SYN packets might be originated by deliberate DoS/DDoS or from large-scale infections of scanning worms. In both cases, the source(s) can be hosts in the Internet (exogenous traffic) or other MS in the RAN (endogenous traffic). In general, exogenous traffic can be blocked at the external firewall as for any other private network. The first element to inspect the IP packets sent by the MSs is the GGSN. The latter generally embeds full router capabilities , therefore it can be configured with the same stateless / stateful firewalling policies and/or throttling mechanisms (see e.g. [12]) to filter endogenous uplink traffic. For an improved robusteness against residual unblocked SYNs, all stateful elements should be designed to resist massive SYN storms rather than just rely on external filtering elements. Wastage of logical resources. The UMTS radio bearer channels (called Dedicated Channel, DCH) are assigned dy-namically to active MSs. The assignment policy is implemented in the RNC and is generally based on a combination of timeouts from the last data packet and thresholds on the recent sending / receving rates. The exact algorithm is vendor-dependent, with parameters configurable by the operator . Let us consider here the simplest case of a purely timeout-based DCH assignment policy: the DCH is assigned to the MS at the time of the first packet (sent or received), and is released after T DCH seconds from the last packet, T DCH being the holding timeout for DCH. Note that when the MS does not have an assigned DCH, packets are ex-changed on the common channels FACH or RACH (see [8, Ch. 7]). Note also that each channel switch operation involves a signaling procedure at the radio interface, contributing to the total transfer delay for the arriving packet. The value of T DCH must be tuned carefully. Too short values causes a high frequency of channel switch cycles, and consequently (i) a higher consumption of signaling resources on the radio link and (ii) longer packet delays and hence worse user experience. On the other hand, too long values will lead to wastage of logical resources, i.e. DCHs, whose available number if limited in each cell. Therefore, the optimal value of T DCH must be chosen according to the distribution of idle-period duration for "typical users". Given such framework, consider what happen when a number of infected terminals are scanning the local address space. Each active MS (not necessarely infected) will be visited by scanning probes at an average rate of R v pkt/sec. The exact value of R v depends on several factors like number of scanning MSs, scanning rate, etc. (see [6] for more details) and can typically be in the order of few seconds or below. In case that the average probe interarrival time is smaller than the DCH holding timer, i.e. v = (R v ) -1 &lt; T DCH , the incoming unwanted traffic will keep the DCH channel assigned to the target MSs indefinitely, until the user switches off the terminal or explicitely close the PDP-context 4 . Note that the volume in byte count of such incoming background traffic is extremely low and would pass unnoticed by the user. No assumption is made about the vulnerability of the 4 The "PDP-context" is the logical connection to the 3G network, conceptually similar to a wired modem dial-up. ACM SIGCOMM Computer Communication Review 55 Volume 36, Number 2, April 2006 target MS to the specific exploit, the only condition being that it is reachable by probing packets, i.e. it has an active PDP-context. Such always-on "spurious" DCH waste resources on the radio interface. Notably, wastage is limited to the logical resources, i.e. DCH, since the physical bandwidth is left largely unused as only sporadic and small packets (probe SYNs) are transmitted over the air. Such phenomenon might lead to logical congestion of some radio cells as soon as the number of active MSs in the cell reaches the number of available DCHs. Signaling overhead. One key assumption in the above scenario is that the average interarrival of background packets is smaller than the DCH holding timeour, i.e. v &lt; T DCH . Other problems arise in case that v is higher but close to T DCH , i.e. v = T DCH + for small , particularly in the case of low T DCH . In this case, a DCH reassignment follows immediately a DCH release at rate 1/T DCH , thus wasting signaling bandwidth in the radio section. Again, the more "victims" are present in the same cell the higher the impact. CONCLUSIONS We warn that unwanted (or "background") traffic can have an impact onto the functionally-complex 3G network, at least under certain conditions of network configuration and setting. Real measurements [6] provide evidence of the presence of such traffic inside a real GPRS/UMTS network. We have speculated on its potential impact under hypothetical network conditions (e.g. MS-to-MS communication enabled , no firewalling set in the GGSNs). The extent to which such conditions are effectively found in a real network is unknown , as mobile operators do not disclose details about the deployment and configuration of their networks. Since the actual impact, if any, depends pointedly on a combination of factors related to the network configuration and equipment features, in many cases the relevant countermeasures and fixes are obvious or anyway simple to implement once that the potential risk has been identified. Often preventive actions are as simple as a careful and informed network engineering and equipment configuration. For instance, stateful firewalling at the GGSN prevents probe packets to reach the target MS thus avoding DCH channels to be "spuri-ously" kept alive by background traffic. Alternatively, a more sophisticated DCH assignment strategy (e.g. based on thresholds on the packet rate) would alleviate the problem. However, such features might never be activated unless an explicit recognition of the problem of unwanted traffic and its consequences. In summary, the very first problem is to recognize and assess the potential risks, which might be hidden in the intricate web of interactions and dependencies embedded within the functionally-complex 3G network. The potential risks due to the presence of unwanted traffic must be taken into account in the design of the network setting , so as to avoid the emergence of hazardous conditions. A coherent process of risk assessment should be considered as a natural component of the network engineering process. In turn, risk recognition must be based on a thorough understanding of the specific traffic environment, which is conti-nously evolving following the emerging of new services, new types of terminals, new forms of infections, new attacks, etc. Automatic or semi-automatic methods can be implemented to detect drifts in the macroscopic composition of the traffic, including the raise of new components of unwanted traffic, borrowing concepts and tools from the recent achievements in the field of anomaly detection in the Internet. The prerequisite for all that is a continuous (always-on) process of large-scale traffic monitoring and analysis from inside the network, i.e. on the internal interfaces like Gn. REFERENCES [1] DARWIN home page http://userver.ftw.at/ ricciato/darwin. [2] A. Bavosa. Attacks and Counter Measures in 2.5G and 3G Cellular IP Networks. Juniper White Paper, June 2004. Online at www.juniper.net/solutions/lit-erature/white papers/200074.pdf. [3] C.C. Zou, W. Gong, D. Towsley. Code Red Worm Propagation Modeling and Analysis. 9th ACM Conf. on Computer and Comm. Security (CCS'02), 2002. [4] Cisco. Dealing with mallocfail and High CPU Utilization Resulting From the "Code Red" Worm. www.cisco.com/warp/public/117/ts codred worm.pdf. [5] E. Barkan, E. Biham, N. Keller. Instant Ciphertext-Only Cryptanalysis of GSM Encrypted Communications . Crypto 2003, Santa Barbara, CA, August 2003. [6] F. Ricciato, P. Svoboda, E. Hasenleithner, W. Fleischer. On the Impact of Unwanted Traffic onto a 3G Network. Technical Report FTW-TR-2006-006, February 2006. Available online from [1]. [7] G. M. Koien. An Introduction ro Access Security in UMTS. IEEE Wireless Communications, 11(1), 2004. [8] H. Holma, A. Toskala. WCDMA for UMTS. Wiley. [9] H. Yang, F. Ricciato, S. Lu, L. Zhang. Securing a Wireless World. Proceedings of the IEEE, 94(2), 2006. [10] J. Bannister, P. Mather, S. Coope. Convergence Technologies for 3G Networks. Wiley, 2004. [11] J. M. Carlson, J. Doyle. HOT: Robustness and design in complex systems. Phys. Rev. Let., 84(11), 2000. [12] J. Twycross, M. M. Williamson. Implementing and testing a virus throttle. Tech. Report HPL-2003-103, May 2003. Online www.hpl.hp.com/techreports/2003. [13] K. Ahmavaara, H. Haverinen, R. Pichna. Interworking Architecture Between 3GPP and WLAN systems. IEEE Communications Magazine, November 2003. [14] L. Kleinrock. The Internet: History and Future. Lectio Magistralis at Politecnico di Torino, October 2005. Online at www.tlc.polito.it/ nordio/seminars. [15] O. Whitehouse. GPRS Wireless Security: Not Ready For Prime Time. Research Report by stake, June 2002. Online at www.atstake.com/research/reports. [16] R. Pang et al. Characteristics of Internet Background Radiation. IMC'04, Taormina, Italy, October 2004. [17] S. Keshav. Why Cell Phones Will Dominate the Future Internet. Computer Communication Review, 35(2), April 2005. [18] W. Enck, P. Traynor, P. McDaniel, T. La Porta. Exploiting Open Functionality in SMS Capable Cellular Networks. 12th ACM Conf. on Computer and Comm. Security (CCS'05), November 2005. ACM SIGCOMM Computer Communication Review 56 Volume 36, Number 2, April 2006

Unwanted traffic;Cellular networks;3G

206

Use of Contextualized Attention Metadata for Ranking and Recommending Learning Objects

The tools used to search and find Learning Objects in different systems do not provide a meaningful and scalable way to rank or recommend learning material. This work propose and detail the use of Contextual Attention Metadata, gathered from the different tools used in the lifecycle of the Learning Object, to create ranking and recommending metrics to improve the user experience. Four types of metrics are detailed: Link Analysis Ranking, Similarity Recommendation, Personalized Ranking and Contextual Recommendation. While designed for Learning Objects, it is shown that these metrics could also be applied to rank and recommend other types of reusable components like software libraries.

INTRODUCTION One of the main reasons to capture and analyze the information about the interaction between a user and a tool is to improve the user experience. For example, a online library system could record the subject of the books that a client has bought before in order to recommend him new books about a similar subject the next time he/she logs in, saving him/her the hassle to search for them [1]. A news web site could record the topic of the news articles that a user normally read in order to filter out news that do not interest such user [2]. A collaborative browser could use the information recollected from the browsing patterns of a given community to improve the rank of different pages on the searches of an individual user, member of that community [3]. The generic name of Attention Metadata[4] has been applied to describe the information about these interactions. When the information stored does not only contain the reference to the user and the action that it performs, but also register when the action took place, through which tool the action was performed, what others thing was doing the user at the same time, what is the profile of the user performing the action, to what community he/she belongs, etc, it leads to an improved and more useful form of record, called Contextualized Attention Metadata [5] (CAM). AttentionXML [6] and its extensions [5] are an effort to standardize the way in which CAM is stored. This standardization will lead to the opportunity to share attention records between different applications. For example, a second generation of an Attention-Sharing online library could know which news topics the user is interested in and it could recommend him/her books related to those topics. The authors believe that one group of applications that could greatly benefit from CAM information is the search and find of Learning Objects. These applications have suffered from an under-par performance compared to similar applications in other fields [7] [8]. The main reason for this is the lack of a meaningful and scalable way to rank or recommend the objects to the users. Currently, two main methods are used to rank (not even recommend) Learning Objects: Manual Rating or Metadata Content Rating. In the first approach, Manual Rating, each Learning Objects should be rated by a group of experts and/or the user community. For each search, the returned objects are ranked based on their average rate. While this is bound to provide meaningful ordering, it does not scale. For example MERLOT use this approach, but only 10% of the total content of the database has ever be rated [9]. The other approach, use only the information contained in the metadata record to perform ranking based on the similarity with the query terms. The most common method used for this is the TFIDF metric[10] that measure in a Vector Space the distance between the query vector and the vector composed from the text contained in the metadata record. Given than TFIDF was designed to work over full text documents and that metadata records contain very few textual descriptions [11], normally the ordering is not meaningful for the user. SILO (Search and Indexing Learning Objects) tools from the ARIADNE [12] repository use this approach. CAM could be used to generate a third approach, one in which the human attention (meaningful) is processed to construct an automated (scalable) rating and recommending procedure. The following sections of this work describe in detail what information should be stored in the CAM record of Learning Object Applications (Section 2) and the mechanisms by which such information could be used to generate rating and recommending metrics (Section 3). It is also discussed how these mechanisms and metrics could be applied to related contexts (Section 4) and which research questions need to be addressed in further work (Section 5). The work finalize with an overview of related research (Section 6) CAM FOR LEARNING OBJECTS APPLICATIONS Users interact with a Learning Object through the object's whole lifecycle. CAM recorders capture and timestamp all those interactions in order to provide the information needed to calculate useful metrics to be used in a next generation breed of learning object management tools. According to the AttentionXML extension proposed by Najjar et al at [5], these interactions are stored inside an Action record. This work will briefly list the different Actions that should be recorded through the Learning Object lifecycle. The lifecycle phases are taken from the enumeration done by Collins and Strijker in [13]. Also, it is suggested which applications should generate the attention records. Creation: In this phase the author creates or assembles the learning object in its digital form using some sort of authoring tool. The Creating Action should be captured and it must include the identity of the created object, its author(s), the authoring tool used and the list of component-objects [14] reused through the creation process. This record should be created by the authoring tool, for example Microsoft Power Point. Labeling: At this stage the author, an indexer or even an automated system could add a metadata record that describes the Learning Object. The Labeling Action must include information that identify the object, the labeler, the origin of the metadata (Automated, Semi-Automated, Manual), the metadata format used, the level of confidence of the information (how sure the autor is that metadata values are correct) and a unique identifier for the metadata record. Normally this record should be also created by the authoring tool at the end of the creation of the objects, but could also be created by metadata editors as [15] or automated metadata generators as [16]. Offering: At this stage the author or indexer inserts the object in a repository or other system that allow the object to be shared with others. The Inserting Action must include the following information: Object Unique Identifier, Inserter, Tool Used and Learning Object Unique Identifier inside the sharing tool. This record should be created by the sharing tool, being it a Learning Object Repository or a Peer to Peer sharing application. Selecting: In this stage the user search, find and select Learning Objects in the Sharing System. Several Actions should be captured during this phase. A Searching Action when a query is performed to find relevant objects. It must include information that describe the query performed and the objects returned. A Recommending Action when the system suggests relevant objects without the user performing a query. It must contain information a list of the object(s) recommended, the user action that trigger the recommendation and the tool used to perform the recommendation. A Browsing Action when the user reviews the metadata or description of an object. It must store information that identifies the metadata record browsed and the time expend in the review. Finally, A Selecting Action when the user chooses an object by downloading it or accepting the recommendation. It must contain information that identifies the selected object. All this actions should also contain information about the user that performs the action. These records should be generated by the sharing or recommending tool. Using: This stage comprehends all the actions that the final user (instructor or learner) performs with the learning object during its normal utilization in a learning environment. There are several actions to be registered. A Publicating Action when the instructor inserts the object into a Course belonging to some kind of Learning Management System. It must contain information that identifies the published object and the context (course, lesson) where it was published. A Sequencing Action when one or more objects are included in an instructional design or sequenced package. It must contain information about the identification (in an ordered form) of the integrated objects. A Viewing Action, the object is read or viewed by learners. It must contain information about the time spent reviewing the material. An Annotating Action when the instructor or the learner add a comment or rate the learning object. It must include information about the comment or the rate given and the identifier of the object. All these action should also store information about the user that performs them. Different tools should be in charge of the generation of the attention records, a LMS for the Publishing Action, a Learning Activity Management System [17] or SCORM [18] Packager for the Sequencing Action, a Web browser or document reader for the Viewing Action and a Rating or Review system for the Annotating Action. Lifecycle Actions Main Information Source Creation Creating author, components Authoring tool, Components Labeling Labeling metadata format, origin, confidence Authoring tool or Metadata generator Offering Inserting inserter LOR or Sharing app. Searching query, results LOR's search tool Recommending objects recommended Recommender Browsing Time LOR or Recommender Selecting Selecting object identifier LOR or Recommender Publicating LMS context LMS Sequencing list of sequenced objects ID tool or Packager Viewing Time, tool used Browser or Reading app. Using Annotating rate or review LMS Retaining Retaining decision to keep or delete LMS Table 1. Proposed CAM information to be stored for Learning Object Applications 10 Retaining: In this phase, the instructor check for the validity of the learning object and decides if it is still useful or if it should be replaced / updated. The Retaining Action should contain information that identifies the object and the decision taken (keep, update, delete). This attention record will normally be generated by the LMS where the object has been published. A summary of the Actions that CAM should record is presented in Table 1. Some of these CAM Actions (Creating, Inserting, Selecting, Viewing) are already produced and stored in different tools [5]. The others are easy to implement in existing tools taking in account that most of them (LMS, Metadata Generators, etc) already produce a log with the user's interactions. In the next section, metrics to exploit this Action records to improve tools to search and find Learning Objects are proposed. RANKING AND RECOMMENDING METRICS USING CAM Several ranking and recommending metrics will be proposed. These metrics will use only two sources of information to be calculated: the first one is the Learning Object Metadata (LOM) [19] record that describe each Learning Object; the second one is the CAM Actions described in the previous section. 3.1 Link Analysis Based Ranking One of the most famous and successful ranking algorithms at the present is PageRank [20]. PageRank use the information contained in the network of links between web pages to calculate the relative "importance" of a page. It could be summarized as: a page is important if it is linked by a high number of pages. Also, the importance increases if the pages linking to it have also a high importance rank. Unfortunately, this algorithm could not be applied directly to Learning Objects. While LOM records have a linking field, it is rarely populated [11]. Also, LOM linking reflect just a semantic relationship; it does not imply a "vote" for that object as it is assumed for Web pages. As an alternative to the explicit linking structure that the web posses, CAM allow us to create an implicit linking between Learning Objects and other entities related to them: Authors, Users, Courses, Learners, etc. For example: Creating Actions can be converted into a link between an author and an object, Selecting Actions can be converted into a link between a user and an object, Publicating Actions can be converted into a link between a course and an object and also between a user and the same object. Viewing Actions can be converter into a link between a learner and an object. As result of this conversion of CAM to links between different entities, a K-partite graph is created (a graph with different partitions, where there are not links between nodes of the same partition). In this graph each type of entity (Learning Object, User, Course, and Learner) is considered a partition Figure 1 present diagram of an example of such a graph. Once CAM information is represented as a graph, it is easy to use basic graph algorithms to calculate ranking metrics. Following there are some metrics that could be developed this way: Popularity Rank (PR): Using the information contained in the Selecting Action (converted already in a 2-partite graph), it is easy to obtain the number of times that an object has been downloaded. To calculate just count the number of incident links that each Learning Object node has from nodes in the User Partition. This metric is a just a basic way to put most downloaded objects first in the result list. ) ( ) ( object inDegree object PR = Figure 1. K-Partite Graph representation of CAM Author-Corrected Popularity Rank (ACP): Combining the Creating and Selecting Actions, it could be calculated how popular an object is based on the number of downloads and the popularity of the Author. The first step is to create a 3-partite graph with Users, Objects and Author partitions. Then the PopularityRank (PC) is calculated for all the objects. Next, the Author Popularity (AP) is calculated adding the PC of all the Learning Objects nodes that are linked to the author node. Finally, the AP is multiplied by a weighting factor and added to the also weighted PC. This metrics enable new objects (that do not have any downloads) from a well downloaded author, appear higher in the list. = i i object PR author AP ) ( ) ( ; if object i is linked to author ) ( ) ( ) ( author AP object PR object ACP + = Weighted Popularity (WP): Selecting, Publicating and Retaining Actions can be combined to generate a 2-partite graph between Users and Learning Objects. The links of this graph will be weighted: if the link is made using the Retaining information (inDegree R ) it will have more weight as if the link was made using the Publication information (inDegree P ). In the same way, Publication links will weight more than Selection links (inDegree S ). The rationale behind this metric is that different actions mean different level of "preference" for an object. If the instructor has use the object and she is happy with it to keep it for the next semester is a stronger vote of support than just using it for the first time or that just downloading it. That difference of importance is represented in the weight given to each kind of link. ) ( ) ( ) ( ) ( object inDegree object inDegree object inDegree object WP R P S + + + = with &gt; &gt; O 1 O 2 O 3 C1 C2 U1 U 2 A1 A2 User Partition Course Partition Author Partition Object Partition 11 Rate of Reuse Rank (RRR): Using the Selecting Action (or also the Publicating and Retaining Actions as in the previous metric), the number of times that an object has been downloaded during a given period of time P (last week, month, year) can be calculated. The 2-partite graph (Users and Objects partitions) can be constructed but only taking in account the Actions occurred in a given period of time. For example: if the last week is selected as P, This rank will calculate how often the object has been downloaded (inserted or retained) on the last 7 days. This value could be normalized using the age of the object, obtained from the related Creation action. This metric will help to rank higher object that have been reused frequently and are relatively new. ) ( ) ( ) ( object age object inDegree object RRR = ; for links inside period P Manual Rank (MR): Using the information that is stored in the Annotation Action, the number of times that an object has been positively (or negatively) rated or reviewed could be considered to calculate a metric. A 2-partite graph (Users and Objects partitions) is created. The procedure will weight the link as 1 if it is a positive rate or review, -1 if it is a negative rate or review. The actual value of the rate is only used to evaluate if the rate is a positive or negative "vote", because different users and system have different scales to grade. The reviews can only be considered if their positiveness or negativeness value is included in the Annotation Action or could be automatically inferred from the text. ) ( ) ( ) ( object inDegree object inDegree object MR Negative Positive = These metrics can be calculated off-line because they are not user or query specific. They calculate an average importance or relevance of the learning objects based in the agglutination of attention information. These metrics, and others that can be developed afterwards, could be integrated in a final ranking metric. This Compound Popularity Metric (CP) can be calculated as the weighted sum of the values of the individual metrics. For example, Google integrates more than 100 of different simple metrics in order to provide its results [21]. MR RRR WP ACP PR CP + + + + = The weighted coefficients ( , , etc) should be estimated (not trivial procedure) to provide an optimal result ordering. Methods to make these estimates are described in [22] and [23]. Also, manual rates should be used carefully because the Annotation Information is optional and could not exist for all the objects involved in the calculation. 3.2 Similarity Metrics for Recommendation One property of a 2-partite graph is that it can be folded over one of its partitions, generating a normal graph with just one entity and links between its nodes. For example if we have a 2-partite graph of Users that have download Learning Objects, we can fold over the Learning Object partition and we will end up with a graph where the Users are linked between them. Each link mean that those two users have download the same object at least once. This new graph could be used to calculate similarity between the users based on the download patterns. In Figure 2 we can see a representation of the folding result. The first part of the figure represents a 2-partite graph with the User and Objects partitions. The graph shows that, for example, that User 5 had downloaded Object 2 and Object 3 and User 1 had only downloaded Object 1. The second part of the figure illustrates the folded version of the graph. In this new graph, the users have a link between them if they linked to the same object in the unfolded graph. The more objects the users have in common, the thicker the line. For example User 1 and User 4 are linked because they both have downloaded Object 1. User 2 and User 5 have a stronger links because they both have downloaded Object 2 and Object 3. This technique is similar to the one applied in scientometrics to obtain relations between different authors, based on the papers that have co-authored[24]. U1 U2 U3 O1 O2 O3 U4 U5 U6 U1 U2 U3 U4 U6 U5 2-Partite Graph (User and Objects) Folded Normal Graph (Users) Figure 2. Unfolded and Folded 2-Partite Graph We present several similarity metrics that can be calculated using the information contained in CAM Actions detailed in the previous section. Object Similarity based on Number of Downloads: Create a 2-partite graph with the information of the Select Actions (when a User download a Learning Object), and fold over the User Partition. A link between two Objects in the final graph means that those objects have been downloaded by the same user. The strength of the similarity is number of users that have downloaded both objects. Object Similarity based on Re-Use: Create a 2-partite graph with the information from the Publish Actions (when a Learning Object is inserted into a Course), and fold over the Course Partition. A link between two Objects in the final graph means that two objects have been inserted in the same course. The strength of the similarity is number of courses that include both objects. Users similarity based on Downloads: Create a 2-partite graph with the information from the Select Actions, and fold over the Object Partition. A link between two Users means that they have downloaded the same object. The strength of the similarity is the number objects that the users have in common. Author similarity based on Re-Use of Components: The Creation Action information could be use to identify re-use of learning object components. For example several authors could use the same picture or diagram inside they presentations. As the Creation Action store information about which existing components have been reused (see Section 2), a 2-partite graph between Authors and Components can be created and then folded over the 12 Components partition. The new graph will represent relationship between different authors. More components those authors have used in common, the stronger their similarity. The similarity metric obtained from the graph could be then applied in recommendation tools. For example: If a user finds an object useful, a link to similar objects could be provided (similar to what Amazon does with books [1]). Also, the similarity between users can be exploited to recommend Learning Objects to a user, based on what other users that are in the same community have recently download (similar to collaborative browsing applications [3]). To automatically extract the communities from the graph, an algorithm like EdgeBetweeness [25] can be applied. The same procedure could be applied to the Author Similarity graph. The communities of authorship can be automatically extracted from the graph. The author then can be recommended with components that have been created by other authors in the same authorship community. Beside recommendation systems for Learning Objects, these similarity metrics could be considered as distance metrics. The distance metric can be used inside clustering algorithms to automatically find groups of similar objects. These clusters could be used to improve the presentation of results of a search, much as Vivisimo [26] does for Web Pages. 3.3 Personalized Ranking To be able to personalize the search result order for a given user, the application should have a representation of such user in a profile. While this profile could be created explicitly by the user, CAM information could help the application to learn it form the user interaction with the tool. For example, the information about stored in the Select, Publish and Retain from a user could help us to determine in which objects is he/she interested, and rank higher objects that are similar to those. This work proposes the creation of a fuzzy profile that could easily account for the evolving and not fixed behavior of an instructor downloading learning objects. Instead of having a crisp preference for one type of object, this profile will provide different grades of likeness for several characteristics of the learning object. This profile is constructed with several Fields. The Fields could be a subset of the fields considered in the LOM standard, especially the ones that use a vocabulary or represent a classification. Each field will contain 2 or more fuzzy sets that represent the values that the field could have (from the vocabulary or the classification values). A user could "prefer" in different degrees 1 or more of the values of a Field. The preference of the user for each one of the values is calculated based on the number of objects that the user have download before that contained that value in the corresponding LOM field. This fuzzy profile has been derived from research done to produce automatic TV recordings for PVRs [27] like TIVO. The fuzzy profile could be easily operationalized to provide a personalized rank for Learning Objects. First, each field should have a weighting value (that express how important that field is). That value could be assigned by an expert or could be calculated automatically for entropy of the distribution of the field values for that user. For example if a user downloads objects from a wide variety of topics, the weight of topic as a good ranking measurement is low. Instead, if the user only downloads objects in one language, the weight of that field should be high. Second, each LOM record from the result list is converted to a similar representation, using the same fields and a preference value of 1.0 for the value found in the metadata. Finally, the object representation is operated with the profile in order to obtain how well the object fits the preference of the user. This operation is described in the following equation: j i i j j i i value object Field value user Field user object nk PersonalRa ). ( ). ( ) , ( = For example, lets consider a user that have download 20 objects, 16 with topic Computer Sciences and 4 with topic Physics. Of those 20, also, 12 are in English, 4 are in French, 4 are in Spanish. A fuzzy profile that represents that user could be expressed as: U1 = {(0.8/ComputerScience + 0.2/Physics), (0.6/English + 0.2/Spanish + 0.2/French)} The fields weighting terms are 0.9 for Topic and 0.6 for Language. Lets now considered 2 objects represented also as fuzzy sets: O1 = {(1.0/ComputerScience), (1.0/Spanish)} O2 = {(1.0/Physics, 1.0/English)} The calculated rank for both objects is: O1 = 0.9*0.8 + 0.6*0.2 = 0.84 O2 = 0.9*0.2 + 0.6*0.6 = 0.54 O1 will be ranked higher than O2 as it is more similar to the user profile. The personalized calculation could be combined with the popularity ranking described before to create a better ranking algorithm, the same way as Google personalized Search mix the standard Popularity measure with information from the user profile to order the results. 3.4 Contextual Recommending If the CAM is considered not only as a source for historic data, but also as a continuous stream of contextualized attention information, we can use very recent CAM (in the order of seconds or minutes) to generate recommendations based on what the user is focusing his/her attention at the moment. For example, the recommender system could use the information stored in the if the user has inserted an object inside a Course in a Learning Management System (LMS), the LMS will generate a CAM record with contextual information about which object was inserted and in which lesson of the course. The recommending system could use that information to present the user with similar objects to the one inserted or others that have been used in similar courses, based on the topic of the course or in similarity metrics as the one explained in the Section 3.2. The recommending system could also present objects that suit the application that the user is using at a given moment, based on the information about the object (LOM record). For example, if the user is working Microsoft Power Point authoring tool, presentations, slides, small texts, images and diagrams will be recommended. If he/she is working with a SCORM Packager, complete learning objects will be presented instead. Contextual recommending techniques have been tried before in several fields [28] [29]. Blinkx [30] is an example of this kind of applications. It recommends web pages, videos and news based 13 on the present content of the screen of the user. A similar application could be developed in a LMS for example, where the system could recommend to the instructor materials to add to each lesson, or could recommend the learner with similar or complementary materials to the one that the instructor has added to the course. APPLICATION IN OTHER CONTEXTS While the CAM based metrics proposed in this work were designed for Learning Objects, they could be easily extended or adapted to work for other kind of reusable components where CAM could be collected. For example, given the exponential grow of open source software libraries that could be reused inside software projects, programmers are sometimes overwhelmed with the amount of available choices. It makes sense to develop some kind of ranking or recommending system that could help the developers to select the right tools. To construct the ranking application we can use the same methods proposed for learning objects. The k-partite graph used to calculate the popularity metric could be constructed using the metadata information about the library (who is the author of a software library) and contextual attention information about how and when the programmers interact with the library (which programmers have download it, in which software project they have been used). Most of this information could be obtained from open source project repositories like SourceFourge [31]. The rationale behind the ranking would be: A library that have been downloaded more often / at a higher rate is more useful. A library produced by authors with highly useful libraries could also be useful. A library re-used in many projects is probable highly useful. This metrics are parallel to the ones described for learning objects. Recommending systems for software libraries could also be constructed in a similar way to the ones proposed for learning objects. For example, we can fold the Libraries-Programmers 2-partite graph over the Libraries Partition, creating a graph that relate Programmers between them based on the Libraries that they have downloaded/used. Communities could be extracted form the resulting graph and could be used for recommending a programmer with new libraries that other members of his/her community have used in their projects. The precaution to have when applying this metrics to other domains is the semantics of the relations that are created with the graphs. For example, if two learning objects are used in the same course, those two learning objects must have something in common (same topic for example), while if two libraries are used inside the same project, that does not mean that the libraries are related (you could use a database access library and graphical interface library inside the same project). Other contexts where CAM information could be exploited to rank and recommend elements with a similar strategy as the one presented in this work are music mixes (component songs or loops) and news aggregators. FURTHER WORK This work is just an introduction to how CAM information could be used to rank and recommend Learning Objects. Several topics should be solved before a big scale application that use the metrics presented could be built: Collection and Integration of the different CAM sources: While today exist several applications that generate CAM, there is not an established multi-application CAM repository that could be used to collect and integrate attention information. Combination of different ranking strategies: When different ranking strategies are combined, some weighting coefficient must be applied. The calculation of those coefficients is not trivial and should be made using extensive user feedback. Critical mass vs. Closed Community: To be useful, the metrics should be calculated over a significant amount of CAM data. But if we integrate data from different communities to obtain a bigger amount of CAM (for example attention from different LORs), there will probably not exist common objects, users or courses that could be used to generate relations between the communities. RELATED WORK Broisin et al in [32] propose a framework to capture usage information about Learning Object from different Learning Management Systems and Repositories in order to analyze the usage patterns of the users through a Management Application. The approach of this paper goes a step further, using the attention information to calculate metrics that could be used to improve existing tools. Broisin's work also uses a simplified format of attention (basically usage information) in a non-standard format, limiting the possible use of the information by other systems, because existing applications should be reprogrammed to produce that format. This work proposes the use of an extension of AttentionXML standard to be able to capture the CAM from a variety of systems that already produce it. In a related area, digital libraries, Nicholson in [33] propose the fusion of bibliometrics analysis with user-related data mining to generate a new field of study, bibliomining. His proposal could be compared with the one presented in this work: Using the information about the book and the usage information generated by the interaction of the users with the digital library system to improve the user experience. While Nicholson mentions several ways in which the attention metadata could be used, he does not detail any specific metric to improve digital library systems. CONCLUSIONS The current immaturity of the tools to search and find Learning Objects could be overcome if CAM information is store through the lifecycle of the Learning Object and used to compute metrics for ranking and recommendation. These metrics should generate a meaningful and automated way in which Learning Object could be ranked. This work presented detailed methods to calculate various metrics and propose several uses for those metrics. The proposed calculations could also be applied to rank and recommend other reusable components from which CAM could be gathered, as it was shown for the case of open source software libraries example. While the metrics are easy to calculate, and some initial data is also present, more research is needed to be able to assemble a large scale system that could gather the necessary amount of CAM in order to render the calculations meaningful. 14 REFERENCES [1] Linden, G.; Smith, B. and York, J. Amazon.com Recommendations: Item-to-Item Collaborative Filtering. IEEE Internet Computing, 7, 1 (2003), 76-80. [2] Shepherd, M.; Watters, C. and Marath, A. Adaptive User Modeling for Filtering Electronic News. In Proceedings of the 35th Annual Hawaii International Conference on System Sciences, 2002. HICSS. (2002), 1180- 1188. [3] James, S. Outfoxed Collaborative Browsing, http://www.getoutfoxed.com. Retrieved on May, 2006. [4] Najjar, J., Meire, M. and Duval, E. Attention Metadata Management: Tracking the use of Learning Objects through Attention.XML. In Proceedings of World Conference on Educational Multimedia, Hypermedia and Telecommunications. (2005). 1157-1161. [5] Najjar, J., Wolpers, M. and Duval, E., Attention Metadata:Collection and Management", WWW2006 workshop on Logging Traces of Web Activity, Edinburgh, Scotland, (2006). [6] AttentionXML, AttentionXML specifications, http://developers.technorati.com/wiki/attentionxml. Retrieved on June, 2006 [7] Duval, E. and Hodgins, W., A LOM research agenda. In Proceedings of WWW2003: Twelfth International World Wide Web Conference, (2003), 659-667. [8] Ochoa, X. Learning Object Repositories are Useful, but are they Usable? In Proceedings of IADIS International Conference Applied Computing. (2005), 138-144 [9] Duval, E. LearnRank: the Real Quality Measure for Learning Materials. Policy and Innovation in Education - Quality Criteria, (2005) [10] Aizawa, A. An information-theoretic perspective of tfidf measures. Information Processing and Management, 39, (2003), 45-65. [11] ISO/IEC JTC1 SC36. International LOM Survey: Report. http://mdlet.jtc1sc36.org/doc/SC36_WG4_N0109.pdf (2004). [12] Ariadne Foundation. Ariadne Foundation. http://www.ariadne-eu.org (2005). [13] Collis, B. and Strijker, A. Technology and Human Issues in Reusing Learning Objects. Journal of Interactive Media in Education, 4, (2004). [14] Verbert, K. Jovanovic, J. Gasevic, D. and Duval, E. Repurposing Learning Object Components. OTM 2005 Workshop on Ontologies, Semantics and E-Learning, (2005). [15] IEEE-LOM Editor, http://www-i5.informatik.rwth-aachen .de/i5new/staff/chatti/LOMEditor/index.html. Retrieved June 2006. [16] Cardinels, K., Meire, M., and Duval, E. Automating metadata generation: the simple indexing interface. In Proceedings of the 14th WWW conference, (2005), 548-556 [17] Dalziel, J. Implementing Learning Design: The Learning Activity Management System (LAMS), ASCILITE (2003) [18] ADL, SCORM Standard, http://www.adlnet.gov/index.cfm, Retrieved March, 2006 [19] IEEE. IEEE Standard for Learning Object Metadata. http://ltsc.ieee.org/doc/wg12/ (2002). [20] Page, L., Brin, S., Motwani, R. and Winograd, T. The PageRank Citation Ranking: Bringing order to the Web. Technical Report, Computer Science Department, Stanford University (1998) [21] Google Technology, http://www.google.com/technology/. Retrieved, August 2006. [22] Radlinski, F. and Joachims, T. Query Chains: Learning to Rank from Implicit Feedback, Proceedings of the ACM Conference on Knowledge Discovery and Data Mining. (2005). [23] Fan, W., Gordon, M. and Pathak, P. A generic ranking function discovery framework by genetic programming for information retrieval, Information Processing and Management. 40 (2004) 587602 [24] Nascimento, M., Sander, J. and Pound, J. Analysis of SIGMOD's co-authorship graph. ACM SIGMOD Record, 32, 3. (2003). 8-10 [25] Girvan, M. and Newman, M. Community structure in social and biological networks. Proc. Natl. Acad. Sci. 11. (2002). [26] Vivisimo Clustering Engine. http://www.vivisimo.com. Retrieved August 2006. [27] Pigeau, A., Raschia, G., Gelgon, M., Mouaddib, N. and Saint-Paul, R. A fuzzy linguistic summarization technique for TV recommender systems. The 12th IEEE International Conference on Fuzzy Systems, 2003. FUZZ '03. 1 (2003) 743-748. [28] Google AdSense, https://www.google.com/adsense/. Retrieved August 2006. [29] Fan, W., Gordon, M. and Pathak, P. Incorporating contextual information in recommender systems using a multidimensional approach. Information Processing and Management. 40, 4. (2004). 587-602. [30] Blinkx Contextual Search. http://www.blinkx.com. Retrieved August 2006. [31] Sourceforge, Open Software Repository. http://www.sourceforge.net. Retrieved August 2006. [32] Broisin, J., Vidal, P. and Sibilla, M. A Management Framework Based On A Model Driven Approach For Tracking User Activities In A Web-Based Learning Environment. EDMEDIA, (2006) 896-903 [33] Nicholson, S. The basis for bibliomining: Frameworks for bringing together sage-based data mining and bibliometrics through data arehousing in digital library services. Information Processing and Management. 42, 3 (2006). 785-804 . 15

Learning Objects;Ranking;Attention Metadata;Recommending

207

Use of Relative Code Churn Measures to Predict System Defect Density

Software systems evolve over time due to changes in requirements, optimization of code, fixes for security and reliability bugs etc. Code churn, which measures the changes made to a component over a period of time, quantifies the extent of this change. We present a technique for early prediction of system defect density using a set of relative code churn measures that relate the amount of churn to other variables such as component size and the temporal extent of churn. Using statistical regression models, we show that while absolute measures of code churn are poor predictors of defect density, our set of relative measures of code churn is highly predictive of defect density. A case study performed on Windows Server 2003 indicates the validity of the relative code churn measures as early indicators of system defect density. Furthermore, our code churn metric suite is able to discriminate between fault and not fault-prone binaries with an accuracy of 89.0 percent.

INTRODUCTION A "reliability chasm" often separates the quality of a software product observed in its pre-release testing in a software development shop and its post-release use in the field. That is, true field reliability, as measured by the number of failures found by customers over a period of time, cannot be measured before a product has been completed and delivered to a customer. Because true reliability information is available late in the process, corrective actions tend to be expensive [3]. Clearly, software organizations can benefit in many ways from an early warning system concerning potential post-release defects in their product to guide corrective actions to the quality of the software. We use code churn to predict the defect density in software systems. Code churn is a measure of the amount of code change taking place within a software unit over time. It is easily extracted from a system's change history, as recorded automatically by a version control system. Most version control systems use a file comparison utility (such as diff) to automatically estimate how many lines were added, deleted and changed by a programmer to create a new version of a file from an old version. These differences are the basis of churn measures. We create and validate a set of relative code churn measures as early indicators of system defect density. Relative churn measures are normalized values of the various measures obtained during the churn process. Some of the normalization parameters are total lines of code, file churn, file count etc. Munson et al. [17] use a similar relative approach towards establishing a baseline while studying code churn. Studies have shown that absolute measures like LOC are poor predictors of pre- and post release faults [7] in industrial software systems. In general, process measures based on change history have been found be better indicators of fault rates than product metrics of code [9]. In an evolving system it is highly beneficial to use a relative approach to quantify the change in a system. As we show, these relative measures can be devised to cross check each other so that the metrics do not provide conflicting information. Our basic hypothesis is that code that changes many times pre-release will likely have more post-release defects than code that changes less over the same period of time. More precisely, we address the hypotheses shown in Table 1. Our experiments on Windows Server 2003 (W2k3) support these four hypotheses with high statistical significance. We analyzed the code churn between the release of W2k3 and the release of the W2k3 Service Pack 1 (W2k3-SP1) to predict the defect density in W2k3-SP1. The relative code churn measures are statistically better predictors of defect density than the absolute measures. They also they are indicative of increase in system defect density and can accurately predict the system defect density with a high degree of sensitivity. Our metric suite is able to discriminate between fault and not fault-prone binaries in W2k3-SP1 with an accuracy of 89.0 percent. Table 1. Research Hypotheses Hypothesis H 1 Increase in relative code churn measures is accompanied by an increase in system defect density H 2 Using relative values of code churn predictors is better than using direct (absolute) values to explain the system defect density H 3 Relative code churn measures can be used as efficient predictors of system defect density. H 4 Relative code churn measures can be used to discriminate between fault and not fault-prone binaries. The organization of this paper is as follows. Section 2 describes the related work. Section 3 explains data collection and section 4 the relative code churn measures. Section 5 presents the case study and the observed results. Section 6 discusses our conclusions and future work. RELATED WORK Prior analyses on predicting defect density used code churn measures as part of a larger set of metrics. Code churn measures have not been studied in isolation as predictors of software defect density. The background work presented below is from studies that involved industrial software systems. The source code base of W2k3 is two orders of magnitude larger than the largest example considered below. Munson et al. [17] observe that as a system is developed, the relative complexity of each program module that has been altered (or churned) also will change. The rate of change in relative complexity serves as a good index of the rate of fault injection. They studied a 300 KLOC (thousand lines of code) embedded real time system with 3700 modules programmed in C. Code churn metrics were found to be among the most highly correlated with problem reports [17]. Khoshgoftaar et al.[13] define debug churn as the number of lines of code added or changed for bug fixes. Their objective was to identify modules where debug code churn exceeds a threshold, in order to classify the modules as fault-prone. They studied two consecutive releases of a large legacy system for telecommunications. The system contained over 38,000 procedures in 171 modules. Discriminant analysis identified fault-prone modules based on 16 static software product metrics. Their model when used on the second release showed a type I and II misclassification rate of 21.7%, 19.1% respectively and an overall misclassification rate of 21.0%. Ohlsson et al. [19] identify fault-prone modules by analyzing legacy software through successive releases. They use a total of 28 measures, twelve of which are based on size and change measures. These measures were used to identify 25 percent of the most fault-prone components successfully. Karunanithi [12] uses a neural network approach for software reliability growth modeling in the presence of continuous code churn, which he shows improves over the traditional time-domain based models. Similarly Khoshgoftaar et al. [15] use code churn as a measure of software quality in a program of 225,000 lines of assembly language. Using eight complexity measures, including code churn, they found neural networks and multiple regression to be an efficient predictor of software quality, as measured by gross change in the code. They suggest that using neural networks may not work in all environments and the results obtained are environment specific. Neural networks can be used for improving software maintenance [15]. Ostrand et al. [20] use information of file status such as new, changed, unchanged files along with other explanatory variables such as lines of code, age, prior faults etc. as predictors in a negative binomial regression equation to predict the number of faults in a multiple release software system. The predictions made using binomial regression model were of a high accuracy for faults found in both early and later stages of development. [20] Closely related to our study is the work performed by Graves et al. [9] on predicting fault incidences using software change history. Several statistical models were built based on a weighted time damp model using the sum of contributions from all changes to a module in its history. The most successful model computes the fault potential by summing contributions from changes to the module, where large and/or recent changes contribute the most to fault potential [9]. This is similar to our approach of using relative measures to predict fault potential. Drawing general conclusions from empirical studies in software engineering is difficult because any process depends to a large degree on a potentially large number of relevant context variables. For this reason, we cannot assume a priori that the results of a study generalize beyond the specific environment in which it was conducted [2]. Researchers become more confident in a theory when similar findings emerge in different contexts [2]. Towards this end we intend that our case study contributes towards strengthening the existing empirical body of knowledge in this field [7, 9, 13, 15, 17, 19, 20]. DATA COLLECTION The baseline used for measuring the code churn and other measures described below is Windows Server 2003 (W2k3). We measured churn between this baseline and Windows Server 2003 Service Pack 1 (W2k3-SP1). We sometimes refer to W2k3-SP1 as the "new version" of the code. Service packs are a means by which product updates are distributed 1 . Service packs contain updates for system reliability, program compatibility, security, etc. that are conveniently bundled for easy downloading. The size of the code base analyzed is 44.97 million LOC (44,970 KLOC). This consisted of 2465 binaries which were compiled from 96,189 files. Some files contribute to more than one binary. As defects for W2k3-SP1 are reported at the binary level, we relate churn to defects at the level of binaries. 1 http://support.microsoft.com/ 285 The absolute measures and methods of data collection are described below: Total LOC is the number of lines of non-commented executable lines in the files comprising the new version of a binary. Internal Microsoft tools were used to compute this measure. Churned LOC is the sum of the added and changed lines of code between a baseline version and a new version of the files comprising a binary. Deleted LOC is the number of lines of code deleted between the baseline version and the new version of a binary. The churned LOC and the deleted LOC are computed by the version control systems using a file comparison utility like diff. File count is the number of files compiled to create a binary. Weeks of churn is the cumulative time that a file was opened for editing from the version control system. Churn count is the number of changes made to the files comprising a binary between the two versions (W2k3 and W2k3-SP1). Files churned is the number of files within the binary that churned. RELATIVE CODE CHURN MEASURES In this section we describe our relative code churn measures. The churn measures are denoted by the elements M1-M8. The elements and their relationship to defect density are explained below (these relationships are verified in section 5.1): M1: Churned LOC / Total LOC. We expect the larger the proportion of churned (added + changed) code to the LOC of the new binary, the larger the magnitude of the defect density for that binary will be. M2: Deleted LOC / Total LOC. We expect the larger the proportion of deleted code to the LOC of the new binary, the larger the magnitude of the defect density for that binary will be. M3: Files churned / File count. We expect the greater the proportion of files in a binary that get churned, the greater the probability of these files introducing defects. For e.g. suppose binaries A and B contain twenty files each. If binary A has five churned files and binary B has two churned files, we expect binary A to have a higher defect density. M4: Churn count / Files churned. Suppose binaries A and B have twenty files each and also have five churned files each. If the five files in binary A are churned twenty times and the five files in binary B are churned ten times, then we expect binary A to have a higher defect density. M4 acts as a cross check on M3. M5: Weeks of churn / File count. M5 is used to account for the temporal extent of churn. A higher value of M5 indicates that it took a longer time to fix a smaller number of files. This may indicate that the binary contains complex files that may be hard to modify correctly. Thus, we expect that an increase in M5 would be accompanied by an increase in the defect density of the related binary. M6: Lines worked on / Weeks of churn: The measure "Lines worked on" is the sum of the churned LOC and the deleted LOC. M6 measures the extent of code churn over time in order to cross check on M5. Weeks of churn does not necessarily indicate the amount of churn. M6 reflects our expectation that the more lines are worked on, the longer the weeks of churn should be. A high value of M6 cross checks on M5 and should predict a higher defect density. M7: Churned LOC / Deleted LOC. M7 is used in order to quantify new development. All churn is not due to bug fixes. In feature development the lines churned is much greater than the lines deleted, so a high value of M7 indicates new feature development. M7 acts as a cross check on M1 and M2, neither of which accurately predicts new feature development. M8: Lines worked on / Churn count: We expect that the larger a change (lines worked on) relative to the number of changes (churn count), the greater the defect density will be. M8 acts as a cross check on M3 and M4, as well as M5 and M6. With respect to M3 and M4, M8 measures the amount of actual change that took place. M8 cross checks to account for the fact that files are not getting churned repeatedly for small fixes. M8 also cross checks on M5 and M6 to account for the fact that the higher the value of M8 (more lines per churn), the higher is the time (M5) and lines worked on per week (M6). ). If this is not so then a large amount of churn might have been performed in a small amount of time, which can cause an increased defect density. Figure 1 illustrates the cross check relationships of these relative code churn measures. As discussed above M1, M2 and M7 cross check on each other and M8 cross checks on the set of M3, M4 and M5, M6. All these measures triangulate on their respective dependent measures with the goal of providing the best possible estimate of defect density with a minimum inflation in the estimation. CASE STUDY We now describe the case study performed at Microsoft. Section 5.1 presents the correlation analysis between the relative code churn measures and system defect density. Section 5.2 details the model building activities and Section 5.3 the predictive ability of the models. Section 5.4 discusses the discriminative power of the relative code churn measures and Section 5.5 the limitations of the study. 286 Figure 1. Relative Churn Measure Cross Check Relationships Table 2. Cross Correlations. All correlations are significant at the 0.01 (99%) level (2-tailed). M1 M2 M3 M4 M5 M6 M7 M8 Defects /KLOC M1 1.000 .834 .795 .413 .707 .651 .466 .588 .883 M2 1.000 .645 .553 .747 .446 .219 .492 .798 M3 1.000 .186 .749 .434 .445 .269 .868 M4 1.000 .531 .429 .210 .631 .288 M5 1.000 .263 .201 .390 .729 M6 1.000 .701 .843 .374 M7 1.000 .507 .288 M8 1.000 .262 Defects/ KLOC 1.000 As mentioned before, the system defect density for W2k3-SP1 was collected at the level of binaries. That is, for each binary we have a count of the number of defects assigned to that binary. Throughout the rest of the paper we assume a statistical significance at 99% confidence (level of significance ( = 0.01)). 5.1 Correlation Analysis Our goal is to verify that with an increase in the code churn measures (M1-M8) there is a statistically significant increase in the defects/KLOC. Table 2 shows the Spearman rank correlation () among the defects/KLOC and the relative code churn measures. Spearman rank correlation is a commonly-used robust correlation technique [8] because it can be applied even when the association between elements is non-linear. Table 2 shows that there exists a statistically significant (at 99% confidence) positive relationship between the measures and the defects/KLOC (shown in bold). Thus, with an increase in the relative churn measures there is a corresponding positive increase in the defects/KLOC. This is indicated by the statistically significant positive Spearman rank correlation coefficient . From the above observations we conclude that an increase in relative code churn measures is accompanied by an increase in system defect density (H 1 ). In order to illustrate the cross checks better consider the measures M1, M2 and M7 in Figure 2 with their Spearman rank correlation coefficients from Table 2. Figure 2: Cross Correlation Relationships The Spearman correlation coefficient of 0.834 between M1 and M2 indicates that there is a very strong correlation between the two measures. But this might not be the case when there is a higher proportion of churned code compared to deleted code (as measured by M7 for new feature development). Since this cannot be measured by M1 or M2, M7 acts as a cross check on them. The correlation between M1 and M7 (0.466) indicates when there is a M1 M2 M7 0.834 0.466 0.219 M7 M2 M1 M6 M3 M4 M5 M8 M1: Churned LOC / Total LOC M2: Deleted LOC / Total LOC M3: Files churned / File count M4: Churn count / Files churned M5: Weeks of churn/ File count M6: Lines worked on / Weeks of churn M7: Churned LOC / Deleted LOC M8: Lines worked on / Churn count Cross check 287 new feature addition there is a corresponding increase in the churned code. For M2 and M7 this correlation is not as strong (but is statistically significant) because there were relatively fewer new feature additions compared to other changes in the W2k3-SP1 source base. 5.2 Model Fitting We now compare predictive models built using absolute measures against those built using the relative churn measures. For the absolute model, defects/KLOC is the dependent variable and the predictors are the absolute measures described in Section 3. For the relative model, defects/KLOC is the dependent variable and the predictors are the relative measures described in Section 4. R 2 is a measure of variance in the dependent variable that is accounted for by the model built using the predictors [4]. R 2 is a measure of the fit for the given data set. (It cannot be interpreted as the quality of the dataset to make future predictions). The adjusted R 2 measure also can be used to evaluate how well a model will fit a given data set [5]. Adjusted R 2 explains for any bias in the R 2 measure by taking into account the degrees of freedom of the predictor variables and the sample population. The adjusted R 2 tends to remain constant as the R 2 measure for large population samples. The multiple regression model fit for absolute measures using all the predictors has an R 2 value of 0.052 (F=16.922, p&lt;0.0005). (The F-ratio is used to test the hypothesis that all regression coefficients are zero). This is a poor fit of the data and irrespective of other transformations (like for e.g. log) we cannot get a marked improvement in R 2 . The adjusted R 2 value for the absolute measures is 0.49. Throughout the rest of this paper we present the adjusted R 2 values in addition to the R 2 measures in order to eliminate any bias in model building. But with respect to the large sample size (2465 binaries) the adjusted R 2 and R 2 value show only minor variation, not sufficient enough to drop the R 2 value and employ the adjusted R 2 value. There are different ways in which regression models [16] can be built. Three common regression methods [16] are forward, backward and step-wise regression. In forward regression, one adds a single predictor at a time to the model based on the strength of its correlation with the dependent variable. The effect of adding each predictor is evaluated based on the results of an F-ratio test [16]. Variables that do not significantly add to the success of the model are excluded. In backward regression, a model is built using all the predictors. The weakest predictor variable is removed and the strength of the overall built model is assessed similar to the forward regression procedure. If this significantly weakens the model then the predictor is put back (and otherwise removed). Step-wise regression [16] is the more robust technique of these methods. The initial model consists of the predictor having the single largest correlation with the dependent variable. Subsequently, new predictors are selected for addition into the model based on their partial correlation with the predictors already in the model. With each new set of predictors, the model is evaluated and predictors that do not significantly contribute towards statistical significance in terms of the F-ratio are removed so that, in the end, the best set of predictors explaining the maximum possible variance is left. A step-wise regression analysis using the absolute set of predictors does not lead to any significant change in the R 2 values (=0.051) (adjusted R 2 = 0.050). Only the LOC and the number of times a file is churned are kept as predictors. This further confirms the fact that using the absolute measures is not an appropriate method for assessing the system defect density. Several empirical studies use Principal Component Analysis (PCA) [10] to build regression models [6]. In PCA a smaller number of uncorrelated linear combinations of metrics, which account for as much sample variance as possible, are selected for use in regression. PCA is not a possible solution when using absolute measures because the correlation matrix is not positive definite. We still use the two principal components generated to build a multiple regression equation. The multiple regression equation constructed has an even lower value of R 2 =0.026, (F=33.279, p&lt;0.0005). Based on the three results discussed above (multiple regression using all the predictors, step-wise regression and PCA) we conclude that the absolute measures are not good predictors of system defect density. As outlined in Section 3 we calculate the relative code churn measures (M1-M8) and build regression models using all the measures, step-wise regression and PCA. Table 3 shows the R 2 value of the regression equation built using all the measures. We also present the adjusted R 2 value and the root MSE (Mean Squared Error). Table 3. Regression Fit Using All Measures Model R 2 Adjusted R 2 Root MSE All Measures .811 .811 1.301215 Table 4 shows how the R 2 value changes in step-wise regression for all the models built during that process. In the step-wise regression model the measure M7 is dropped. The best R 2 value in Table 4 (without M7) is the same as that of Table 3 (.811) but there is a change in the third decimal place of the standard error of the estimate. M7 probably was dropped because there were relatively fewer new feature additions compared to other changes in the W2k3-SP1 source base. The adjusted R 2 values are also shown but are not significantly different from the R 2 values due to the large sample size used to build the models. 288 Table 4. Step-wise Regression Models Model R-Square Adjusted R-Square Root MSE (a) .592 .592 1.908727 (b) .685 .685 1.677762 (c) .769 .769 1.437246 (d) .802 .801 1.331717 (e) .808 .807 1.312777 (f) .810 .809 1.305817 (g) .811 .811 1.300985 a Predictors: (Constant), M2 b Predictors: (Constant), M2, M3 c Predictors: (Constant), M2, M3, M8 d Predictors: (Constant), M2, M3, M8, M1 e Predictors: (Constant), M2, M3, M8, M1, M6 f Predictors: (Constant), M2, M3, M8, M1, M6, M5 g Predictors: (Constant), M2, M3, M8, M1, M6, M5, M4. The PCA of the eight relative code churn measures yields three principal components. PCA can account for the multicollinearity among the measures, which can lead to inflated variance in the estimation of the defect density. But for PCA to be applicable the KMO (Kaiser-Meyer-Olkin) measure[11] of sampling adequacy should be greater than 0.6 [4]. The KMO measure of sampling adequacy is a test of the amount of variance within the data that can be explained by the measures. The KMO measure of the eight relative code churn measures is 0.594 which indicates that PCA might not be an appropriate method to apply. We still perform the analysis to investigate and present those results as well on a comparative basis. The results for all three models are summarized in Table 5. Table 5. Relative Measures Model Fits Model R 2 Adjusted R 2 F-Test sig. All measures 0.811 0.811 1318.44, (p&lt;0.0005) Step-wise regression 0.811 0.811 1507.31, (p&lt;0.0005) PCA 0.749 0.748 2450.89, (p&lt;0.0005) From the above results we can see that using relative values of code churn predictors is better than using absolute values to explain the system defect density (H 2 ). Figure 3: Actual vs. Estimated System Defect Density 289 5.3 Defect Density Prediction We use the technique of data splitting [18] to measure the ability of the relative code churn measures to predict system defect density. The data splitting technique was employed to get an independent assessment of how well the defect density can be estimated from a population sample. We randomly select two thirds of the binaries (1645) to build the prediction model and use the remaining one third (820) to verify the prediction accuracy. We constructed models using all the measures, step-wise regression and PCA (for purpose of completeness). Table 6 shows the results for these models. Table 6. Regression Data Fit Model R 2 Adjusted R 2 F-Test sig. All measures 0.821 0.820 938.304, (p&lt;0.0005) Step-wise regression (M7 dropped) 0.821 0.820 1072.975, (p&lt;0.0005) PCA 0.762 0.761 1749.113, (p&lt;0.0005) Using the fitted regression equation we estimate the system defect density for the remaining 820 binaries. Figure 3 shows the estimated and actual defect density using the regression equation constructed using all the measures (sorted by estimated defect density). The estimated defect density is shown by the thicker continuous line. From the graph we can see that the estimated defect density is similar to the actual defect density. The axes on the graphs are removed in order to protect proprietary data To quantify the sensitivity of prediction, we run a correlation analysis between the estimated and actual values. A high positive correlation coefficient indicates that with an increase in the actual defect density there is a corresponding positive increase in the estimated defect density. We perform Pearson and Spearman correlations to indicate their sensitivity. The Pearson correlation indicates a linear relationship. The Spearman correlation is a more robust correlation technique. Table 7 shows that the correlations are all positive and statistically significant. The magnitude of the correlations indicates the sensitivity of the predictions (the stronger the correlations the more sensitive are the predictions). The models built using all the measures and the step-wise method have the same sensitivity and are better than the model built using PCA. Table 7. Correlation Results Model Pearson (sig.) Spearman (sig.) All measures 0.889 (p&lt;0.0005) 0.929 (p&lt;0.0005) Step-wise regression 0.889 (p&lt;0.0005) 0.929 (p&lt;0.0005) PCA 0.849 (p&lt;0.0005) 0.826 (p&lt;0.0005) Analyses that are based on a single dataset that use the same data to both estimate the model and to assess its performance can lead to unreasonably negative biased estimates of sampling variability. In order to address this we repeat the random sampling with 3 different random samples to verify if the above results are repeatable. For each sample the model is fit with 1645 binaries to build the model. Table 8 shows the fit of the various models built for each sample. Table 8. Random Splits Data Fit Model R 2 Adjusted R 2 F-Test (Sig.) Random 1: All 0.836 0.835 1045.07, (p&lt;0.0005) Random 1: Stepwise (drop none) 0.836 0.835 1045.07, (p&lt;0.0005) Random 1: PCA 0.757 0.756 1701.98, (p&lt;0.0005) Random 2: All 0.822 0.821 941.86, (p&lt;0.0005) Random 2: Stepwise (drop M4) 0.821 0.820 1074.05, (p&lt;0.0005) Random 2: PCA 0.765 0.764 1776.87, (p&lt;0.0005) Random 3: All 0.799 0.798 813.12, (p&lt;0.0005) Random 3: Stepwise (drop M7) 0.799 0.798 927.54, (p&lt;0.0005) Random 3: PCA 0.737 0.736 1529.25, (p&lt;0.0005) Using each of the above predictive models we calculate the estimated defect density for the remaining 820 binaries. Table 9 shows the correlation between the estimated and the actual defect density. Table 9. Correlation Between Actual and Estimated Defects/KLOC Model Pearson Correlation (sig.) Spearman Correlation (sig.) Random 1: All 0.873 (p&lt;0.0005) 0.931 (p&lt;0.0005) Random 1: Stepwise 0.873 (p&lt;0.0005) 0.931 (p&lt;0.0005) Random 1: PCA 0.858 (p&lt;0.0005) 0.836 (p&lt;0.0005) Random 2: All 0.878 (p&lt;0.0005) 0.917 (p&lt;0.0005) Random 2: Stepwise 0.876 (p&lt;0.0005) 0.906 (p&lt;0.0005) Random 2: PCA 0.847 (p&lt;0.0005) 0.825 (p&lt;0.0005) Random 3: All 0.899 (p&lt;0.0005) 0.892 (p&lt;0.0005) Random 3: Stepwise 0.901 (p&lt;0.0005) 0.893 (p&lt;0.0005) Random 3: PCA 0.880 (p&lt;0.0005) 0.818 (p&lt;0.0005) Based on the consistent positive and statistically significant correlations, indicating the sensitivity of predictions obtained in Table 9 we can say that relative code churn measures can be used as efficient predictors of system defect density (H 3 ). Our results demonstrate it is effective to use all eight measures rather than dropping any of them from the predictive equation. Each of these measures cross check on each other and any 290 abnormal behavior in one of the measures (for e.g. like a file getting churned too many times) would be immediately highlighted. By interchanging the measures in a model equation we can get estimated values for all the relative measures independently. For example, in order to determine the maximum allowable code churn with respect to the file size (i.e. M1), say for a particular software model we fix the maximum allowable system defect density. We then can build a regression model with M2-M8 and defect density as predictors and M1 as the dependent variable. 5.4 Discriminant Analysis Discriminant analysis, is a statistical technique used to categorize programs into groups based on the metric values. It has been used as a tool for the detection of fault-prone programs [13, 14, 18]. The ANSI-IEEE Std. [1] defines a fault as an accidental condition that causes a functional unit to fail to perform its required function. We use discriminant analysis to identify binaries as fault-prone or not fault-prone. To classify if a binary is fault-prone or not we use the system defect density in a normal confidence interval calculation as shown in equation 1. LB = x -z /2 *Standard deviation of defect density... (1) n where LB is the lower bound on system defect density; x is the mean of defect density; Z /2 is the upper /2 quantile of the standard normal distribution; n is the number of observations. We conservatively classify all binaries that have a defect density less than LB as not fault-prone and the remaining as fault-prone. Table 10 shows the eigenvalue and overall classification ability of using the eight measures and the three principal components. The eigenvalue is a measure of the discriminative ability of the discriminant function. The higher the eigenvalue the better is the discriminative ability. For all measures, the function correctly classifies nearly nine out of every ten binaries. Table 10. Overall Discriminant Function Fit Model Eigenvalue Classification ability All Measures 1.025 2188/2465 (88.8%) PCA 0.624 2195/2465 (89.0%) As before, we split the data set into 1645 programs to build the discriminant function and the remaining 820 binaries to verify the classification ability of the discriminant function. We perform this analysis using all the measures and the principal components. The results of this fit and classification are shown below in table 11. Table 11. Discriminant Analysis For Model Fit (for 1645 binaries to build the model) For Test Data (820 binaries) Model Eigen value Classification ability Classification ability All Measures 1.063 1464/1645 (90.0%) 735/820 (89.6%) PCA 0.601 1461/1645 (88.8%) 739/820 (90.1%) Table 11 shows that the relative code churn measures have effective discriminant ability (comparable to prior studies done on industrial software [13]). We conclude that relative code churn measures can be used to discriminate between fault and not fault-prone binaries (H 4 ). 5.5 Limitations of Study Internal validity. Internal validity issues arise when there are errors in measurement. This is negated to an extent by the fact that the entire data collection process is automated via the version control systems. However, the version control systems only records data upon developer check-out or check-in of files. If a developer made many overlapping edits to a file in a single check-out/check -in period then a certain amount of churn will not be visible. A developer also might have a file checked out for a very long period of time during which few changes were made, inflating the "weeks of churn" measure. These concerns are alleviated to some extent by the cross check among the measures to identify abnormal values for any of the measures and the huge size and diversity of our dataset. In our case study we provide evidence for using all the relative churn measures rather than a subset of values or principal components. This is case study specific and should be refined based on further results. External validity. External validity issues may arise from the fact that all the data is from one software system (albeit one with many different components) and that the software is very large (some 44 million lines of code) as other software systems used for a similar analysis might not be of comparable size. CONCLUSIONS AND FUTURE WORK We have shown how relative code churn metrics are excellent predictors of defect density in a large industrial software system. Our case study provides strong support for the following conclusions: Increase in relative code churn measures is accompanied by an increase in system defect density; Using relative values of code churn predictors is better than using absolute values to explain the system defect density; Relative code churn measures can be used as efficient predictors of system defect density; and Relative code churn measures can be used to discriminate between fault and not fault-prone binaries. We plan to validate our approach on other products developed inside Microsoft like SQL Server and Office. We also plan to develop standards for all the measures to provide guidance to the developers on the maximum allowable change. We also plan to investigate how testing can more effectively be directed towards churned code. ACKNOWLEDGEMENTS We would like to express our appreciation to Brendan Murphy of Microsoft Research for providing the Windows Server 2003 SP1 data set. We would like to thank Madan Musuvathi of Microsoft Research, for critical feedback on the relative churn measures. We would like to thank Jim Larus of Microsoft Research, Laurie Williams, Jason Osborne of North Carolina State University for 291 reviewing initial drafts of this paper and the anonymous referees for their thoughtful comments on an earlier draft of this paper. REFERENCES [1] ANSI/IEEE, &quot;IEEE Standard Glossary of Software Engineering Terminology, Standard 729,&quot; 1983. [2] Basili, V., Shull, F.,Lanubile, F., &quot;Building Knowledge through Families of Experiments,&quot; IEEE Transactions on Software Engineering, Vol. Vol. 25, No.4, No., 1999. [3] Boehm, B. W., Software Engineering Economics. Englewood Cliffs, NJ: Prentice-Hall, Inc., 1981. [4] Brace, N., Kemp, R., Snelgar, R., SPSS for Psychologists: Palgrave Macmillan, 2003. [5] Brito e Abreu, F., Melo, W., &quot;Evaluating the Impact of Object-Oriented Design on Software Quality,&quot; Proceedings of Third International Software Metrics Symposium, 1996, pp. 90-99. [6] Denaro, G., Pezze, M., &quot;An Empirical Evaluation of Fault-Proneness Models,&quot; Proceedings of International Conference on Software Engineering, 2002, pp. 241 - 251. [7] Fenton, N. E., Ohlsson, N., &quot;Quantitative analysis of faults and failures in a complex software system,&quot; IEEE Transactions on Software Engineering, Vol. 26, No. 8, pp. 797-814, 2000. [8] Fenton, N. E., Pfleeger, S.L., Software Metrics. Boston, MA: International Thompson Publishing, 1997. [9] Graves, T. L., Karr, A.F., Marron, J.S., Siy, H., &quot;Predicting Fault Incidence Using Software Change History,&quot; IEEE Transactions on Software Engineering, Vol. 26, No. 7, pp. 653-661, 2000. [10] Jackson, E. J., A User's Guide to Principal Components: John Wiley & Sons, Inc., 1991. [11] Kaiser, H. F., &quot;An Index of Factorial Simplicity,&quot; Psychometrika, Vol. 39, No., pp. 31-36, 1974. [12] Karunanithi, N., &quot;A Neural Network approach for Software Reliability Growth Modeling in the Presence of Code Churn,&quot; Proceedings of International Symposium on Software Reliability Engineering, 1993, pp. 310-317. [13] Khoshgoftaar, T. M., Allen, E.B., Goel, N., Nandi, A., McMullan, J., &quot;Detection of Software Modules with high Debug Code Churn in a very large Legacy System,&quot; Proceedings of International Symposium on Software Reliability Engineering, 1996, pp. 364-371. [14] Khoshgoftaar, T. M., Allen, E.B., Kalaichelvan, K.S., Goel, N., Hudepohl, J.P., Mayrand, J., &quot;Detection of fault-prone program modules in a very large telecommunications system,&quot; Proceedings of International Symposium Software Reliability Engineering, 1995, pp. 24-33. [15] Khoshgoftaar, T. M., Szabo, R.M., &quot;Improving Code Churn Predictions During the System Test and Maintenance Phases,&quot; Proceedings of IEEE International Conference on Software Maintainence, 1994, pp. 58-67. [16] Kleinbaum, D. G., Kupper, L.L., Muller, K.E., Applied Regression Analysis and Other Multivariable Methods. Boston: PWS-KENT Publishing Company, 1987. [17] Munson, J. C., Elbaum, S., &quot;Code Churn: A Measure for Estimating the Impact of Code Change,&quot; Proceedings of IEEE International Conference on Software Maintenance, 1998, pp. 24-31. [18] Munson, J. C., Khoshgoftaar, T.M., &quot;The Detection of Fault-Prone Programs,&quot; IEEE Transactions on Software Engineering, Vol. 18, No. 5, pp. 423-433, 1992. [19] Ohlsson, M. C., von Mayrhauser, A., McGuire, B., Wohlin, C., &quot;Code Decay Analysis of Legacy Software through Successive Releases,&quot; Proceedings of IEEE Aerospace Conference, 1999, pp. 69-81. [20] Ostrand, T. J., Weyuker, E.J, Bell, R.M., &quot;Where the Bugs Are,&quot; Proceedings of the 2004 ACM SIGSOFT International Symposium on Software Testing and Analysis (ISSTA), 2004, pp. 86-96. 292

principal component analysis;Relative code churn;defect density;fault-proneness;multiple regression

208

Using Case-Based Reasoning in Traffic Pattern Recognition for Best Resource Management in 3G Networks

With the underlying W-CDMA technique in 3G networks, resource management is a very significant issue as it can directly influence the system capacity and also lead to system QoS. However, the resource can be dynamically managed in order to maintain the QoS according to the SLA. In this paper, CBR is used as part of an intelligent-based agent management system. It uses information from previously managed situations to maintain the QoS in order to meet the SLA. The results illustrate the performance of an agent in traffic pattern recognition in order to identify the specific type of problem and finally propose the right solution.

INTRODUCTION The third generation (3G) cellular system has been developed to satisfy increasing customer demands for higher bit-rate access in order to provide wireless Internet access anytime and anywhere. In addition, 3G networks will integrate different type of services like voice, data, and video. With W-CDMA, all users share the same spectrum and use codes to identify themselves. Hence the whole bandwidth can be reused in every cell. The system is considered a soft capacity system as all users simultaneously transmit so increasing the interference seen by others. The system capacity is, therefore, limited by the total interference that occurs from other users (in the case of the network being uplink-capacity limited) or other base stations (in the case of the network being downlink-capacity limited) and the background noise. The benefit of this technique is therefore providing the flexible, higher bandwidth services, and maintaining the best system capacity. On the other hand, it leads to more complexity in resource management. Previous work [1] introduced the use of intelligent agents in managing the resources to meet the service level agreement (SLA) when congestion occurs. It shows that by using intelligent agents together with the assignment and admission scheme, the system environment can be monitored and the policy that is suitable for that particular situation will be selected and applied to the system. Also the quality of service (QoS) for each particular class of customer can be monitored and controlled according to the SLA. In [2], Case-Based Reasoning (CBR) is introduced as a mean of giving the agent more "intelligence". The aim of using CBR is so that the problem can be automatically solved by referring to a similar traffic pattern that the system has seen before and kept in the case library. The end solution from the previous case can then be applied immediately to give a fast and efficient response. In this paper, a wider range of traffic situations will be illustrated, which will also show the benefit of using CBR in order to identify different traffic patterns and to propose the best solution. In addition, results show the outcome of system flexibility in giving different priority patterns to customers according to the system requirements. The paper is organised as follows. The agent system and architecture for the multi-agent system are described in section 2. In section 3, the implementation of CBR in SLA-based control by the agent will be introduced. The assignment and admission scheme is presented in section 4 and section 5 covers the simulation model. Traffic pattern recognition and numerical results are illustrated and discussed in section 6. Lastly, the conclusions of the paper are in section 7. AGENT SYSTEM AND ARCHITECTURE Critical in a radio network is the allocation of bandwidth to radio cells in order to avoid local congestion or degradation of the QoS and it is generally the capacity of the wireless link to the user that limits the overall system capacity, rather than any back-haul part of the network. In [3], an agent approach for a distributed resource management system is introduced. The main reason for using intelligent agents is to give greater autonomy to the base stations; this gives an increase in flexibility to deal with new situations in traffic load and to decrease the information load (the messaging resulting from taking, or determining control actions) on the network. In the past, mobile network operators have generally restricted the customer to only one service provider. With the influence of the Internet, more widespread choice of service providers (SPs) will be available to 3G users. By using an agent, it would be possible to allow selection of SP by offering on price, QoS, or value added service. In this work, each agent uses three layers taking action and decisions on different timescales: reactive, local planning and cooperative planning. As an individual connection must have the decision made in real-time , the reactive layer is designed for a very fast response. More complex functions have been implemented at the planning layers. Generally the local planning layer is concerned with long-term actions within its own instance, whereas the co-operative layer is concerned with long-term actions between peer agents, or with other types of agent. The reactive layer is, therefore very simple, implementing policies being passed down by the higher layer. This is discussed in more detail later in the paper. CBR IN SLA-BASED CONTROL CBR is an Artificial intelligence (AI) approach that can allow the agent to learn from past successes. It is a method that finds the solution to the new problem by analysing previously solved problems, called cases, or adapting old solutions to meet new demands. Figure 1 Case-based reasoning process model (Based on the CBR cycle in [4]) Figure 1 shows the process model of the case-based reasoning. The process of CBR starts when there is a new problem or new case happening. The first step is case retrieval, which uses the characterizing indexes of the event to find the best-match solved case(s) from the case library. The solution from the retrieved case(s) will be reused. However, the solution might need to be modified to fit the new situation as the new situation will rarely match the old one exactly: this step is called "revising". Once the new solution is proposed, the next step is to test it with the real environment. The result is either success or failure. If the solution fails, a monitoring process will analyse the failure, repair the working solution, and test again. If the solution succeeds, this new solution will be indexed and retained in the case library to use for future problem solving. The work shown in [5] gives an example of using CBR in network traffic control by using it to control traffic flow in the standard public switched telephone network of the Ile de France. In another work in [6], CBR is used to correct the error estimation of the required bandwidth computed by conventional connection admission control schemes. In the work described in this paper, CBR is used to recognise traffic patterns as congestion occurs in a 3G network and to define the policies to respond to that congestion in the reactive layer of the resource agent. Congestion here means the situation where system could not maintain the QoS required by the SLA. (This is explained in more detail in section 5) 3.2 Resource agent In this work the resource agent is the focus of attention as it is an important agent in managing the resource within the network. The architecture of the resource agent is illustrated in figure 2. Figure 2 Resource agent internal architecture The reactive layer is designed to be fast, performing the same function that would be in a conventional RNC (Radio Network Controller), assigning the connection a Node B, and performing Previous Cases General Knowledge Suggestion Solution Confirmed Solution RETRIEVE Problem Retrieve Case Tested / Repaired Case Solved Case Case New New Case Learned Case RETAIN REUSE REVISE Co-operative Planning Layer - Action between cells Local Planning Layer - Act on changing QoS within cell Case - Based Reasoning Reactive Layer Assignment OK OK CAC Exception handle No Yes No Yes Assigned request Modified request Set up connection Reject Request pol i cy pol i cy pol i cy S t at us re p o rti ng 253 CAC (Connection Admission Control) but it does this according to policies assigned by the planning layer. The connection request (containing information about the service provider, QoS, type of connection) is first considered for assignment to a Node B using an algorithm or set of rules passed down from the planning layer. As a result, the system performance can be monitored at all times. Any congestion occurring can be detected and reported to the planning layer which, will then find the best solution using the CBR approach in order to maintain the SLA. ASSIGNMENT AND ADMISSION SCHEME Assignment and admission control together determine which base station will have power control over a mobile, which means that base station must have available bandwidth to support the new call, and also must make sure that none of the existing connection will be dropped. A great deal of work has been done in this area. In [7], a comparison is made between a transmitted power-based call admission control (TPCAC) that protects the ongoing calls and a received power-based call admission control (RPCAC) that blocks new calls when the total received power at a base station exceeds a threshold. The result shows that the RPCAC scheme is found to offer significant performance benefits. In [8], the number-based CAC and interference-based CAC are compared. SIR-based CAC (signal-to-interference based CAC) has been proposed in [9], the benefit being in improving the system performance at traffic hot-spots . In this paper, a combination between the ideal scheme and SIR-based CAC has been chosen with uplink capacity limitation (which means the signal-to-interference of the received signal from mobile to base station is calculated) As for the ideal scheme, the system has to make sure none of the existing connections will be dropped when accepting a new connection request. Hence, two perfect power control loops are run to verify that the new request can really be accepted; otherwise it would be blocked or put into the buffer. The admission process is as follow : - With the new connection request, the new mobile's transmitted power is estimated in order to get the target SIR. (the open power control in section 5.4) - If the estimated transmitted power is in the accepted range, it means the new mobile can make a connection. Otherwise, it will be blocked or held in the buffer. - Set up the new connection and perform the first perfect power control loop. With this, the new transmitted power that is supposed to give each connection the target SIR can be determined. - The second perfect power control loop is performed to achieve the actual SIR for each connection as a result of accepting a new connection request. - If any existing connection would be dropped (by having SIR less than the threshold), the new connection is still rejected otherwise it is accepted and the connection can be made. The rejected connection request will be put into a queue until the next calculation or new call arrival and it will be blocked at the expiry of a timer: setting the timer to zero means that a request is immediately accepted or rejected. Furthermore, the base station serving the mobile can be reassigned at anytime during the connection if the current base station cannot provide the required link quality. SIMULATION MODEL The simulation model has been implemented in MatLab. The system used for the results in this paper consists of 9 hexagonal cells (25 cells have been used for other work but the large model suffers from an excessively long run time) and each cell has its own base station with an omni-directional antenna placed at the centre of the cell. A number of mobiles have been generated randomly according to the input traffic. When considering different classes of user, it is quite common to use three classes: bronze, silver, and gold. In the results described here, 50% of the users are bronze, 30% are silver and 20% are gold. It is assumed that the gold customers will pay the highest service charge followed by silver and bronze customers, so that the gold customer is paying for the best service and more flexibility than the others. 5.1 Radio Propagation Model In cellular systems, radio propagation is crucially influenced by the path loss according to the distance, log-normal shadowing, and multipath fading. The relationship between the transmitted power and received power can be expressed as [9]. 0 10 / 10 ) ( P r r P = (1) where P(r) is the received power, P0 is transmitted power, r is the distance from the base station to mobile, in decibels has a normal distribution with zero mean and standard deviation of (typical value of 8 dB), and 2 represents the gain (typical values of in a cellular environment are 2.7-4.0). 5.2 Traffic Model The model consists of two traffic types, voice and video. The model has been simplified from the three type traffic model that also included data traffic, which was used in [2]. The reason in simplifying the traffic model is because modelling data traffic to the level of packet results in unrealistically long simulation times. 5.2.1 Voice traffic Voice traffic is considered to be real-time traffic. The common model for a single voice source is illustrated by the ON-OFF process. It consists of two stages, active (ON) and silent (OFF) stage, with a transition rate from ON to OFF and from OFF to ON stage. Figure 3 illustrates the ON-OFF model. The silent period is assumed to be the period that cannot be used to transmit data message or voice call. Figure 3 Traffic model for voice call (ON-OFF model) Silent Active Duration in silent state : exponential distribution Duration in active state : exponential distribution 1 1 254 To simplify the simulation, the approach of [9] is used with an activity factor of 0.45 is used. The transmission rate for voice traffic is assumed to be 8kbit/s and mean holding time is 180second. 5.2.2 Video traffic Video traffic is also considered as real-time traffic. The common model for video source is illustrated by the discrete-state continuous time Markov process illustrated in figure 4. The bit rate of video traffic is quantized into finite discrete level (0 A 2A... MA). Transitions between levels occur with exponential transition rates that may depend on the current level. [11] Figure 4 Video source model (Discrete-state continuous time Markov process) The and are the state transitions and they are obtained by: + = M N 04458 . 5 1 9 . 3 (2) = 9 . 3 (3) where N is a Number of aggregated video sources (typical assumption 1) and M is a number of quantization levels (typical assumption 8). Implementing this video traffic model in the simulation causes simulation times to be very long. Many authors therefore simplify this by using activity factor. [9][12] Here, an activity factor of 1 has been assumed for the real-time video source as use in [9]. The transmission rate for video traffic is assumed to be 64, 144, or 384kbit/s and mean holding time is 300 second. 5.3 Receiver Model For the uplink capacity limited, the SIR of each transmission is calculated at the base station and it can be expressed as follows, (based on [13]) thermal er ra N I I R W SIR + + = int int Pr (4) where, (W/R) is the processing gain, Pr is the received signal strength, I intra is the sum of the received signal powers of other transmissions within the same cell, I inter is the sum of the received signal powers from the other cells, and N thermal is the thermal noise power. 5.4 Power Control Model Power control is the crucial part in the system since it is necessary to minimise the interference in the system by minimising the level of transmitted power to the optimum level, which means just enough to maintain the link quality. Power control in UMTS consists of three main functions: (i) open-loop power control, (ii) inner-loop power control, and (iii) outer-loop power control [14]. As the simulation focuses on the uplink-limited capacity, power control for the uplink is applied. In this work, the first two types are applied in the simulation since they have the major effect on the simulation result. Without outer-loop power control, the target SIR has to be fixed; here, it is assumed to be 6 dB and the threshold is 4 dB. [10] The power control step is assumed to be 1 dB at each power control cycle. [15][16] 5.4.1 Open-Loop Power Control Open-loop power control is applied when new connection requests arrive in the system as initial step of the admission process (section 4). The total interference at the base station is calculated as it is the parameter that User Equipment (UE) needs to use in the estimation of its initial transmit power. According to the parameter and the target SIR, UE estimates the transmit power and uses it as an initial transmit power. 5.4.2 Inner-Loop Power Control This is done periodically to allow the transmitted power of each connection to be kept as low as possible, yet maintain the target SIR. Firstly, the base station calculates the received SIR from the UE. If the SIR is less than the target SIR , the TPC (Transmit Power Control) command "up" is sent to the UE which increases the transmitted power by one step. If the SIR is more than the target SIR+1, the TPC command "down" is sent to the UE which decreases the transmitted power by one step. Otherwise, the UE maintains the same transmitted power. After the power control cycle has been performed, the new SIR for each mobile can be calculated. Any mobile that has an SIR less than the threshold will not be dropped immediately; instead the system will try to reallocate that mobile to another base station nearby that still has available bandwidth and can provide the link quality. If it is possible, the mobile will be handed over to the next base station, otherwise the mobile will be dropped. The transmitted power has a maximum of 21dBm; if the calculated transmit power is more than the maximum power, the maximum transmitted power will be applied. In this simulation, the inner-loop power control is performed every 10ms. Experiments have been done by varying the power control time step and the results show that the blocking rate becomes erratic as the timing gets too high. From the experiment, the time step of 10ms has been chosen as it is the highest value that gives consistent results. From the simulation point of view, it is preferable to use the highest time as this reduces the length of the simulation. The power control is essentially that used in 3G without including outer-loop power control. 5.5 Verification and Validation One of the most important aspects in developing the simulation model is its credibility; therefore the validation and verification of any simulation model are essential. The simulation model was validated by comparing the result with the relevant result from [8]: this is discussed in [1] . 0 MA (M-1)A (M-2)A A 2A 2 M M ) 1 ( M 2 ) 1 ( M 255 5.6 CBR Model According to [17], there are several proposed schemes of organizational structure and retrieval algorithms for CBR. In this work, the hierarchical memory with parallel search is used as it provides an efficient retrieval that is less time consuming, as the matching and retrieving happen in one step, which also give less complexity. The monitoring process of the system performed every 10 seconds. This means the monitoring parameters will be collected for 10 seconds and sent to the local planning layer of an agent where the CBR model is located as shown in figure 1. The parameters will then be compared with the SLA requirements and any deviation from the SLA can be reported. The CBR model will then be used to find the best solution for the situation. Base on the process model in figure 1, a solution will be proposed, or where the best matched case cannot be found or the evaluating process fails, a calculation might be used instead in order to find the solution according to certain rules. As the parallel search has been chosen for the CBR model, the whole library will be searched for each characterizing index in one step. If the new case is to be retained in the library, the library indexes have to be re-sorted according to the priority of the characterizing index of the new case. 5.7 Monitoring and case matching process As explained above, the monitoring process is done every 10 seconds. The call blocking, call dropping and the accumulative value of blocking rate are calculated and by comparing them with the SLA requirements, the error can be detected. If the error being reported is significant, the CBR model will be called. There are seven characterising indexes used to describe the case at the moment. Currently, they are obtained by matching the actual monitoring factors into a suitable range where the value belongs. Therefore, the characterising indexes will be in form of small integer numbers. The seven monitoring factors are as follows-: Total throughput for the whole system Offered traffic for the whole system Offered traffic for silver customer for the whole system Offered traffic for gold customer for the whole system Cell identity where offered traffic exceeds limit Accumulative blocking rate for silver class Accumulative blocking rate for gold class In case that there is not an exact match, there needs be a way to identify whether the closest match is acceptable for the situation. However, in the current model, only the best match will be chosen. In future work, the acceptable level for each case will be determined by the distance to the seven-dimensional coordinates defining the individual point of the case. If it is within the tolerable range, the case will be used. TRAFFIC PATTERN RECOGNITION AND NUMERICAL RESULTS The previous work [1] was done to support the basic idea that the reactive layer of the agent system can be controlled by the planning layer in order to ensure system compliance with the SLA. Here, the SLA assumptions made for the maximum acceptable level of call blocking rate. - Maximum acceptable call blocking rate for gold : 0.03 - Maximum acceptable call blocking rate for silver : 0.05 The gold customer pays the highest rate for the least elasticity of service level. These rates can either be instantaneous values or measured over a period of time; naturally the numerical limits would be different. In [1], the random overload situation has been tested with the traffic load being increased after the system reached stability. The call dropping rate is acceptable before high traffic load was applied, but after changing the load, the call dropping rate increases, then slowly declines to about the same level as before, because of the implementation of ideal assignment and admission control. On the other hand, the call blocking rate increases as a greater number of mobiles attempt to get into the system and the system tries not to drop any existing connection, so more will be blocked. The call buffering time for all classes of customer and all types of service has been set to zero to give immediate accept or reject decisions. Figure 5 shows the comparison between the result from the conventional system that does not chance the policy with the dashed lines and the result of policy chance as the solid lines. Without the SLA based control, the call blocking rate for all customer classes raises as the traffic load increases. For the SLA based control, the implementation here uses a buffer mechanism so the a call request that cannot be served immediately is held for a short time in case resources become available. The buffering time is configurable. From solid lines at the point when the level of call blocking rate for gold customer reaches the maximum level, policy 2 is applied which allocates a short buffering time to call requests from gold and silver customers, with that for bronze customers still being set to zero. The result shows that the call blocking rate of gold and silver customers stabilises, but does not go below the limit set for gold. After waiting for a short period (here 2 minutes) to ensure the trend is stable, a further change in policy is applied; this gives longer buffering time for gold customers and slightly longer buffering time for silver customers, so increasing still further the probability of gold and silver customers (especially gold) being accepted at the expense of bronze. Figure 5 Comparison between the result from the conventional system and SLA based control system Silver maximum level Gold maximum level Mean interarrival 100ms Mean interarrival 25 ms. Policy 1 Policy 2 Policy 3 Bronze Silver Gold With SLA control Without SLA control Gold Bronze Silver 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 18000 90000 2E+05 2E+05 3E+05 4E+05 5E+05 5E+05 6E+05 Simulation time (ms.) Rate of call blocking over total avtice call 256 A simple case library has been generated partly from this previous work and the knowledge from the work under current study. By using the simulation model mentioned before, a few traffic patterns can be implemented to test the system performance. The two main environments being tested here are the random overload situation and the hot spot situation. As the system detects the congestion, the CBR model is called to analyse the situation and the simulation results will be divided in two sections. 6.1 Random overload case For this case, the simulation repeated the previous work explained before by adding the CBR model and also use the simulation model illustrated in section 5. (In previous work [1] the less detailed simulation model has been used.) Figure 6 shows the simulation results of the call blocking across the simulation time as the traffic load increases in a conventional system that does not change policy. The call buffering time for all classes of customers and all types of services has been set to zero to give immediate accept or reject decisions. Figure 6 The simulation result from conventional system for the random overload situation Figure 7, 8 and 9 show the effect of using the CBR approach to identify the current traffic pattern and manage the reactive layer policies accordingly. It might be thought that these results are simply the normal result of applying priorities, but the technique is more powerful. In many SLAs, it is not short-term violations that are important: an SLA might specify for instance that the blocking rate must not exceed a certain value during a day or a month. The new policy has been applied to the reactive layer as soon as the system recognises congestion, in this example using the accumulative error rate over a period of 10s. The implementation here again uses a buffer mechanism to give short buffering time to call request that cannot be served immediately, especially for the higher priority customer. The buffering time is configurable. It can be seen from the result that CBR keeps the call blocking rate for gold and/or silver customers within the SLA bounds, according to the congestion pattern. In figure 7, the traffic reaches overload when the accumulative call blocking rate for gold exceeds the limit, at that point silver is still in an acceptable range. In this case the chosen policy gives the highest buffering time to gold and lower value for silver with that for bronze still at zero. Figure 7 Simulation results showing the effect of SLA-based control by CBR approach for the first random overload case It can be seen that the system detects the overload situation at the point where the traffic load increases and generates the appropriate policy. As the new policy gives priority to gold, the call blocking for gold customer is maintained within an acceptable range at the expense of both silver and bronze. Figure 8 shows the result from the second case, where the traffic is overloaded with the accumulative call blocking rate for gold and silver exceeding the maximum value. In this case, both silver and gold QoS need to be handled. By giving highest buffering time to silver and slightly lower for gold, the blocking for both can be kept within the range. As the buffer in this implementation uses the priority arrangement, gold customers are always in the top of the queue, so, in order to also give priority to silver customers, their buffering time has to be higher. Figure 8 Simulation results showing the effect of SLA-based control by CBR approach for the second random overload case Silver Gold maximum level Silver maximum level Normal traffic Overload traffic Gold Bronze Silver Gold maximum level Silver maximum level Normal traffic Overload traffic Gold Bronze Silver Gold maximum level Silver maximum level Normal traffic Overload traffic Gold Bronze 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 5 35 65 95 125 155 185 215 245 275 305 335 365 395 Simulation time (s.) Call blocking rate 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 5 35 65 95 125 155 185 215 245 275 305 335 365 395 Simulation time (s.) Call blocking rate 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 5 35 65 95 125 155 185 215 245 275 305 335 365 395 Simulation time (s.) Call blocking rate 257 In Figure 9 the situation is that the long-term value for gold customers has been met, but that for silver is at the limit. When congestion occurs then, silver customers have to be given priority in order that their long-term blocking is not exceeded, but gold customers can be allowed to have worse service since there is still "slack" in their SLA. Figure 9 Simulation results showing the effect of SLA-based control by CBR approach for the last random overload case The SLA monitoring here is looking at the long-term blocking, has detected that silver needs priority and has applied that priority. These results show the flexibility of the control system which assigns different policies to different scenarios and also shows that the highest priority can also be a sacrifice in order to maintain the customer who normally has the overall long-term values. In fact any SLA that can be evaluated numerically can be used as the basis for controlling the policy: the system is that flexible. 6.2 Hot spot case With hot spots, the monitoring process is able to identify the congestion from the individual blocking and dropping parameters of each cell. CBR model will then match the pattern with the cases in the library. The proposed mechanism here can be seen in figure 10. The bronze and silver users near the boundary will be transferred to the neighbouring cells that have normal traffic: effectively then by controlling the cell size in a more comprehensive manner than simple cell breathing from power control. By doing this, some of the capacity will be released for the hot spot cell in order to maintain the users nearer to the centre and high priority users. Figure 10 Hot spot situation and the proposed solution In the initial work for this hot spot case, the transferring process or handover will be done every 10s, which is the frequency of the monitoring process. An example of results from the initial work in this case are shown in figure 11 and 12. In figure 11, after the traffic load has increased in the hot spot cell, the call blocking rate for the hot spot call rise up while the other cells still have low blocking rate as the traffic was controlled within normal level. Figure 11 Result from the conventional system for the hot spot situation Gold Silver Bronze Silver Gold maximum level Silver maximum level Normal traffic Overload traffic Gold Bronze Silver Gold maximum level Silver maximum level Normal traffic Overload traffic Gold Bronze Hot spot cell Normal traffic cell 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 5 35 65 95 125 155 185 215 245 275 305 335 365 395 Simulation model (s.) Call blocking rat e 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 5 35 65 95 125 155 185 215 245 275 305 335 365 395 Simulation Time (s.) Call block rate 258 Figure 12 shows the result of using the CBR model which instructs the system to perform the handover for bronze and silver users near boundary to the neighbouring cells every 10s. The blocking rate for the hot spot cell still increases after the traffic load has increased but by comparing with the result in figure 11, the call blocking rate is lower. Further work is being done on evaluating more complex scenarios. Figure 12 Result from SLA based control system with CBR model for the hot spot situation CONCLUSIONS This paper has introduced the concept of combining CBR with an intelligent agent layered architecture to manage SLAs in W-CDMA networks. The simulation results show that the CBR system has been able to detect congestion occurring and then apply the appropriate policy to manage the behaviour of the CAC to block those customers who, at that time, are perceived as less important to the operator. The scenarios illustrated are fairly simple but further work is evaluating the approach over a much more complex range of situations. REFERENCES [1] Chantaraskul, S. and Cuthbert, L.G. SLA Control for Congestion Management in 3G Networks, in Proceeding of the IASTED International Conference on Wireless and Optical Communications (WOC2003), Banff, Alberta, Canada, 2003, pp. 447-452. [2] Chantaraskul, S. and Cuthbert, L.G. Introducing Case-Based Reasoning in SLA Control for Congestion Management in 3G Networks, in Proceeding of the IEEE Wireless Communications and Networking Conference 2004 (IEEE WCNC2004), Atlanta, Georgia, USA, 2004. [3] Cuthbert, L.G., Ryan, D., Tokarchuk, L., Bigham, J., Bodanese, E. Using intelligent agents to manage resource in 3G Networks, Journal of IBTE, vol. 2 part 4, Oct.-Dec. 2001, pp. 1-6. [4] Admodt A. and Plaza E., Case-Based Reasoning: Foundational Issues, Methodological Variations, and System, AI Communications, The European Journal of Artificial Intelligence, vol. 7:1, pp. 39-59. 1994. [5] Caulier, P. and Houriez, B. A Case-Based Reasoning Approach in Network Traffic Control, in Proceeding of the IEEE International Conference on Systems, Man and Cybernetics, 1995. Intelligent Systems for the 21st Century, Volume 2, 1995, pp. 1430-1435. [6] Hassanein, H., Al-Monayyes, A. and Al-Zubi,M. Improving Call Admission Control in ATM Networks Using Case-Based Reasoning, in Proceeding of the IEEE International Conference on Performance, Computing, and Communications, 2001, pp. 120-127. [7] Huang, C. Y. and Yates, R. D. Call Admission in Power Controlled CDMA Systems, in Proceeding of the IEEE Vehicular Technology Conference, 1996, pp.227-231. [8] Capone, A., Redana S. Call Admission Control Techniques for UMTS, in Proceeding of the IEEE Vehicular Technology Conference, 2001, pp.925-929. [9] Liu, Z. and Zarki, M. E. SIR Based Call Admission Control for DS-CDMA Cellular System, IEEE Journal on Selected Areas in Communications, vol. 12, issue 4, May 1994, pp. 638-644. [10] Kuri, J. and Mermelstein, P. Call Admission on the Uplink of a CDMA System based on Total Received Power, in Proceeding of the IEEE International Conference on Communications, vol. 3, 1999, pp. 1431-1436. [11] So, J.W. and Cho, D.H. Access Control of Data in Integrated Voice/Data/Video CDMA Systems, in Proceeding of the VTC Spring 2002, IEEE 55 th vol. 3, 2002, p.1512-1516. [12] Angelou, E.S., Koutsokeras, N.Th, Kanatas, A.G. and Constantinou, Ph. SIR-Based Uplink Terrestrial Call Admission Control Scheme with Handoff for Mixed Traffic W-CDMA Networks , in Proceeding of the 4 th International Workshop on Mobile and Wireless Communications Network, 2002, pp. 83-87, 2002. [13] Radio Frequency (RF) system scenarios, 3GPP TR 253942, Avialable: htpp://www.sgpp.org [14] Laiho, J., Wacker, A. and Novosad, T. Radio Network Planning and Optimisation for UMTS, John Wiley & Sons, Ltd., 2002. [15] Baker, M.P.J., Moulsley, T.J. Power Control in UMTS Release '99, 3G Mobile Communication Technologies, 2000. First International Conference on (IEE Conf. Publ. No.471),27-29, March 2000, pp.3640 . [16] Thong, W.S., Bigham, J. Hierachical Managament of CDMA Network Resources, in Proceeding of the Third International Conference on 3G Mobile Communication Technologies, 2002 (Conf. Publ. No. 489), 8-10 May 2002 pp. 216 220. [17] Kolodner, J. Case-Based Reasoning, Morgan Kaufmann Publishers, Inc., 1993, pp. 289-320. Silver Gold maximum level Silver maximum level Normal traffic Overload traffic Gold Bronze Hot spot cell Normal traffic 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 5 35 65 95 125 155 185 215 245 275 305 335 365 395 Simulation Time (s.) Call block rate 259

Service Level Agreement;Intelligent agent and Case-based reasoning;3G Resource Management

209

Using Roles and Business Objects to Model and Understand Business Processes

Business process modeling focus on describing how activities interact with other business objects while sustaining the organization's strategy. Business objects are object-oriented representations of organizational concepts, such as resources and actors, which collaborate with one another in order to achieve business goals. These objects exhibit different behavior according to each specific collaboration context. This means the perception of a business object depends on its collaborations with other objects. Business process modeling techniques do not clearly separate the multiple collaborative aspects of a business object from its internal aspects, making it difficult to understand objects which are used in different contexts, thus hindering reuse. To cope with such issues, this paper proposes using role modeling as a separation of concerns mechanism to increase the understandability and reusability of business process models. The approach divides a business process model into a business object model and a role model. The business object models deals with specifying the structure and intrinsic behavior of business objects while the role model specifies its collaborative aspects.

INTRODUCTION Representing and keeping the alignment between the multiple elements of an organization is fundamental to understand how it operates and how it can be adapted to cope with a changing business environment [5]. This requires understanding how business activities interact and are aligned with other organizational elements while supporting the operation of the business. In the past years, significant work, particularly in the area of business process modeling has been proposed, ranging from general modeling concepts to business automation languages [10, 16, 17, 18]. Business process modeling can be used for multiple purposes , such as facilitating human understanding and communication [29], supporting process improvement and re-engineering through business process analysis and simulation [8, 17] and automating the execution of business processes [1, 22]. A business process model captures the relationships that are meaningful to the business between different organizational concepts , such as activities, the resources used by activities and the human or automated actors who perform these activities. Identifying the properties and relationships of these concepts is fundamental to help understanding and evolving the business since it facilitates the communication between stakeholders, business specialists and support system specialists. We model business concepts as classes of business objects in a consistent object-oriented glossary of business concepts from where objects can be composed, specialized and reused. However, fully characterizing the type of a business object, its properties and relationships is not straightforward. This results from a business object generally being used in different contexts and relating to several other business objects in the organization. For example, a business object modeling a Product may be brought into play in several processes, such as Manufacturing , Logistics and Selling . In each of these contexts, it relates with different activities and resources, displaying different and possibly overlapping properties and behavior that are context-dependent. This means the object acts as a multi-dimensional concept. If business objects are modeled as one-dimensional concepts, i.e. without its properties and behavior being described as dependent on the context, then the objects will not have explicit information on how to guide the design of a business support system that is able to cope with evolution. For example, if the Manufacturing process changes, there may be changes to the Product object. However, if the Product object does not explicitly represent the aspects related to its manufacture, then there will be no information on the properties requiring modifications. This paper focus on describing on how to break up the universe of process modeling and its business objects into different aspects or areas of concern, of which each can then be handled independ-ently and later composed to synthesize a complete model. To do so, we propose defining two complementary conceptual models, a role model and a business object model. The role model describes business object collaborations and the properties of business objects that are concerned with each role, being each role a type on its own. The business object model describes the structure and the properties of a business object that are independent of a specific context. The relationships between business objects are specified by roles the objects play while collaborating. We argue that using roles and business objects to model business processes improves the understandability of the individual business objects and of the process model. It also improves model reengineering since it promotes reuse and makes explicit the dependencies between the model elements. The remainder of this paper is structured as follows: next section reviews some of the research on business process modeling. Section 3 reviews role modeling, describes how roles can be identified and defines the concepts of business objects and role. Section 4 presents how the business object and the role model can describe a business process, followed by an example of application in section 5. Finally, section 6 sets out the conclusions and future work. MODELING BUSINESS PROCESSES The Workflow Reference Model [31] defines a business process as a set of one or more connected activities, which collectively realize a business objective or policy goal, normally within the context of an organizational structure defining functional responsibilities and relationships. This definition extends the definition proposed by Davenport and Short [7] stating that a business process is a set of logically related tasks performed to achieve a defined business outcome. Most approaches to business process modeling concentrate on some sort of process map or diagram, which shows how activities are scheduled in the course of a business process. Indeed, there is little disagreement about the key elements process diagrams. There are usually ways to represent decision points and to express various activity coordination patterns , such as sequential flow, branching and parallel execution. Some techniques introduce swim-lanes to indicate the responsibilities of participants, such as departments or individuals. This allows representing the activities performed by actors in the context of a process. Two representative coordination-oriented business process modeling techniques that make use of actors, activities and swim-lanes are Role Interaction Networks [24] and Role Activity Diagrams [21]. Role Activity Diagrams provide the means to identify roles and interactions. Roles organize a process' activities into sets of operations associated with a given participant in the process. Interactions show the dependencies between those participants. While this approach improves the understandability of a process model since it depicts what a participant does in a process, it falls short to explain the behavior of the business objects in a specific context of interaction. Additionally, roles are defined as groups of activities and not as types, so they cannot be explicitly composed or specialized . Business process modeling is not limited to process diagrams. The focus of this paper is not on process diagrams but on describing the roles that are used to specify the responsibilities of business objects. A business object is the model of a concept in the business universe of discourse. It plays roles in a business process by means of participating in different activities. Business objects participate in different business processes in different contexts, thus playing multiple roles. It is important to note that process diagrams do not fully describe the business object structure and relationships, and do not emphasize why activities are performed or roles are enacted . Besides, they only identify actor roles, i.e. the roles of the performer of an activity. This means, for example, that the properties of a resource that is used by multiple activities are not separated according to its usage context. The next section introduces the fundamental concepts behind role theory and role modeling. ROLE MODELING In the late 1920s, role theory started to generated interest among social scientists from many backgrounds, such as psychology and sociology. Its central concern has been with patterns of human conduct, context and social structure as well as with individual response. The motivation for roles is to allow particular viewpoints regarding the factors presumed to be influential in governing behavior . It lies on a theatrical analogy of actors playing parts or roles in a play. As Biddle and Thomas [4] have stated: "When actors portray a character in a play, their performance is determined by the script, the director's instructions, the performances of fellow actors, and reactions of the audience as well as by the acting talents of the players. Apart from differences between actors in the interpretation of their parts, the performance of each actor is pro-grammed by all of these external factors; consequently, there are significant similarities in the performances of actors taking the same part, no matter who the actual actors are." There are many complementary definitions for the concept of role but still there is no consensus on the properties to represent it. In the late 1970s, sociological role theorists defined a role as "a comprehensive pattern for behavior and attitude" [26] or as "be-havioral repertoire characteristic of a person or a position" [3]. Nonetheless, the concept of role is used in computer science and software engineering as a modeling technique that deals with separation of concerns, i.e. the separation of the behavioral repertoire characteristics of some concept. It is used in methodologies such as RM-ODP [14] and in several object-oriented frameworks [10, 12, 14, 15, 25]. 3.1 Business Objects and Roles Modeling is an abstraction technique that consists of identifying concepts of interest in some universe of discourse and representing its essential features for a specific purpose in a model. In business modeling, the universe of discourse corresponds to what is per-ceived of an organization as being reality by business domain experts . Ontologies typically distinguish entities (nouns) from activities (verbs). Entities are things that exist in the business, either concrete (e.g. a person) or abstract (e.g. an organization). Activities are things that happen in the business. Activities make use of the business entities. We model both of these concepts as business objects. A business object is then the super type of all objects that represent business concepts with a well-defined boundary and identity. It encapsulates the definition, attributes, behavior and relationships to other business objects [20]. The state of a business object is characterized by the values of its attributes. The behavior is given by the actions that the business object is capable of performing to fulfill its purpose, including changing its intrinsic attributes and collaborating with other business objects. Business objects have intrinsic and extrinsic features. Intrinsic features describe it in isolation, while extrinsic features arise from the relationships or collaborations with other business objects. For example, a Person has intrinsic features such as Age and Sex , and extrinsic features such as Job Position and Salary, which derive from a transitory relationship between the Person and some Organization or Company . Intrinsic features may change over time (e.g. Age ) but always characterize the object. However, extrinsic features may become inappropriate (e.g. the Job Position property is not relevant when characterizing an unemployed person). One way to separate the intrinsic features from the extrinsic features of an object is by means of roles [4, 15, 23]. Roles, as a modeling construct, aim at separating the concerns that arise from 1309 business object collaborations. We define a role as the observable behavioral of a business object defined in a specific collaboration context. Thus, a role represents the extrinsic features of a business object when it collaborates with other business objects. 3.2 Identifying Roles To distinguish roles from entities, Guarino et al. proposed two criteria [11]. A role is a type that (1) is founded and (2) lacks semantic rigidity. Something is founded if it is defined in terms of relationships with other things in a given context. For instance, the concept of Reader is founded since for a Person to be a Reader there must be something being read. Conversely, a Person is not founded for the reason that its intrinsic properties are defined on their own regardless of the collaborations with other things. Something is semantically rigid if its identity directly depends on being kind of some class. A Book is semantically rigid since its identity is still that of a Book regardless someone is reading it or not. In contrast, Reader is not rigid because an entity filling the role of Reader retains its identity outside the context of that role. For example, a Person is a Reader while reading a Book , but when it stops reading it, it is still a Person . Therefore, roles are founded, semantically non-rigid types while entities are non-founded, semantically rigid types. ROLE-BASED PROCESS MODELING The proposed approach deals with decomposing the business process modeling universe into two complementary models, the business object model and the role model, and later binding these two models into an integrated specification of the business process . The business object model deals with the structure and intrinsic properties of business objects. Here, a process is modeled as a network of business objects. However, business objects relate to other business objects in specific contexts and are often used in more than one context, where they may play different roles. So, the roles for a business object only need to be included in its definition when the object acts in the collaboration contexts described by the roles. It is also impossible to forecast all of the possible roles of a business object. Thus, adding superfluous roles to the object impairs several design quality attributes such as understandability , maintainability and reusability. To deal with such a concern , roles and business objects should be dealt with separately and later bound together. The concept of role allows a system to be decomposed into a set of business objects capable of clearly separating core parts and collaboration-dependent parts and then to abstract and compose such objects. Consequently, a set of roles helps business objects to be defined to be more reusable and extensible. Roles may also be reused as an independent unit encapsulating specific collaborations . Roles are organized into role models, which deal with specifying the network of related roles required for a collaboration to happen. We propose defining and represented both of these models using the Unified Modeling Language [19] since its graphical syntax and semantics is well-know by software specialists and, although to a lesser scale, by business specialists. However, the standard UML does not have explicit constructs to represent the required business domain concepts. We make use of the UML extensibility package to define such concepts. The extensibility package specifies how UML model elements can be extended and customized with new graphical representations and new semantics by specifying stereotypes , tagged values and constraints. A coherent set of such extensions defined for a specific purpose makes up a UML profile [2, 19]. The next subsections describe how the business object models and role models are represented. 4.1 The Business Object Model The business object model specifies the structure and intrinsic properties of business objects. Business objects are coordinated towards the achievement of goals that describe why actions occur . A business process describes how objects are coordinated. Figure 1. Classes in the business object model profile. Figure 1 is a class diagram describing the UML stereotypes (classes in white) that are used in the business object model. A Business Object is a UML Class and it is specialized as a noun or verb by means of the Entity and Activity class stereotypes. Business object models are represented as UML class diagrams and the intrinsic behavior of its objects is represented using UML's behavioral diagrams. Note that collaborations between business objects are not represented in this model but in the role model. The stereotypes within the business object model can be summa-rized as follows: Business Object : an abstraction of a concept of interest in the organization. It is a UML Class . Activity : a specialization of Business Object . It is a verb describing how a piece of work is performed. Activities are performed by Actors , and operate over Business Objects , especially those acting as Resources . Entity : a specialization of Business Object . It is a noun describing a concrete or abstract business concept. Resource : a specialization of Entity , which is the input or output of an Activity . It represents things such as materials or information . Actor : a specialization of Entity . It is someone (a human actor) or something (an automated actor, such as an information system or a production machine) that can perform the actions required by an Activity . Goal : a specialization of Entity that represents a measurable state that the organization intends to achieve. Goals are achieved by Business Objects , especially Activities . A business process is composed of Activities that use input Resources , such as materials or information, to produce output Resources . Nevertheless, the input of an Activity may be any other Business Object or a composition of Business Objects . For instance, changing or reengineering a business process is in itself a process. This process takes as input a business object model (i.e. a network of relationships between business objects) and produces a modified model. Therefore, the composed business object model is being used as a resource in this context. Activities are performed to achieve specific business Goals . Analyzing Goals and their relationships with the Activities produces an alignment measure between the processes and the organization's operational strategy. The Activities of a business process are not 1310 autonomous in the sense they require one or more Actor or Business Support Systems to perform them. Actors represent people, systems (mechanical or computer based) or a combination of both. At a large scale, business processes are aggregated into value chains (which are also business processes) that produce a measurable value that is visible to external customers. Figure 2. Example of activity composition and specialization. Business objects are classes conforming to a type. They can be specialized and composed just like ordinary objects. Figure 2 shows an example of a class diagram depicting composition and specialization. Each chevron icon represents an activity or process as previously defined. The Sell Product activity is composed by a set of sub-activities such as Identify Customers and Handle Order . These activities can be further decomposed into actions that are more refined. The activity Sell Product is specialized as Sell by Mail Order and Sell Online . Note that composition and specialization do not imply any collaboration constraints between the activities. 4.2 The Role Model Roles are a separation of concerns mechanism that allows business objects to be observed from different perspectives. Role models identify roles as types and describe the network of roles required for a specific collaboration to happen. As a player of a collaboration, a role defines the set of extrinsic properties and behavior necessary to realize its participating collaborations . Figure 3. Representation of a role model package (left). Pair or related roles (right). Role models are represented as UML packages with two compartments (v. Figure 3, left). The bottom compartment of the role model is a standard UML activity or interaction diagram describing how the roles are orchestrated. The top compartment of the package depicts the roles within the role model. Roles are represented by rounded rectangles, connected by a navigable collaboration relationship between the roles. The representation of a role always shows its name. Optionally, it also depicts in parenthesis the name of the role model to where the role belongs so that its scope is clearly defined (v. Figure 3, right). Figure 4 show an example of three role collaborations contained in two role models. The Tutorship role model defines a collaboration pattern between two roles, Tutor and Student . In this Course role model defines two pairs of collaborations: Participant / Taken Course and Lecturer / Given Course . Figure 4. Example of role collaborations. Roles are modeled as classes and represented in class diagrams Methods and attributes concerning the specific collaboration context can be specified in the class diagram. Roles can also be constrained . A constraint asserts conditions between the roles in a role model. It can be expressed informally or formally (e.g. in plain text or OCL). An example of a constraint is disallowing two roles to be played simultaneously by the same player, such as forbidding an object playing the role of Tutor and that of Student simultaneously and in the same context. Figure 5. Example of role specialization. Figure 5 is a class diagram that shows how the Teacher role is specialized as Tutor and Lecturer . Role specialization means that if a business object is able to play a child role, then it is also able to play the super role. We have not yet found the need to define abstract roles, i.e., a role that may only have its non-abstract speciali-zations instantiated. 4.3 Binding Roles to Objects Roles are bound to business objects pertaining to a given business object model. The binding is accomplished via the play relationship stereotype, which links a business object to a role. It means the business object is able to exhibit or play the behavior specified by the target role. Figure 6. Binding roles to business objects. Figure 6 shows a class diagram where the pairs of roles Tutor / Student , Lecturer / Given Course and Participant / Taken Course defined earlier in Figure 4 and Figure 5, are bound to two different business objects, Person and Course . The binding between objects and roles is depicted as a strong arrow. The light arrow represents the relationships between roles. The model also defines a constraint in the Tutorship role model. It asserts that the instances actually playing the Tutor and Student role must be distinct. In this example, it means the Person acting as a Tutor and the Person acting as Student must be different objects, as expected. 1311 EXAMPLE Figure 7 shows two base role models, Supply and Pay and a composed role model, Purchase . Each role is a class and has methods and attributes concerning the specific collaboration context (e.g. the Supplier role in the Supply role model has the inquire and order methods). The Purchase role model describes the collaborations between a Client and a Supplier , while the Pay role model, specifies Payer and Payee . Figure 7. Supply , Pay and Purchase role models. Purchase is composed of the role model Supply and role model Pay . The Purchase role model is a composition of the Supply and Pay role models. A purchase results from supplying a product and paying for it. Figure 8 shows the binding of the roles within the Purchase role model to a set of business objects. In the first case, a Retailer acts as the Client and Payer to a Producer who is a Supplier and a Payee to the Retailer . However, the Retailer also acts as a Supplier (and a Payee ) to a customer. Client (BankClient) play play play play play play Supplier (Purchase) Client (Purchase) Supplier (Purchase) Client (Purchase) business object Customer Payer (Purchase) Payee (Purchase) business object Bank B Banker (BankClient) Payer (Purchase) play Payee (Purchase) business object Bank A Banker (BankClient) Client (BankClient) play business object Producer business object Retailer Figure 8. Binding roles to business objects. CONCLUSIONS This paper has presented the fundamental concepts towards a conceptual object-oriented framework for role-based business process modeling. It relies on defining two distinct models. The business object model focus on describing the components of a business process (activities, goals, resources and actors) as business objects. This model depicts the type of each business object, its intrinsic behavior and properties but does not address the representation of the object's features that are related to its collaborations with other objects. The role model depicts the collaborative behavior between roles and the constraints that regulate them. Roles are bound to business objects in a specific business object model, thus defining their usage context. This model describes roles as types on their own that can be specialized and aggregated. Role reuse is possible whenever the semantics of the interaction pattern is the same, regardless of the interaction context. The proposed approach separates the specification of the intrinsic features of a business object from its extrinsic features, meaning that the properties and behavior that arise from the collaborations with other objects are separated from the properties concerning the object. This separation results in an increase of the understandability of the business process since each different aspect of the business object may be discussed, analyzed and dealt with separately. Additionally roles also contribute to keep the alignment between the multiple organizational levels where a business process is defined . When a business object specified at business level is mapped to a component at business process support systems level, roles provide information on how to design the component so that changes to other levels can be traced and managed. Since the collaborative aspects of a business object are specified outside the object as roles, changes to a business process only interfere with the roles which derive from the corresponding activities, leaving the intrinsic properties of the object and its remaining roles un-changed . This means that only the implementation of the concerned roles needs modifications. The same reasoning applies the other way around. When the implementation of a specific role or business object is changed due to technical modifications or to the evolution of the software, these changes can be traced up to the processes and goals depending on it. The value of using role modeling increases with the need of making explicit the patterns of interaction between business objects . This is the case of processes where its business objects relate to several other business objects. In this case, understanding and reengineering such a process is often difficult due to the number of dependencies between objects, which are not separated or organized according to the interaction context. This also makes difficult to abstract common behavior patterns so that the business process elements may be reused in other contexts. We are currently extending this framework to enhance the representation of the interaction between business objects and the corresponding business support systems. The goal is to analyze the gap between the existing human skills and information system services of an organization and the requirements imposed by the as-is and to-be business models so that the alignment between these two levels may be improved. REFERENCES 1. W.Aalst, K.Hee, Workflow Management, MIT Press, 2002. 2. S.Alhir, Unified Modeling Language Extension Mechanisms, Distributed Computing, 1998. 3. C. Bachman, M. Daya. The role concept in data models, Proceedings of the 3 rd International Conference on VLDB, 1977. 4. B. Biddle, E. Thomas, Role Theory, Concepts and Research, Kluwer Publishers, 1979. 1312 5. Y. Chan, Why Haven't We Mastered Alignment?: The Importance of the Informal Organization Structure, MISQ Executive, Vol.1, No.2, 2002. 6. B. Curtis, M. Kelner, J. Over, Process Modeling, Communications of the ACM, Vol. 35, No. 9, 1992. 7. T. Davenport, J. Short, The New Industrial Engineering: Information Technology and Business Process Redesign. Sloan Management Review, 1990. 8. H. Eertink, W. Janssen, P. Luttighuis, W. Teeuw, C. Vissers, A Business Process Design Language, World Congress on Formal Methods, Springer, 1999, pp. 76-95. 9. H. Eriksson, M. Penker, Business Modeling with UML, OMG Press, 2001. 10. G. Gottlob, M. Schrefl, B. Rck, Extending Object-Oriented Systems with Roles, ACM Transactions on Information Systems, Vol, 14, 1996 pp. 268-296. 11. N. Guarino, M. Carrara, and P. Giaretta. An ontology of meta-level categories. Proceedings of the Fourth International Conference on Knowledge Representation and Reasoning, pages 270280. Morgan Kaufmann, 1994. 12. T. Halpin, Augmenting UML with Fact-orientation, 34th Hawaii International Conference on System Sciences, IEEE Press, Hawaii, USA, 2001. 13. ISO, ISO/IEC 10746 ODP Reference Model, International Standards Organization, 1995. 14. E. Kendall, Agent Roles and Role Models, New Abstractions for Multiagent System Analysis and Design, International Workshop on Intelligent Agents, 1998. 15. B. Kristiansen, Object-Oriented Modeling with Roles, 1 st Conference on Object Information Systems, 1996. 16. M. Madhavji, The Process Cycle, Software Engineering Journal, Vol. 6, No. 5, 1991. 17. C. McGowan, L. Bohmer, Model-based business process improvement, 15thInternational Conference on Software Engineering, IEEE Computer Society Press, 1993. 18. D. Miers, Business Process Engineering, C-T Colin, Kogan Page, London, 1996. 19. OMG, Unified Modeling Language Specification, Version 1.5, formal/03-03-01, 2003. 20. OMG, Business Object Management Special Interest Group (BOMSIG) Glossary of Terms, 1995. 21. M. Ould, Business Processes, Modeling and Analysis for Reengineering and Improvement, John Wiley & Sons, 1995. 22. A. Scheer, ARIS Business Process Modeling, 2 nd edition, Springer, 1999. 23. T. Reenskaug et al., Working With Objects: The OOram Software Engineering Method. ManningPublication Co., 1996. 24. B. Singh, G. Rein. Role Interaction Nets (RINs): A Process Description Formalism, MCC, 1992 25. D. Taylor, Business Engineering with Object Technology, John Wiley & Sons, 1995. 26. R. Turner, Strategy for Developing an Integrated Role Theory. Humboldt Journal of Sociology and Religion 7: 123-139, 1979. 27. M. Uschold, M. King, S. Moralee, Y. Zorgios, The Enterprise Ontology, The Knowledge Engineering Review, Vol. 13, 1998. 28. E. Verharen, A Language-Action Perspective on the Design of Cooperative Information Agents, CIP-Gegevens Koninklijke Biibliotheek, 1997. 29. T. Walford, Business Process Implementation for IT Professionals and Managers, Arthech House, MA, 1999. 30. E. Yourdon, Modern Structured Analysis, Prentice-Hall, Englewood Cliffs, NJ, 1989. 31. Workflow Management Coalition, The Workflow Reference Model, 1995 1313

Business Object;Business Process Modeling;Role Modeling;Organizational Engineering

21

A Taxonomy of Ambient Information Systems: Four Patterns of Design

Researchers have explored the design of ambient information systems across a wide range of physical and screen-based media. This work has yielded rich examples of design approaches to the problem of presenting information about a user's world in a way that is not distracting, but is aesthetically pleasing, and tangible to varying degrees. Despite these successes, accumulating theoretical and craft knowledge has been stymied by the lack of a unified vocabulary to describe these systems and a consequent lack of a framework for understanding their design attributes. We argue that this area would significantly benefit from consensus about the design space of ambient information systems and the design attributes that define and distinguish existing approaches. We present a definition of ambient information systems and a taxonomy across four design dimensions: Information Capacity, Notification Level, Representational Fidelity, and Aesthetic Emphasis. Our analysis has uncovered four patterns of system design and points to unexplored regions of the design space, which may motivate future work in the field.

INTRODUCTION From the very first formulation of Ubiquitous Computing, the idea of a calmer and more environmentally integrated way of displaying information has held intuitive appeal. Weiser called this "calm computing" [35] and described the area through an elegant example: a small, tangible representation of information in the world, a dangling string that would wiggle based on network traffic. When information can be conveyed via calm changes in the environment, users are more able to focus on their primary work tasks while staying aware of non-critical information that affects them. Research in this sub-domain goes by various names including "ambient displays", "peripheral displays", and "notification systems". The breadth of the systems in these broad categories is quite large. We seek to disentangle the terminology used to describe and categorize the wide array of systems in order to provide a common language for discussing research therein. An ambient display can represent many types of data, from stock prices, to weather forecasts, to the presence or absence of colleagues. Maintaining awareness of co-located and distant work and social groups has been a long-term research thread in the area of Computer Supported Cooperative Work (CSCW) [5, 8]. The Tangible Media Group at the MIT Media Lab, directed by Ishii, also helped shape the field of ambient computation. They coined the term "tangible media," citing inspiration from Weiser's vision [35] and from Pederson and Sokoler's AROMA system [29] and developed AmbientROOM [17] and Ambient Fixtures [6, 18]. These systems use ambient displays to make people aware of both group activity and other information such as network traffic. Recent work in Ambient Intelligence has brought techniques from Artificial Intelligence to ambient systems, spearheaded by the Disappearing Computer initiative of the European Union [31]. This research thrust seeks to imbue ambient systems with contextual knowledge about the environment. The Roomware project has resulted in smart architectural spaces that support information conveyance (and group collaboration) [33]. Researchers have developed systems that use a multitude of everyday objects to display information. Examples include lights of various sorts [2, 17], sounds [25], shadows [8], artificial flowers [18], mobiles [24], and office-dcor water fountains [12, 16]. Further research has sought to use framed photographs [26] and larger artistic pictures to represent information from the world in an art-like manner [14, 30, 32]. There are also peripheral display "modes" of a user's main desktop, including screensavers like What's Happening [36], information bars and menus such as those leveraged in Sideshow and Irwin [6, 22], and alternate panes, like Apple's Dashboard [3]. As one can see, the design space is large. All these systems provide a rich history of system design principles, approaches, and decisions, but accumulating theoretical and craft knowledge has been stymied by the lack of a unified vocabulary to define and describe these systems. In this paper we propose a set of design choices that developers of ambient information systems must confront to build successful and compelling systems. First we set out a definition of an ambient information system that is a synthesis of the varied definitions given in published research. We hone the intuitive set of 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. AVI '06, May 23-26, 2006, Venezia, Italy. Copyright 2006 ACM 1-59593-353-0/06/0005. $5.00. 67 characteristics that distinguish ambient systems from other ubiquitous computing research systems. Next, we propose a set of design dimensions for ambient information systems. The four dimensions of system design elucidate the main decisions one confronts when designing an effective ambient system. Finally, we explore the clusters across dimensions to uncover four coherent combinations of system designs, which work as design patterns for the field. The results also identify new ways of combining the design attributes to explore new possibilities for ambient information systems. AMBIENT INFORMATION SYSTEMS Many different terms have been used to describe the types of systems we discuss in this paper. Three of the most commonly used terms are "ambient display," "peripheral display," and "notification system." But how does one differentiate these terms? Based on general understandings, we claim that: all ambient displays are peripheral displays, some notification systems are peripheral displays (some notification systems are not peripheral but are instead the object of focused work and attention) The words of researchers themselves likely best explain their conceptions of the systems that they have built. Below, we present germane definitional quotes. Ishii et al: "[In Ambient Displays] information is moved off the screen into the physical environment, manifesting itself as subtle changes in form, movement, sound, color, smell, temperature, or light. Ambient displays are well suited as a means to keep users aware of people or general states of large systems, like network traffic and weather." [17] Matthews et al: Peripheral displays, then, are displays that show information that a person is aware of, but not focused on. [24] Matthews et al: "Ambient displays might be defined as those that are &quot;minimally attended&quot; (e.g. just salient enough for conscious perception) while alerting displays are &quot;maximally divided&quot; (e.g. slightly less salient than focal tasks). [24] Stasko et al: Ambient displays typically communicate just one, or perhaps a few at the most, pieces of information and the aesthetics and visual appeal of the display is often paramount. Peripheral displays refer to systems that are out of a person's primary focus of attention and may communicate one or more pieces of information." [32] Mankoff et al: "Ambient displays are abstract and aesthetic peripheral displays portraying non-critical information on the periphery of a user's attention... They generally support monitoring of non-critical information." "Ambient displays have the ambitious goal of presenting information without distracting or burdening the user." [20] Rounding and Greenberg: "The [notification collage] is designed to present info[rmation] as lightweight and peripheral objects. It does not demand the full attention of its users: rather it can be attended to in passing, where people collaborate should the need or desire arise." [14] McCrickard et al: "Often implemented as ubiquitous systems or within a small portion of the traditional desktop, notification systems typically deliver information of interest in a parallel, multitasking approach, extraneous or supplemental to a user's attention priority." [21] McCrickard et al: Notification systems are defined as interfaces that are typically used in a divided-attention, multitasking situation, attempting to deliver current, valued information through a variety of platforms and modes in an efficient and effective manner [21]. The easiest way to explain the differences between systems is to look at the design motivations that informed them. Ambient displays are those that have pointed aesthetic goals and present a very small number of information elements. These systems are a proper subset of peripheral displays, which can appear either in the environment or on secondary or even primary computer displays. Notification systems' design motivation results from divided attention situations. As such, they can be equal to a primary work task in their attentional needs or be secondary. When notification systems are designed to be secondary to a primary task, the systems are appropriately defined as peripheral. In this paper, we propose the term ambient information system as the unit of study and define the behavioral characteristics of such as systems as follows: Display information that is important but not critical. Can move from the periphery to the focus of attention and back again. Focus on the tangible; representations in the environment. Provide subtle changes to reflect updates in information (should not be distracting). Are aesthetically pleasing and environmentally appropriate. PREVIOUS TAXONOMIES A small number of research papers that describe ambient information systems also include extended discussions of the design dimensions that motivate and contextualize their work. The authors provide dimensions to compare and contrast their systems to others in order to explain their design rationales. Matthews et al use the dimensions notification level, transition, and abstraction to characterize systems in this space [24]. They developed the Peripheral Display Toolkit [23] that helps people to develop ambient information displays more easily. Their concept of notification level means the relative importance of a particular data stream. Transitions are the programmatic changes to the display, based on the data. Transitions include fading, scrolling, or animation effects. They define abstraction as the mapping that takes a piece of numerical or ordinal data and turns it into something that the ambient display can use, something "more easily interpreted with less [user] attention." Matthews et al segregate notification level into five levels: Ignore, Change Blind, Make Aware, Interrupt, and Demand Attention. The gradations run from low, a system ignoring the change in the data, to high, a system demanding attention in a way that must also be explicitly dismissed. They propose categories of transition: interrupt, make aware, and change blind. Finally, they bifurcate abstraction into feature abstraction or degradation. McCrickard et al introduce a different set of three dimensions to classify notification systems: interruption, reaction, and comprehension [21]. Interruption is defined psychologically, similar to Matthews' notion, "as an event prompting transition and reallocation of attention focus from a [primary] task to the notification." Reaction is defined as the rapid response to a given stimulus, while comprehension is the long-term notion of remembering and sense-making. 68 McCrickard et al then plot the design space as a 3-tuple of interaction, reaction, and comprehension (IRC). Each dimension is assigned a rating of high (1) or low (0), creating models like 0-1-0. They label these models with meaningful names like "Ambient Media, 0-0-1" "Indicator, 0-1-0" and "Critical Activity Monitor, 1-1-1." Eight models serve as the corners of a design space. The resulting space, it should be noted, is larger than the design space of ambient information systems as we discuss in this paper because it contains games, secondary displays, and critical activity monitors (which by our definition, are notification systems that are not also peripheral systems). McCrickard also classifies a set of 14 extant systems in the design space on the three dimensions. Both of these taxonomies deal thoroughly with interruption and detail some of the criteria for categorizing systems along this design dimension. We extend this analysis to other dimensions of data representation, flexibility, and aesthetics. This more holistic view points out design trade-offs between aesthetic emphasis and and flexibility, and between a system's information display style and display capacity. Mankoff et al proposed a set of heuristics for evaluating ambient systems [20], which may also assist system builders. The heuristics attempt to give guidance for the formative evaluation of ambient systems, but they also can be viewed as high-level design guidelines, such as "The display should be designed to give `just enough' information. Too much information cramps the display, and too little makes the display less useful." DESIGN DIMENSIONS OF AMBIENT SYSTEMS Designers of ambient information systems make decisions about how much information to display, what specific aspects to depict, and how exactly to display it, transparently or abstractly, on a monitor or via a decorative sculpture. We present four design dimensions that capture the space of ambient information systems. The dimensions can be thought of as design choices or design questions that system builders must answer. The dimensions are: information capacity notification level representational fidelity aesthetic emphasis We rank 19 research systems and three consumer ambient information systems on each of the four axes. Each axis is divided into 5 bands, from low to high. We place systems into groups based on information from published conference and journal proceedings, including images and videos of systems in use if available. The 19 systems we chose are not intended to be an exhaustive list of all ambient information systems in the research literature. The 19 systems are representative of the breadth of the field and we feel that attempting an exhaustive list, while amplifying completeness, would not significantly alter the design dimensions. Research systems that we analyzed include: Bus Mobile [24], Dangling String [35], Digital Family Portrait [26], InfoCanvas [33], Informative Art [30], Information Percolator [16], Irwin [22], Kandinsky [11], Kiumra [19], Lumitouch [5], Notification Collage [14], Scope [34], Sideshow [7], Table Fountain [12], Water Lamp [8], and What's Happening [36]. We include three consumer systems that fit our definition of ambient information systems, Ambient Devices Ambient Orb [2], the My Yahoo! web portal [27] and Apple's Dashboard [3]. Figure 1 shows the four dimensions for our analysis, and each of the 19 systems placed into a group along each. Thin colored lines trace the rankings of systems on each axis, similar to a parallel coordinates plot. Each axis has values that range from low to high through five grades. The dimensions of notification level and representational fidelity have more descriptive axis labels that will be explained in detail below. 4.1 Information Capacity Ambient information systems are created to convey information to users--information that typically is important to a user's sense of wellbeing and general awareness, but not critical to their work or personal life. Information capacity represents the number of discrete information sources that a system can represent. Some systems are capable of displaying a single piece of data such as the current price of a stock index. Others can display the value of 20 (or more) different information elements on one screen. We rank systems from "Low" to "High" on this design dimension. Information elements are discrete information "nuggets". For example, if a system monitors campus shuttle buses, each bus is a single nugget. If the system can represent both the time to a location and a direction of travel, then there are two nuggets of information for each bus that is monitored. Information capacity makes visible the design trade-off between space and time. A designer can increase the information capacity of a display by increasing the space for information to be presented or by creating a display that transitions through a set of views over time. If a system is designed with multiple views or uses scrolling, we rank it in the top tier, since the number of pieces of information that it could display is arbitrarily large. A further caveat about information capacity is necessary. Some of the analyzed systems such as InfoCanvas, Sideshow, and Dashboard are user-configured and user-customizable. This means that these and other systems could potentially be made to display hundreds of elements. Instead of attempting to calculate a theoretical maximum throughput for the display in these cases, we use the system designer's naturalistic portrayal in their published work to determine the "everyday maximum." Each of these systems is also in the top tier of information capacity. The design dimension of information capacity has a barbell distribution. Five of the 19 systems display a single information element and are ranked "Low". Conversely, there are eight systems that display from ten to 20 information elements, with some systems having the potential to display more and these are ranked "High." Only a few systems take a middle-ground approach, attempting to display a small number (from two to ten) of information elements. The systems with low ratings on the attribute of information conveyance are those that are physical displays. Fountains, glowing lights, and office-decoration sculptures afford designers only so much flexibility for changes. 69 Figure 1: Parallel Coordinate plot of 19 existing ambient information systems across four design dimensions. Colored lines trace each system's ranking along the design dimensions. Different colors are used to denote groups of systems which are similar as explained more fully in Section 5. Since the number of changes possible is small, the total number of information nuggets that can be represented is correspondingly small. The systems with high information conveyance are those that are presented on LCD screens. The systems that run at full screen (instead of as a small section of a focused main monitor) are ranked the highest. 4.2 Notification Level Notification level is the degree to which system alerts are meant to interrupt a user. Notification level is a design attribute that is present in the two taxonomies of ambient and peripheral information systems we reviewed earlier. Matthews et al subdivides notification level into five categories: ignore, change blind, make aware, interrupt, and demand attention. For our analysis we adopt those categories but replace the lowest level of system alert function, ignore (a degenerate case) with user poll. Systems such as Apple Dashboard and My Yahoo! do not always appear in a user's environment and must be explicitly called to the fore. Notification level can be thought of as the "ambience" of the systems in question. Some systems in the ambient space are quiet, and afford opportunistic glances to the information, while others provide more strident alerts by blinking, flashing, beeping, or even opening dialog windows. Systems that provide unobtrusive change blind or make aware notifications to the user are at the core of the ambient information system design space. Systems that interrupt users with alarms or that demand attention (by launching system dialog windows) are not subtle, so are further from the core concept of ambient information systems, though, as Matthews et al argues, the smooth transition from more subtle to more jarring is an interesting design direction for ambient system designers. Notification level is the designer-intended level of alert. We do not take pains to distinguish between systems that are proven to be "change blind" through user experimentation versus those that merely claim change blindness. We remain agnostic here about the techniques used for ensuring subtlety including slow animation, scrolling, and fading (these implementation details are at a lower level of design rationale). Once the decision has been made to produce a system with change blind transitions, the designer must then produce system transitions that meet the goal in the specifics of the system. Our analysis focuses on the high level decision on the part of the designer or design team. The distribution of systems here shows a good fit to our definition of ambient information systems. It is apparent that most ambient information systems adhere to the central notion of subtle visual or representational changes. The vast majority of ambient information systems fall into the change blind and make aware transition categories (somewhat low and medium). Few systems are designed to interrupt users or demand attention. 70 Two that do however are Scope and Sideshow. Note that most systems that are physical displays do not have make-aware or interruption-level alerts, much less demand attention alerts. The Bus Mobile does enable make-aware transitions, when, for example, the last bus of the day approaches. 4.3 Representational Fidelity Representational fidelity describes a system's display components and how the data from the world is encoded into patterns, pictures, words, or sounds. Some systems reproduce the information being monitored in a very direct way, while others are much more abstract in their representation. Matthews et al's taxonomy characterizes this design choice as abstraction, but only distinguishes two sub-types, feature degradation and feature abstraction. We consider this design dimension to be rich and complex, so we will try to tease apart the many different types of abstraction that appear in ambient information systems. Representational fidelity can be described in the language of Semiotics, the branch of Philosophy that deals with signs, sign systems (such as natural languages) and their meanings. As such it has an accepted vocabulary for the elements of a symbolic representation. Semiotics can help analyze the way that particular signifiers--words, pictures, sounds, and other things--stand for the things they represent. A semiotic sign is made up of three parts [28]. The object is called the signified; it is the physical thing or idea that the sign stands for. The signifier is the representation of the object, which could be a word, a picture, or a a sound. The sense is the understanding that an observer gets from seeing or experiencing either the signified or its signifier. The signifier and the signified need not have any direct relationship. However, both the signified and the signifier create the same sense in the head of an observer; seeing a log aflame and seeing the word "fire" create the same meaning for a person. Ambient information systems, in the vocabulary of semiotics, contain one or more signs. Each sign has its object, information in the world, and its representation, the lights, pictures, or sounds used to signify that information. Many ambient information systems contain multiple signs--each picture element standing for a different piece of information. The theory of Semiotics also helps to explain the notion that some signs are transparent, easily understood, while others are metaphorical and still others are abstract. Signs can be symbolic, iconic, or indexical. Symbolic signs are those that are completely arbitrary. For example languages are arbitrary, for the word "bachelor" has no more natural relation to an unmarried man than does the word "foobar." Symbolic signs are those signs for which a code, or rule-following convention, is required to understand. Language characters and numbers are all symbolic, as are abstract visual representations (the color red standing for "danger"). Iconic signs are those signs that have an intermediate degree of transparency to the signified object. Iconic signs include metaphors as well as doodles, drawings, and caricatures. Icons represent their objects by having some similarity or resemblance to the object or to an essential aspects of the object. Indexical signs are those that are directly connected to the signified. Examples include measuring instruments, maps, and photographs. We have subdivided the three main categories of representational fidelity to distinguish between ambient information systems. We propose five groups, ranked from indexical (high) to symbolic (low): INDEXICAL: measuring instruments, maps, photographs ICONIC: drawings, doodles, caricatures ICONIC: Metaphors SYMBOLIC: language symbols (letters and numbers) SYMBOLIC: abstract symbols Some ambient information systems have displays that do not afford representational flexibility, because of the constraints of the display. For example, the LiveWire system and the Ambient Orb cannot represent language symbols, nor can they convey indexical forms like photographs. However, some flexibility is present. The systems might map information in an arbitrary way, remaining fully abstract (representing stock increases with the color green and losses with the color red), or it could map information more metaphorically, as would be the case if LiveWire were connected to information from a seismograph or ocean tides. As one can see, the question concerning representational flexibility requires one to consider both the display and the information that is displayed. The InfoCanvas is a very flexible system when considering representational fidelity. The InfoCanvas uses all five types of representational fidelity. It uses abstract symbols, such as the color red standing for traffic being stopped, metaphors, like a cartoon drawing of a cloud representing cloudy conditions, and also photographs and words of news stories, which are fully indexical. We show this ability for a system to straddle multiple representational forms by duplicating the system in each category and noting them with an asterisk (see Figure 1). Systems which are designed to represent information at multiple levels of fidelity are: Apple's Dashboard, InfoCanvas, Informative Art, Notification Collage, Sideshow, and What's Happening. In these cases, we draw the parallel coordinate plot to the top-most tier of representational fidelity for each system. The majority of systems however, only afford a single level of representational fidelity. Many of the sculptural displays only afford symbolic, that is abstract, representations, while a smaller number afford text and photographic representations. 4.4 Aesthetic Emphasis The final dimension concerns the relative importance of the aesthetics of the display. Some system designers seek to build displays and artifacts with sculptural or artistic conventions. For these systems, being visually pleasing is a primary objective. Others however place relatively little focus on aesthetics and typically focus more on information communication ability. Since aesthetic judgment is at its core a subjective phenomenon, we do not judge systems on their relative artistic merits. Instead we attempt to rank ambient information systems by our perception of the importance given to aesthetics. There is often a tradeoff made between communication capacity, representational fidelity, and aesthetics, a relationship that we explore in this section. Ambient information systems are intended to be visible; positioned on a shelf, hung on the wall, or placed as a small sculpture on a desk, the systems are seen not just by a user, but also by co-workers, colleagues, or family members. There are a 71 multitude of approaches when it comes to building aesthetically pleasing devices. One approach is to build systems that mirror existing artworks by a particular artist, as is the case in Kandinsky and Informative Art. A second approach is to design a display that is representative of a particular style or art movement. InfoCanvas, through its use of themes, allows the display to take on characteristics of Asian water-color paintings, for example. We rank systems on the design dimension of aesthetic emphasis as low, somewhat low, medium, somewhat high and high. Note again that we are not assessing the degree to which the systems are successful as art. We are providing a subjective measure of how much the system designers focused on aesthetics and how much they emphasized aesthetic considerations in their research and design decisions. Most systems that we analyzed had medium or somewhat high degrees of aesthetic emphasis (12 of 19). The decisions of designers to strive for visually pleasing displays is most clear in the cases where the display is intended to leverage the work of existing artists. The physical ambient information displays are often sculptural in their design decisions. They attempt to set themselves off from the rest of the environment, often on pedestals or stands. Their capability to display much information (information capacity) is often limited by their design clarity and austerity. We consider this design trade-off in the next section. Systems that we ranked at the middle of the spectrum of aesthetic emphasis are those which are not intended by their designers to be art worthy of contemplation as art objects. But they are explicitly intended to be viewed as calm pleasing objects and displays. Apple's Dashboard widgets have a clean design sense about them, as does Kimura, What's Happening and the Information Percolator. The systems that are ranked low on aesthetic emphasis are Scope, Sideshow, Bus Mobile, Elvin, and My Yahoo!. These systems put information conveyance at a higher priority than being aesthetically pleasing. They are still calm and environmentally appropriate, but their designers did not emphasize their aesthetic qualities. Cleary, some systems that are early-stage prototypes like Bus Mobile, may not have the aesthetic polish of more finished systems. FOUR DESIGN PATTERNS In this section, we introduce four design patterns for ambient information systems, after Alexander's pattern language for architectural studies [1]. The design patterns illustrate four coherent combinations of the four design dimensions previously presented. We have already pointed out trends and clusters that are present in each particular design dimension. However, there are fruitful conclusions for system designers as we consider the interaction between the design dimensions to form design patterns. Considering the clusters of systems in each dimension and the correspondences that are visible in the parallel coordinate plot, we find four main archetypes in existing ambient information system design: Symbolic Sculptural Display, Multiple-Information Consolidators, Information Monitor Display, and High Throughput Textual Display. Figure 2 shows the pattern of each archetype across the dimensions. Figure 2: a-d System design archetypes shown in the context of the design space. Heavy boxes indicate core design decisions, while light boxes show alternate choices. Symbolic Sculptural Displays are ambient information systems that display very few pieces of information, usually a single element. They represent information in an abstract sculptural way with light, water, or moving objects. They are intended to be decorative objects for a home or office setting and as such are highly aesthetic in their design (see Figure 2a). This design pattern is a core of ambient system design, and accounts for six of our analyzed systems: Ambient Orb, Dangling String, Digital Family Portrait, Information Percolator, Lumitouch, Table Fountain, and Water Lamp. The Digital Family Portrait combines multiple information sources and so truly represents more information than the other members of this type. 72 Multiple Information Consolidators are ambient systems that display many individual pieces of information in a consolidated manner. They are typically screen-based in order to convey much information and make users aware of changes to that information (usually by blinking the visual representation of a certain element). They are reasonably aesthetically motivated, but all clearly demonstrate the trade-off between aesthetics and customization and information capacity (see Figure 2b). Systems which illustrate this design pattern are: Kandinsky, Kimura, InfoCanvas, Notification Collage, and What's Happening. Kandinsky departs from the other systems in that it is explicitly modeled on the fine art of Kandinsky, and as such is highly stylized and design-focused. It does so at the expense of flexibility, since it can only display photographs in its slots. Information Monitor Displays are displays that are a peripheral part of a user's computer desktop. As such, they afford different interactions and design choices. They display multiple sources of information, and do so usually by visual metaphors. They are capable of notifying users in multiple ways about changes in the source data, including subtle awareness, interrupting, and even demanding user attention when necessary (i.e., requiring the user to switch focus to dismiss a notification). The systems achieve aesthetics, but their primary purpose is not good looks (see Figure 2c). Examples of this design archetype include: Scope, and Sideshow. High Throughput Textual Display systems are those that use text and very simple graphics (icons) to denote information. They are capable of representing voluminous information, but do not draw attention with interruption-level notifications. These systems are not primarily as concerned with aesthetics as they are with information conveyance (see Figure 2d). These systems are simple but efficient for certain types of tasks. Examples of this design archetype are: Elvin, and My Yahoo!. The four design archetypes cover nearly all of the analyzed systems, but do not cleanly categorize three systems. Apple's Dashboard system is most similar to a Multiple Information Consolidator. It fails being a pure example of this archetype because of its inability to alert users to changes in information it requires users poll the system by calling up the transparent pane via a hot key. The Bus Mobile is an early stage prototype, and as such is not concerned with aesthetics to a large degree. With a higher degree of aesthetic emphasis, it might be closer to a Information Monitor Display (albeit a physical instead of screen-based system). Informative Art is quite unlike the four design archetypes. Informative Art has high aesthetic emphasis, but low information capacity (e.g. 5 or 6 city's weather forecast information). It is metaphorical and abstract in its information mapping fidelity. EXTENDING THE PATTERNS The four patterns for system design can help designers to make appropriate choices as they develop new ambient information systems. The design patterns can be used as models so a designer can decide to build "an information monitor display for a home health awareness application", or "a set of symbolic sculptural displays for work-group collaboration". Further, the designer may be depart from the pattern, by building up a system's range of possible notification levels, or by choosing to trade aesthetics for increased information capacity. However, our analysis also points at what has not yet been explored. The four design patterns show four coherent combinations, but they are not the only possibilities for building useful ambient systems. Combined with longer-term trends in the fields of Ambient Intelligence and Ubiquitous Computing, new archetypes for system design are emerging. We note possibilities here, which change both the dimensions and the four design patterns. We do not expect the information capacity for ambient systems to increase by dramatically. Though scrolling or time-divided ambient systems (What's Happening, Elvin) can already display data elements numbering in the hundreds, simultaneous visual displays are usually limited to 25 or 30 elements by readability and user learnability. Ambient information systems will not turn into information visualization systems showing thousands of data points. However, contextual sets of information may be useful for ambient systems in specialized environments. Systems which display contextual sets of information like that of the Bus Mobile (all of the buses on a college campus) or Scope (email and calendar data) would increase the number of systems in the middle portion of this design dimension. We also expect to see changes to the design dimension of representational flexibility. Designers have begun to explore the affordances of abstract and symbolic mappings between information sources and their representations. We see this continuing, with new systems focusing on personally relevant symbolic representations, and utilizing metaphors from the natural and built worlds. Another shift that we foresee is the designers creating systems where multiple information sources and aspects interact to affect a single part of the representation. This is apparent already in Digital Family Portrait where the size of the butterflies represents "activity," even though activity is not the reading from a single sensor, but it instead a reading from multiple sensors in a home. Informative Art also has aspects of this approach, changing both the color and dimensions of squares based on two different aspects of weather. As regards aesthetic emphasis, we foresee a more radical change. We predict further exploration of the space of truly artistically motivated ambient information systems. These generative artworks use information from the world to drive their behavior and ask (and answer) art questions as well as technology questions. Though most of these works are outside the academy (they are shown in galleries instead of computer science conferences), Bolen and Mateas' Office Plant #1 [4] is a sculpture that characterizes the mood of a user's email stream and conveys it via transformations of a robotic plant. These systems are going to create a new design space above the top tier that we depict in this work. CONCLUSIONS In this work we synthesize a definition that distinguishes research in ambient information systems from that of notification systems and peripheral displays. We propose four design dimensions, rank systems to show clusters, and uncover four design patterns on which system developers may model their system designs. Future work will expand the four dimensions to include aspects of the social interaction and impact that system have on the behavior of individuals and groups. In this work we point toward open areas in the design space, and we point to new design directions that may fill these gaps. Future work may also turn this taxonomy into an evaluation framework for ambient information systems. 73 REFERENCES 1. Alexander, C., A Pattern Language: Towns, Buildings, Construction. Oxford University Press, 1977. 2. Ambient Orb. http://www.ambientdevices.com/ 3. Apple Mac OS X Dashboard. http://www.apple.com/ macosx/features/dashboard/index.htm 4. Bohlen, M., and Mateas, M. Office Plant #1. Leonardo 31:5. pp. 345-349. 5. Chang, A., Resner, B., Koerner B., Wang, X and Ishii, H., Lumitouch: An emotional communication device. Extended Abstracts of CHI 2001, pp. 371-372. 6. Cadiz, J ., Fussell, S., Kraut, R., Lerch, J., and Scherlis, W. The Awareness Monitor: A Coordination Tool for Asynchronous, Distributed Work Teams. Unpublished manuscript. Demonstrated at CSCW 1998. 7. Cadiz, J., Venolia, G., Janke, G., ans Gupta, A. Designing and deploying an information awareness interface. Proceedings of CSCW 2002, pp. 314 - 323. 8. Dahley, A., Wisneski, C., and Ishii, H. Water lamp and pinwheels: Ambient projection of digital information into architectural space. CHI Conference Summary 1998, pp. 269270. 9. Espinosa, A., Cadiz, J., Rico-Gutierrez L., Kraut, R., Sherlis, W., and Lautenbacher, G. Coming to the Wrong Decision Quickly: Why Awareness Tools Must be Matched with Appropriate Tasks. Proceedings of CHI 2000, pp. 392-399. 10. Fitzpatrick, G., Kaplan, S., Arnold, D., Phelps, T., and Segall, B. Augmenting the Workaday World with Elvin. Proceedings of ECSCW 1999, pp. 431-450. 11. Fogarty, J., Forlizzi, J., and Hudson, S. Aesthetic Information Collages: Generating Decorative Displays that Contain Information. Proceedings of the UIST 2001, pp. 141-150. 12. Gellersen, H.-W., Schmidt, A. and Beigl. M. Ambient Media for Peripheral Information Display. Personal Technologies 3, 4 : 199-208. 1999. 13. Greenberg, S., and Fitchett, C. Phidgets: Easy development of physical interfaces through physical widgets. Proceedings of UIST 2001. pp 209-218. 14. Greenberg, S., and Rounding, M. The Notification Collage: Posting Information to Public and Personal Displays. Proceedings of CHI 2001, pp. 515-521. 15. De Guzman, E., Yau M, Park, A., and Gagliano, A. Exploring the Design and Use of Peripheral Displays of Awareness Information. Extended Abstracts of CHI 2004, pp. 1247-1250. 16. Heiner, J. M., Hudson, S., and Kenichiro, T. The Information Percolator: Ambient information display in a decorative object. In Proc. of UIST 1999, pp. 141-148. 17. Ishii, H.,Wisenski, C., Brave, S., Dahley, A., Gorbet, M., Ullmer, B., and Yarin, P. AmbientROOM: Integrating Ambient Media with Architectural Space. Summary of CHI 1998, pp.173-174. 18. Ishii, H., Ren, S., and Frei, P. Pinwheels: visualizing information flow in an architectural space. Extended Abstracts of CHI 2001, pp. 111-112. 19. MacIntyre, B., Mynatt, E., Voida, S., Hansen, K., Tullio, J., and Corso, G. Support For Multitasking and Background Awareness Using Interactive Peripheral Displays. Proceedings of UIST 2001, pp. 41-50. 20. Mankoff, J., Dey, A., Heish, G., Kientz, J., Lederer, S., and Ames, M. Heuristic evaluation of ambient displays. Proceedings of CHI 2003, pp. 169-176. 21. McCrickard, D. S., Chewar, C., Somervell, J., and Ndiwalana, A. A Model for Notification Systems Evaluation--Assessing User Goals for Multitasking Activity. ACM Transactions on CHI 10,4 : 312 338. 2002 22. McCrickard, D.S., Catrambone, R., and Stasko, J. Evaluating animation in the periphery as a mechanism for maintaining awareness. Proceedings of INTERACT 2001, pp. 148-156. 23. Matthews, T., Dey, A.., Mankoff, J., Carter S., and Rattenbury, T. A Toolkit for Managing User Attention in Peripheral Displays. Proceedings of UIST 2004, pp. 247-256 . 24. Matthews ,T., Rattenbury, T., Carter, S., Dey, A., and Mankoff, J. A Peripheral Display Toolkit. Tech Report IRB-TR -03-018. Intel Research Berkeley. 2002. 25. Mynatt, E.D., Back, M., Want, R., and Ellis, J.B. Designing audio aura. Proceedings of CHI 1998, pp. 566-573. 26. Mynatt, E.D., Rowan, J., Jacobs, A., and Craighill, S. Digital Family Portraits: Supporting Peace of Mind for Extended Family Members. Proceedings of CHI 2001, pp. 333-340. 27. My Yahoo!. http://my.yahoo.com/index.html 28. Ogden, C., and Richards I. The Meaning of Meaning. Routledge & Kegan. London, England. 1923. 29. Pederson, E. R., and Sokoler, T. AROMA: Abstract Representation of Presence Supporting Mutual Awareness. Proceedings of CHI 1997, pp.51-58. 30. Redstrom, J., Skog, T., and Hallanas, L. Informative Art: Using Amplified Artworks as Information Displays. Proceedings of DARE 2000, pp. 103-114. 31. Russel, D., Streitz, N., and Winograd, T. Building Disappearing Computers. Communications of the ACM. 48(3):42-48. 2005. 32. Stasko, J., Miller, T., Pousman Z., Plaue, C., and Ullah, O. Personalized Peripheral Information Awareness through Information Art. Proceedings of UbiComp 2004, pp. 18-35. 33. Streitz, N., Tandler, P., Muller-Tomfelde, C., and Konomi, S. Roomware: Towards the Next Generation of Human-Computer Interaction based on an Integrated Design of Real and Virtual Worlds. In: J. Carroll (Ed.): Human-Computer Interaction in the New Millennium, Addison-Wesley. pp. 553-578. 2001. 34. Van Dantzich, M., Robbins, D., Horvitz, E., and Czerwinski, M. Scope: Providing Awareness of Multiple Notifications at a Glance. Proceedings of AVI 2002. pp. 157-166. 35. Weiser, M. and Brown, J.S. Designing Calm Technology. PowerGrid Journal, 1:1, 1996. 36. Zhao, A., and Stasko, J. What's Happening?: Promoting Community Awareness through Opportunistic, Peripheral Interfaces. Proceedings of AVI 2002, pp. 69-74. 74

Peripheral Display;four main design patterns;calm computing;symbolic Sculptural display;high throughput textual display;Notification System;information monitor display;ambient information system;Taxonomy;framework to understand design attributes;user interface;notification systems and peripheral displays;Design Guidelines;multiple-information consolidators;Ambient Display;definition and characteristics of ambient systems;Ubiquitous Computing

210

Using Web Helper Agent Profiles in Query Generation

ABSTRACT Personalized information agents can help overcome some of the limitations of communal Web information sources such as portals and search engines. Two important components of these agents are: user profiles and information filtering or gathering services. Ideally, these components can be sep-arated so that a single user profile can be leveraged for a variety of information services. Toward that end, we are building an information agent called SurfAgent;in previous studies, we have developed and tested methods for automatically learning a user profile [20]. In this paper, we evaluate alternative methods for recommending new documents to a user by generating queries from the user profile and submitting them to a popular search engine. Our study focuses on three questions: How do different algorithms for query generation perform relative to each other? Is positive relevance feedback adequate to support the task? Can a user profile be learned independent of the service? We found that three algorithms appear to excel and that using only positive feedback does degrade the results somewhat. We conclude with the results of a pilot user study for assessing interaction of the profile and the query generation mechanisms.

INTRODUCTION The Web has become an indispensable source of information for many people. Based on surveys of the most popular Web sites [14], users deal with the overwhelming amount and constantly updating nature of the information by routinely visiting hub sites (e.g., Netscape, Yahoo, CNN) and making copious use of search engines (e.g., AltaVista, Excite, Magellan ). Users have derived tremendous leverage from shared information resources such as those just mentioned. Hubs or portals provide communally useful information about perennial (e.g., financial management, child rearing) and timely (e.g., September 11 events, stock quote) topics. Search engines satisfy specific, spontaneous information needs. As our appetite for information increases, so does its availability on the Web. Studies (e.g., [21, 12]) have identified limitations with these tools for satisfying users' needs;for example, users appear to lack the motivation to learn how to formulate complex queries or to view long lists of potential matches. Meta-search engines, such as Meta-Crawler [18], SavvySearch [6], and NECI [11], propose to overcome the Web coverage problem by combining the indexing power of multiple stand-alone search engines. However, because they leverage the capabilities of many search engines, they tend to generalize the search task: limiting the access to search-engine -specific advanced search capabilities and, perhaps, introducing even more noise into the return results. One promising approach to compensating for the limitations is to personalize the tools. Pretschner and Gauch divide personalization into two types: personalized access to resources and filtering/ranking [15]. For example, My Yahoo (http://my.yahoo.com) provides personalized access by allowing users to design their own Yahoo page with pertinent information;many search and meta-search engines support some customization (e.g., types of search, return amount, and search engine selection). "Softbot"s or information agents augment searching and information gathering (filtering/ranking). Personalized information agents, such as Letizia [13], WebWatcher [1, 10], and WebMate [5], can provide a range of services from automatically retrieving Web pages to assisting in formulating queries. These agents generally comply with the architecture presented in Figure 1. The agent intercedes between the user and their Web access, monitoring the user's activities to construct a model of user requests (the profile) to be used for specific information services (e.g., modifying requests and/or filtering results). In our research, we adopt the principle that user overhead needs to be minimized: The profile should be learned by asking the user to simply click on a feedback button positioned on the bottom of each page to indicate interest. Learning should track changes in user interests. The profile should support multiple information services . In previous papers, we have assessed some alternative approaches to learning user profiles [20, 19]. In this paper, we examine alternative approaches to one of the services: query generation to find new documents (i.e., automatically retrieving Web pages that match a user's interests by submitting queries to a Web search engine). In particular, we are interested in answering the following questions: 1. How do different algorithms for query generation perform relative to each other? For our case, query generation involves constructing queries from a user profile that are submitted to a search engine for the purpose of harvesting documents. 2. Is positive relevance feedback adequate to support the task? To minimize user overhead, we have solicited only positive relevance feedback. Obviously, this provides relatively weak knowledge, requiring the profiling mechanism to self-organize the categories of interest and to trade-off precision. 3. Can a user profile be learned independent of the service ? If so, then user overhead can be minimized and multiple services provided as part of the same agent. This paper describes a framework for evaluating alternative algorithms for information gathering agents and a study that was designed to address the three questions above. In summary , we found: Three algorithms perform best in terms of extracting sufficient numbers of documents from the search engine and in terms of the relevance of the links returned. We did find evidence that soliciting only positive feedback hampers query generation;however, it is not clear that the degradation in performance is worth the cost of obtaining the negative feedback. As often happens, the study raised some issues that are still to be resolved (particularly about the evaluation criteria and the interaction of profiling and query generation);we conclude with a pilot study in which we investigate how to resolve these issues. SURFAGENT SurfAgent [19] is a personalized Web information agent, which follows the basic architecture in Figure 1. It is designed as a testbed for expediting plug-and-play and evaluation of alternative algorithms, front-ends, and service tasks. Its two primary components are the user profile and the module which generates requests for document streams. Monitoring should be simple and unobtrusive. Filtering depends on the representation and construction of the user profile, forcing a relatively tight coupling of those two components. This section provides an overview of its user profiling and document stream generation. 2.1 Building User Profiles and Filtering The user profile maintained by a Web helper agent is a model of what documents the user finds relevant. Like most other personal Web helper agents, SurfAgent uses TF-IDF vectors [17] as the basis of its user profile representation. One such vector is used to represent each of the several different topics of interest associated with each user. Over time, topic descriptions are learned from user-supplied examples of relevant documents, which are added to the existing TF-IDF vectors in what is known as relevance feedback [16]. Associated with each vector is a dissemination threshold , which is used when new documents are filtered by the agent: if the similarity between the new document's vector and a vector in the profile exceeds the associated dissemination threshold, the document is considered relevant and shown to the user. We found that learning the appropriate dissemination threshold was critical to filtering performance and that one could be learned with relatively little feedback (i.e., 10 relevance judgments) [19]. TF-IDF vectors and their associated dissemination thresholds are known in the Information Retrieval (IR) literature as filtering queries. This type of query is distinguished from a typical retrieval query (used with search engines or at a library) by a few characteristics. Filtering queries tend to be used repeatedly over a long period of time, during which they can be improved and maintained through learning and relevance feedback, whereas retrieval queries are typically used only once. Also, filtering queries typically contain lots of terms with complex weighting schemes, whereas retrieval queries tend to be a boolean combination of relatively few terms, with no weighting at all. Each filtering query in SurfAgent's user profile corresponds to a distinct topic of interest. User profiles are learned in one of two ways. First, relevant documents provided as training by the user can be explicitly associated with the topic of interest. Alternatively, to minimize overhead to the user, incremental clustering can be used by the agent to automatically classify relevant examples into internally learned topics [20, 5]. In the latter situation, the user only needs to prompt the agent when a relevant document has been encountered, without having to associate it with a particular topic of interest . To minimize user disruption, we request only positive examples. We augmented existing IR clustering techniques to accommodate Web needs (i.e., avoid storing the documents themselves, require minimal user overhead and be associated with a user). In our earlier study, we found that a tuned version of the Doubling algorithm [4] achieved high recall, without a great sacrifice on precision. 2.2 Incoming Document Streams Personal information agents use a wide range of techniques to generate incoming streams. Letizia pre-fetches Web pages by exploring the links in the Web page cur-rently being viewed. Similarly, WebWatcher analyzes text in and around links in the current page to predict relevant links worth following. Fab builds a list of likely to be relevant documents through best-first search;documents that pass the filtering phase are then included in a list of recom-813 User User Profile Extract Query Search Engine Filter Documents Figure 2: Incoming document streams generated by querying a search engine mended links. Finally, WebMate filters articles found on a list of well-known news sources in order to compile a personal newspaper for its user. Our goals are to maximize the quality of the incoming document stream generated for SurfAgent, while at the same time minimizing effort. For this purpose, a promising technique appears to be the construction of queries that are suitable for a large-scale search engine such as Google [3]. Well-formulated queries have the potential to off-load significant portions of the filtering task to the search engine, which should provide the agent with a document stream consisting of more relevant documents. In this paper, we explore several methods of generating search engine queries from the user profile of a personal Web helper agent. We wish to find both the method and the query length that would optimize the relevance of the documents returned by the search engine with respect to the user profile. This process is illustrated in Figure 2. TECHNIQUES FOR QUERY GENERATION Filtering queries are not directly suitable for being submitted to a search engine. They are complex models representing a possibly large collection of documents;they contain a large number of terms with associated weights, which would overly restrict the range of documents a search engine might return. Query generation techniques have evolved from the more general query refinement mechanism of relevance feedback [16]. For instance, in [9, 7] search engine queries are extended with features extracted from positive and negative examples in order to bias them toward a more relevant sub-topic . Several other researchers have been concerned with extracting only a few highly representative terms from the representation of a large document cluster. For WebACE [2], the authors propose a mechanism for generating queries from a cluster of documents based on the average word count (term frequency, TF) of the terms contained in the documents of the cluster, and the document frequency (DF), i.e., the number of documents in the cluster that contain a specified term. A set of k terms with the highest TF, and another set of k terms with the highest DF are selected from the cluster. The intersection of these two sets was submitted to Yahoo search as a query, yielding a total of 2280 results, which were later refined to 372 by augmenting the query with terms resulting from the difference between the TF and DF sets. CorpusBuilder [8] uses automatic query generation to collect documents in a specified language (Slovenian) from the Web, with the purpose of building a corpus in that language. The point is to preserve computation power by avoiding a brute-force crawl of the Web where an expensive classifier would have to be run on each encountered document. By generating queries from the already existing corpus, the authors hope to significantly increase the likelihood that the resulting documents would already be in Slovenian, thus speeding up document collection. Several methods for generating queries are used interchangeably: uniform select n terms from the relevant documents with equal probability; term-frequency select n most frequent terms from the relevant documents; probabilistic term-frequency select n terms from the relevant documents with probability proportional to their term frequency; odds-ratio select n terms with highest odds-ratio scores, as given by the formula: OR t = log 2 P (t|relevant) (1 - P (t|nonrelevant)) P (t|nonrelevant) (1 - P (t|relevant)) where t is the term, and P (t|relevant) and P (t|nonrelevant) are the probabilities of t appearing in a relevant and non-relevant document, respectively; probabilistic odds-ratio select n terms with probability proportional to their odds-ratio score; The authors report best results with n = 4 and the simple odds-ratio method. However, this method is not necessarily applicable to our task because identifying relevance with respect to a query cluster is somewhat more subtle than determining whether a returned document is in a particular language such as Slovenian. OUR STUDY OF QUERY GENERATION The purpose of this study is to examine the role of query generation technique for a Web information agent: what techniques work well, how much user overhead is warranted and how query generation interacts with profiling. These three factors correspond to the three questions articulated in the Introduction. 4.1 Experiment Design The basic protocol for our study was as follows: 1. construct user profiles, 2. generate queries from those profiles using each of the query generation methods, 3. submit queries of different lengths to Google, 4. evaluate the results. This led to a factorial experiment in which the independent variables were query generation method and query length and the dependent variables were two evaluation metrics: return count and relevance. 814 4.1.1 Constructing the profiles To expedite experimental control and data collection, we constructed two user profiles from TREC 1 disk #5. TREC data consist of a large number of articles (over 250,000), of which a large portion have been judged as either relevant or non-relevant with respect to a set of 450 different topics. Our first topic is on airport security measures, and was constructed based on 46 relevant documents which appeared in the Los Angeles Times over the course of two years (1989-1990 );this topic will be referred to as LATIMES. The second topic is on genetic research, and was constructed based on 55 relevant documents from the Foreign Broadcasting Information Service appeared in 1996;this topic will be referred to as FBIS. One topic was picked randomly from each of the two document collection on the TREC disk. We used synthetically generated topics in order to test the hypothetical scenario in which negative feedback is available. By default, SurfAgent does not collect negative feedback in the form of documents which are non-relevant to a given topic. Thus, we are interested in how much performance might be sacrificed by restricting feedback to only positive examples. The number of positive documents used in the construction of each topic (46 and 55, respectively) is realistic compared to what a human user would have been capable of providing while building her profile in real life. 4.1.2 Generating Queries We implemented several methods, including both methods which use such negative examples (e.g., odds-ratio) against methods which do not (e.g., Boley's method [2] and term frequency based methods). In addition to the methods mentioned in Section 3, we add two methods: deterministic extraction of highest weight terms for SurfAgent's TF-IDF profile vectors and probabilistic weighted extraction from the TF-IDF profile vectors. The complete list of methods used is given below: Uniform (Unif ) baseline case, select n terms with uniform probability; Boley select the intersection of the k top ranking terms according to term frequency in one set, and document frequency in the other; TF select n top ranking terms according to term frequency; Probabilistic TF (P-TF) select n terms with probability proportional to their term frequency; OR select the top ranking n terms according to their odds-ratio score; Probabilistic OR (P-OR) select n terms with probability proportional to their odds-ratio score; TFIDF select n terms with the highest TF-IDF weight from the profile vector; Probabilistic TF-IDF (P-TFIDF) select n terms with probability proportional to their TF-IDF weights; 1 Text REtrieval Conference: TREC benchmark disks are publicly available and can be ordered from the conference homepage at http://trec.nist.gov The probabilistic versions were included because injection of small amounts of randomness has been found to be helpful in other search problems. Code from SurfAgent was modified to collect the data required by all query generation methods employed. For each topic, we collected the following data: average term frequencies for terms in relevant documents ; document frequencies for terms in relevant documents; TF-IDF vector built from relevant documents; odds-ratio scores for terms in relevant documents (odds-ratio scores are based on both relevant and non-relevant documents related to the topic). From these data, we generated queries of seven lengths (two to eight terms) for each of the eight methods. For the four probabilistic methods, we generated 10 queries of each length, which means their reported results will be averages over those 10 queries. For Boley's method, we repeatedly computed the intersection of the sets consisting of the top k ranking terms w.r.t. TF and DF, while incrementing k. We retained all distinct queries of length between 2 and 8. For FBIS, no value of k generated a query of length 6. Similarly, for LATIMES, there was no value of k resulting in a query of length 7. 4.1.3 Submit the Queries The previous step resulted in 614 queries (307 for each topic). We submitted these queries to the Google search engine and collected back the first page of results. By default, a page can contain up to 10 responses. 4.1.4 Collect the results The results of the queries (the URLs returned) were parsed out of the page returned by Google, and their corresponding documents were retrieved from the Web and stored locally. We discarded (and decremented our hit count accordingly) all dead links and all hits that were in a format other than ASCII 2 or PDF: a total of 312 out of 2917 hits were discarded . The PDF documents were converted into ASCII using the pdftotext utility. 4.2 Results For each valid hit, we computed the similarity between the document's TF-IDF vector and the TF-IDF vector of the appropriate topic, which is a measure of the document's relevance. For each combination of query generation method and query length, we recorded the number of hits received, the relevance of the best hit, and the average relevance over all hits received. For the probabilistic methods, these measurements represent average values obtained over the ten repetitions for each combination of parameters. The results are summarized in Table 1 for the FBIS topic, and Table 2 for the LATIMES topic. The three rows corresponding to each method indicate average relevance (top), maximum relevance , and number of returned hits (bottom). All methods return at least seven documents with query lengths of 2, but most taper off in the number returned 2 ASCII includes all variants and versions of HTML, XML, etc. 815 query length method 2 3 4 5 6 7 8 avg: .022 .046 .059 .018 .011 Unif max: .051 .077 .101 .019 .011 cnt: 7.7 4.3 3.2 0.4 0 0 0.1 avg: .026 .054 .044 .053 .069 .091 .082 P-TF max: .059 .096 .079 .102 .120 .192 .138 cnt: 8.9 7.7 7.9 5.2 6.5 7.2 6.3 avg: .039 .047 .019 -P -OR max: .099 .090 .019 -cnt : 9.0 3.2 0 0.1 0 0 0 avg: .045 .058 .088 .069 .035 .034 .030 P-TFIDF max: .100 .110 .178 .097 .054 .035 .055 cnt: 9.1 6.1 8.4 2.4 2.7 0.6 1.4 avg: .053 .077 .090 .081 .111 .088 Boley max: .120 .112 .136 .168 .239 .239 cnt: 9 9 10 7 0 8 9 avg: .036 .031 .048 .082 .081 .087 .083 TF max: .065 .059 .129 .134 .103 .130 .135 cnt: 10 9 10 9 9 10 9 avg: .123 .186 .102 -OR max: .155 .361 .190 -cnt : 9 8 2 0 0 0 0 avg: .100 .144 .160 .176 .214 .278 .242 TFIDF max: .377 .377 .377 .279 .399 .404 .242 cnt: 9 10 10 7 10 4 1 Table 1: Average relevance, Maximum relevance, and count of returned hits for the FBIS topic on genetic technology query length method 2 3 4 5 6 7 8 avg: .012 .012 .012 .013 .004 -Unif max: .024 .024 .028 .019 .006 -cnt : 8.0 5.3 3.9 1.0 0.6 0 0 avg: .017 .026 .025 .028 .032 .024 .010 P-TF max: .042 .073 .062 .061 .046 .042 .011 cnt: 9.1 9.5 6.0 6.5 2.0 4.0 0.7 avg: .017 .018 .016 .011 .007 -P -OR max: .052 .039 .029 .013 .007 -cnt : 8.2 8.3 4.0 0.9 0 0.1 0 avg: .026 .036 .064 .063 .059 .020 .010 P-TFIDF max: .058 .103 .125 .156 .106 .036 .014 cnt: 9.2 8.1 8.1 5.7 5.3 1.3 0.2 avg: .040 .098 .135 .193 .199 .167 Boley max: .086 .199 .299 .343 .359 .299 cnt: 8 9 8 8 8 0 7 avg: .107 .058 .030 .048 .041 .069 -TF max: .222 .093 .051 .075 .069 .069 -cnt : 7 10 10 7 6 1 0 avg: .048 .036 .348 -OR max: .122 .096 .402 -cnt : 9 9 2 0 0 0 0 avg: .115 .144 .155 .171 .144 .153 .143 TFIDF max: .331 .331 .357 .299 .276 .349 .349 cnt: 9 7 8 8 9 9 9 Table 2: Average relevance, Maximum relevance, and count of returned hits for the LATIMES topic on airport security with longer query lengths. For the deterministic methods, the relevance increases as the query length increases (until 7 or 8), but the relevance for the probabilistic methods tends to plateau early. All methods consistently outperform the baseline uniform term selection. Probabilistic methods are outperformed by Figure 3: Box plot of relevance by method for FBIS topic at query length 2 Figure 4: Box plot of relevance by method for FBIS topic at query length 3 the non-probabilistic ones, which is consistent with the observations in [8]. The best results for the FBIS topic were obtained using TFIDF at query length 7: a total of 4 hits were returned, with an average relevance of .278, and a maximum relevance of .404. The best results for the LATIMES topic were obtained using OR at query length 4: two hits were returned, with average relevance of .348 and maximum relevance of .402. Query lengths 2 and 3 were the only ones where all methods lead to non-empty returns for both topics. To test whether the differences between the methods were significant , we ran an analysis of variance (ANOVA) study on each topic at query lengths 2 and 3, with the query generation method as the independent variable, and relevance as the dependent. The effects were significant in each case: for FBIS, we obtained F = 14.007, p &lt; 0.001 at query length 2, and F = 8.692, p &lt; 0.001 at query length 3;for LATIMES, we obtained F = 24.027, p &lt; 0.001 at query length 2, and F = 20.277, p &lt; 0.001 at query length 3. Box plots of relevance by method for query lengths 2 and 3 are given in Figures 3 and 4 for FBIS, and Figures 5 and 816 Figure 5: Box plot of relevance by method for LATIMES topic at query length 2 Figure 6: Box plot of relevance by method for LATIMES topic at query length 3 6 for LATIMES. Note that medians rather than means are marked on the graph. These plots illustrate several points obscured by the previous table. First, while TFIDF returns the best match and second best median in all cases, overall better results are obtained using OR for FBIS, and TF and Boley for LATIMES. Second, each method can show high variance in the results, although TFIDF tends generally to higher variance. Additionally, the results for query length 3 have higher variance than those for query length 2. Finally, the distributions are often skewed. For example, Boley consistently has a higher median than mean. Because relevance for all returned documents is measured against the TF-IDF vector of the corresponding topic, the experiments are slightly biased in favor of the TFIDF query generation method. Our experiments cannot prove that TFIDF query generation is sufficient, but its good performance coupled with the not always good performance of OR suggest that we do not incur a significant loss by leaving out negative feedback. Collecting other information based on positive feedback in addition to TF-IDF topic vectors may be required with SurfAgent: e.g., straight TF vectors and query length method 2 3 4 5 6 7 8 TFIDF nrel: 8 3 4 4 4 -cnt : 10 10 10 10 9 0 0 P-TFIDF nrel: 1.0 1.8 0.7 0.4 0.0 0.5 -cnt : 9.5 8.1 7.7 4.1 1.3 1.3 0.0 Table 3: Number of relevant documents and count of returned hits for the user-generated topic on stress relief topic-specific document frequency information would allow us to use TF and Boley query generation in addition to the TFIDF method. As the results show, sometimes TF and Boley perform better than OR and TFIDF. Both Boley and TFIDF consistently result in many links being returned, even for long query lengths. Many hits are desirable when the agent will be pruning out previously visited sites and filtering the returned links according to a filtering threshold. The computation time required for query generation by each of the studied methods is negligible when compared to the network-related delays incurred while downloading the actual documents. PILOT USER STUDY To gain an understanding of the performance of TFIDF query generation without the bias present in our experiments with synthetically generated topics, we have also performed a pilot study with a user-generated topic, containing 34 documents on stress relief. Since SurfAgent only collects TF-IDF information at this time, query generation was limited to the TFIDF and P-TFIDF methods. We followed the same protocol as with the synthetically generated topics: query lengths were between 2 and 8, and results were aver-aged over ten trials for the probabilistic version P-TFIDF. A total of 343 distinct URLs were returned by Google. We shuffled these URLs and asked the user to examine each one of them, and mark the ones she found to be relevant to her topic. 56 documents out of the total of 343 were found relevant . Table 3 presents the number of relevant documents and the number of hits returned for each parameter combination . This pilot study supports the hypothesis that TFIDF based queries will generate an adequate incoming stream: queries of length up to six returned at least nine hits from Google. Unlike the previous study, the shorter queries yielded lower relevance, which could be due to the way the user was judging relevance or to the nature of the topic. As a followup, we will be designing a user study that includes the three apparently best methods (TF, TFIDF, and Boley). We will focus on three issues: Does method performance vary among users and topics (as is suggested by our current study)? Should profile construction incorporate more information? Does relevance assessment change as profiles become more mature? Can the best query length be determined a priori? CONCLUSIONS We studied several methods for generating queries from a user profile to answer three questions related to the design of Web information agents. First, how do different query 817 generation algorithms perform relative to each other? In fact, we observed significantly different performance among the eight methods tested. Overall, Boley, TFIDF and to a lesser extent TF provided a good number of hits and relatively high relevance. Second, is positive relevance feedback adequate to support the task? We found that leaving out negative training examples does not incur a significant performance loss. Odds-Ratio was found to excel on one topic, but its competitive advantage does not appear to be worth the additional overhead expected from the user. TFIDF and Boley, requiring only positive relevance feedback, generated queries that resulted in relevant hits. Third, can user profiles be learned independently of the service? The results from TFIDF and the pilot experiment do suggest it. However, the pilot study also suggests that either user relevance judgments may be a bit lower (harsher) than the automated method or that the profile may not adequately reflect the users' interests. In fact, the good performance of Boley and TF indicates that in some cases it might be worthwhile to collect more than TF-IDF information from the user-supplied positive training examples. This last question will be examined in more detail in the future. Our study confirmed that additional user burden in the form of negative feedback appears unwarranted to support document generation and that queries generated based on automatically learned profiles can guide harvesting of new documents of interest. This last result is excellent news for the development of agents that leverage a single learned profile to personalize a multitude of web information services. ACKNOWLEDGMENTS This research was supported in part by National Science Foundation Career Award IRI-9624058. The United States Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright notation herein. REFERENCES [1] R. Armstrong, D. Freitag, T. Joachims, and T. Mitchell. WebWatcher: A Learning Apprentice for the World Wide Web. In Proceedings of the AAAI Spring Symposium on Information Gathering from Heterogeneous, Distributed Resources, Stanford, CA, 1995. [2] D. Boley, M. Gini, R. Gross, E. Han, K. Hastingsand G. Karypis, V. Kumar, M. Bamshad, and J. Moore. Document Categorization and Query Generation on the World Wide Web Using WebAce. AI Review, 13(5-6):365391, 1999. [3] S. Brin and L. Page. The Anatomy of a Large-scale Hypertextual Web Search Engine. Computer Networks and ISDN Systems, pages 107117, 1998. [4] M. Charikar, C. Chekuri, T. Feder, and R. Motwani. Incremental Clustering and Dynamic Information Retrieval. Proceedings of the 29th ACM Symposium on Theory of Computing, 1997. [5] L. Chen and Katia Sycara. WebMate: A Personal Agent for Browsing and Searching. In Proceedings of the Second International Conference on Autonomous Agents, Minneapolis, MN, 1998. [6] D. Dreilinger and A.E. Howe. Experiences with selecting search engines using meta-search. ACM Transactions on Information Systems, 15(3):195222, 1997. [7] 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 (WWW 2002), Honolulu, HI, U.S.A., 2002. [8] R. Ghani, R. Jones, and D. Mladenic. On-line learning for query generation: Finding documents matching a minority concept on the web. In Proc. of the First Asia-Pacific Conference on Web Intelligence, 2001. [9] E.J. Glover, G.W. Flake, S. Lawrence, W.P. Birmingham, A. Kruger, C.L. Giles, and D. Pennock. Improving Category Specific Web Search by Learning Query Modifications. In Proceedings of the IEEE Symposium on Applications and the Internet (SAINT 2001), San Diego, CA, U.S.A., 2001. [10] T. Joachims, D. Freitag, and T. Mitchell. WebWatcher: A Tour Guide for the World Wide Web. In Proc. of the 15th International Joint Conference on Artificial Intelligence, Nagoya, Japan, 1997. [11] S. Lawrence and C.L. Giles. Context and page analysis for improved web search. IEEE Internet Computing, 2(4):3846, 1998. [12] S. Lawrence and C.L. Giles. Searching the world wide web. Science, 280:98100, April 3 1998. [13] H. Lieberman. Letizia: An agent that assists web browsing. In Proceedings of the 14th International Joint Conference on Artificial Intelligence (IJCAI-95), Montreal, Canada, 1995. [14] Inc. Netmation. 100 most popular web sites. http://netmation.com/list100.htm. [15] A. Pretschner and S. Gauch. Personalization on the web. Technical Report ITTC-FY2000-TR-13591-01, Dept. of Electrical Engineering and Computer Science, University of Kansas, December 1999. [16] J.J. Rocchio. Relevance feedback in information retrieval. In G. Salton, editor, The SMART Retrieval System: Experiments in Automatic Document Processing. Prentice-Hall, 1971. [17] G. Salton. Automatic Text Processing: The Transformation, Analysis, and Retrieval of Information by Computer. Addison-Wesley, 1988. [18] E. Selberg and O. Etzioni. The metacrawler architecture for resource aggregation on the web. IEEE Expert, 12(1):814, 1997. [19] G.L. Somlo and A.E. Howe. Adaptive lightweight text filtering. In Proceedings of the 2001 Conference on Intelligent Data Analysis (IDA '01), Lisbon, Portugal, September 2001. [20] Gabriel L. Somlo and Adele E. Howe. Incremental clustering for profile maintenance in information gathering web agents. In Proceedings of the 2001 International Conference on Autonomous Agents (AGENTS'01), Montreal, Canada, May 2001. [21] A. Spink, J. Bateman, and B.J. Jansen. Searching heterogeneous collections on the web: Behavior of excite users. Information Research: An Electronic Journal, 5(2), 1998. http://www.shef.ac.uk/~is/publications/infers 818

query generation;information agents;user modeling

211

Very Low Complexity MPEG-2 to H.264 Transcoding Using Machine Learning

This paper presents a novel macroblock mode decision algorithm for inter-frame prediction based on machine learning techniques to be used as part of a very low complexity MPEG-2 to H.264 video transcoder. Since coding mode decisions take up the most resources in video transcoding, a fast macro block (MB) mode estimation would lead to reduced complexity. The proposed approach is based on the hypothesis that MB coding mode decisions in H.264 video have a correlation with the distribution of the motion compensated residual in MPEG-2 video. We use machine learning tools to exploit the correlation and derive decision trees to classify the incoming MPEG-2 MBs into one of the 11 coding modes in H.264. The proposed approach reduces the H.264 MB mode computation process into a decision tree lookup with very low complexity. The proposed transcoder is compared with a reference transcoder comprised of a MPEG-2 decoder and an H.264 encoder. Our results show that the proposed transcoder reduces the H.264 encoding time by over 95% with negligible loss in quality and bitrate.

INTRODUCTION During the past few years, technological developments, such as novel video coding algorithms, lower memory costs, and faster processors, are facilitating the design and development of highly efficient video encoding standards. Among the recent works in this area, the H.264 video encoding standard, also known as MPEG-4 AVC occupies a central place [1]. The H.264 standard is highly efficient by offering perceptually equivalent video quality at about 1/3 to 1/2 of the bitrates offered by the MPEG-2 format. However, these gains come with a significant increase in encoding and decoding complexity [2]. Though H.264 is highly efficient compared to MPEG-2, the wide and deep penetration of MPEG-2 creates a need for co-existence of these technologies and hence creates an important need for MPEG-2 to H.264 transcoding technologies. However, given the significant differences between both encoding algorithms, the transcoding process of such systems is much more complex compared to the other heterogeneous video transcoding processes [3-6]. The main elements that require to be addressed in the design of an efficient heterogeneous MPEG-2 to H.264 transcoder are [7]: the inter-frame prediction, the transform coding and the intra-frame prediction. Each one of these elements requires to be examined and various research efforts are underway. In this paper, we focus our attention on a part of the inter-frame prediction: the macroblock mode decision, one of the most stringent tasks involved in the transcoding process. A video transcoder is comprised of a decoding stage followed by an encoding stage. The decoding stage of a transcoder can perform full decoding to the pixel level or partial decoding to the coefficient level. Partial decoding is used in compressed domain transcoding where the transform coefficients in the input format are directly transcoded to the output format. This transformation is straightforward when the input and output formats of the transcoder use the same transform (e.g., MPEG-2 to MPEG-4 transcoding) [5]. When these transforms differ substantially, the compressed domain transcoding becomes computationally expensive. The utility of this compressed domain transcoding is limited to intra MB transcoding. For predicted MBs, the transcoding in compressed domain becomes prohibitively expensive. The substantial differences in MPEG-2 and H.264 make even intra transcoding in the compressed domain relatively expensive [8]; pixel domain transcoding is shown to produce better results [9]. Pixel domain transcoders have a full decoding stage followed by a reduced complexity encoding stage. The complexity reduction is achieved by reusing the information gathered from the decoding stage. It is assumed that the input video is encoded with reasonable RD optimization. The MPEG-2 to H.264 complexity reduction techniques reported in the literature fall into two categories: 1) MB mode mapping in H.264 based on the MB modes of the incoming video [10] and 2) Selective evaluation of MB modes in H.264 based on heuristics [11]. Because of the large number of inter and intra MB coding modes supported by H.264, there is no one-to-one mapping between MPEG-2 and H.264 MB modes. A direct mapping leads to either a sub-optimal decision if the mapped mode is the final MB mode or an increase on complexity if additional evaluations have to be made to improve the mode decision. Selective evaluation is based on the observation that certain MB modes are less likely to occur for a class of videos and bitrates. If the selective evaluation is aggressive in limiting the number of allowed modes, the performance is sub-optimal. On the contrary, increasing the number of allowed modes increases the complexity. We have developed an innovative approach that is not limited by the inefficiencies of mode mapping or selective evaluation approaches. The proposed approach is based on the hypothesis that MB coding mode decisions in H.264 video have a correlation with the distribution of the motion compensated residual in MPEG-2 video. Exploiting this correlation together with the MB coding modes of MPEG-2 could lead to a very low complexity transcoder. Figure 1 shows a plot of the mean and variance of the MPEG-2 MB residual in the input video and the H.264 MB coding mode of the corresponding MB in the transcoded video. As the coding mode changes, the shift in the mean and variance of the corresponding MB can be clearly seen. This correlation can be effectively exploited using machine learning approaches. Thus, the H.264 MB mode computation problem is posed as a data classification problem where the MPEG-2 MB coding mode and residual have to be classified into one of the several H.264 coding modes. The proposed transcoder is developed based on these principles and reduces the H.264 MB mode computation process into a decision tree lookup with very low complexity. Figure 1. Relationship between MPEG-2 MB residual and H.264 MB coding mode. The rest of the paper is organized as follows. Section 2 reviews the principles of operation of the prediction of inter-coded macroblocks in p-slices used by the H.264 encoding standard. Section 3 introduces our macroblock mode decision algorithm for inter-frame prediction based on machine learning techniques, specifically designed for MPEG-2 to H.264 transcoders. In Section 4, we carry out a performance evaluation of the proposed algorithm in terms of its computational complexity and rate-distortion results. We compare the performance of our proposal to the reference transcoder with the encoding stage using the H.264 reference implementation. Finally, Section 5 draws our conclusions and outlines our future research plans. MACROBLOCK MODE DECISION AND MOTION ESTIMATION IN H.264 In the H.264 standard, the macroblock decision mode and motion estimation are the most computationally expensive processes. H.264 uses block-based motion compensation, the same principle adopted by every major coding standard since H.261. Important differences from earlier standards include the support for a range of block sizes (down to 4x4) and fine sub-pixel motion vectors (1/4 pixel in the luma component). H.264 supports motion compensation block sizes ranging from 16x16 to 4x4 luminance samples with many options between the two. The luminance component of each macroblock (16x16 samples) may be split up in 4 ways: 16x16, 16x8, 8x16 or 8x8. Each of the sub-divided regions is a macroblock partition. If the 8x8 mode is chosen, each of the four 8x8 macroblock partitions within the macroblock may be further split in 4 ways: 8x8, 8x4, 4x8 or 4x4 (known as sub-macroblock partitions). These partitions and sub-partitions give rise to a large number of possible combinations within each macroblock (see Figure 2). This method of partitioning macroblocks into motion compensated sub-blocks of varying size is known as tree structured motion compensation. Figure 2. Macroblock partitions, sub-macroblock partitions and partition scans. A separate motion vector (previously calculated in the motion estimation module) is required for each partition or sub-partition. Each motion vector must be coded and transmitted; in addition, the choice of partition(s) must be encoded in the compressed bitstream. Choosing a large partition size (e.g. 16x16, 16x8, 8x16) means that a small number of bits are required to signal the choice of motion vector(s) and the type of partition; however, the motion compensated residual may contain a significant amount of energy in areas with high detail. Choosing a small partition size (e.g. 8x4, 4x4, etc.) may give a lower-energy residual after motion compensation but requires a larger number of bits to signal the motion vectors and choice of partition(s). The choice of partition size therefore has a significant impact on compression performance. In general, a large partition size is appropriate for homogeneous areas of the frame and a small partition size may be beneficial for areas with high detail . Va ri a n ce MPEG-2 Res. MB Var. H.264 MB Mode MB Number Mea n MPEG-2 Res. MB Mean H.264 MB Mode 932 The resolution of each chroma component in a macroblock (Cr and Cb) is half that of the luminance (luma) component. Each chroma block is partitioned in the same way as the luma component, except that the partition sizes have exactly half the horizontal and vertical resolution (an 8x16 partition in luma corresponds to a 4x8 partition in chroma; an 8x4 partition in luma corresponds to 4x2 in chroma; and so on). The horizontal and vertical components of each motion vector (one per partition) are halved when applied to the chroma blocks. Each partition in an inter-coded macroblock is predicted from an area of the same size in a reference picture. The offset between the two areas (the motion vector) has -pixel resolution (for the luma component). If the video source sampling is 4:2:0, 1/8 pixel samples are required in the chroma components (corresponding to -pixel samples in the luma). The luma and chroma samples at sub-pixel positions do not exist in the reference picture and so it is necessary to create them using interpolation from nearby image samples. Sub-pixel motion compensation can provide significantly better compression performance than integer-pixel compensation, at the expense of increased complexity. Quarter-pixel accuracy outperforms half-pixel accuracy. Encoding a motion vector for each partition can take a significant number of bits, especially if small partition sizes are chosen. Motion vectors for neighboring partitions are often highly correlated and so each motion vector is predicted from vectors of nearby, previously coded partitions. The method of forming the prediction MVp depends on the motion compensation partition size and on the availability of nearby vectors. In H.264, the macroblock mode decision is the most computationally expensive process. Mode decision is a process such that for each possible block-size a cost is evaluated. The encoder selects the coding-modes for the macroblock, including the best macroblock partition (sub-macroblock partition) and mode of prediction for each macroblock partition, such that the cost is optimized. In the JM reference code (version 10.2) [12], the motion estimation and the mode decision are executed together. This implies that for each macroblock partition (sub-macroblock partition) within the MB, motion estimation is done first and the resulting cost is used for the mode decision. In the H.264, two methods have been defined to evaluate the cost for MB mode decision: RD-cost and SAE-cost. In the following, we describe these two methods. 2.1 The RD-Cost The Rate-Distortion (RD) optimization method is based on a Lagrange multiplier [13] [14]. The H.264 standard can make use of this optimization method to choose the best macroblock mode decision. Different from evaluating the cost of coding a macroblock on a pixel by pixel basis (SAE cost), the RD-cost consists of making the selection based on a Lagrange function. In this way, the H.264 standard selects the macroblock mode exhibiting the minimum Lagrange cost. This implies that for each existing macroblock partition (sub-partition) within the MB, bitrate and distortion are calculated by actually encoding and decoding the video. Therefore, the encoder can achieve the best Rate-Distortion performance results, at the expense of calculation complexity. For evaluating the RD-cost, the standard has to obtain the encoding rate, R, and the distortion, D, of each macroblock partition (sub-macroblock partition). The former is obtained by first computing the difference between the original macroblock and its predictor. Thereafter, a 4x4 Hadamard Transform (HT) has to be applied followed by a quantization process. The distortion, D, is obtained by performing an inverse quantization process followed by its inverse HT and then comparing the original macroblock to the reconstructed one. The H.264 standard chooses then the decision mode having the minimum cost, J. The cost is evaluated using the Lagrange function J=D + x R, where is the Lagrange multiplier. Figure 3 depicts the overall process. One of the main drawbacks of this method is its excessive computational cost. On the contrary, the encoder can achieve the best Rate-Distortion performance results. However, for many applications, the use of the Lagrange multiplier may be prohibitive. This is the case when developing a transcoding architecture aimed to work in real-time. HT + QP Encoder H.264/AVC with loop Rate-Distorsion QP -1 IHT + Compute rate Prediction Frame + Determine distorsion + + Compute cost (J = D+ x R) Decision R D Figure 3. RD-cost method in the H.264 encoder. 2.2 The SAE-Cost In this method, the H.264 encoder selects the best macroblock mode by using the Sum of Absolute Errors (SAE). This implies that for each existing macroblock partition (sub-partition) within the MB, a predictor within the pixel-domain is created from the motion estimation of the current partition and the SAE costs is evaluated. For each MB and for each color component (Y,Cr,Cb), one prediction mode have to be obtained. The best mode is determined corresponding to the mode exhibiting the minimum SAE cost. One of the main advantages of this method is its low computational cost. On the contrary, the Rate-Distortion performance results are sub-optimal. 2.3 The Fast Motion Estimation Option Motion estimation is one of the most important tools in H.264 encoder for exploiting the high temporal redundancy between successive frames to improve video coding efficiency. And motion estimation is also the most time consuming part in the H.264 encoder (since it is also used for mode decision). Generally motion estimation is conducted into two steps: first is integer pel motion estimation; and the second is fractional pel motion estimation around the position obtained by the integer pel motion estimation. Algorithms on Fast Motion Estimation (FME) are always hot research spot, especially fast integer pel motion estimation has achieved much more attention because traditional fractional pel 933 motion estimation only take a very few proportion in the computation load of whole motion estimation. Fast motion estimation algorithms such as EPZS [15], UMHexagonS [16], and SEA [17] have been proposed to reduce the number of searching points in motion estimation. The UMHexagonS algorithm proposed by Tsinghua University was adopted by the H.264/MPEG-4 Part 10 (AVC) reference software implementation [12]. This algorithm uses the hybrid and hierarchical motion search strategies. It includes four steps with different kinds of search pattern: 1) Predictor selection and prediction mode reordering; 2) Unsymmetrical-cross search; 3) Uneven multi-hexagon-grid search; 4) Extended hexagon-based search. With the second and third step, the motion estimation accuracy can be nearly as high as that of full search. But the computation load and operations can be reduced even more. Unsymmetrical-cross search uses prediction vector as the search center and extends in the horizontal and vertical directions respectively. Uneven multi-hexagon-grid search includes two sub-steps : first a full search is carried out around the search center. And then a 16-HP multi-hexagon-grid search strategy is taken. Extended hexagon-based search is used as a center based search algorithm, including hexagon search and diamond search in a small range. In the H.264 reference software, the Fast Motion Estimation (FME) algorithm (based in the UMHexagonS algorithm) can be employed for the motion estimation in addition to the original Full Search (FS) algorithm. MACHINE LEARNING Machine learning refers to the study of algorithms and systems that "learn" or acquire knowledge from experiences. Deductive machine learning deduces new rules/knowledge from existing rules and inductive machine learning uses the analysis of data sets for creating a set of rules to take decisions. These rules can be used, in the machine learning, to build a tree decision using a set of experiments or examples, named the training data set. This set of data must have the following properties [18]: 1. Each attribute or variable can take nominal or numerical values, but the number of attributes cannot vary from an example to another. This is to say, all the samples in the training data set used for training the model must have the same number of variables. 2. The set of categories that the examples can be assigned to must a priori be known to enable supervised learning. 3. The set of categories must be finite and must be different from one another. 4. Since the inductive learning consists of obtaining generalization from examples, it is supposed the existence of a sufficiently great number of examples. Machine learning uses statistics with different kinds of algorithms to solve a problem by studying and analyzing the data. Machine learning has been used in an extensive range of applications including search engines, medical diagnosis, stock market analysis, classifying DNA sequences, speech and handwriting recognition, object recognition in computer vision, game playing and robot motion, etc. In this paper, we describe the process of using machine learning to build a decision tree for very low complexity transcoding. The decision tree will be used to determine the coding mode of an MB in P frames of the output H.264 video, based on the information gathered during the MPEG-2 decoding stage. Figure 4 depicts the process for building the decision trees to be used in the MPEG-2 to H.264 transcoding process. The incoming MPEG-2 video is decoded and during the decoding stage, the MB coding mode, the coded block pattern (CBPC), and the mean and variance of the residual information for this MB (calculated for its 4x4 sub-blocks resulting in 16 means and 16 variances for each MB) are saved. The decoded MPEG-2 video is then encoded using the standard H.264 encoder. The coding mode of the corresponding MBs in H.264 is also saved. Based on the MPEG-2 data and the corresponding H.264 coding mode decision for each MB, a machine learning algorithm is used to create decision trees that classify an MB into one of the 11 H.264 MB coding modes. Figure 4. Process for building decision trees for MPEG-2 to H.264 transcoding. 3.1 Creating the Training Files A decision tree is made by mapping the observations about a set of data to a tree made of arcs and nodes. The nodes are the variables and the arcs the possible values for that variable. The tree can have more than one level; in that case, the nodes (leafs of the tree) represent the decisions based on the values of the different variables that drive the decision from the root to the leaf. These types of trees are used in the machine learning processes for discovering the relationships in a set of data. The tree leafs are the classifications and the branches are the features that lead to a specific classification. A tree decision is a classifier based on a set of attributes allowing us to determine the category of an input data sample. The decision tree for the transcoder was made using the WEKA data mining tool [18]. The files that are used for the WEKA data mining program are known as Attribute-Relation File Format (ARFF) files. An ARFF file is written in ASCII text and shows the relationship between a set of attributes. Basically, this file has two different sections:1) the header which contains the name of the relation, the attributes that are used, and their types; and 2) the section containing the data. The training sets were made using MPEG-2 sequences encoded at higher than the typical broadcast encoding rates for the same quality, since the B frames are not used. The H.264 decisions in the training set were obtained from encoding the MPEG-2 934 decoded sequence with a quantization parameter of 25 and RD optimization enabled. After extensive experimentation, we found that sequences that contain regions varying from homogenous to high-detail serve as good training sets. Good sample sequences could be Flower and Football. The goal is to develop a single, generalized, decision tree that can be used for transcoding any MPEG-2 video. Figure 5 shows the decision trees built using the process depicted in Figure 4. As shown in Figure 4, the Decision Tree for the proposed transcoder is a hierarchical decision tree with three different WEKA trees: 1) classifier for Intra, Skip, Inter 16x16, and Inter 8x8, 2) classifier to classify inter 16x16 into one of 16x16, 16x8, and 8x16 MBs and 3) classifier to classify inter 8x8 into one of 8x8, 8x4, 4x8, or 4x4. This paper focuses on the Inter MB mode computation and the further classification and processing for Intra MBs is not discussed in this paper. For creating the first WEKA tree (Figure 5 node 1), the first training data set uses the mean and variance of each one of the sixteen 4x4 residual sub-blocks, the MB mode in MPEG-2 (skip, intra, and three non-intra modes, labeled as 0, 1, 2, 4 and 8 in the code shown below), the coded block pattern (CBPC) in MPEG-2, and the corresponding H.264 MB coding mode decision for that MB as determined by the standard reference software. The header section of the ARFF files has the attribute declaration depicted herein: The supposed dependent variable, namely class in the example, is the variable that we are trying to understand, classify, or generalize. The other attributes are the variables that determine the classification. The ARFF data section has the instance lines, which are the samples used to train our model. Each macroblock sample is represented on a single line. In this case the variable class can take four values (skip, 16x16, 8x8 or Intra labeled as 0, 1, 8 and 9 in the code). The second training data set, used for creating the second WEKA tree (Figure 5 node 2), was made using the samples (MBs) that were encoded as 16x16 MBs in the H.264 reference encoder. It uses the mean and variances of each one of the sixteen 4x4 residual sub-blocks, the MB mode in MPEG-2 (in this case only the three non-intra modes), the coded block pattern (CBPC) in MPEG-2, and the corresponding H.264 MB coding sub-mode decision in the 16x16 mode, as determined by the standard reference software: 16x16, 16x8 or 8x16. This tree determines the final coding mode of the MBs classified as inter 16x16 by the first tree. The third and last training data set, was used to create the third WEKA tree (Figure 5 node 3) and was made using the samples (MBs) that were encoded as inter 8x8 MBs in the H.264 reference encoder. It uses four means and four variances of 4x4 residual sub-blocks, the MB mode in MPEG-2 (the three non-intra modes), the coded block pattern (CBPC) in MPEG-2, and the corresponding H.264 MB sub-partition decision in the 8x8 mode, as determined by the standard reference software: 8x8, 8x4, 4x8 or 4x4. Since this decision is made separately for each 8x8 sub-block , only the four means and four variances of 4x4 residual sub-blocks are used in each sample for training the model. Based on these training files, the J48 algorithm implemented in the WEKA data mining tool was used to create the three decision trees. The J48 algorithm is an implementation of the C4.5 algorithm proposed by Ross Quinlan [19]: the algorithm widely used as a reference for building decision trees. The decision tree, that is proposed to solve the inter-prediction problem, is a model of the data that encodes the distribution of the class label in terms of the attributes. The final goal of this decision tree is to help find a simple structure to show the possible dependences between the attributes and the class. 3.2 The Decision Tree This sub-section discusses the proposed macroblock mode decision algorithm aiming to accelerate the inter-frame prediction. This goal is achieved by making use of the MPEG-2 MB coding modes, the coded block pattern (CBPC), and the mean and variance of the residual information for this MB calculated for its 4x4 sub-blocks. MPEG-2 uses 16x16 motion compensation (MC) and does not temporally decorrelate an image fully. The MC residual can thus be exploited to understand the temporal correlation of variable block sizes in H.264. The open source WEKA data mining tool is used to discover a pattern of the mean, variance, MPEG-2 coding modes, and the coded block pattern in MPEG-2 (CBPC) for H.264 coding mode decisions. Figure 5 shows the decision tree used in the proposed transcoder. The decision tree consists of three WEKA decision trees, shown in Figure 5 with grey balls. The first WEKA tree is used to check for the skip, Intra, 8x8 and 16x16 MBs modes. If an MB is 8x8 or 16x16, a second and a third decision tree is used for selecting the final coding mode of the MB. The WEKA tool determined the mean and variance thresholds for each of the three WEKA trees in the decision tree. Due to space constraints we cannot show all the rules being evaluated in the WEKA decision nodes. The process described in herein should be sufficient for interested people to develop the decision trees and repeat these experiments. The decision tree works as follows: Node 1. The inputs for this node are all the MPEG-2 coded MBs. In this node a tree decision generated with WEKA is used to decide whether the MB should be coded in H.264. This tree examines whether the MB has a very high residual or a medium residual. The output of this node is a first level decision mode that should be used for coding the MB: skip, Intra, 8x8 or 16x16. The intra decision process is not discussed in this paper. In the other cases, the algorithm has to make a second level decision based in the first decision. For example, the following rules were given by WEKA: If the MPEG-2 MB was "MC not coded", (non-zero MV present, none of the 8x8 block has coded coefficients), then @RELATION mean-variance_4x4 @ATTRIBUTE mean0 Numeric @ATTRIBUTE variance0 Numeric @ATTRIBUTE mean1 Numeric @ATTRIBUTE variance1 Numeric ................................................................................. @ATTRIBUTE mean15 Numeric @ATTRIBUTE variance15 Numeric @ATTRIBUTE mode_mpeg2 {0,1,2,4,8} @ATTRIBUTE CBPC0 {0,1} ................................................................................. @ATTRIBUTE CBPC6 {0,1} @ATTRIBUTE class {0,1,8,9} 935 the MB will be coded as 16x16 in H.264. Again, a second decision level will be made to select the best choice in this case (see node 2). If the MPEG-2 MB was coded in intra mode, the MB will be coded as intra or inter 8x8 mode in H.264. In some cases the algorithm will propose Intra, and the algorithm will end, and in other cases the algorithm will propose 8x8 mode, so a second level decision will be done (see node 3). If the MPEG-2 MB was coded in skip mode, then the H.264 decision mode should be skip. The decision will be made in node 4. Figure 5. The Decision Tree. Node 2. The inputs for this node are the 16x16 MBs classified by the node 1. In this node we use again a decision tree generated with WEKA to decide whether the MB should be coded in H.264 (16x16, 16x8 or 8x16). This tree examines if there are continuous 16x8 or 8x16 sub-blocks that might result in a better prediction. The output of this node is the 16x16 sub-mode decision mode that should be used for coding the MB: 16x16, 16x8 or 8x16. When the node decision is 16x8 or 8x16 the coding mode is finalized. In the other case, the evaluation continues in node 4, where the final decision will be made. Node 3. The inputs for this node are the MBs classified by the node 1 as 8x8. This node evaluates only the H.264 8x8 modes using the third WEKA tree and selects the best option: 8x8, 8x4, 4x8 or 4x4. As explained in the previous section, this tree is run 4 times, once for each of the four sub-macroblocks in the MB. This tree is different from the others because this one only uses four means and four variances to make the decision. Node 4. The inputs for this node are skip-mode MBs in the MPEG-2 bitstream classified by the node 1, or the 16x16 MBs classified by the node 2. This node evaluates only the H.264 16x16 mode (without the sub-modes 16x8 or 8x16). Then, the node selects the best option, skip or inter 16x16. Since the MB mode decision, and hence the thresholds, depend on the quantization parameter (QP) used in the H.264 encoding stage, the mean and variance threshold will have to be different at each QP. The two solutions here are: 1) develop the decision trees for each QP and use the appropriate decision tree depending on the QP selected and 2) develop a single decision tree and adjust the mean and variance threshold used by the trees based on the QP. The first option is complex as we have to develop and switch between 52 different decision trees resulting in 156 WEKA trees in a transcoder. Since the QP used by H.264 is designed to change the quantization step size and the relationship between the QPs is well defined, this relationship can be used to adjust the mean and variance thresholds. The proposed transcoder uses a single decision tree developed for a mid-QP of 25 and then adjusted for other QPs. Since the quantization step size in H.264 doubles when QP increases by 6, the thresholds are adjusted by 2.5% for a change in QP of 1. For QP values higher than 25, the thresholds are decreased and for QP values lower than 25 thresholds are proportionally increased. Figure 6 shows an example of the results obtained by applying our proposed algorithm. Figure 6a illustrates the residual for the MPEG-2 encoded Tempete sequence. Figures 6b and 6c show the mean and variance of the residual. Figures 6.e and 6.f show the differences between the inter mode selection made by the H.264 standard (with the RD-optimized option enabled), and the proposed algorithm, with a value of 10 for QP. From these figures, it is clear that our algorithm obtains very similar results to those obtained using the full estimation of the H.264 standard. (a) MPEG-2 residual (+128) (b) Mean of the MPEG-2 residual (+128) (c) Variance of the MPEG-2 residual (d) Different kinds of Macroblocks in the grid pictures (e) H.264 Rd opt, first frame P, Tempete (CIF) QP= 10. Inter mode selected by H.264 (f ) H.264 Rd opt, first frame P, Tempete (CIF) QP= 10. Inter mode selected by our proposal Inter 16x16 Macroblock Skip Macroblock Intra Macroblock Inter 8x16 Macroblock Inter 16x8 Macroblock Inter 8x8 Macroblock Inter 4x8 Sub-macroblock Inter 8x4 Sub-macroblock Inter 4x4 Sub-macroblock Inter 8x8 Sub-macroblock Figure 6. Macroblock partitions generated by the proposed algorithm for the first P-frame in the Tempete sequence. PERFORMANCE EVALUATION The proposed low complexity MB coding mode decision algorithm is implemented in the H.264/AVC reference software, version JM 10.2 [12]. Figure 7 shows the overall operation of the proposed transcoder. The MPEG-2 video is decoded and the information required by the decision trees is gathered in this stage. The additional computation here is the computation of the mean and variance of the 4x4 sub-blocks of the residual MBs. The MB coding mode decision determined by the decision trees is used in the low complexity H.264 encoding stage. This is an 936 H.264 reference encoder with the MB mode decision replaced by simple mode assignment from the decision tree. The H.264 video encoder takes as input the decoder MPEG-2 video (pixel data) and the MB mode decision from the decision tree and encodes the H.264 video. The MPEG-2 motion vectors are not used and the encoder performs the motion estimation just for the final MB mode determined by the decision tree. MPEG-2 Video H.264 Video Figure 7. Proposed transcoder. The performance of the proposed very low complexity transcoder is compared with a reference transcoder comprised of a full MPEG-2 decoder followed by a full H.264 encoder. We compare the performance of our proposal to the full H.264 encoder when the RD-cost (with and without FME option enabled) and the SAE-cost (with and without FME option enabled) are used. The metrics used to evaluate the performance are the reduction in the computational cost and rate distortion function. The time results reported are for the H.264 encoding component as the MPEG-2 decoding cost is the same for both the proposed and reference encoders. We have conducted an extensive set of experiments with videos representing wide range of motion, texture, and color. Experiments were conducted to evaluate the performance of the proposed algorithm when transcoding videos at commonly used resolutions: CCIR-601, CIF, and QCIF. The input to the transcoder is a high quality MPEG-2 video. Since the proposed transcoder addresses transcoding P frames in MPEG-2 to H.264 P frames, MPEG-2 bitstreams were created without B frames. Since the B frames, which are much smaller than P frames, are not used in the input video, the video has to be encoded at higher than the typical encoding rates for equivalent broadcast quality. Table 1 shows the bitrates used for the input MPEG-2 video. The experiments have shown that the proposed approach performs extremely well across all bitrates and resolutions. Table 1. Bitrates for the input sequences Format Bitrate CCIR-601 (720x480) 5 Mbps CIF (352x288) 1.15 Mbps QCIF (176x144) 0.768 Mbps The sequences have been encoded with H.264 using the QP factors ranging from 5 up to 45 in steps of 5. This corresponds to the H.264 QP range used in most practical applications. The size of the GOP is 12 frames; where the first frame of every GOP was encoded as I-frame, and the rest of the frames of the GOP were encoded as a P-frames. The rate control was disabled for all the simulations. The ProfileIDC was set to High for all the simulations, with the FRExt options enabled. The simulations were run on a P4 HT at 3.0 GHz Intel machine with 512 MB RAM. The results are reported for six different sequences: two for each of the three resolutions shown in Table 1. CCIR Sequences (720x480, 200 Frames, 25 Hz) 30 35 40 45 50 55 60 0 5000 10000 15000 20000 25000 30000 35000 Bit rate [kbits/s] PSN R [ d B ] H.264 (Rd opt) Proposed (Rd opt) Ayersroc Martin (a) CIF Sequences (352x288, 200 Frames, 25 Hz) 30 35 40 45 50 55 60 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 Bit rate [kbits/s] PSN R [d B] H.264 (Rd opt) Proposed (Rd opt) Paris Tempete (b) QCIF Sequences (176x144, 200 Frames, 25 Hz) 30 35 40 45 50 55 60 0 500 1000 1500 2000 2500 3000 Bit rate [kbits/s] P S NR [db ] H.264 (Rd opt) Proposed (Rd opt) Foreman News (c) Figure 8. RD results for RD-cost without FME option. Figure 8 shows the RD results for the reference and proposed transcoder with RD optimization enabled and fast motion estimation (FME) disabled. Figure 9 shows the RD results for the reference and proposed transcoder with RD optimization enabled and fast motion estimation (FME) enabled. As seen from the figures, the PSNR obtained with the proposed transcoder deviates slightly from the results obtained when applying the considerable 937 more complex reference transcoder. Compared with the reference transcoder, the proposed transcoder has a PSNR drop of at most 0.3 dB for a given bitrate and bitrate increase of at most 5% for a given PSNR. This negligible drop in RD performance is more then offset by the reduction in computational complexity. Tables 2 and 3 show the average encoding time per frame given in milliseconds. As shown in Table 2 and Table 3, the transcoding time reduces by more than 80% with RD optimization, and more than 90% with FME enabled for both the reference and proposed transcoders. CCIR Sequences (720x480, 200 Frames, 25 Hz) 30 35 40 45 50 55 60 0 5000 10000 15000 20000 25000 30000 35000 Bit rate [kbits/s] P S N R [ d B] H.264 (Rd opt, Fast ME) Proposed (Rd opt, Fast ME) Ayersroc Martin (a) CIF Sequences (352x288, 200 Frames, 25 Hz) 30 35 40 45 50 55 60 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 Bit rate [kbits/s] P S NR [ d B] H.264 (Rd opt, Fast ME) Proposed (Rd opt, Fast ME) Paris Tempete (b) QCIF Sequences (176x144, 200 Frames, 25 Hz) 30 35 40 45 50 55 60 0 500 1000 1500 2000 2500 3000 Bit rate [kbits/s] P S NR [ db] H.264 (Rd opt, Fast ME) Proposed (Rd opt, Fast ME) Foreman News (c) Figure 9. RD results for RD-cost with FME option. Figure 10 shows the RD results for the reference and proposed transcoder with SAE-Cost (RD optimization disabled) and fast motion estimation (FME) disabled. Figure 11 shows the RD results for the reference and proposed transcoder with SAE-Cost (RD optimization disabled) and fast motion estimation (FME) enabled. As seen from the figures, in some cases the proposed transcoder have better results than the reference transcoder. This happens because the best solution is obtained by enabling the RD optimization, and in the experiments reported in the figures we are comparing the faster configuration of a H.264 encoder (SAE cost) with our proposed reduced-complexity transcoder. With SAE based encoding (RD-optimization disabled), the proposed transcoder continues to outperform the reference transcoder computationally (Tables 2 and 3). The transcoder still maintains a PSNR drop of less than 0.3 dB and bitrate increase of less than 5%. The computational cost is reduced by over 38% for the SAE case and by over 82% with FME enabled for both the reference and proposed transcoders. Table 2. Mean encoding time (milliseconds) per frame with the reference transcoder Sequence RD Opt RD Opt + FME SAE SAE + FME Martin 7370 6420 2110 940 Ayersroc 7650 6820 2095 1030 Paris 2305 2020 590 235 Tempete 2360 2050 605 290 Foreman 565 495 155 68 News 550 470 150 55 Table 3. Mean encoding time (milliseconds) per frame with the proposed transcoder Sequence RD Opt RD Opt + FME SAE SAE + FME Martin 1460 470 1190 170 Ayersroc 1620 670 1160 190 Paris 415 95 360 45 Tempete 445 135 360 53 Foreman 102 24 93 12 News 103 21 92 11 Table 4. Mean Time Reduction (%) per frame with the proposed transcoder Sequence RD Opt RD Opt + FME SAE SAE + FME Martin 80,19 92,68 43,60 81,91 Ayersroc 78,82 90,18 44,63 81,55 Paris 82,00 95,30 38,98 80,85 Tempete 81,14 93,41 40,50 81,72 Foreman 81,95 95,15 40,00 82,35 News 81,27 95,53 38,67 80,00 Based on the results shown in the Tables 2 and 3, the proposed transcoder with SAE and FME has the lowest complexity. The proposed transcoder with RD optimization and FME is still faster than the fastest case of the reference transcoder (SAE + FME). Using FME reduces the complexity substantially. Selecting RD optimization with the proposed transcoder doubles the complexity compared with SAE+FME case. The decision to enable RD optimization can be based on the operating bitrates and sensitivity to the PSNR drop. At higher bitrates, RDOPT + FME option give about 0.6 dB better than the SAE + FME option; this is doubling 938 the complexity for a gain of 0.6 dB. However, at lower bitrates, the PSNR gain reduces to about 0.3 dB. CCIR Sequences (720x480, 200 Frames, 25 Hz) 30 35 40 45 50 55 60 0 5000 10000 15000 20000 25000 30000 35000 Bit rate [kbits/s] P S NR [ d B] H.264 (SAE) Proposed (SAE) Ayersroc Martin (a) CIF Sequences (352x288, 200 Frames, 25 Hz) 30 35 40 45 50 55 60 0 2000 4000 6000 8000 10000 12000 Bit rate [kbits/s] P S N R [dB] H.264 (SAE) Proposed (SAE) Paris Tempete (b) QCIF Sequences (176x144, 200 Frames, 25 Hz) 30 35 40 45 50 55 60 0 500 1000 1500 2000 2500 3000 Bit rate [kbits/s] PS N R [ d b ] H.264 (SAE) Proposed (SAE) Foreman News (c) Figure 10. RD results for SAE-cost without FME option. Table 4 summarizes the reduction in the computational cost due to the proposed machine learning based mode decision algorithms in the proposed transcoder. With RD optimization and FME, the computational cost is reduced by over 90%. The cost reduction reaches as high as 95.5% for QCIF sequences. With SAE and FME, the computational cost is reduces by over 80%. The computational cost reduction come at a cost of reduced quality. The quality reduction, however, is very small and negligible for most video applications. Table 5 shows the quality variation versus time reduction of the proposed transcoder with respect the reference transcoder for the same input bitrates shown in Table 1, showing over 96% reduction in the computational complexity characterizing our proposed scheme. As shown in the table, using the proposed transcoder reduces the PSNR by at most 0.3dB with RD optimization enabled and by at most 0.1 dB with SAE cost based transcoder. Our results show that the proposed algorithm is able to maintain a good picture quality while considerably reducing the number of operations to be performed in all the scenarios. CCIR Sequences (720x480, 200 Frames, 25 Hz) 30 35 40 45 50 55 60 0 5000 10000 15000 20000 25000 30000 35000 Bit rate [kbits/s] P S NR [ d B ] H.264 (SAE, Fast ME) Proposed (SAE, Fast ME) Ayersroc Martin (a) CIF Sequences (352x288, 200 Frames, 25 Hz) 30 35 40 45 50 55 60 0 2000 4000 6000 8000 10000 12000 Bit rate [kbits/s] PSN R [d B ] H.264 (SAE, Fast ME) Proposed (SAE, Fast ME) Paris Tempete (b) QCIF Sequences (176x144, 200 Frames, 25 Hz) 30 35 40 45 50 55 60 0 500 1000 1500 2000 2500 3000 Bit rate [kbits/s] P S N R [ db] H.264 (SAE, Fast ME) Proposed (SAE, Fast ME) Foreman News (c) Figure 11. RD results for SAE-cost with FME option. 939 Table 5. Quality Variation vs Time Reduction (for transcoding rate) Quality Variation from Reference Transcoder (dB) Time Reduction from Reference Transcoder (%) Sequence MPEG-2 Bit Rate (Mbps) RD OPT RD FME SAE SAE FME RD OPT RD FME SAE SAE FME Ayersroc 5.0 - 0.3 - 0.3 0.0 - 0.1 80.0 90.5 43.3 82.3 Martin 5.0 - 0.2 - 0.2 - 0.1 - 0.1 80.5 92.8 42.1 82.0 Tempete 1.15 - 0.2 - 0.2 0.0 0.0 80.0 93.8 41.1 82.5 Paris 1.15 - 0.3 - 0.3 0.0 - 0.1 81.6 95.6 38.5 80.7 Foreman 0.768 - 0.3 - 0.3 0.0 0.0 83.5 95.5 37.4 82.6 News 0.768 - 0.2 - 0.2 0.0 0.0 84.1 96.0 35.1 81.1 CONCLUSIONS In this paper, we proposed a novel macroblock partition mode decision algorithm for inter-frame prediction to be used as part of a high-efficient MPEG-2 to H.264 transcoder. The proposed algorithms use machine learning techniques to exploit the correlation in the MPEG-2 MC residual and the H.264 coding modes. The WEKA tool was used to develop decision trees for H.264 coding mode decision. The proposed algorithm has very low complexity as it only requires the mean and variance of the MPEG-2 residual and a set of rules to compare the mean and variance against a threshold. The proposed transcoder uses a single decision tree with adaptive thresholds based on the quantization parameter selected in the H.264 encoding stage. The proposed transcoder was evaluated using MPEG-2 videos at CCIR, CIF, and QCIF resolutions. Our results show that the proposed algorithm is able to maintain a good picture quality while considerably reducing the computational complexity by as much as 95%. The reduction in computational cost has negligible impact on the quality and bitrate of the transcoded video. The results show that the proposed transcoder maintains its performance across all resolutions and bitrates. The proposed approach to transcoding is novel and can be applied to develop other transcoders as well. Our future plans will focus on further reducing the complexity of the proposed transcode by reusing the MPEG-2 motion vectors followed by a motion vector refinement. By reusing the motion vector, we believe, real-time transcoding of CIF resolution video at 30 FPS is within reach. REFERENCES [1] ITU-T RECOMMENDATION H.264 "Advanced Video Coding for Generic Audiovisual Services". May 2003. [2] Implementation Studies Group, "Main Results of the AVC Complexity Analysis". MPEG Document N4964, ISO/IEC JTC11/SC29/WG11, July 2002. [3] T. Shanableh and M. Ghanbari, "Heterogeneous Video Transcoding to Lower Spatio-Temporal Resolutions and Different Encoding Formats," IEEE Transactions on Multimedia, vol.2, no.2, June 2000. [4] A. Vetro, C. Christopoulos, and H.Sun "Video Transcoding Architectures and Techniques: An Overview". IEEE Signal Processing Magazine, vol. 20, no. 2, pp.18-29, March. 2003. [5] H. Kalva, A. Vetro, and H. Sun, "Performance Optimization of the MPEG-2 to MPEG-4 Video Transcoder". Proceeding of SPIE Conference on Microtechnologies for the New Millennium, VLSI Circuits and Systems, May 2003. [6] S. Dogan, A.H. Sadka and A.M. Kondoz, "Efficient MPEG-4/H .263 Video Transcoder for Interoperability of Heterogeneous Multimedia Networks," IEE Electronics Letters, Vol. 35, No.11. pp. 863-864. [7] H. Kalva. &quot;Issues in H.264/MPEG-2 Video Transcoding&quot;. Proceedings of Consumer Communications and Networking Conference, January 2004. [8] Y. Su, J. Xin, A. Vetro, and H. Sun, "Efficient MPEG-2 to H.264/AVC Intra Transcoding in Transform-Domain," IEEE International Symposium on Circuits and Systems, 2005. ISCAS 2005. pp. 1234- 1237 Vol. 2, 23-26 May 2005. [9] B. Petljanski and H. Kalva, &quot;DCT Domain Intra MB Mode Decision for MPEG-2 to H.264 Transcoding&quot; Proceedings of the ICCE 2006. January 2006. pp. 419-420. [10] Y.-K. Lee, S.-S. Lee, and Y.-L. Lee, "MPEG-4 to H.264 Transcoding using Macroblock Statistics," Proceedings of the ICME 2006, Toronto, Canada, July 2006. [11] X. Lu, A. M. Tourapis, P. Yin, and J. Boyce, "Fast Mode Decision and Motion Estimation for H.264 with a Focus on MPEG-2/H.264 Transcoding," Proceedings of 2005 IEEE International Symposium on Circuits and Systems (ISCAS), Kobe, Japan, May 2005. [12] Joint Video Team (JVT) of ISO/IEC MPEG and ITU-T VCEG, Reference Software to Committee Draft. JVT-F100 JM10.2. 2006. [13] G. Sullivan and T. Wiegand, "Rate-Distortion Optimization for Video Compression," IEEE Signal Processing Magazine, vol. 15, no. 6, pp. 74-90, November. 1998. [14] T. Wiegand et al., "Rate-Constrained Coder Control and Comparison of Video Coding Standards," IEEE Transactions on Circuits Systems and Video Technology, vol. 13, no. 7, pp. 688-703 , July 2003. [15] A.M. Tourapis, O.C. Au, M.L. Liou, "Highly Efficient Predictive Zonal Algorithms for Fast Block-Matching Motion Estimation," IEEE Transactions on Circuits and Systems for Video Technology, Vol. 12, Issue 10, Oct. 2002. [16] Z. Chen, P. Zhou, and Y. He, "Fast Integer Pel and Fractional Pel Motion Estimation for JVT", 6th Meeting. Awaji, December 2002 [17] M. Yang, H. Cui, K. Tang, "Efficient Tree Structured Motion Estimation using Successive Elimination," IEE Proceedings-Vision , Image and Signal Processing, Vol. 151, Issue 5, Oct. 2004. [18] Ian H. Witten and Eibe Frank, "Data Mining: Practical Machine Learning Tools and Techniques", 2nd Edition, Morgan Kaufmann, San Francisco, 2005. [19] J.R. Quinlan, "C4.5: Programs for Machine Learning", Morgan Kaufmann, 1993. 940

H.264;Inter-frame;Machine Learning;Transcoding;MPEG-2

212

Video-Streaming for Fast Moving Users in 3G Mobile Networks

The emergence of third-generation (3G) mobile networks offers new opportunities for the effective delivery of data with rich content including multimedia messaging and video-streaming. Provided that streaming services have proved highly successful over stationary networks in the past, we anticipate that the same trend will soon take place in 3G networks. Although mobile operators currently make available pertinent services, the available resources of the underlying networks for the delivery of rich data remain in-herently constrained. At this stage and in light of large numbers of users moving fast across cells, 3G networks may not be able to warrant the needed quality-of-service requirements. The support for streaming services necessitates the presence of content or media servers properly placed over the 3G network; such servers essen-tially become the source for streaming applications. Evidently, a centralized approach in organizing streaming content might lead to highly congested media-nodes which in presence of moving users will certainly yield increased response times and jitter to user requests . In this paper, we propose a workaround that enables 3G networks to offer uninterrupted video-streaming services in the presence of a large number of users moving in high-speed. At the same time, we offer a distributed organization for the network's media-servers to better handle over-utilization.

INTRODUCTION The third generation ( 3G) mobile phone system UMTS enables better quality and allows for more convenient use of multimedia messaging and video-services by offering higher bandwidth and lower latency than its GSM and GPRS predecessors [15, 1]. UMTS furnishes upto 2Mbps rates for indoor and upto 384Kbps for outdoor environments. Clearly, much improvement in terms of allo-cated resources has been made for the handling of "rich" data including multimedia messages and video-services. Nevertheless, the available resources still present significant limitations for the scale up of services when very large number of clients are present in a cell of a network. Perhaps, the most daunting challenge comes from moving users who access multimedia data and video feeds with the help of their mobile phones and PDAs while traveling on either private vehicles or mass transportation means such as commuter trains and public busses. Evidently, a large number of concurrent connections soliciting data resources in a cell and being handled in real-time pose significant capacity problems for the underlying 3G network. The situation becomes even more challenging when users attempt to follow streaming sources while on the move. We consider streaming services to be of key importance as they will ultimately offer information on demand for the moving user in any geographic position and at any time. In order to facilitate video streaming over UMTS networks a number of issues have to be addressed so that users do not experience delays and discontinuities during playback. The two core aspects that require attention are the variations of the available bandwidth as users enter and leave cells as well as the effective management of handovers as roaming users attach to different base-stations along their trajectory. The problem of graceful transition when moving between base-stations becomes more critical when the users are on high-speed motorways . In this case, handovers become more frequent and traffic load at successive base-stations may vary. In this paper, we outline the above emerging problem and propose a scheme that allows for not only improved network resource sharing but also for enhanced management of streaming sources to the mobile user. It is expected that base-stations are to transmit in different bitrates throughout the journey of an individual as cells will undoubtedly present diverse levels of congestion and availability of connections. When considering vehicular users in general, one can exploit the fact that the user's trajectory can be predicted in a satisfactory manner . An early method to attain this goal is to keep an aggregate history of observations made regarding the movement of users within each cell [9]. Based on this information, probability density functions for the prediction for the next-cell-to-move can be derived and used. Traffic authorities imposed speed limits and road signals can 65 also assist in the more accurate estimation of a user's average speed. In addition, the user's direction can be predicted reasonably well by keeping track of her trajectory thus far. Although a model for precise prediction is beyond the scope of this paper, we can assume that there are already techniques that can offer a good estimation of a moving user path. For instance, an individual's geographic location could be tracked with the assistance of UMTS Location Service (LCS) [2] that can identify the cell that a user presently appears in. If a user is moving along a highway, one could easily estimate not only the direction of his movement but also his average speed along a given trajectory. Finally, the soon anticipated incorporation of Global Positioning System (GPS) receivers into mobile phones through A-GPS features [30] will help in the very accurate user positioning and extraction of their movement characteristics. It is our conjecture that at this stage, simply knowing the overall direction of a user's trajectory in conjuction with the highway that she travels on can ensure timely video-streaming and playout continuity for users. In our streaming environment, there exist three distinct types of synergistic computing systems: media-servers, base-stations, and user equipment. These systems are organized in three functional layers as Figure 1 depicts. The role of the media-servers is to predominantly manage content in a highly distributed fashion where fast content retrieval and data resilience can be guaranteed. Base-stations handle all user initiated connections to the 3G network and through their channels offer users requested data. The last tier of Figure 1 consists of cellular phones and PDAs equipped with appropriate video-players and featuring minimal buffer capabilities to support streaming. Base Station Media Server Base Station Base Station Media Server Base Station WCDMA radio interface Figure 1: Three-Tier Organization for Streaming Provision. Media and base-stations are internetworked via high-speed wired links while UMTS offer wireless connections between user equipment and base-stations. This distributed media-server architecture provides for dividing of large video streams into multiple segments [34, 14]. A media-server initially retrieves a solicited video object either from its storage units or from another remote media-server . In this paper, we take the approach that instead of transmitting the entire object to a single base-station, we first segment it and then forward its segments to the base-stations along the user's path. Our rationale is that an individual base-station handles only a section of the video file; the size of the section in discussion is commensurate to the duration of a user's trip inside the cell. Clearly, video-object segmentation reduces both network transmission costs between media and base-stations and start-up latencies that users experience upon cell entrance. On the other hand, segmentation might get more complex if a user remains longer in a cell than her estimated time and so she may face delays in the reception of frames. We address this issue by continual monitoring of both user speed and position and by doing so giving the base-station the option to receive additional video increments and sufficiently feed a user at all times. Our work requires minimal buffer capabilities for mobile stations so that a sufficient number of frames can be accommodated. Buffer presence assures that the playout does not stop if the base-station emits at lower bitrate due to 3G network congestion. We propose a rate adaptation scheme which allows a base-station to adjust its transmitting bitrate at any time according to base-station load and the states of the client-side buffer. The UMTS streaming class defines that the bitrate assigned to a moving user is guaranteed even though it might be less than the maximum video bitrate [15]. The base-station's decision of accepting a video streaming session has an impact on all subsequent base-stations that follow up on the video delivery process. While a streaming session is in progress, the load of base-stations-to-service may dynamically change and potentially lead to session drops. Such drops are highly undesirable and we adopt a policy to address this issue. Our proposed scheme gives a base-station the opportunity to appropriately alter the transmission bitrate by taking into account the current base-station load and simultaneously ensuring that the client buffer does not starve. The rest of the paper is organized as follows: Section 2 presents the overall system architecture and examines the interaction between media-servers and base-stations. Section 3 proposes our bitrate adaptation scheme and Section 4 discusses the results of our preliminary experimentation. Finally, related work and conclusions can be found in Sections 5 and 6 respectively. MEDIA-SERVERS/BASE-STATIONS INTERACTION The fast movement of users via different cells of the 3G network imposes a set of new requirements for the entire video delivery system. As the user relocates rapidly, she faces a high number of handovers during her journey and, as a consequence, a large video stream has to be fetched from different base-stations in order to warrant continuous playback. As suggested earlier, we assume the deployment of dedicated media-servers which undertake both the storage and distribution of video-streams to underline base-stations . It is imperative that media-servers, base-stations, and user-equipment involved in a streaming session must cooperate in order to guarantee QoS for the video reception of the moving individual . In this section, we outline our overall architecture, discuss the content delivery process that media-servers carry out, and present specific algorithms for video segmentation and content distribution. 2.1 Architecture The three distinct types of cooperative computing systems (shown in Figure 1) organized in a multi-tier architecture constitute our proposed streaming environment. We assume that base-stations communicate with the mobile stations through the WCDMA radio interface [15]. Each cell of the UMTS network is served by a different base-station whose responsibility is to deliver the video streams to its constituent mobile users. A streaming service necessitates the use of media servers that will handle the storage and delivery of video files [3]. Although we could adopt a centralized approach to accommodate the media in delivery, high contention and resource over-utilization would impact user request response times greatly. Clearly a distributed approach that webs media-servers together is required. High-speed wired communication means link these servers and all share required meta-data structures. 66 Due to incurred costs and the fact that users move at high-speeds, having a dedicated media server for each base-station would be a poor decision. If the mobile user is traveling at a speed of 100 km/h and the cell radius is 0.7 km, then he will pass through the given cell in 50.4 seconds at maximum, assuming an hexagonal shape. This implies that the number of frequent handovers taking place increase the interaction among media-servers that will have to be involved throughout the streaming session. Furthermore, in order to avoid under-utilization of media-servers and strike a balance in the aggregate use for facilitating streaming, we group 3G cells into groups as Figure 2 depicts. This assignment is expected to happen in a static manner although it could be modified to reflect emerging new realities in the core network. In this regard, Figure 2 shows a network layout in which sets of sixteen cells are configured to function as a group. In this example, the mobile user is currently in a cell of group A and is heading towards group F . The media-servers that can be involved in the delivery of video objects are A, D, E and F . The server A is expected to interact with server D , D with E , and E with F . In this chained-fashion, we anticipate that the media servers notify each other about the streaming session of the oncoming mobile user. In addition, the media-servers send and receive in pipelined fashion the video object under transmission. ! &quot; # $ %'& (0)1324 576 86 98 #9@ A B 5 Figure 2: Grouping of Cells There may be other interactions as well. For instance, there must be cooperation between media server A and F if the requested video object is initially located at F . A media server accepts video requests from base-stations residing in its group; if it does not currently have the object it is responsible for locating using the meta data structure and fetching it. Due to the location awareness of our approach, we assume that each server predominantly stores streams specific to its own geographic area. For instance, if the route drawn in Figure 2 crosses a county, video clips with showing traffic conditions ahead in specific points may be requested. Similarly in a city setting, such requests may entail multimedia location-based virtual presentations. 2.2 Content Distribution Provided that a media server has a video clip in place, a straightforward approach would be to transmit the object into all base-stations operating in cells located on or near-by the user's trajectory . This is not only wasteful in both network and base-station resources but also increases the user-perceived playout latency. Therefore, we resort to using video segmentation [14], [21] in order to reduce network transmission costs between the video-holding media server and its subordinates base-stations. Segmentation also decreases the start-up latency that users experience upon cell entrance . The length of a video-segment, denoted S t , sent to each base-station is proportional to the average time that the user is expected to stay in the specific cell. The process of segmenting the video streams into chunks of frames of specific duration assumes that the media servers are aware of the underlying cell configuration. In particular, a media-server has to be aware of the precise manner with which a motorway cuts across its subordinate cells, the direction as well as the speed of moving users. With this data available, the media-server in discussion can approximate the time that a user spends in a cell. For example, if a motorist moving at 100km/h has just entered a cell and departs after traversing a 1km route portion, the server can compute the duration of the user's stay to be at ap-proximately 36 seconds. The media-server can capitalize on this very information to appropriately segment the streamed video; it only dispatches enough frames for a playout period of 36 seconds. The duration of a user's presence within a cell may vary according to speed changes with clearly lower speeds leading to elongated stays in the cell and vice versa. Should the speed be decreasing, the base-station will ultimately require more frames from the media server than the number predicted once the user appeared in the cell. Such a request constitutes a "cache miss" which will not be noted by the user if detected on time and acted upon by the coordinating base-station. Imposing a minimum threshold in the number of frames always available for delivery at a base-station may help overcome such "cache-misses". Therefore, when the number of frames awaiting transmission on a base-station falls below the abovementioned threshold, the base-station signals its need for additional frames to its overseeing media-server; should the latter act upon this request, additional frames arrive on time at the base-station for delivery. On the other hand, as soon as a user increases speed, she will depart the cell extend earlier than initially anticipated . The drawback here is that the media-server has provided the base-station with more frames than those eventually needed. During the handover process, the coordinating media-server has to generate a video-fragment which in its opening contains frames that have already been transmitted to the previous base-station but not-yet -seen by the user. Our approach allows a base-station to dynamically alter the transmission bitrate according to its current load. Under light load, a base-station may opt to increase the transmission rate for a specific video-stream thus leading to potential frame shortage. To easily avoid such shortage, we use the minimum allowed vehicle speed to compute the size of the video-segment S t to be transported to base-stations: S t = Distance M inimumSpeed (1) In most freeways there are authority-posted limits for minimum allowed speed in each road segment. As media-servers are aware of the geographic area that they serve, such minimum speed rates are statically designated for each cell in their jurisdiction. Evidently, the video-segment size that we potentially use as safety against frame shortage is: S t = Distance M inimumSpeed Distance AverageSpeed (2) 67 Algorithms 1 and 2 depict the video segmentation and distribution that we follow in our media distribution. Upon a new video-streaming request, we assume that the media-server can efficiently retrieve the corresponding video-file either from local storage options or remote servers via its low-latency/high-bandwidth wired networking infrastructure. The identification of the user's current location, the precompiled knowledge of the traveled distance within a cell, in conjunction with the minimum allowed speed on pertinent highway segments, permit for the estimation of the maximum user stay S t in a specific cell. Subsequently, the media-server can create the first segment of frames needed for transmission via the base-station to the requesting client. The size of a video-segment is given by V = P N i=1 F i , where F i is the size of the i-th frame and N is the number of frames in the segment; we can easily compute N by multiplying the frame-rate (frames/second) with the duration of stay in a cell S t . Algorithm 1 Video Segmentation at Media-Server 1: Get Minimum User Speed MinSpeed 2: Get PathLength in cell range 3: p last frame transmitted 4: if (New Session) then 5: S t = P athLength M inSpeed 6: else 7: // Shortage of Frames 8: S t = P athLength M inSpeed 9: // with &lt;&lt; 1 10: end if 11: V i P p+N i=p+1 F i , where F i is size of i-th frame, N =F rameRate S t In light of frame shortage, our video-segmentation algorithm dispatches into the base-station with need an increment of frames. This is defined as a fixed fraction of the length of S t in the current cell (line 7 of Algorithm 1). Requests of such increments may occur multiple times before a motorist leaves a cell due to congestion . A handover might find a moving user either serviced by a base-station in the realm of the current media-server or under the authority of a completely new media-server along the motorist's path. In the first case, the media-server initiates the delivery of the next video-fragment to the next-base-station encountered. The just departed base-station can help in determining the appropriate stream position p from which the segmentation will have to resume. The duration of the video-segment is computed anew using the same algorithm that now takes into consideration the data points from the new cell. Clearly, the length of the route as well as designated minimum speed limits may be different from those encountered in the previous cell. In the second scenario, a handover may force a user to operate in an entirely new group of cells supported by a new media-server. In general, the portion of the "not-yet-seen" stream has to be forwarded from the previous to the new media-server unless the latter already maintains its own copy. If we are not interested in reducing the transmission costs, we can transport the entire video object to the new media-server using the assumed high-speed link. The media-server now in charge takes over the session identifier of the moving user and along with user state data from its previous location can help coordinate the delivery of the video in the new cell. To enhance coordination among media-servers in the highest-level of Figure 1, prefetching could be used [7, 34]. We could de-Algorithm 2 Video Distribution from Media-Server to Base-Station (s) 1: while (OutstanindRequests) do 2: if New Session then 3: Start session (user's location, video stream, cell ID) 4: if Video Stream Not in Storage then 5: Get Video Stream from corresponding Media-Server 6: end if 7: end if 8: if (not(Handover)) then 9: Apply Video Segmentation Algorithm 10: Send Video-Segment V to Base-Station 11: else 12: if (new Base-Station within Media-Server realm) then 13: Apply Video Segmentation Algorithm 14: Send Segment V to Base-Station 15: Send playback position p to new Base-Station 16: else 17: Send Video Stream to next Media-Server 18: Send playback position p to next Media-Server 19: end if 20: end if 21: end while ploy prefetching of video-segments to media-servers and/or base-stations to facilitate playout continuity and minimize the start-up latencies. Users moving with similar speed and nearby to the streaming user can benefit from the already segmented video stream and start the playout immediately. Caching efficiency is limited by the fact that only users with similar traveling behavior may use the video segments. We can overcome this limitation if the starting point of the video segment at the next cell corresponds to users traveling at the maximum speed within the current cell. At the same time, the total size of the segment caters for users that travel at minimum speed within the next cell, thus remaining longer in the cell's range. This ensures that successive base stations hold a sufficient amount of frames to serve users traveling at different speeds. RATE ADAPTATION In this section, we propose a rate adaptation scheme whose objective is to better serve the overall needs of fast-roaming users. More specifically, we present a mechanism used by base-stations to control the rate at which they transmit video to each user when the cell becomes overloaded and the transmission bitrate eventually needs to be decreased. In light of this reduction, we seek ways to avoid discontinuities in user playback and cell bandwidth over-utilization lowering so the probability for a session drop. While focusing on bitrate management between the second and last tier of Figure 1, we assume that pertinent video-segment data is available at a base-station. Upon session initiation, the size Q of the mobile device buffer becomes known to the managing base-station . In general, we assume that a video-object is divided into frames of constant duration. Frames that belong to the same file vary in size depending on the encoding rate and the scene content. A time domain perspective allows us to control the transmission rate examining the time interval between transmission of successive frames rather than their respective sizes. If the inter-departure time corresponds to the rate instructed by the file's frame rate 1 , 1 Typical frame rate values are 25 frames/sec for the PAL color sys-68 Figure 3: Rate Adaptation Modules within a base-station the transmission bitrate corresponds to the file's encoding bitrate. Alterations in the inter-departure times result in the inversely proportional changes in transmission bitrate. Let {X i k } k 1 denote the departure process of the video frames from the base-station, for the i-th user. If i k is the departure time for frame k, X i k = i k+1 i k is the inter-departure interval for the k-th frame. In the absence of buffering capabilities on the mobile device, the smoothness of {X i k } is critical for the smoothness of playback at the user's end. Thus, the following should hold: P {X i k = T } 1 (3) where T is the inter-departure interval specified by the video-object frame rate. The buffer support that we assume available at the user-end enables the modification of the {X k } process reflecting modifications to the actual transmitting bitrate of the base-station. Video Streaming modules are integral parts of the base-station configuration and each such module handles the transmission of a video-stream. Hence, a segment of a video-stream is assigned to an instance of a Video Streaming module for final delivery to the user's equipment. A Rate Adaptation (RA) element is assigned to each user session for the specified video stream. An RA is aware of the user's buffer size and is responsible for the forwarding of video frames to the actual Transmitter of the base-station. As information about the station's load is fed-back by the Transmitter, the Rate Adaptation element regulates the inter-departure process of video frames from the station to a user, in a way that preserves playback continuity. Figure 3 depicts the interaction among these two elements and the Transmitter at a base-station that serves k concurrent sessions for the same video-object. The operation of the Rate Adaptation element is governed by periodic time intervals of constant duration, termed Control Cycles. Operating in the time domain, the module is aware of the exact number of frames the media player at the user-end will need over a specific period of time to ensure smooth playback. Let A denote the duration of the control cycle. Also, let Q A &gt; 0 be the occupancy (i.e., number of frames) of the buffer at the beginning of the control cycle and N A the number of frames that will be reproduced at the user-end during the control cycle. Since N A frames are requested tem that corresponds to an inter-departure time of 40 msec, and 30 fps for the NTSC system which corresponds to inter-departure time of 33.3 msec. from the media-player and Q A frames are accumulated the rate adapter needs only transmit N A - Q A frames at minimum over the control cycle. The video frame rate instructs that a frame be transmitted every T = A N A mseconds. Each one of the N A -Q A frames transmitted at minimum during the control cycle will depart the base-station at longer intervals equal to T = A N A -Q A . The initial inter-departure time has been spread by a tolerance factor ( 0) where: A N A - Q A = (1 + ) A N A = Q A N A - Q A (4) The tolerance factor, , is a parameter of the control cycle; may turn negative only when Q A &gt; N A . Thus, a more specific definition of would be = Q A N A - Q A 1 {Q A N A } + 1 {Q A &gt;N A } (5) If the RA element forwards frames at the rate instructed by the tolerance factor, the transmission bitrate over the control cycle will be equal to B/(1+), where B is the encoding bitrate of the video-stream . A control cycle during which the base-station transmits at the minimum bitrate instructed by is called a degraded cycle. A degraded cycle will lead to zero buffer occupancy at the end of the control cycle and the tolerance factor for the next control cycle will be equal to zero. Therefore, no two successive degraded cycles may occur. Non-zero buffer occupancy at the beginning of a control cycle will be present only if the overall transmission rate over the previous control cycles exceeded B. This can be achieved if the RA element forwards frames at a higher rate when the station is un-derutilized . Let denote the speed-up factor, the factor by which the bitrate increases in this case. An expression for the speed-up factor may be obtained if we consider that the maximum transmission bitrate will lead to a full user buffer at the end of the control cycle. If Q A is the buffer occupancy at the beginning of the cycle , then the station may transmit at maximum Q - Q A frames over the control cycle. In this case, each frame will be transmitted every A Q -Q A mseconds, with inter-departure interval having been decreased by : A Q - Q A = (1 - ) A N A = 1 N A Q - Q A (6) An upgraded cycle will transmit at a rate of B/(1 - ). The speed-up factor may turn negative only when the free buffer space is less than the frames that will be played back during the control cycle. In this case, the cycle is forced to operate in degraded mode, so that we can avoid buffer overflow. It is clear that the n-th control cycle may forward frames at a rate in the range of: B max{(1 - ), (1 + )} B n B (1 - ) (7) The respective inter-departure process, {X n } will be in the range of: (1 - )T {X n,k } max{(1 - )T, (1 + )T } (8) The RA element knows at any time the exact number of frames that have been transmitted to the user, and it also knows the time that has passed since the session initiation, which corresponds to the number of frames the playback process has consumed. The difference between the two values denotes the user buffer occupancy, 69 so no feedback mechanism is required as far as the user buffer occupancy is concerned. The algorithm followed by each Rate Adaptation element in the Video Streaming module is outlined in Algorithm 3. Algorithm 3 Rate Adaptation element operation 1: // Executed at the beginning of every control cycle 2: // for user i 3: Q i A = F ramesT ransmit i - F ramesP layed i 4: = Q i A /(F ramesCycle - Q i A ) 5: = 1 - F ramesCycle/(Q - Q i A ) 6: M inInterval i = (1 - ) T 7: if &lt; 0 then 8: M axInterval i = M inInterval i 9: else 10: M axInterval i = (1 + ) T 11: end if 12: Interval i = M inInterval i + (CellLoadP erc/100) (M axInterval i - M inInterval i ) Since the duration of the control cycle is constant, multiple control cycles may occur during a user's presence in the range of a single cell, depending on the size of the cell and the user's speed. We assume that the each cell handover always initiates a new control cycle. Algorithm 3 allows for alteration in the transmission bitrate by providing upper and lower bounds (i.e., M inInterval i and M ax Interval i ) to ensure the smoothness of the playout process. The choice of the actual bitrate within the specified range, at which the base-station transmits during a control cycle, is ultimately a function of the station's load at the time. This load is continually estimated with the help of the Transmitter module. This feedback enables each Rate Adaptation element to cater for buffer occupancy increase, taking advantage of low system load periods. At the same time, by detecting high system load, the Rate Adaptation element lowers the transmission bandwidth, allowing for more sessions to be accommodated, while at the same time the playback process is not distorted. EVALUATION RESULTS In order to reproduce and experiment with the behavior of our proposed architecture and bitrate adaptation scheme, we have setup a simulation testbed. We have assumed a user trajectory with duration of 200 seconds. The user traverses numerous cells of different sizes. Each base-station is equipped with the Video Streaming module as described earlier. A Control Cycle of 5 seconds is adopted by all elements. The buffer size at the user-equipment is assumed to be large enough to store 10 seconds which is readily met by modern cellular phones and/or PDAs. We designate ten levels of base-station load with load changing at random times. The maximum duration of each load state is 30seconds. The PAL color system is assumed for the video being transmitted, so the default inter-departure time for each frame is set at 40 mseconds. At the beginning of each control cycle, the Rate Adaptation element applies the proposed Algorithm 3. The testbed initially computes the tolerance and speed-up factors thus generating the acceptable inter-departure times range. The actual inter-departure time for the control cycle is proportional to the station's load at the time. If the station is lightly loaded, the minimum interdeparture time (maximum bandwidth) is applied. Conversely when the base-station is heavily loaded, the frames are forwarded to the transmitter at the minimum rate instructed by the maximum interdeparture time. For intermediate load levels an appropriate value from the inter-departure times range is selected according to Algorithm 3 in a uniform fashion. Figure 4 shows the evolution of the base stations' load during the user's trajectory, over all cells that the individual travels in. Having defined load of value 5 as "normal" load, the base-station load remains relatively high through the user trajectory with a few very short periods of low load. 0 2 4 6 8 10 0 50 100 150 200 Load Time Base station load Figure 4: Base-station load during user trajectory Figure 5 shows the applied transmission bitrate, along with the minimum and maximum bitrates allowed for every control cycle. The y-axis represents the percentage of the actual transmission bitrate to the video encoding bitrate. The bitrate is inversely proportional to the inter-departure times which are illustrated in figure 6. If we compare the curves of Figures 5 and 4, we can easily establish that the actual bitrate applied is a function of the base-station load and the calculated allowed bitrate range. At times of high load, the applied bitrate is closer to the minimum acceptable bitrate and conversely at times of light load, the applied bitrate is closer to the maximum acceptable bitrate. Figure 6 shows that the interdeparture times are indeed proportional to the base-station load. 0 50 100 150 200 0 50 100 150 200 Transmission bitrate (%) Time Transmission bitrate Figure 5: Allowed transmission bitrate limits and applied transmission bitrate. 70 20 25 30 35 40 45 50 55 60 65 70 0 50 100 150 200 Interdeparture time Time Frame interdeparture time Figure 6: Applied interdeparture interval. We show the user buffer occupancy throughout the trajectory in Figure 7. The buffer does not starve at any time suggesting the usefulness of our proposed scheme. The occupancy increases at times where the base station load falls below the normal load. Under overall higher-than-normal base-station load settings, our tests show that the playback continuity is preserved by only taking advantage of relatively small periods of station underutilization to increase the transmission bitrate. The range of the acceptable transmission rate includes the bandwidth already guaranteed by the network upon session acceptance at all times. Therefore, the network was never forced to transmit at a higher-than-agreed bitrate. At times of station over-utilization, the decreased transmission bitrate allows for more calls to be accommodated significantly decreasing the probability of session drop. 0 50 100 150 200 250 0 50 100 150 200 Occupancy Time Occupancy Figure 7: User buffer occupancy. A system that adopts no rate adaptation scheme would constantly require a transmission bitrate equal to the 100% of the video encoding bitrate throughout the session. Although the bitrate of real-time video streaming sessions is guaranteed by the UMTS specifications , at near-capacity situations, the network would have to either drop a session or be forced to transmit with lower bitrate. The former case is clearly undesirable and the latter generates jitter effects for the end-user. This would happen even if the station gets overloaded only for a period of time equal to the proposed scheme's control cycle duration. Discontinuities in the playback process may be observed at a system adopting the proposed rate adaptation scheme as well, however only in the case when the base-station is constantly load saturated, thus not allowing for any upgraded cycles to take place. RELATED WORK There has been a large amount of reported work in related areas that include caching for video systems, management of moving objects , and rate adaptation for streaming systems on wired networks. Data caching and mirroring has been proposed as a way to help the scalability of video delivery systems [32]. By placing content closer to the consuming clients not only network costs can be curtailed but also the load of streaming servers can be reduced. Various aspects of the use of proxy servers for video objects has been examined in [14, 26, 23, 10]. In [14], the segmentation and caching of streaming objects is proposed and the merging of requests that are temporally related is investigated. This merging idea has been used in [8, 18, 11] to save bandwidth in light of requests for the same video object that arrive closely in time. The partial caching of two successive intervals of a video stream is proposed in [10] as a way to speed-up the serving of follow-up requests for the the same video object. In [26], the caching of initial frames of a video object is used in a proxy-setting to reduce startup latency. In the same direction, the storage of the bursty parts of a video-stream in a proxy is advocated in [34]; the remaining parts of the video are directly retrieved from the source helping significantly reduce peak bandwidth requirements in the backbone. A caching mechanism for layered encoded multimedia streams is suggested in [23]; the objective of the technique is to selectively deliver stream quality by differentiating on the client network connection. Stream quality differentiation is also addressed in [24], in conjuction with a seamless handoff mechanism for mobile users. A formal spatiotemporal framework for modeling moving objects and a query language is discussed in [27]. Efficient techniques for indexing moving objects in one and two dimensions are proposed in [12] while in [5] the trade-offs for indexing schemes to answer interval, rectangle, approximate nearest-neighbor, approximate farthest-neighbor and convex-hull queries are examined. The indexing of current and anticipated positions of moving objects in the context of location based services is examined in [25]. Much work has been also reported in data broadcast and dissemination over wireless networks during the last decade [4, 22, 33, 19, 6]. Rate adaptation for wired streaming systems has been exten-sively studied in [28, 20, 31, 29, 16, 17]. These studies assume an adaptable video encoding system that changes the encoding parameters on the fly based on feedback information about the channel state. The notion of cycle-based operation is used in [13] with cycles being successively alternating between good and bad cycles. Our work differs in that it does not require a prefetching period so that an initial occupancy is built up in the buffer before the playback begins. Also, our algorithm functions in a graceful manner when the base-station load does not allow for aggressive use of channel resources. CONCLUSIONS In this paper, we address the problem of efficient video delivery in real-time to high-speed users roaming a 3G network. We propose a network of media servers handling the content distribution on top of the mobile environment that closely cooperates with the 71 base-stations and user-equipment for the provision of continuous video playout. We segment video streams into variable-sized parts according to the user's speed and traversal path through different cells. In this manner, we minimize the transmission costs between media-servers and base-stations as well as the start-up latency experienced by users during handover. We adopt the use of Video Streaming modules along with their Rate Adaptation elements with the infrastructure of base-stations to ensure smoothness of the playout process. Preliminary experimentation results through simulation show that the proposed scheme rapidly adapts to changes in load conditions at base-stations, thus minimizing the probability of buffer starvation or even session drops. The low complexity of the proposed mechanism makes it suitable for real-time applications. REFERENCES [1] 3rd Generation Partnership Project. Universal Mobile Telecommunication System/IMT2000 Spectrum. Technical Report 6, UMTS Forum, 1998. [2] 3rd Generation Partnership Project. Stage 2 Functional Specification of Location Services in URAN. Technical Report (3G TR 25.923 version 1.4.0), UMTS Forum, 1999. Technical Specification Group(TSG) RAN, Working Group 2 (WG2). [3] 3rd Generation Partnership Project. Transparent End-to-End Packet Switched Streaming Service (PSS) General Description (Release 4). Technical Report 3GPP-TS-26.233-V4.0.0, UMTS Forum, 2000. Technical Specification Group Services and System Aspects. [4] S. Acharya, M.J. Franklin, and S. B. Zdonik. Balancing Push and Pull for Data Broadcast. In Proceedimgs of SIGMOD 1997, Tucson, AZ, May 1997. [5] P.K. Agarwal, L. Arge, J. Erickson, and H. Yu. Efficient Tradeoff Schemes in Data Structures for Querying Moving Objects. In 12th Annual European Symposium on Algorithms (ESA), pages 415, Bergen, Norway, September 2004. [6] D. Aksoy, M. Altinel, R. Bose, U. Cetintemel, M. Franklin, J. Wang, and S. Zdonik. Research in Data Broadcast and Dissemination. In Proceedings of International Conf. on Advanced Multimedia Content Processing (AMCP), Osaka, Japan, November 1998. [7] P. Cao, E. W. Felten, A. R. Karlin, and K. Li. A Study of Integrated Prefetching and Caching Strategies. In Proceedings of ACM SIGMETRICS Conf., pages 188197, Ottawa, Canada, May 1995. [8] S. Chan and F. Tobagi. Caching schemes for distributed video services. In Proceedings of the 1999 IEEE International Conference on Communications (ICC'99), Vancouver, Canada, June 1999. [9] S. Choi and K. G. Shin. Predictive and Adaptive Bandwidth Reservation for Hand-Offs in QoS-Sensitive Cellular Networks. In Proceedings of ACMSIGCOMM, pages 155166, 1998. [10] A. Dan and D. Sitaram. A Generalized Interval Caching Policy for Mixed Interactive and Long Video Environments. In Proceedings of IST/SPIE Multimedia Computing and Networking Conference, San Jose, CA, January 1996. [11] A. Dan, D. Sitaram, and P. Shahabuddin. Dynamic Batching Policies for an On-demand Video Server. Multimedia Systems, 4(3):112121, 1996. [12] G. Kollios and D. Gunopulos and V.J. Tsotras. On Indexing Mobile Objects . In Proceedimgs of 1999 ACM SIGACT-SIGMOD-SIGART Symposium on Principles of Database Systems (PODS), Philadephia, PA, 1999. [13] M. Hassan, L. Atzori, and M. Krunz. Video Transport over Wireless Channels: A Cycle-based Approach for Rate Control. In Proceedings of the ACM Multimedia 2004 Conference. ACM Press, October 2004. [14] M. Hofmann, E. Ng, K. Guo, S. Paul, and H. Zhang. Caching Techniques for Streaming Multimedia over the Internet. Technical report, Bell Laboratories, April 1999. BL011345-990409-04TM. [15] H. Holma and A. Toskala. WCDMA for UMTS Radio Access for Third Generation Mobile Communications. John Wiley & Sons Inc., New York, NY, 2nd edition, 2002. [16] C.-Y. Hsu, A. Ortega, and A.R. Reibman. Joint Selection of Source and Channel Rate for VBR Video Transmission Under ATM Policing Constraints. IEEE Journal of Selected Areas in Communications, 15(6):10161028, 1997. [17] P.-C. Hu, Z-L. Zhang, and M. Kaveh. Channel Condition ARQ Rate Control for Real-time Wireless Video Under Buffer Constraints. In Proceedings of the IEEE International Conf. on Image Processing, Vancouver BC, Canada, September 2000. [18] K.A. Hua, Y. Cai, and S. Sheu. Patching: a Multicast Technique for True Video-on-demand Services. In Proceedings of the 6th ACM International Conference on Multimedia, pages 191200. ACM Press, 1998. [19] T. Imielinski, S. Viswanathan, and B. R. Badrinath. Data on Air: Organization and Access. IEEE Transactions on Knowledge and Data Engineering, (3):353372, 1997. [20] H.-J. Lee, T. Chiang, and Y.-Q. Zhang. Scalable Rate Control for MPEG-4 Video. IEEE Transactions On Circuits and Systems for Video Technology, 10(9):878894, September 2000. [21] S.-J. Lee, W.-Y. Ma, and B. Shen. An Interactive Video Delivery and Caching System Using Video Summarization. Computer Communications, 25(4):424435, March 2002. [22] E. Pitoura and P.K. Chrysanthis. Multiversion Data Broadcast. IEEE Transactions on Computers, 51(10):12241230, 2002. [23] R. Rejaie, H. Yu, M. Handley, and D. Estrin. Multimedia Proxy Caching Mechanism for Quality Adaptive Streaming Applications in the Internet. In Proceedings of INFOCOM(2), pages 980989, 2000. [24] S. Roy and B. Shen abd V. Sundaram. Application Level Handoff Support for Mobile Media Transcoding Sessions. In 12th International Workshop on Network and Operating System Support for Digital Audio and Video, Miami, FL, 2002. [25] S. Saltenis and C.S. Jensen. Indexing of Moving Objects for Location-Based Services. In Proceedings of the IEEE International Conference on Data Engineering (ICDE), pages 463472, 2002. [26] S. Sen, J. Rexford, and D. F. Towsley. Proxy Prefix Caching for Multimedia Streams. In Proceedings of INFOCOM(3), pages 13101319, New York, NY, 1999. [27] A.P. Sistla, O. Wolfson, S. Chamberlain, and S. Dao. Modeling and Querying Moving Objects. In Proceedings of the 13th IEEE International Conf. on Data Engineering, Birmingham, UK, April 1997. [28] H. Song and C.-C.-J. Kuo. Rate control for Low Bit Rate Video via Variable Encoding Frame Rates. IEEE Transactions On Circuits and Systems for Video Technology, 11(4):512521, April 2001. [29] W. Tawbi, F. Horn, E. Horlait, and J.-B. Stefani. Video Compression Standards and Quality of Service. The Computer Journal, 36(1):4354, 1993. [30] Texas Instruments. Mobile Connectivity: Assisted-GPS. http://focus.ti.com/general/docs/wtbu, 2004. [31] T. Wiegand, M. Lightstone, T. Campbell, D. Mukherjee, and S. Mitra. Rate-Distortion Optimized Mode Selection for Very Low Bit Rate Video Coding and the Emerging H.263 Standard. URL: citeseer.ist.psu.edu/wiegand95ratedistortion.html, 1999. [32] D. Wu, Y.T. Hou, W. Zhu, Y.-Q. Zhang, and J.M. Peha. Streaming Video over the Internet: Approaches and Directions. IEEE Transactions on Circuits and Systems for video Technology, 11(1):120, February 2001. [33] X. Yang and A. Bouguettaya. Broadcast-Based Data Access in Wireless Environments. In Proceedings of the EDBT International Conference, Prague, Czech Republic, 2002. [34] Z.-L. Zhang, Y. Wang, D.H.C. Du, and D. Shu. Video staging: a proxy-server-based approach to end-to-end video delivery over wide-area networks. IEEE/ACM Transactions on Networking, 8(4):429442, 2000. 72

mobile multimedia services;rate adaptation;real-time streaming;Streaming for moving users

213

Web Taxonomy Integration through Co-Bootstrapping

We address the problem of integrating objects from a source taxonomy into a master taxonomy. This problem is not only currently pervasive on the web, but also important to the emerging semantic web. A straightforward approach to automating this process would be to learn a classifier that can classify objects from the source taxonomy into categories of the master taxonomy. The key insight is that the availability of the source taxonomy data could be helpful to build better classifiers for the master taxonomy if their categorizations have some semantic overlap. In this paper, we propose a new approach, co-bootstrapping , to enhance the classification by exploiting such implicit knowledge. Our experiments with real-world web data show substantial improvements in the performance of taxonomy integration.

INTRODUCTION A taxonomy, or directory or catalog, is a division of a set of objects (documents, images, products, goods, services, etc.) into a set of categories. There are a tremendous number of taxonomies on the web, and we often need to integrate objects from a source taxonomy into a master taxonomy. This problem is currently pervasive on the web, given that many websites are aggregators of information from various other websites [2]. A few examples will illustrate the scenario. A web marketplace like Amazon may want to combine goods from multiple vendors' catalogs into its own. A web portal like NCSTRL may want to combine documents from multiple libraries' directories into its own. A company may want to merge its service taxonomy with its partners'. A researcher may want to merge his/her bookmark taxonomy with his/her peers'. Singapore-MIT Alliance , an innovative engineering education and research collaboration among MIT, NUS and NTU, has a need to integrate the academic resource (courses, seminars, reports, softwares, etc.) taxonomies of these three universities. This problem is also important to the emerging semantic web [4], where data has structures and ontologies describe the semantics of the data, thus better enabling computers and people to work in cooperation. On the semantic web, data often come from many different ontologies, and information processing across ontologies is not possible without knowing the semantic mappings between them. Since taxonomies are central components of ontologies, ontology mapping necessarily involves finding the correspondences between two taxonomies, which is often based on integrating objects from one taxonomy into the other and vice versa [10, 14]. If all taxonomy creators and users agreed on a universal standard, taxonomy integration would not be so difficult. But the web has evolved without central editorship. Hence the correspondences between two taxonomies are inevitably noisy and fuzzy. For illustration, consider the taxonomies of two web portals Google and Yahoo : what is "Arts/Music/Styles/" in one may be "Entertainment/Music/Genres/" in the other, category "Computers_and_Internet/Software/Freeware" and category "Computers/Open_Source/Software" have similar contents but show non-trivial differences, and so on. It is unclear if a universal standard will appear outside specific domains, and even for those domains, there is a need to integrate objects from legacy taxonomy into the standard taxonomy. Manual taxonomy integration is tedious, error-prone, and clearly not possible at the web scale. A straightforward approach to automating this process would be to formulate it as a classification problem which has being well-studied in machine learning area [18]. Normally the classifier would be constructed using objects in the master taxonomy as training examples, and the source taxonomy would be completely ignored during learning. However, the availability of the source taxonomy data could be helpful to build better classifiers for the master taxonomy if their categorizations have some semantic overlap, particularly when the number of training examples is not very large. Possible useful semantic relationships between a master category C and a source category S include: C S = (identical): an object belongs to C if and only if it belongs to S ; C S = (mutual exclusion): if an object belongs to S it cannot belong to C ; C S (superset): any object that belonging to S must also belong to C ; C S (subset): any object not belonging to S also cannot belong to C ; C and S overlap but neither is a superset of the other. In addition, semantic relationships may involve multiple master and source categories. For example, a master category C may be a subset of the union of two source categories a S and b S , so if an object does not belong to either a S or b S , it cannot belong to C . The real-world semantic relationships are noisy and fuzzy, but they can still provide valuable information for classification. For example, knowing that most (80%) objects in a source category S belong to one master category a C and the rest (20%) examples belong to another master category b C is obviously helpful. The difficulty is that knowledge about those semantic relationships is not explicit but hidden in the data. In this paper, we propose a new approach, co-bootstrapping, to enhance the classification by exploiting such implicit knowledge. Our experiments with real-world web data show substantial improvements in the performance of taxonomy integration. The rest of this paper is organized as follows. In 2, we give the formal problem statement. In 3, we describe a state-of-the-art solution. In 4, we present our approach in detail. In 5, we conduct experimental evaluations. In 6, we review the related work. In 7, we make concluding remarks. PROBLEM STATEMENT Taxonomies are often organized as hierarchies. In this work, we assume for simplicity, that any objects assigned to an interior node really belong to a leaf node which is an offspring of that interior node. Since we now have all objects only at leaf nodes, we can flatten the hierarchical taxonomy to a single level and treat it as a set of categories [2]. Now we formally define the taxonomy integration problem that we are solving. Given two taxonomies: a master taxonomy M with a set of categories 1 2 , ,..., M C C C each containing a set of objects, and a source taxonomy N with a set of categories 1 2 , ,..., N S S S each containing a set of objects, we need to find the categories in M for each object in N. To formulate taxonomy integration as a classification problem, we take 1 2 , ,..., M C C C as classes, the objects in M as training examples, the objects in N as test examples, so that taxonomy integration can be automatically accomplished by predicting the classes of each test example. Such a classification problem is multi-class and multi-label, in the sense that there are usually more than two possible classes and one object may be relevant to more than one class. A STATE-OF-THE-ART SOLUTION Agrawal and Srikant recently proposed an elegant approach to taxonomy integration by enhancing the Nave Bayes algorithm [2]. The Nave Bayes (NB) algorithm is a well-known text classification technique [18]. NB tries to fit a generative model for documents using training examples and apply this model to classify test examples. The generative model of NB assumes that a document is generated by first choosing its class according to a prior distribution of classes, and then producing its words independently according to a (typically multinomial) distribution of terms conditioned on the chosen class [15]. Given a test document d , NB predicts its class to be arg max Pr[ | ] C C d . The posterior probability Pr[ | ] C d can be computed via Bayes's rule: Pr[ | ] C d Pr[ , ] Pr[ ] C d d = Pr[ ] Pr[ | ] Pr[ ] C d C d = Pr[ ] Pr[ | ] C d C ( ) ( , ) Pr[ ] Pr[ | ] n d w w d C w C = , where ( , ) n d w is the number of occurrences of w in d . The probability Pr[ ] C can be estimated by the proportion of training documents in C . The probability Pr[ | ] w C can be estimated by ( ) ( , ) ( , ) i i w V n C w n C w + + , where ( , ) n C w is the number of occurrences of w in training documents in C , V is the vocabulary of terms, and 0 1 &lt; is the Lidstone's smoothing parameter [1]. Taking logs, we see that NB is actually a linear classifier: log Pr[ | ] C d ( ) ( ) ( , ) log Pr[ ] Pr[ | ] n d w w d C w C ( ) log Pr[ ] ( , ) log Pr[ | ] w d C n d w w C = + . The enhanced Nave Bayes (ENB) algorithm [2] uses the categorization of the source taxonomy to get better probability estimations. Given a test document d that is know to be in category S in N, ENB predicts its category in M to be arg max Pr[ | , ] C C d S . The posterior probability Pr[ | , ] C d S can 411 be computed as Pr[ | , ] C d S Pr[ , , ] Pr[ , ] C d S d S = Pr[ ] Pr[ , | ] Pr[ , ] S C d S d S = Pr[ , | ] C d S . ENB invokes a simplification that assumes d and S are independent given C , therefore Pr[ , | ] C d S Pr[ | ] Pr[ | , ] C S d S C = Pr[ | ] Pr[ | ] C S d C = ( ) ( , ) Pr[ | ] Pr[ | ] n d w w d C S w C = . The probability Pr[ | ] w C can be estimated in the same way of NB. For the probability Pr[ | ] C S , ENB estimates it by ( ) i i i C C C S C C S , where C is the number of documents in C , C S is the number of documents in S classified into C by the NB classifier, and 0 is a parameter reflecting the degree of semantic overlap between the categorizations of M and N. The optimal value of can be found using a tune set (a set of objects whose categories in both taxonomies are known). The tune set can be made available via random sampling or active learning [2]. Taking logs, we see that ENB is still a linear classifier: log Pr[ | , ] C d S ( ) ( ) ( , ) log Pr[ | ] Pr[ | ] n d w w d C S w C ( ) log Pr[ | ] ( , ) log Pr[ | ] w d C S n d w w C = + . Comparing the classification functions of NB and ENB, it is obvious that all ENB does is to shift the classification threshold of its base NB classifier, no more and no less. To achieve multi-class multi-label classification that is required by taxonomy integration, we use the "one-vs-rest" method to create an ensemble of binary (yes/no) NB or ENB classifiers, one for each category C in M. OUR APPROACH Here we present our approach in detail. In 4.1, we introduce the boosting technique. In 4.2, we propose the co-bootstrapping method. In 4.3, we discuss the advantages of our approach. 4.1 Boosting In our approach to taxonomy integration, we utilize a powerful machine learning method, boosting [17, 23], to build classifiers. The main idea of boosting is to combine many weak hypotheses (simple and moderately accurate classification rules), into a highly accurate classifier. In this paper, we focus on boosting for text classification. Generalization to other kinds of data and learning algorithms would be straightforward. 4.1.1 Term-Features Text objects (documents) can be represented using a set of term-features 1 2 { , ,... } T T T T n F f f f = . The term-feature Th f (1 ) h n of a given object x is a binary feature indicating the presence or absence of h w (the h-th distinct word in the document collection) in x , i.e., 1 if 0 if h Th h w x f w x = . 4.1.2 Weak Hypotheses Let X denote the domain of possible objects, and let Y be a set of k possible classes. A labeled example is a pair ( , ) x Y where x X is an object and Y Y is the set of classes which x belongs to. We define [ ] Y l for l Y to be 1 if [ ] 1 if l Y Y l l Y + = . A hypothesis is a real-valued function : h R X Y . The sign of ( , ) h x l is a prediction of [ ] Y l for x , i.e., whether object x is contained in class l . The magnitude of ( , ) h x l is interpreted as a measure of confidence in the prediction. Based on a binary feature f , we are interested in weak hypotheses h which are simple decision stumps of the form 1 0 if 1 ( , ) if 0 l l c f h x l c f = = = , where 1 0 , l l c c . 4.1.3 AdaBoost Algorithm The most popular boosting algorithm is AdaBoost introduced in 1995 by Freund and Schapire [12]. Our work is based on a multi-class multi-label version of AdaBoost, AdaBoost.MH [24, 25], which is described in Figure 1. Given m training examples 1 1 ( , ),...,( , ) m m x Y x Y where each i x X , i Y Y , AdaBoost.MH dynamically maintains a distribution t D over all objects and classes. Initially this distribution 1 D is uniform. In the t-th round, the optimal weak hypothesis t h is selected based on the set of training examples and the current distribution t D . Then a parameter t is chosen, and the distribution t D is updated in a manner that puts more weights on "difficult" examples (object-class pairs) that are misclassified by t h . Please be referred to [24, 25] for the details on computing optimal t h and t . This procedure repeats for T rounds. The final hypothesis ( , ) H x l is actually a weighted vote Given: 1 1 ( , ),...,( , ) m m x Y x Y where each i x X , i Y Y . Initialize 1 ( , ) 1 ( ) D i l mk = . for 1,..., t T = do Pass distribution t D to weak learner. Get weak hypothesis : ht R X Y . Choose t . Update: 1 ( , ) exp( [ ] ( , )) ( , ) t t i t i t t D i l Y l h x l D i l Z + = where t Z is the normalization factor end for Output the final hypothesis: 1 ( , ) ( , ) T t t t H x l h x l = = . Figure 1: The boosting algorithm AdaBoost.MH. 412 of weak hypotheses 1 ( , ) T t t t h x l = , and the final prediction can be computed according to the sign of ( , ) H x l . 4.2 Co-Bootstrapping Thus far we have completely ignored the categorization of N. Although M and N are usually not identical, their categorizations often have some semantic overlap. Therefore the categorization of N contains valuable implicit knowledge about the categorization of M. Hereby we propose a new approach, co-bootstrapping , to enhance the classification by exploiting such implicit knowledge. 4.2.1 Category-Features If we have indicator functions for each category in N, we can imagine taking those indicator functions as features when we learn the classifier for M. This allows us to exploit the semantic relationship among the categories of M and N without explicitly figuring out what the semantic relationships are. More specifically, for each object in M, we augment the ordinary term-features with a set of category-features 1 2 { , ..., } N F f f f = N N N N derived from N. The category-feature j f N (1 ) j N of a given object x is a binary feature indicating whether x belongs to category j S (the j-th category of N), i.e., 1 if 0 if j j j x S f x S = N . In the same way, we can get a set of category-features 1 2 { , ..., } M F f f f = M M M M derived from M to be used for supplementing the features of objects in N. The remaining problem is to obtain these indicator functions, which are initially not available. 4.2.2 Co-Bootstrapping Algorithm When building the classifier for M, the training examples are the objects in M and the test examples are the objects in N. To leverage the categorization of N to reinforce classification, our classifier uses term-features T F as well as category-features F N . However, we do not know the exact values of F N of the training examples. Our proposed algorithm overcomes the above obstacle by utilizing the bootstrapping idea. Let ( ) r F T B denote a boosting-classifier for taxonomy T's categorization based on feature set F at step r . Initially we build a classifiers 0 ( ) T F B N based on only term-features, then use it to classify the objects in M (the training examples) into the categories of N, thus we can predict the value of each category-feature j f F N N for each object x M . At next step we will be able to build 1 ( ) T F F M N B using the predicted values of F N of the training examples. Similarly we can build 0 ( ) T F M B and 1 ( ) T F F N M B . The new classifier 1 ( ) T F F N M B ought to be better than 0 ( ) T F B N because 1 ( ) T F F N M B leverages more knowledge. Hence we can predict the value of each category-feature j f F N N for each object x M more accurately using 1 ( ) T F F N M B instead of 0 ( ) T F B N , and afterwards we can build 2 ( ) T F F M N B . Also 2 ( ) T F F M N B is very likely to be better than 1 ( ) T F F M N B because 2 ( ) T F F M N B is based on a more accurate prediction of F N . This process can be repeated iteratively in a "ping-pong" manner. We name this approach co-bootstrapping since the two classifiers ( ) r T F F B M N and ( ) r T F F B N M collaborate to bootstrap themselves together. Figure 2 presents the co-bootstrapping algorithm, and Figure 3 depicts its process. 4.3 Discussion 4.3.1 Why Choose Boosting We have selected to employ the boosting technique to build classifiers in our co-bootstrapping approach to taxonomy integration because of its following virtues. Boosting has shown outstanding classification performance on many kinds of data such as text documents [17, 23, 24]. Boosting finds the optimal combination of heterogeneous weak hypotheses automatically, therefore alleviates the problem of how to weight ordinary features (e.g. term-features) and category-features appropriately. In contrast, approaches based on other machine learning algorithms like Support Vector Machines (SVMs) [9] would require to adjust relative combination weights, which is a non-trivial problem. Boosting generates descriptive and human-readable hypotheses as the final classifier, and the learned classifier is usually sparse despite the large feature set. Although boosting looks an ideal choice, other machine learning algorithms can also be utilized in the co-bootstrapping approach. We have not investigated this issue yet. 4.3.2 Comparison with ENB Although ENB [2] has been shown to work well for taxonomy integration, we think that a more general approach is still attractive. It has been experimentally shown that AdaBoost is more promising than NB for text classification [24]. The co-bootstrapping approach allows more powerful machine learning algorithms like AdaBoost to be utilized. Both ENB and our co-bootstrapping approach exploit the categorization of N to enhance classification. While all ENB does is to shift the classification threshold of its base NB classifier (see 3), co-bootstrapping has the ability to achieve more complex adjustments on the classification function of its base classifier. 413 Furthermore, ENB needs a stand-alone tune set to find the optimal value of parameter which controls the influence of source categorization information on classification, whereas co-bootstrapping based on boosting does not have such burdens. Although co-bootstrapping looks more effective, ENB still holds an advantage in efficiency. EXPERIMENTS We have collected 5 datasets from Google and Yahoo. One dataset includes the slice of Google's taxonomy and the slice of Yahoo's taxonomy about websites on one specific topic, as shown in Table 1. In each slice of taxonomy, we take only the top level directories as categories, e.g., the "Movie" slice of Google's taxonomy has categories like "Action", "Comedy", "Horror", etc. For each dataset, we show in Table 2 the number of categories occurred in Google and Yahoo respectively. In each category, we take all items listed on the corresponding directory page and its sub-directory pages as its objects. An object (list item) corresponds to a website on the world wide web, which is usually described by its URL, its title, and optionally a short annotation about its content. Here each object is considered as a text document composed of its title and annotation. All documents are pre-processed by removal of stop-words and stemming. For each dataset, we show in Table 3 the number of objects occurred in Google (G), Yahoo (Y), either of them (G Y), and both of them (G Y) respectively. The set of objects in GY covers only a small portion (usually less than 10%) of the set of objects in Google or Yahoo alone, which suggests the great benefit of automatically integrating them. This observation is consistent with [2]. Figure 3: The co-bootstrapping process. Given: two taxonomies M and N . Build classifier 0 ( ) T F B M , then use it to predict the value of each category-feature i f F M M for each object x N . Build classifier 0 ( ) T F B N , then use it to predict the value of each category-feature j f F N N for each object x M . for 1,..., r R = do Build classifier ( ) r T F F B M N , then use it to predict the value of each category-feature i f F M M for each object x N . Build classifier ( ) r T F F B N M , then use it to predict the value of each category-feature j f F N N for each object x M . end for For each object x N , if the value of its category-feature i f F M M is positive, then we classify it into i C M . For each object x M , if the value of its category-feature j f F N N is positive, then we classify it into j S N . Figure 2: The co-bootstrapping algorithm. Table 1: The datasets. Google Yahoo Book / Top/ Shopping/ Publications/ Books/ / Business_and_Economy/ Shopping_and_Services/ Books/ Bookstores/ Disease / Top/ Health/ Conditions_and_Diseases/ / Health/ Diseases_and_Conditions/ Movie / Top/ Arts/ Movies/ Genres/ / Entertainment/ Movies_and_Film/ Genres/ Music / Top/ Arts/ Music/ Styles/ / Entertainment/ Music/ Genres/ News / Top/ News/ By_Subject/ / News_and_Media/ Table 3: The number of objects. Google Yahoo G Y GY Book 10,842 11,268 21,111 999 Disease 34,047 9,785 41,439 2,393 Movie 36,787 14,366 49,744 1,409 Music 76,420 24,518 95,971 4,967 News 31,504 19,419 49,303 1,620 Table 2: The number of categories. Google Yahoo Book 49 41 Disease 30 51 Movie 34 25 Music 47 24 News 27 34 414 The number of categories per object in these datasets is 1.54 on average. This observation justifies the necessity of building multi-class multi-label classifiers. 5.2 Tasks For each dataset, we pose 2 symmetric taxonomy integration tasks: G Y (integrating objects from Yahoo into Google) and Y G (integrating objects from Google into Yahoo). As described in 2, we formulate each task as a classification problem. The objects in G Y can be used as test examples, because their categories in both taxonomies are known to us [2]. We hide the test examples' master categories but expose their source categories to the learning algorithm in training phase, and then compare their hidden master categories with the predictions of the learning algorithm in test phase. Suppose the number of the test examples is n . For G Y tasks, we randomly sample n objects from the set G-Y as training examples. For Y G tasks, we randomly sample n objects from the set Y-G as training examples. This is to simulate the common situation that the sizes of M and N are roughly in same magnitude. For each task, we do such random sampling 5 times, and report the classification performance averaged over these 5 random samplings. 5.3 Measures As stated in 2, it is natural to accomplish taxonomy integration tasks via building multi-class multi-label classifiers. To measure classification performance for each class (category in M), we use the standard F-score (F 1 measure) [3]. The F-score is defined as the harmonic average of precision (p) and recall (r), 2 ( ) F pr p r = + , where precision is the proportion of correctly predicted positive examples among all predicted positive examples, and recall is the proportion of correctly predicted positive examples among all true positive examples. The F-scores can be computed for the binary decisions on each individual category first and then be averaged over categories. Or they can be computed globally over all the M n binary decisions where M is the number of categories in consideration (the number of categories in M) and n is the number of total test examples (the number of objects in N). The former way is called macro-averaging and the latter way is called micro-averaging [27]. It is understood that the micro-averaged F-score (miF) tends to be dominated the classification performance on common categories, and that the macro-averaged F-score (maF) is more influenced by the classification performance on rare categories [27]. Providing both kinds of scores is more informative than providing either alone. 5.4 Settings We use our own implementation of NB and ENB. The Lidstone's smoothing parameter is set to an appropriate value 0.1 [1]. The performance of ENB would be greatly affected by its parameter . We run ENB with a series of exponentially increasing values of : (0, 1, 3, 10, 30, 100, 300, 1000) [2] for each taxonomy integration task, and report the best experimental results. We use BoosTexter [24] for the implementation of AdaBoost, taking single words as terms. We set the boosting rounds 1000 T = and the co-bootstrapping iteration number 8 R = (see Figure 1 & 2). In the following sections, we denote the normal AdaBoost approach by AB, and denote the co-bootstrapping approach based on AdaBoost algorithm by CB-AB. 5.5 Results The experimental results of NB and ENB are shown in Table 4. We see that ENB really can achieve much better performance than NB for taxonomy integration. 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 1 2 3 4 5 6 7 8 maF(G Y) miF(G Y) maF(Y G) miF(Y G) Figure 4: The taxonomy integration performance increases along with the number of co-bootstrapping iterations, on the Book dataset. Table 5: Experimental Results of AB and CB-AB. AB CB-AB maF miF maF miF Book 0.1740 0.4499 0.2540 0.6030 Disease 0.5375 0.6674 0.6533 0.7703 Movie 0.1930 0.4892 0.3172 0.6716 Music 0.3316 0.5025 0.4851 0.6826 G Y News 0.2150 0.4625 0.3083 0.6218 Book 0.2436 0.3853 0.3516 0.6341 Disease 0.3719 0.6350 0.4371 0.7287 Movie 0.2559 0.5214 0.3922 0.7154 Music 0.4369 0.6397 0.5799 0.7994 Y G News 0.3774 0.4942 0.4340 0.6421 Table 4: Experimental Results of NB and ENB. NB ENB maF miF maF miF Book 0.1286 0.2384 0.1896 0.5856 Disease 0.4386 0.5602 0.5230 0.6895 Movie 0.1709 0.3003 0.2094 0.5331 Music 0.2386 0.3881 0.2766 0.5408 G Y News 0.2233 0.4450 0.2578 0.5987 Book 0.1508 0.2107 0.2227 0.5471 Disease 0.2746 0.4812 0.3415 0.6370 Movie 0.2319 0.4046 0.2884 0.5534 Music 0.3124 0.5359 0.3572 0.6824 Y G News 0.2966 0.4219 0.3639 0.6007 415 The experimental results of AB and CB-AB are shown in Table 5. Obviously AB beats NB, which is consistent with the conclusion of [24]. Also we find that CB-AB works better than AB for taxonomy integration, which suggests that co-bootstrapping makes effective use of the categorization of N to enhance classification for M. Figure 4 shows that the taxonomy integration performance increases along with the number of co-bootstrapping iterations, on the Book dataset. This implies that the two boosting-classifiers learned from two taxonomies do mutually boost each other until they become stable. The experimental results of ENB and CB-AB are compared in Figure 5 and 6. It is clear that CB-AB outperforms ENB consistently and significantly. RELATED WORK Most of the recent research efforts related to taxonomy integration are in the context of ontology mapping on semantic web. An ontology specifies a conceptualization of a domain in terms of concepts, attributes, and relations [11]. The concepts in an ontology are usually organized into a taxonomy: each concept is represented by a category and associated with a set of objects (called the extension of that concept). The basic goal of ontology mapping is to identify (typically one-to-one) semantic correspondences between the taxonomies of two given ontologies: for each concept (category) in one taxonomy, find the most similar concept (category) in the other taxonomy. Many works in this field use a variety of heuristics to find mappings [7, 16, 19, 21]. Recently machine learning techniques have been introduced to further automate the ontology mapping process [10, 13, 14, 20, 26]. Some of them derive similarities between concepts (categories) based on their extensions (objects) [10, 13, 14], therefore they need to first integrate objects from one taxonomy into the other and vice versa (i.e., taxonomy integration). So our work can be utilized as a basic component of an ontology mapping system. As stated in 2, taxonomy integration can be formulated as a classification problem. The Rocchio algorithm [3, 22] has been applied to this problem in [14]; and the Nave Bayes (NB) algorithm [18] has been applied to this problem in [10], without exploiting information in the source taxonomy. To our knowledge, the most advanced approach to taxonomy integration is the enhanced Nave Bayes (ENB) algorithm proposed by Agrawal and Srikant [2], which we have reviewed and compared with our approach. In [6], AdaBoost is selected as the framework to combine term-features and automatically extracted semantic-features in the context of text categorization. We also choose AdaBoost to combine heterogeneous features (term-features and category-features ), but it is for a different problem (taxonomy integration) and it works in a more complex way (through co-bootstrapping). In [8], an approach called co-boosting is proposed for named entity classification. Essentially co-boosting is a co-training [5] method that attempts to utilize unlabeled data to help classification through exploiting a particular form of redundancy in data: each instance is described by multiple views (disjoint feature sets) which are both compatible and uncorrelated (conditionally independent). However, the multi-view assumption does not hold in the context of taxonomy integration: the set of category features should not be considered as a view because category features alone are not sufficient for classification and they are strongly correlated with term features. In contrast to co-boosting (co-training), co-bootstrapping works with two taxonomies but not two views. CONCLUSION Our main contribution is to propose a new approach, co-bootstrapping , that can effectively exploit the implicit knowledge in the source taxonomy to improve taxonomy integration. The future work may include: theoretical analysis of the co-bootstrapping approach, incorporating commonsense knowledge and domain constraints into the taxonomy integration process, and so forth. ACKNOWLEDGMENTS We would like to thank the anonymous reviewers for their helpful comments and suggestions. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 B ook Di s e a s e Mo v i e Mu s i c Ne ws B ook Di s e a s e Mo v i e Mu s i c Ne ws G Y Y G ENB CB-AB Figure 5: Comparing the macro-averaged F-scores of ENB and CB-AB. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 B ook Di s e a s e Mo v i e Mu s i c Ne ws B ook Di s e a s e Mo v i e Mu s i c Ne ws G Y Y G ENB CB-AB Figure 6: Comparing the micro-averaged F-scores of ENB and CB-AB. 416 REFERENCES [1] Agrawal, R., Bayardo, R. and Srikant, R. Athena: Mining-based Interactive Management of Text Databases. in Proceedings of the 7th International Conference on Extending Database Technology (EDBT), Konstanz, Germany, 2000, 365-379. [2] Agrawal, R. and Srikant, R. On Integrating Catalogs. in Proceedings of the 10th International World Wide Web Conference (WWW), Hong Kong, 2001, 603-612. [3] Baeza-Yates, R. and Ribeiro-Neto, B. Modern Information Retrieval. Addison-Wesley, New York, NY, 1999. [4] Berners-Lee, T., Hendler, J. and Lassila, O. The Semantic Web Scientific American, 2001. [5] Blum, A. and Mitchell, T. Combining Labeled and Unlabeled Data with Co-Training. in Proceedings of the 11th Annual Conference on Computational Learning Theory (COLT), Madison, WI, 1998, 92-100. [6] Cai, L. and Hofmann, T. Text Categorization by Boosting Automatically Extracted Concepts. in Proceedings of the 26th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), Toronto, Canada, 2003, 182-189. [7] Chalupsky, H. OntoMorph: A Translation System for Symbolic Knowledge. in Proceedings of the 7th International Conference on Principles of Knowledge Representation and Reasoning (KR), Breckenridge, CO, 2000, 471-482. [8] Collins, M. and Singer, Y. Unsupervised Models for Named Entity Classification. in Proceedings of the Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora (EMNLP), College Park, MD, 1999, 189-196. [9] Cristianini, N. and Shawe-Taylor, J. An Introduction to Support Vector Machines. Cambridge University Press, Cambridge, UK, 2000. [10] Doan, A., Madhavan, J., Domingos, P. and Halevy, A. Learning to Map between Ontologies on the Semantic Web. in Proceedings of the 11th International World Wide Web Conference (WWW), Hawaii, USA, 2002. [11] Fensel, D. Ontologies: A Silver Bullet for Knowledge Management and Electronic Commerce. Springer-Verlag, 2001. [12] Freund, Y. and Schapire, R.E. A Decision-theoretic Generalization of On-line Learning and an Application to Boosting. Journal of Computer and System Sciences, 55 (1). 119-139. [13] Ichise, R., Takeda, H. and Honiden, S. Rule Induction for Concept Hierarchy Alignment. in Proceedings of the Workshop on Ontologies and Information Sharing at the 17th International Joint Conference on Artificial Intelligence (IJCAI), Seattle, WA, 2001, 26-29. [14] Lacher, M.S. and Groh, G. Facilitating the Exchange of Explicit Knowledge through Ontology Mappings. in Proceedings of the Fourteenth International Florida Artificial Intelligence Research Society Conference (FLAIRS), Key West, FL, 2001, 305-309. [15] McCallum, A. and Nigam, K. A Comparison of Event Models for Naive Bayes Text Classification. in AAAI-98 Workshop on Learning for Text Categorization, Madison, WI, 1998, 41-48. [16] McGuinness, D.L., Fikes, R., Rice, J. and Wilder, S. The Chimaera Ontology Environment. in Proceedings of the 17th National Conference on Artificial Intelligence (AAAI), Austin, TX, 2000, 1123--1124. [17] Meir, R. and Ratsch, G. An Introduction to Boosting and Leveraging. in Mendelson, S. and Smola, A.J. eds. Advanced Lectures on Machine Learning, LNCS, Springer-Verlag , 2003, 119-184. [18] Mitchell, T. Machine Learning. McGraw Hill, 1997. [19] Mitra, P., Wiederhold, G. and Jannink, J. Semi-automatic Integration of Knowledge Sources. in Proceedings of The 2nd International Conference on Information Fusion, Sunnyvale, CA, 1999. [20] Noy, N.F. and Musen, M.A. Anchor-PROMPT: Using Non-Local Context for Semantic Matching. in Proceedings of the Workshop on Ontologies and Information Sharing at the 17th International Joint Conference on Artificial Intelligence (IJCAI), Seattle, WA, 2001, 63-70. [21] Noy, N.F. and Musen, M.A. PROMPT: Algorithm and Tool for Automated Ontology Merging and Alignment. in Proceedings of the National Conference on Artificial Intelligence (AAAI), Austin, TX, 2000, 450-455. [22] Rocchio, J.J. Relevance Feedback in Information Retrieval. in Salton, G. ed. The SMART Retrieval System: Experiments in Automatic Document Processing, Prentice-Hall, 1971, 313-323. [23] Schapire, R.E. The Boosting Approach to Machine Learning: An Overview. in MSRI Workshop on Nonlinear Estimation and Classification, Berkeley, CA, 2002. [24] Schapire, R.E. and Singer, Y. BoosTexter: A Boosting-based System for Text Categorization. Machine Learning, 39 (2/3). 135-168. [25] Schapire, R.E. and Singer, Y. Improved Boosting Algorithms Using Confidence-rated Predictions. Machine Learning, 37 (3). 297-336. [26] Stumme, G. and Maedche, A. FCA-MERGE: Bottom-Up Merging of Ontologies. in Proceedings of the 17th International Joint Conference on Artificial Intelligence (IJCAI), Seattle, WA, 2001, 225-230. [27] Yang, Y. and Liu, X. A Re-examination of Text Categorization Methods. in Proceedings of the 22nd ACM International Conference on Research and Development in Information Retrieval (SIGIR), Berkeley, CA, 1999, 42-49. 417

Taxonomy Integration;Bootstrapping;Semantic Web;Classification;Ontology Mapping;Machine Learning;Boosting

214

WebKhoj: Indian language IR from Multiple Character Encodings

Today web search engines provide the easiest way to reach information on the web. In this scenario, more than 95% of Indian language content on the web is not searchable due to multiple encodings of web pages. Most of these encodings are proprietary and hence need some kind of standardization for making the content accessible via a search engine. In this paper we present a search engine called WebKhoj which is capable of searching multi-script and multi-encoded Indian language content on the web. We describe a language focused crawler and the transcoding processes involved to achieve accessibility of Indian langauge content. In the end we report some of the experiments that were conducted along with results on Indian language web content.

INTRODUCTION India is a multi-language, multi-script country with 22 official languages and 11 written script forms. About a billion people in India use these languages as their first language. English, the most common technical language, is the lingua franca of commerce, government, and the court system, but is not widely understood beyond the middle class and those who can afford formal, foreign-language education. Not only is there a large societal gap between the rich and poor, but that gap appears to be widening due the dominance of English in the society. About 5% of the population (usually the educated class) can understand English as their second language . Hindi is spoken by about 30% [5] of the population, but it is concentrated in urban areas and north-central India , and is still not only foreign but often unpopular in many other regions. Computability of Indian languages could help bridge the societal gaps in education, economy and health-care . However the research and development, availability of standards, support from operating systems and applications in these directions moved very slow due to language heterogeneity. Today this phenomenon can also be observed on the world wide web. The percentage of Indian language content is very less compared to the official languages of United Nations [7]. Even within the available content, majority is not searchable and hence not reachable due to multiple encodings used while authoring such websites. Web publishers of such content were hesitant to use any available standards such as Unicode due to very delayed support from operating systems and browsers in rendering Indic scripts. Even today Hindi is rendered properly only on Windows XP and beyond. Linux has very little support for these languages. Indian languages had barely any support till Windows 2000 operating system. This creates a major bottleneck for web publishers in these languages to get viewership. Despite all these issues, we found considerable amount of content being published on the web. However such content gets unnoticed or gets very less viewership since most of such content is not accessible through search engines due to nonstandard encodings being rendered using proprietary fonts. This paper is organized into seven sections. In the next sub-section we give an introduction to characters, glyphs and fonts in order to appreciate the complexity involved in rendering complex scripts. We then introduce to the complexity of Indic scripts in the sub-section 1.2. In Section 2 we make the problem statement and explain an implementation to solve this problem in Section 3. We report some experiments and results in Section 4, followed by a conclusion in Section 5. 1.1 Fonts, characters and glyphs In the history of mankind the act of writing has always been considered as an activity producing visual results, namely text. The computer has brought a more abstract layer to 801 it, by storing and transmitting textual data. The atomic unit of this abstract representation of text, as defined in the Unicode standard [8], is called a character. And indeed, characters prove to be useful for obtaining alternative (non-visual ) representations of text such as Braille, speech synthesis , etc. The visual representation of a character is called a glyph [8]. Displaying textual contents, whether on screen or on paper, involves translating characters into glyphs, a non-trivial operation for many writing systems. Going in the opposite direction (from glyphs to characters) is known as OCR when done by a machine, or as reading when done by a human [8]. The technology trend over the last few years has been to use characters for most of the text processing and to limit glyph issues to the last stage, namely rendering. At that level, character to glyph translation is handled by increasingly "intelligent" (cf. OpenType and AAT technologies ) fonts and font encodings. Unicode is an effort in this direction. At the same time, restoring the original character stream from a rendered electronic document output for operations such as searching, indexing, or copy-pasting, no general solution exists in today's popular document formats yet. Despite the problems involved, web authors tend to use proprietary encodings due to the complex characteristics of Indic scripts as described in the following section. 1.2 Characteristics of Indic Scripts Indic scripts are phonetic in nature. There are vowels and consonant symbols. The consonants become a syllable after the addition of a vowel sound to it. Further to compound the problem there are `compound syllables' also referred as ligatures. For instance, if we consider `tri' in `triangle', there are three letters corresponding to three sounds `ta', `ra', `yi'. But in the case of Indic Scripts the three are built together to make a single compound consonant having a non-linear structure unlike Latin based languages. The main problem with display of Indic scripts is dealing with their non-linear structures. Glyphs have variable widths and have positional attributes. Vowel signs can be attached to the top, bottom, left and right sides of the base consonant. Vowel signs may also combine with consonants to form independent glyphs. Consonants frequently combine with each other to form complex conjunct glyphs. Although the encoding encapsulates only the basic alphabetic characters , the number of glyphs and their combinations required for the exhaustive rendering of these scripts can be quite large [11]. Since the character to glyph mappings have to be achieved using a 256 character address space, web authors come up with an intelligent way of representing all the characters in the language using some 256 glyphs. Most of these glyphs do not have any semantic significance in the language by themselves. However when displayed together using some positional parameters, they achieve human readable characters . This situation makes the Indian language web content inaccessible for machine processing. PROBLEM STATEMENT Many information seekers use a search engine to begin their Web activity. In this case, users submit a query, typically a list of keywords, and receive a list of Web pages that may be relevant, typically pages that contain the keywords. Today though considerable amount of content is available in Indian languages, users are unable to search such content. Because Indian language websites rely on unique encodings or proprietary extensions of existing standard encodings [11]. This plurality of encodings creates a problem for information retrieval to function as desired. Also many research groups in information retrieval and natural language processing feel the need to collect corpora in these languages from the web in the same way they obtain corpora for other languages [14], [7], [1], [10]. Therefore in order to search or process Indian language websites, we should be able to transliterate all the encodings into one standard encoding and accept the user's queries in the same encoding and build the search index. This task involves many steps. First step is to be able to identify the various encodings in Indian languages on the web. Since these encodings are non-standard, there is no one comprehensive list of such possible encodings. Therefore we need to somehow identify all such encodings and also be able to classify these encodings into the existing types. Second step is to build a transliteration mapping for the given encoding into a standard encoding which is UTF-8 and hence convert any page into a standard and index it. Third step is to be able to accept user's queries in the same standard as that of the transliterated documents which is UTF-8. WEBKHOJ ARCHITECTURE In this paper we report a search engine called WebKhoj which can search web pages in the top 10 Indian languages according to the number of native speakers. WebKhoj cur-rently supports Hindi, Telugu, Tamil, Malayalam, Marathi, Kannada, Bengali, Punjabi, Gujarati and Oriya. Before we describe the architecture of WebKhoj, it is useful to understand how a Web search engine is typically put together and then see its extensions for our task. 3.1 General web search engine Figure 1 shows a general web search engine schematically [2]. The major modules of a web search engine are a Crawler, an Indexer, a Query Engine and a Ranking Engine. Every engine relies on a crawler module to provide the grist for its operation (shown on the left in Figure 1). Crawlers are small programs that browse the Web on the search engine's behalf, similar to how a human user follows links to reach different pages. The programs are given a starting set of URLs whose pages they retrieve from the Web. The crawler extracts URLs appearing in the retrieved pages and give this information to the crawler control module. This module determines what links to visit next and feeds these links back to the crawler. (Some of the functionality of the crawler control module may be implemented by the crawlers themselves .) The crawlers also pass the retrieved pages into a page repository. Crawlers continue visiting the Web until local resources, such as storage, are exhausted. The indexer module extracts all the words from each page and records the URL where each word occurred. The result is a generally very large "lookup table" that can provide all the URLs that point to pages where a given word occurs (the text index in Figure 1). The table is of course limited to the pages that were covered in the crawling process. As mentioned earlier, text indexing of the Web poses special difficulties, due to its size and its rapid rate of change. In addition to these quantitative challenges, the Web calls for some special, less common, kinds of indexes. For example, the indexing 802 Figure 1: General web search engine architecture module may also create a structure index, which reflects the links between pages. Such indexes would not be appropriate for traditional text collections that do not contain links. The collection analysis module is responsible for creating a variety of other indexes. During a crawling and indexing run, search engines must store the pages they retrieve from the Web. The page repository in Figure 1 represents this possibly temporary collection. Search engines sometimes maintain a cache of the pages they have visited beyond the time required to build the index. This cache allows them to serve out result pages very quickly, in addition to providing basic search facilities. Some systems, such as the Internet Archive, have aimed to maintain a very large number of pages for permanent archival purposes. Storage at such a scale again requires special consideration. The query engine module is responsible for receiving and filling search requests from users. The engine relies heavily on the indexes , and sometimes on the page repository. Due to the Web's size and the fact that users typically only enter one or two keywords, result sets are usually very large. Hence the ranking module has the task of sorting the results such that results near the top are the most likely to be what the user is looking for. In the rest of this section we describe the additional modules that were used in a general web search engine to make it work for Indian languages. 3.2 Language focused crawling Since our goal is to be able to search web sites of specific languages, we are looking for a relatively narrow segment of the web. Crawlers that fetch pages related to a particular topic of interest are called topic focused crawlers [6]. While our crawler is very similar to the one mentioned in [6], we use a language identification module instead of a classifier and hence call it as language focused crawling. The language identification module returns the name of the language for a given web page. This module is aware of all the proprietary encodings and also uses a bag of words to recognize unknown encodings from meta-tag information that might be found in an HTML page. In many cases web pages contain more than one language, especially one of the languages being English. This happens since many of the website organzi-ation information such as menu items, or disclaimers and other such formatting information. In some websites such as blogs or forums majority of the content might be English , with Indian language content being a minority. The language identifier module returns a language only if the number of words in a web page are above a given threshold value. 3.3 Transcoding Since Indian language content is being published in multiple encodings on the web, transliteration of encodings to a popular standard such as Unicode [15] is needed. In order to transliterate a non-UTF-8 encoding into UTF-8 which is a Unicode based encoding one has to come up with byte sequence mappings between source and target encodings. Such mappings are rarely one to one mappings, and involve many to one, one to many and many to many mappings of byte sequences. As it was explained in the beginning of this paper, a sequence of bytes represent a sequence of glyphs of a font, which in turn could render a single character or a ligature in the Indic script. Ideally mappings are to be created to all the unique characters in the language, which could be a large number in the order of tens of thousands. Since it would be tedious to list out all the characters and ligatures, we make use of the large number of documents collected by the crawler to come up with a semi-automatic process of generating mappings. We use a simple heuristic to identify the potential character boundaries from byte sequences. First the text from the collected web pages is divided into words using a suitable word tokenizer. Then the algorithm lists all the possible word beginning bytes in both the source and target font encodings . Now each word is scanned from left to right until one such byte occurs in the word. Whenever a valid word beginner occurs in the word, we tokenize at that point, and the byte sequence till that point is treated as a potential character. For example in a given encoding if all the possible word beginning bytes are `a', `b' and `c', a new word `cars' is tokenized as `c', `ars', since neither `r' nor `s' are 803 Figure 2: Transcoding from Jagran encoding to UTF-8 valid word beginners. The byte sequences thus obtained by segmentation are potential characters or ligatures in that language. Once such segmentation is done, the frequency of such byte sequences (or potential characters) is calculated. It was found from our experiments that the ranks based on the normalized frequency of such potential characters is highly cor-related (we present more details in our experiments section). Therefore we use this algorithm to come up initial suggested mappings for transcoding, and then the user would manually correct any errors by going through the font mappings as shown in the Figure 2. The transcoding tool sorts the potential characters according to their ranks, so that the user would find the equivalent match in the target encoding among the top few possibilities. Also since the mappings are ordered based on the normalized frequency found in the corpus, mapping source and target bytes in this order ensures optimal precision that can be obtained from a set of mappings. Once such transcoder mappings are generated for all possible encodings in Indian languages, a transcoding module is called during indexing of the web documents. If a web document is encoded in an encoding other than UTF-8, the transcoder module is called to transliterate the encoding of the given web page into UTF-8 standard. In order to do this, the HTML page is parsed to obtain its document object model (DOM) using the JTidy utility 1 . All the nodes of type "font" are extracted from the DOM and the font encoding is checked against a known set of encodings on the web. Based on the font encoding, the appropriate transcoder mappings are used to transliterate the relevant text into UTF-8. One word is transcoded at a time. In order to transcode, the maximum byte sequence available in the mapping table is used to transliterate the encodings and the process is repeated to the remaining substring of the word. This transliterated document is then sent to the indexer to build the inverted index. 3.4 Retrieval Algorithm The score of query q for document d is defined in terms of TFIDF [13] metric as shown below: score (q, d) = c(q, d).q n (q).( X t in q tf (t in d).idf (t)) 1 JTidy is a Java implementation of Dave Raggett's HTML tidy. JTidy can be found at http://jtidy.sourceforge.net 3.4.1 tf (term frequency) `tf' (also known as term frequency) is a score factor based on a term or phrase's frequency in a document. Terms and phrases repeated in a document indicate the topic of the document, so implementations of this score usually return larger values when frequency is large, and smaller values when frequency is small. 3.4.2 idf (inverse document frequency) `idf' is a score factor based on a term's document frequency (the number of documents which contain the term). Terms that occur in fewer documents are better discriminators of topic, so implemenations of this method usually return larger values for rare terms, and smaller values for common terms. 3.4.3 c (coverage of query terms) `c' is a score factor based on the fraction of all query terms that a document contains. This value is multiplied into scores. The presence of a large portion of the query terms indicates a better match with the query, so implemenations of this function usually return larger values when the ratio between these parameters is large and smaller values when the ratio between them is small. 3.4.4 q n (query normalization) This is the normalization value for a query given the sum of the squared weights of each of the query terms. This value is then multiplied into the weight of each query term. This does not affect ranking, but rather just attempts to make scores from different queries comparable. 3.5 User Interface Currently there is no easy means to key-in UTF-8 queries to the search engine using the normal keyboard. So WebKhoj is provided with a soft keyboard which displays the UTF-8 character set of the language on the screen. The layout of the keys is very similar to the Keylekh layout [9]. We also tried providing a roman to local language transliteration keyboard which dynamically renders Indian language text when its phonetic equivalent is typed using roman characters . We had student volunteers from a near by village to try out the keyboards. However, we found that the students who are taught in the local language in schools are not comfortable with English symbols. Also within the local language, the way symbols are taught in schools is much 804 Figure 3: Hindi soft keyboard user interface for WebKhoj search engine Figure 4: Search results being displayed for a Hindi query in UTF-8 805 different from the way UTF-8 characters need to be typed in. However, with some training these students were able to adapt to the soft keyboard. Currently soft keyboards for 10 Indian languages are provided in the searching interface. One language is shown to the user at any given instance. The user can change the keyboard to a different language by clicking on the desired language hyperlink displayed on the interface as shown in Figure 3. After thus framing the query, the user can search for the web documents, and the results are ranked and displayed much like Google as shown in Figure 4. 3.6 Word spelling normalization Indian language words face standardization issues in spelling, thereby resulting in multiple spelling variants for the same word. For example we found widely used spelling variations for the hindi word `angrezi' as shown below The major reasons for this phenomenon can be attributed to unavailability of proper website authoring tools equipped with spell checkers for Indian languages and multiple dialects of spoken language, transliteration of proper names and words borrowed from foreign languages whose spellings are not standardized. While we have to handle Indian language words with spelling variations and errors, we also showed that a considerable percentage of foreign language words mainly English have entered into Indian language usage which cannot be ignored. While such words are being frequently used by people, there is no standardization in spelling for such words thereby resulting in huge variations due to transliteration. Given such variations in spelling it becomes difficult for web Information Retrieval applications built for Indian languages, since finding relevant documents would require more than performing an exact string match. It was shown that normalization rules for specific languages work best with spelling normalization problems. We make use of a set of rules [12] to normalize the words before indexing them or looking them up from the index. These rules are language specific and we describe the rules for Hindi in the next sub-sections. We achieve normalization of word spellings by mapping the alphabet of the given language L into another alphabet L where L L. We use the following rules to achieve such a normalized mapping. 3.6.1 Mapping chandrabindu to bindu Often people tend to use chandrabindu (a half-moon with a dot) and bindu (a dot on top of alphabet) interchangeably. Lots of confusion exists in common language usage on which to use when. In order to equate all such words we convert all occurrences of chandrabindu to bindu, which would equate all the words shown below. 3.6.2 nukta deletion Unicode contains 10 consonant characters with nukta (a dot under consonant) and one nukta character itself. We delete all occurrences of nukta character and replace all consonants with nuktas with their corresponding consonant character . This would equate words like the ones shown below. 3.6.3 halanth deletion Hindi and many other Indian languages face the problems of 'schwa' (the default vowel 'a' that occurs with every consonant) deletion. Lots of spelling variations occur due to 'schwa' deletion. In order to normalize such words we delete all the halanth characters in the given word before making a string match. This operation would normalize words as shown in the example below. 3.6.4 vowel shortening Many times in written script people use shorter vowels instead of longer ones or vice versa. Therefore in our application we convert all the longer vowels to their corresponding shorter ones. Using this feature we can normalize words as shown in this example. 3.6.5 chandra deletion 'chandra' (half-moon) is used for vowel rounding. Usually words borrowed from English at times require vowel rounding operation. For example the word "documentary". But this character is used inconsistently many times. Therefore deleting such a character would normalize the words where vowel rounding has been used. These rules were compared with many approximate string matching algorithms are were found to result in a better f-measure [12]. EXPERIMENTS AND DISCUSSION We report here some experiments that were conducted in transcoding the proprietary encodings and present some statistics from our language focused crawl about the Indian language web. The transcoding tool was designed to generate mappings between two encodings in a semi-automatic fashion. In order to achieve this the tool automatically gives some mapping suggestions based on the rank correlation of the two encodings in question. We found that the byte sequences from two encodings of same language correlate very well, by looking at the Spearman's rank correlation coefficient. In-tuitively this phenomenon can be understood as the convergence of unique lexicon from two encodings from sufficiently large corpus, since they both belong to the same language. To find the amount of correlation, we experimented with two different encodings from Hindi. We ran the character segmentation algorithm and computed the normalized frequencies as mentioned above and ranked the character sequences in both the encodings from a corpus of 2,000 documents from each of these encodings. We manually marked 806 the corresponding frequency based rank positions of a given character or a ligature from these encodings and calculated the Spearman's rank correlation coefficient. We then plotted a graph with the Spearman's correlation coefficient on y-axis and the number of mappings on x-axis as shown in Figure 5. We observed that the rank correlation is 100% for the first 25 mappings that were automatically generated, and are close to 90% for the first 200 mappings which can achieve a transcoding precision of above 90%. 0 0.2 0.4 0.6 0.8 1 0 20 40 60 80 100 120 140 160 180 200 Rank Correlation coefficient Number of byte sequences Rank Correlation of byte sequence frequencies &quot;Spearman Correlation&quot; Figure 5: Spearman's rank correlation for number of byte sequences between Jagran and Webdunia font encodings Since these byte sequences are an ordered set, ordered by their normalized frequency, the precision of transliteration obtained by providing mappings between encodings in the order provided in the ordered set is optimal. We have observed that with about 2,000 encoding mappings for each encoding on average once can achieve around 99% precision. However this number also depends on the language complexity . For instance, the number of encodings required in Telugu transliteration is more than the number of encodings required in Hindi to obtain the same amount of precision. We now report some of our experiments on the Indian language focused crawling. We ran a daily crawl for 6 months period. Our crawler was focused to fetch content in the top 10 spoken languages in India, namely Hindi, Telugu, Tamil, Bengali, Marathi, Gujarati, Kannada, Malayalam, Oriya and Punjabi. In another experiment, in order to find the effectiveness of language focused crawling, we executed the crawler in two modes with a set of 100 seed URLs which constitute popular Indian based web portals, news sites and home pages of people of Indian origin. In the first mode it was executed without language focus restriction using a pure FIFO crawl queue while the second mode was with language focus restriction using a priority queue from which the crawler fetched the next crawl URL. We plotted the number of relevant pages fetched in the first 50,000 URLs in both the runs as shown in the Figure 6. The relevance of the fetched pages was calculated by checking the encoding on that page. It can be clearly seen that language focus restriction on the crawler helps in downloading more relevant pages. From the 6 month crawl, about half a million unique documents were collected from all the languages. Unique web pages were picked after eliminating approximate duplicate pages using shingling technique [4]. These half a million pages were distributed across the 10 languages as shown in the Figure 7. Figure 8 shows the population of people speaking the various Indian languages [3]. It can be observed that even within India there is a divide in the web publishing activity in various languages. For instance it can be observed that the content is actively getting published in south Indian languages like Telugu, Tamil and Malayalam when compared to the northern languages such as Marathi, Gujarati, Oriya, Bengali and Punjabi. Hindi has the majority of content published on the web but Hindi is also the language spoken by majority of Indian population. It can be seen from Figure 10 that a very few websites publish content using a global standard such as Unicode. This explains the reason for most of the Indian language not being indexed or searchable by the present day popular web search engines. On the other hand it can be seen from Figure 9 and Figure 11 that the number of unique encodings found on the web for these languages is almost equivalent to the number of websites. This observation suggests that every web publisher is coming up with their own proprietary encodings to publish web content. We did not consider the websites that publish using images in this study, but our preliminary study suggests that there are a large number of websites that publish content as images as well. Figure 6: Crawl with and without language focus Figure 7: Languages on x-axis and number of unique web pages on y-axis 807 Figure 8: Languages on x-axis and number of native speakers on y-axis Figure 9: Languages on x-axis and number of encodings found on web including UTF-8 on y-axis Figure 10: Languages on x-axis and number of UTF-8 websites on y-axis CONCLUSIONS In this paper we discussed the importance of being able to search the Indian language web content and presented a web search engine which takes the UTF-8 queries from a soft keyboard and capable of searching 10 most spoken Indian languages' web pages encoded in multiple encodings. We presented a language focussed crawler which can fetch web pages of specific languages and also the distribution of Figure 11: Languages on x-axis and number of websites (web servers) on y-axis the Indian language content on web based on the pages that were crawled. This distribution clearly shows the need for processes and algorithms to transcode non-Unicode encodings to Unicode. Hence we have discussed a semi-automatic algorithm to generate the mappings between different encodings . This shows that transcoding of proprietary encodings into a standard encoding makes Indian language web content accessible through search engines. ACKNOWLEDGMENTS We would like to thank the Department of Science and Technology, Ministry of Communications and IT, Government of India for funding this project. REFERENCES [1] J. Allan, J. Aslam, N. Belkin, C. Buckley, J. Callan, B. Croft, S. Dumais, N. Fuhr, D. Harman, D. J. Harper, D. Hiemstra, T. Hofmann, E. Hovy, W. Kraaij, J. Lafferty, V. Lavrenko, D. Lewis, L. Liddy, R. Manmatha, A. McCallum, J. Ponte, J. Prager, D. Radev, P. Resnik, S. Robertson, R. Rosenfeld, S. Roukos, M. Sanderson, R. Schwartz, A. Singhal, A. Smeaton, H. Turtle, E. Voorhees, R. Weischedel, J. Xu, and C. Zhai. Challenges in Information Retrieval and Language Modeling: Report of a Workshop held at the Center for Intelligent Information Retrieval, University of Massachusetts Amherst, September 2002. SIGIR Forum, 37(1):3147, 2003. [2] A. Arasu, J. Cho, H. Garcia-Molina, A. Paepcke, and S. Raghavan. Searching the Web. ACM Trans. Inter. Tech., 1(1):243, 2001. [3] G. B. 14th ed. Ethnologue: Languages of the World. SIL International, Dallas, TX, 2003. [4] S. Brin, J. Davis, and H. Garcia-Molina. Copy Detection Mechanisms for Digital Documents. In SIGMOD '95: Proceedings of the 1995 ACM SIGMOD International Conference on Management of Data, pages 398409, New York, NY, USA, 1995. ACM Press. [5] G. E. Burkhart, S. E. Goodman, A. Mehta, and L. Press. The Internet in India: Better times ahead? Commun. ACM, 41(11):2126, 1998. 808 [6] S. Chakrabarti, K. Punera, and M. Subramanyam. Accelerated Focused Crawling through Online Relevance Feedback. In WWW '02: Proceedings of the 11th International Conference on World Wide Web, pages 148159, New York, NY, USA, 2002. ACM Press. [7] F. Gey, N. Kando, and C. Peters. Cross Language Information Retrieval: A Research Roadmap. SIGIR Forum, 36(2):7280, 2002. [8] Y. Haralambous and G. Bella. Injecting Information into Atomic Units of Text. In DocEng '05: Proceedings of the 2005 ACM Symposium on Document Engineering, pages 134142, New York, NY, USA, 2005. ACM Press. [9] A. Joshi, A. Ganu, A. Chand, V. Parmar, and G. Mathur. Keylekh: a Keyboard for Text Entry in Indic Scripts. In CHI '04: CHI '04 Extended Abstracts on Human Factors in Computing Systems, pages 928942, New York, NY, USA, 2004. ACM Press. [10] L. S. Larkey, M. E. Connell, and N. Abduljaleel. Hindi CLIR in thirty days. ACM Transactions on Asian Language Information Processing (TALIP), 2(2):130142, 2003. [11] D. P. Madalli. Unicode for Multilingual Representation in Digital Libraries from the Indian Perspective. In JCDL '02: Proceedings of the 2nd ACM/IEEE-CS Joint Conference on Digital Libraries, pages 398398, New York, NY, USA, 2002. ACM Press. [12] P. Pingali and V. Varma. Word Normalization in Indian Languages. In ICON05: Proceedings of the 2005 International Conference on Natural Language Processing, 2005. [13] G. Salton and C. Buckley. Term-weighting Approaches in Automatic Text Retrieval. Information Process. Management, 24(5):513523, 1988. [14] S. Strassel, M. Maxwell, and C. Cieri. Linguistic Resource Creation for Research and Technology Development: A Recent Experiment. ACM Transactions on Asian Language Information Processing (TALIP), 2(2):101117, 2003. [15] F. Yergeau. UTF-8, a transformation format of ISO 10646. RFC Editor, United States, 2003. 809

web search;Indian languages;non-standard encodings

215

What's There and What's Not? Focused Crawling for Missing Documents in Digital Libraries

Some large scale topical digital libraries, such as CiteSeer, harvest online academic documents by crawling open-access archives, university and author homepages, and authors' self-submissions. While these approaches have so far built reasonable size libraries, they can suffer from having only a portion of the documents from specific publishing venues. We propose to use alternative online resources and techniques that maximally exploit other resources to build the complete document collection of any given publication venue. We investigate the feasibility of using publication metadata to guide the crawler towards authors' homepages to harvest what is missing from a digital library collection. We collect a real-world dataset from two Computer Science publishing venues, involving a total of 593 unique authors over a time frame of 1998 to 2004. We then identify the missing papers that are not indexed by CiteSeer. Using a fully automatic heuristic-based system that has the capability of locating authors' homepages and then using focused crawling to download the desired papers, we demonstrate that it is practical to harvest using a focused crawler academic papers that are missing from our digital library. Our harvester achieves a performance with an average recall level of 0.82 overall and 0.75 for those missing documents. Evaluation of the crawler's performance based on the harvest rate shows definite advantages over other crawling approaches and consistently outperforms a defined baseline crawler on a number of measures.

INTRODUCTION Digital libraries that are based on active crawling methods such as CiteSeer often have missing documents in collections of archived publications, such as ACM and IEEE. How do such digital libraries find and obtain those missing? We propose using external resources of publication metadata and focused crawlers to search the Web for those missing. The basic concept of a focused crawler (also known as a topical crawlers) [1], is based on a crawling strategy that relevant Web pages contain more relevant links, and these relevant links should be explored first. Initially, the measure of relevancy was based on keywords matching; connectivity-based metrics were later introduced [2]. In [3] the concept of a focused crawler was formally introduced: a crawler that seeks, acquires, indexes, and maintains pages on a specific set of topics that represent a relatively narrow segment of the Web. Today, focused crawling techniques have become more important for building specialty and niche (vertical) search engines While both the sheer volume of the Web and its highly dynamic content increasingly challenge the task of document collection, digital libraries based on crawling benefit from focused crawlers since they can quickly harvest a high-quality subset of the relevant online documents. Current approaches to harvesting online academic documents normally consist of focused crawling of open-access archives, author and institution web sites and directories of authors' self-submissions . A random sample of 150 journals and conferences in Computer Science show that less than 10% have websites that are open to crawlers. Many of the top publishing venues that have their documents electronically available to subscribers such as the ACM Digital Library, the IEEE Digital, Library or the Springer-Verlag Digital Library, normally use access permission techniques and robots.txt to ban crawlers. A recent study indicates that CiteSeer indexes 425, 000 unique research documents related to Computer Science, DBLP contains 500,464 records and there are 141,345 records in the Association for Computing Machinery (ACM) Digital Library and 825,826 records in the more comprehensive ACM Guide [4]. The study also shows that in CiteSeer there is an overlapping portion of 86, 467 documents (20.2% of CiteSeer's total archive) comprising 17.3% of the Digital Bibliography & Library Project (DBLP) archive. This research investigates alternative online resources and focused crawling techniques to build a complete document collection for any given publication venue. We propose to answer the following: Q1 - What are the best focused crawling techniques to maximally exploit online resources, in order to harvest the desired papers effectively and efficiently? Q2 Is it effective to use authors' homepages as alternative online resources to find the missing documents? Q3 How can the above methods be automated to effectively obtain missing documents? The rest of the paper is organized as follows. In section 2 we present a review of related work. In Section 3 we cover in much detail the design rationale of the system. In Section 4 we describe how we collect data and perform the evaluation, and present the results with discussion. Finally, we conclude the paper with future work proposed in Section 5. RELATED WORK The focused crawling literature shows that much has been focused on enhancing the dynamic performance, scalability, effectiveness, and efficiency of the crawler, namely, harvesting higher-quality documents in a shorter period of time. Breadth-first searching is probably the simplest strategy for crawling, i.e. traversing the Web in a way that a directed graph is traveled using a breadth-first search algorithm. Interestingly, a breadth-first crawler is found to be capable of yielding high-quality documents at an early stage of the crawl [5]. Although more sophisticated crawlers tend to retrieve even higher quality pages than their breadth-first counterparts, they are usually computationally more expensive. In our study, we use a multi-threaded breadth-first crawler as a baseline to compare to our own crawling method. Best-first crawling attempts to direct the crawler towards the best (i.e. most relevant in terms of topic relevance) documents. Different heuristics, such as link-based criteria, lexical similarity measures, contextual knowledge, and fine-tuned combinations of such have been explored in a number of studies over the years. In [2], the authors find that PageRank [6] can yield the best performance when ordering seed URLs. However, a more recent study [7] shows that PageRank metrics may just be too general in context without regard to the specific target topic. An updated version of PageRank algorithm which reflects the importance with respect to a particular topic has been proposed [8]. In [3], a Bayesian classifier is used to estimate the probability that a page belongs to the target topic, in a way that a node belongs to a certain position in an existing taxonomy hierarchy. In [9], a keyword-based vector space model is used to calculate the similarity of Web pages to the seed URLs, and if the similarity is above a certain threshold, the pages are downloaded and indexed, and their out-going links are followed. A focused crawler [10] based on context graphs is proposed by so that the crawler can extract information about the context within which desired documents are usually found. A set of classifiers are then trained to classify in-degree Web pages according to an estimation of their relevance to the topic. The relevance estimation then navigates the crawler towards desired documents. Crawlers with a probability model are used for calculating priorities, which combines Web page content-based learning, URL token-based learning, and link-based learning [11]. In a later work, [12] takes into account the users' access behavior and re-tunes the previous model to connect this behavior with the predicate satisfaction probability of the candidate Web pages waiting to be crawled. An interesting "reversed" approach is proposed in [13], which suggests a given scientific document from a digital library be used as an input to the focused crawler. The main title and reference titles of the document are extracted and used to train a classifier to learn topical knowledge. The crawler is then guided by such knowledge to discover other topic-relevant documents on the Web. More up-to-date reviews of focused crawling algorithms are presented in [14] and [15]. In [14], five different methods are implemented and evaluated within a unified evaluation framework on small and large datasets. Here we discuss two studies that bear similarities to ours. The HPSearch and Mops presented in [16] support the search for research papers close to the homepages of certain scientists. However, their system does not investigate the issues of document harvesting for digital libraries for different publishing venues. Furthermore, our system outperforms theirs in terms of the percentage of correct homepages returned. In a more recent study [17], a Paper Search Engine (PaSE) is proposed, which uses citation information to locate online copies of scientific documents. While their study addresses a different research question, the PaSE system employs similar heuristics as we do to favor certain out-going links in order to quickly locate academic papers. SYSTEM DESIGN We develop an automated system in which document metadata is used to automatically locate the homepages of the authors and focused crawl these homepages with the intent of finding missing documents. Our system, shown in Figure 1, consists of a Homepage Aggregator and a smart Focused Crawler. The system accepts a user's request to harvest the desired papers published in a specific venue (e.g. a conference or a journal). The Homepage Aggregator will query a Public Metadata Repository and extract useful metadata heuristics to assist in quickly and accurately locating URLs of the authors' homepages. A list of such URLs will be inserted into the Homepage URL Database. The Crawler uses focused crawling techniques to search the domains for desired publications. It accepts the seed URLs as an input and uses them as starting points for the crawl. The Crawler uses anchor text to determine link priorities and quickly navigates through the websites using to get to the desired academic papers. The harvested documents will be stored in the Document Database. 302 Figure 1. System Architecture 3.2 Using Metadata to Locate Homepages Crawling authors' homepages first requires the system to be able to locate such websites quickly and accurately. A study of the literature indicates that personal website and homepage finding have been studied a lot since the birth of WWW. In [18], the authors present AHOY! as the first working system for personal homepage finding, which can filter irrelevant pages based on pattern matching heuristics. Later, the TREC (Text REtrieval Conference) hosted the task of Web homepage finding in 2001 and its subsequent years, and algorithms based on link analysis, linguistic cues, and machine learning etc. are proposed [19, 20, 21]. Examples of current working systems include HomePageSearch (hpsearch.uni-trier.de) which is a Homepage Aggregator mainly for computer scientists, and compiled directories (e.g. Google Directory) See Figure 2 for the architecture of the Homepage Aggregator component. Figure 2. Architecture of the Homepage Aggregator The goal of the Homepage Aggregator is to look for homepages of the authors and save them as seed URLs to feed the Focused Crawler. First it queries the Metadata Repository and retrieves the document metadata. For each author, it extracts from metadata a value pair of (N, P), where N is the name of the author and P is the name of the venue (with a number of variations) in which the paper is published. A list of such pairs is then submitted to a Web search engine. Pages returned by the search engine will go through a Homepage Filter where we use metadata heuristics to remove false positives (pages that are not likely to be the homepages of the authors) and disambiguate among namesakes, if there is any. Different priority weights are assigned to the remaining pages according to their likelihood of being the homepage of the author. The more likely it's the homepage of the author, the higher priority it receives. Eventually the page with the highest priority weights will be inserted into the Homepage URL Database, and will be crawled later. Recall that we extract from metadata a pair value of (N, P). Now let U be the URL and T be the title of a Web page P returned by the Web search engine. When there are more than two authors for the same paper, assume Ui are the URLs of the homepages of other authors already found by the system. We have incorporated the findings in [16] about major characteristics of personal homepages. The metadata heuristics employed in the Homepage Filter are explained in Table 1. Table 1. Heuristics Employed in Homepage Filter Function Heuristic Rules Remove false positives Remove U if U or T indicates a publisher's website. Remove U if U or T indicates a digital library. Remove U if U points to a file other than .htm/.html Disambiguate between namesakes Choose U among the candidates if U is in the same domain as Ui. Remove U if its parent-domain is already found by the system. Assign priority U receives high priority if T contains N and any of the following: homepage (home page), web (website), research, publication, papers. U receives medium priority if T contains any of the following: homepage (home page), web (website), research, publication, papers. U receives low priority when neither one of the above two rules is fired. 3.3 Crawler Architecture The Focused Crawler crawls web pages, using heuristics to quickly navigate to the publications. The architecture of the component is shown in Figure 3. The crawler accepts two primary sets of inputs that vary for each crawl. The first is a set of seed URLs that are the starting points of the crawl. These are added to the crawl queue at low priority. The second set of inputs is a collection of domain names that the crawler is permitted to crawl. Once the seed URLs are entered into the queue, the crawler threads are started. Each thread gets one URL from the priority queue, and downloads the page that it points to. After a page is downloaded, the out-going links are examined and those matched with the ignored list are removed, either because they are out of the target domain or because their MIME types are not processed by the crawler. At this point, if a PDF/PostScript document is found, it will be inserted into the Document Database. The rest of the out-going links will each be classified as high, medium, or low priority, and inserted into different priority queues. Metadata Extractor Web Search Engine Homepage Filter Homepage Aggregator Public Metadata Repository Homepage URL Database Focused Crawler Metadata Heuristics Public Metadata Repository Document Database Homepage URL Database Homepage Aggregator 303 Figure 3. Architecture of the Focused Crawler In order to concentrate or limit the crawls towards only desirable content, the crawler is provided with three lists for reference. The contents of the lists may be changed depending on the types of domains being crawled. The Ignore List is a set of file types that are to be ignored by the crawler. The most common types of URLs that are ignored by the crawler are links to image files. The list can also include parts of the domain(s) being crawled, which the crawler is not supposed to visit. Table 2 shows a sample Ignore List. Table 2. Sample Ignore List File Types .jpg, .bmp, .gif, .png, .jpeg, .mpg, .mpeg, .avi http://clgiles.ist.psu.edu/picture.html Domains http://clgiles.ist.psu.edu/courses.html Files of type JPG, BMP etc will be ignored during the crawl. Also any outgoing links to pages within the ignored domains will not be considered for crawling. The Allow List on the other hand is a collection of domain names that make up the crawl space of the crawler. Links pointing outside the specified domains are ignored by the crawler (unless they are determined to be research documents). This list is useful to limit the breadth of the crawl to only those domains that are of interest. Table 3 shows a sample Allow List. Table 3. Sample Allow List Domains http://clgiles.ist.psu.edu So the link http://clgiles.ist.psu.edu will be considered for crawling if it's discovered. Priority lists contain a set of keywords and their assigned weights that are used to determine the priorities of the extracted links. The links will be visited by the crawler in the order of their assigned priority. The Crawl Queue holds the discovered URLs that are yet to be crawled. This queue consists of three sub-queues: High-priority, Medium-priority and Low-priority queue. The Low-priority queue is the default queue. The seed URLs are entered into this queue. We adopt a simple yet very effective heuristics to make the priority classification based upon the likelihood of the link eventually leading to academic publications. We first train a classifier with data collected from two publishing venues: the Very Large Data Bases (VLDB) Conference and the Text REtrieval Conference (TREC). Several crawls are carried out with a breadth-first policy. The logs of the crawls are analyzed and a traverse tree is generated for each of the crawl that indicates the URLs visited and the link path that is followed by the crawler to reach the desired publications. Consider a small website having 11 pages as shown in Figure 4. Figure 4. Sample Website The circles represent URL's in the website and the arrows are the hyperlinks from one page to another. The link structure shown is that which is followed by the breadth-first crawler to visit each URL. All other links such as those that may point outside the domain are ignored in the above diagram. The node marked with `S' is the seed or start URL. The nodes marked with `P' are research document files that are detected by the crawler. Now the links that are of interest to us are S A P and S C D P. The anchor text contained in these links `SA', `AP', `SC', `SD', `DP' is extracted and marked as `interesting'. The text in the remainder of the links is also noted, but goes in `not interesting' set. Similar analysis is done on all the logs that are generated by the breadth-first crawl. All the keywords that are commonly occurring in the "interesting" class and not so commonly occurring in the "non-interesting" class are extracted. Weights are assigned to each of these keywords depending on their placement in the link structure. The keywords closer to the documents are given more weight that those closer to the seed URL. For e.g. keyword `SA' has a lesser weight than keyword `DP' as `DP' is closer to P than to S as opposed to `SA'. The formula for calculating keyword weight is: W (OT oq ) = D (Q) / D (P) (I) where OT oq is the anchor text of the out-going link from page O to page Q; P is the desired academic paper found by following the link from O to Q; D(P) denotes the distance (number of hops) between P and the starting URL S on the path S A B C P E D P Depth 1 Depth 2 Depth 3 Depth 0 Homepage URL Database Document Database Ignore List Priority Queues Priority Heuristics Download Pages Link Extractor Link Filter Link Priority Analyzer PDF/PS Documents Crawler Thread Allow List 304 S...OQ...P; D(Q) denotes the distance between Q and the starting URL S on the path S...OQ. Now that a list of anchor texts and their corresponding priority weights has been compiled during the training process, we can classify each of them into different priority categories according to the weights. Table 4 shows a few samples extracted from our list. Table 4. Sample Anchor Texts Priority Anchor Texts p_High volume, pub, paper, conf, journal, content, program, research, list p_Medium topic, faculty, people, group, lab We now need to consider how to prioritize out-going links that are more likely to lead to desired academic publications. The anchor text in these links is compared against the weighted keywords. If any of the weighted keywords are present in the text, the comparison is considered to be successful. There are no keywords having more than one weight. The final priority of the link is calculated by the following function. The priority of a link may also depend on the priority of its parent. This is mainly due to the fact that not all the links that emerge from a page with a medium or high priority may lead to a research document. For e.g. in Figure 4 the node `C' will be crawled with a medium priority, however only node `D' leads to a research document. The priority of the node `E' is thus reduced to low as it will not have a weighted keyword attached to it and that of `D' is increased to high. The priorities of links thus established are used to insert the link in the proper priority queue for crawling. In order to achieve high efficiency, the crawler spawns multiple threads which will be fed with URLs on the descending order of priority. When there is no URL left in the priority queues and no crawler thread is currently running, the crawling task is finished. RESULTS AND DISCUSSION We have collected data from two Computer Science publication venues: the ACM SIGMOD International Workshop on the Web and Databases (WebDB), first held in 1998 and then each year in conjunction with the annual ACM SIGMOD Conference, and Journal of Artificial Intelligence Research (JAIR), which was established in 1993 both as an electronic scientific journals and a hard-copy semiyearly published by AAAI Press. We choose these two venues because both of them are highly selective venues with less than a 25% acceptance rate and we want to observe if there is a major difference of performance between conferences and journals. We have extracted the metadata of WebDB and JAIR from the DBLP repository. By analyzing these metadata, we successfully identify the 593 unique authors who have in total published 289 papers in either one of these two venues during the period from 1998 to 2004. Please see Table 5 for more details of the dataset. Table 5. Statistics of the collected data WebDB JAIR Year Unique Authors Publication Unique Authors Publication 1998 32 13 40 20 1999 51 17 50 28 2000 61 20 33 20 2001 51 18 45 25 2002 47 17 64 27 2003 56 17 72 30 2004 51 16 57 21 Total 285 118 308 171 In order to examine whether our approach is effective in recovering those missing documents from a digital library, we use the CiteSeer Scientific Digital Library as another data source. Cross-referencing the metadata of each of the two venues from DBLP, we successfully identified 30 out of 118 (25.42%) WebDB papers and 46 out of 171 (26.90%) JAIR papers that are not indexed by CiteSeer (see Figure 5 for details). This is done by exact title-matching between the records in the DBLP metadata repository and the CiteSeer document archive. WebDB 8 12 17 13 10 14 14 5 5 3 5 7 3 2 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 1998 1999 2000 2001 2002 2003 2004 NOT Indexed Indexed JAIR 17 25 20 21 17 16 15 3 3 0 4 10 14 6 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 1998 1999 2000 2001 2002 2003 2004 NOT Indexed Indexed Figure 5. Coverage of the two venues by CiteSeer The metadata extracted from DBLP are also used as heuristics to locate the homepages of the 593 authors. The name of the author // Get_Priority(): Returns the priority for link L T with anchor text T which has weight W T . // Low=0, Medium=1, High=2 (for weight and priority) Get_Priority { If W T = 0 and (Priority(Parent(L T )) &gt; 0 then Priority(L T ) = Priority(Parent(L T )) -1; Else if W T &gt; 0 Priority(L T ) = W T ; End IF Return (Priority(L T )); } 305 and the corresponding venue (with a number of variations) are submitted to Google API and the first 10 URLs returned are parsed automatically by the Homepage Filter component. Using the heuristics discussed in the previous section, we assign priority weights to each of the URLs. For each author, URLs with the highest priority weights are inserted into the URL Database and crawled by the Focused Crawler at a later stage. We have manually examined the records in the URL Database in order to evaluate the effectiveness of the Homepage Aggregator. In total, homepages of 539 authors (90.89%) have been found. Details about the 54 authors whose homepages cannot be found by the system are shown in Table 6. Here we define Non-U.S. authors to be those whose affiliations are not currently in the States. Table 6. Number of authors whose homepages are not found WebDB JAIR U.S. Authors 13 6 Non-U.S. Authors 25 10 Total (Percentage) 38 (13.33%) 16 (5.19%) There are only 2 papers ([22], [23]) of which all the authors' homepages are not found by the system, which account for less than 1% of the 289 papers in our data set. In other words, although the system fails to locate the homepages of about 9% of the authors, it is not a major performance impact on the document recall and the crawler should still be able to find 99.31% of all the papers. For the cases where the system fails to locate some of the homepages, we notice that most of the 19 U.S. authors whose homepages are not found were actually in their graduate programs when they co-authored the paper, and their Web presences seem to have disappeared after graduation. In addition, there's a significant difference between the numbers of U.S. and non-U.S. authors whose homepages cannot be found, with non-U.S. almost twice the number of U.S. authors. Since this is our initial attempt limited to only the domain of computer science, whether this difference holds true for other disciplines and the reason behind remain an open question. Finally, there are several cases where the homepages of those with famous names actually show up instead of the desired authors. For example, a search via Google API for the first author in [24] returns the homepage of a comic artist. The top 5 websites for George Russell, the first author of [25], happen to belong to that of a famous Jazz musician. There are also a few cases where the search engine actually returns the homepage of the co-author instead of the author himself, because the author's name is listed on the co-author's page as a collaborator and the co-author's page receives a higher page ranking. All these indicate that the disambiguation capability needs to be improved. 4.1 Finding Desired Academic Publications When the crawl is finished, we manually examine the downloaded PDF/PostScript documents in order to evaluate the performance of the crawler. In total, the crawler has acquired 236 out of the 289 papers (81.66%) published in WebDB (100 out of 118, 84.75%) and JAIR (136 out of 171, 79.53%) from 1998 to 2004. For details of the results for each venue, please see Figure 6 and 7. WebDB, 1998 - 2004 0 5 10 15 20 25 1998 1999 2000 2001 2002 2003 2004 Numb er Papers found Papers published Figure 6. Number of WebDB Papers JAIR,1998 - 2004 0 5 10 15 20 25 30 35 1998 1999 2000 2001 2002 2003 2004 Number Papers found Papers published Figure 7. Number of JAIR Papers Here we adopt one of the performance metrics, recall level, first proposed in [16] and used in [17]. Recall level is defined as: (i) = | S(i) T | / | T | where S(i) is the set of documents downloaded by the crawler during a crawl on the dataset of a calendar year i; T is the set of desired documents, which in this study are the papers published by a specific venue in the same calendar year. This measure represents the capability of the system to capture desired academic papers. Overall, our system has achieved a recall level of 0.8475 for WebDB and 0.7953 for JAIR documents. See Figure 8 for more details. It's interesting to note that while the recall level of WebDB is constantly increasing until reaching 1.0 in the last two years, the recall level of JAIR seems to fluctuate around 0.8 over the 7-years period. We find that 29 out of the 35 (82.86%) JAIR papers not found by the system are actually downloadable via a link from the authors' homepages to the publisher's website. Yet we miss these papers simply because we limit our crawler not to go beyond the domain of authors' homepages. We believe that a more sophisticated domain restriction for the crawler can be easily employed in order to achieve an even higher recall level. 306 0 0.2 0.4 0.6 0.8 1 1.2 1998 1999 2000 2001 2002 2003 2004 Reca ll Level WebDB JAIR Figure 8. Overall Recall Level, 1998 - 2004 We calculate the recall level for the documents published in WebDB and JAIR yet missing from CiteSeer's collection (see Figure 9). In this case, S(i) is the set of missing documents downloaded by the crawler, and T is the set of the papers not indexed by CiteSeer and missing from the collection. On average, the recall level has achieved 0.78 for WebDB and 0.72 for JAIR. Especially WebDB's recall level is constantly increasing, reaching 1.0 for the last three years. This proves that it's practical to harvest the missing documents for a given publishing venue. 0 0.2 0.4 0.6 0.8 1 1.2 1998 1999 2000 2001 2002 2003 2004 Reca l l L e v e l WebDB JAIR Figure 9. Recall Level for the Missing Documents The trends shown in Figure 8 and 9 seem to indicate that a rising number of academic papers have been put online, especially in and after the year 2000. However, it's interesting to note that it seems conference/workshop authors favor putting their publications on their homepages, while journal authors don't. Due to the limited size of our sample, we feel this is an open question to be answered with more data across multiple venues. 4.2 Crawler Comparison: BF Crawler In order to further evaluate the performance of our system, we also compare our work to other crawling approaches. First we crawled three conference websites using our system and a breadth-first (BF) crawler. Figures 10, 11 and 12 show the results of crawls on different conference websites. The BF crawls are shown by the dashed line while the results of the focused crawler are shown by the solid line on the figures. The horizontal axis indicates the number of pages crawled and the vertical axis represents the number of research documents found by searching those pages. The number of documents found is a cumulative sum of all PDF, PS and GZ files found on those sites. Since they may contain duplicate files or the same content in different file types, the numbers shown do not indicate unique papers. The number of pages crawled does not include academic papers. The same crawl restrictions applied to both the crawlers. 0 200 400 600 800 1000 1200 1400 1600 1800 0 10 20 30 Pages Crawled Nu m b e r o f Do c u m e n t s FC BF Figure 10. ACL Conference Crawl Figure 10 shows the crawls done on parts of the Association for Computational Linguistics (ACL) conference website. The total number of pages crawled on this site were less than 30. Both crawls overlap which indicates that there is virtually no difference between the document detection rate of the BF crawler and our focused crawler. For such a small website, both crawlers detect the same number of documents after crawling the same number of pages on the website. 0 1000 2000 3000 4000 5000 6000 7000 0 100 200 300 400 Pages Crawled Nu m b e r o f Do c u m e n t s FC BF Figure 11. TREC Conference Crawl Figure 11 shows the crawls done on the Text Retrieval Conference (TREC) pages. Here the total of pages crawled is about 1000 (only first half of the crawl is shown in the graph). Both crawlers start detecting documents at the same rate. After detecting around 1393 documents (35 pages crawled) the document detection rate of the focused crawler becomes slightly better than the BF crawler. Although the difference is not very significant, the focused crawler does detect the research documents slightly earlier in the crawl as compared to the BF crawler. The BF crawler detects the same amount of documents (4800 documents) as the focused crawler but after crawling 20-30 307 pages more than the focused crawler. The total number of documents found by both the crawlers is around 6000. 0 500 1000 1500 2000 2500 3000 3500 4000 0 500 1000 1500 2000 2500 3000 3500 Pages Crawled Nu m b e r o f Do c u m e n t s FC BF Figure 12. VLDB Conference Crawl The crawls performed on the Very Large Database (VLDB) conference pages as shown in Figure 12 indicate that the focused crawler detects the documents much earlier in the crawl. Here the total number of pages crawled is about 3500. Approximately 28% (1000 out of 3500) of the documents are located by both the crawlers after crawling around 8.5% (300 out of 3500) of the domain. At this point the focused crawler continues to locate more documents while the BF crawler does not uncover any new documents until 28% (1000 out of 3500) of the total crawl. 85% (3000 out of 3500) of the documents are located by the focused crawler after completing just 33% (1189 out of 3500) of the total crawl, while the breadth first crawler locates the same amount of documents after completing 50% (1781 out of 3500) of the total crawl. Towards the end of the crawl the breadth-first crawler detects more papers as compared to the focused crawler. It takes the focused crawler around 1000 more pages of crawls until it makes up the difference. This seems to be due to the lack of keywords associated with the links that eventually led to the documents. The focused crawler evaluates other papers that have a higher priority values before eventually discovering the remaining documents. The behavior of the BF crawler is consistent for all the three crawls. Most of the documents located were in crawl depths 2, 3, 4 and 5. The BF crawler detects them after completing search of the previous crawl depths. As the focused crawler prioritizes the links for crawling, the higher depths with more priority are crawled before the lower depths with less priority. The above experiment indicates that the document harvest rate is almost the same for smaller websites. The difference becomes apparent when the size of the website being crawled is large. The focused crawler is able to detect the documents much earlier in the crawl as compared to the BF crawler. Since the crawls are not terminated early for the focused crawler, the number of documents found and the relevance of documents are same for both the crawlers. Therefore as the size of websites being crawled increases, the focused crawler detects more documents earlier during the crawls as compared to the BF crawler. We assess the crawler's capability of harvesting academic publications in a more general sense which is not only limited to a specific venue. We have manually examined the first 500 PDF/PostScript documents found by the two crawlers, classified the documents into academic publications which are desirable (papers published in conferences and journals; technical reports; degree thesis, etc.), and non-publication documents which are considered noise for a publication collection (course material; presentation slides; project schedule; etc.) Percentage of both categories is compared side-by-side and shown in Figure 13. Our crawler has outperformed the breadth-first counterpart by having much less of this noise. 423 480 77 22 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% BF FC Total: 50 0 Non-Publication Documents Academic Publications Figure 13. Composition of the First 500 PDF/PS Documents 4.3 Crawler Comparison: Nutch Crawler We compare the performance of our system with Nutch (http://www.nutch.org/docs/en/), an open source Web crawler and search engine built upon Lucene. In our experiment, we run the Nutch crawler on the official websites of WebDB and JAIR, and identify those papers published between 1998 and 2004 from the downloaded documents. We then compare the number of papers harvested by Nutch and FC crawler (see Figure 14 for details). Results show that guided by certain heuristics, crawling authors' homepages can actually achieve almost the same recall level as crawling publishers' websites. 0 50 100 150 200 WebDB JAIR Number of Documents Nutch FC Total Publications Figure 14. Comparison between Nutch and Focused Crawler Figure 15 indicates the progress of the crawls conducted by both the Focused Crawler and the Nutch Crawler on the ACL conference website. The documents found are of PDF and PS only. The focused crawler starts discovering documents earlier in the crawl and the process continues gradually. Nutch on the other hand discovers most of the documents after crawling around 84% (22 out of 26) of the website. 308 Figure 6. ACL Conference Crawl 0 200 400 600 800 1000 1200 1400 1600 1800 0 10 20 30 Pages Crawled Nu m b e r o f Do c u m e n t s Nutch FC Figure 15. Crawling ACL Conference Websites Documents found during the ACL conference crawl are classified into two categories: relevant (i.e. academic publications) and non-relevant (non-publication). Figure 16 shows the number of documents in each category. Note that determining documents' relevancy is an offline process. Here R indicates relevant and NR indicated non-relevant documents. R, 1588 R, 1404 NR, 0 NR, 0 0 200 400 600 800 1000 1200 1400 1600 1800 FC Nutch Crawler Nu m b e r o f Do c u m e n t s NR R Figure 16. Relevancy of the ACL Conference Crawl Figure 16 indicates that all the documents (PDF and PS) found by both the crawlers are academic publications (thus NR = 0). However, the 184 documents Nutch failed to detect are determined to be all relevant research publications. The same comparison is also conducted by crawling the official WebDB conference websites. Figure 17 shows that the Focused Crawler starts detecting desired documents at an earlier stage as compared to the Nutch crawler. Yet due to the small number of pages crawled, a rigorous comparison cannot be made in this case. Figure 18 shows that the focused crawler locates two more academic publications than the Nutch crawler, both of which are marked as relevant documents. CONCLUSION AND FUTURE WORK We have shown the feasibility of using authors' homepages as alternative online resources to harvest the academic papers missing from a collection of digital libraries, as well as the techniques to maximize the crawler's performance in doing so. We have designed and implemented a heuristic-based system which utilizes document metadata to accurately locate authors' homepages and performs a focused crawling to quickly navigate to the desired publications. Evaluation has been conducted using a large dataset collected from several publishing venues in the Computer Science domain, and detailed results are presented and discussed. Figure 10. WEDB Conference Crawl 0 20 40 60 80 100 120 0 10 20 30 40 Pages Crawled Nu m b e r o f Do c u m e n t s Nutch FC Figure 17. Crawling WebDB Conference Websites R, 104 R, 106 NR, 1 NR, 1 0 20 40 60 80 100 120 FC Nutch Crawler Nu m b e r o f Do c u m e n t s NR R Figure 18. Relevancy of the WebDB Conference Crawl For the academic venues investigated in this study, we are able to fill many of the missing documents in the CiteSeer digital library. The designed focused crawling technique efficiently locates desired publications on authors' homepages as well as conference websites. The Homepage Aggregator detects homepages well and the Focused Crawler outperforms the baseline crawler in a number of measures. Future work includes a more rigorous disambiguation scheme for the Homepage Aggregator and a more sophisticated weighting scheme for the Focused Crawler. In addition, we are now developing a training process for the crawler to learn the URL patterns of alternative resources other than author homepages, such as institutional archives. Also, the automation of the process cycle of crawling, log analysis, and heuristics generation can help search engine based digital libraries scale and significantly reduce costs. The actual URL of the web pages can also be used to assist in priority assignment instead of just using the anchor text of the link. A comparison of this approach to techniques other than a Breadth-first crawl is currently underway. Furthermore, we plan to evaluate the validity of this approach by expanding our experiment on to disciplines other than the Computer Science 309 domain. We believe our study and its consequents will shed lights on the question of finding missing papers for our digital library, or "what's there and what's not". ACKNOWLEDGEMENTS We gratefully acknowledge P. Mitra and the anonymous reviewers for their comments, I. Councill and P. Teregowda for their work on the CiteSeer metadata, and E. Maldonado and D. Hellar for the crawl list. This work is partially supported by Microsoft. REFERENCES [1] De Bra, P., Houben, G., Kornatzky, Y., and Post, R Information Retrieval in Distributed Hypertexts. In Proceedings of the 4th RIAO (Computer-Assisted Information Retrieval) Conference, pp. 481-491, 1994. [2] Cho J., Garcia-Molina, H., and Page, L. Efficient Crawling Through URL Ordering. In Proceedings of the 7th World Wide Web Conference, Brisbane, Australia, pp. 161-172. April 1998. [3] Chakrabarti, S., Van den Berg, M., and Dom, B. Focused Crawling: A New Approach to Topic-Specific Web Resource Discovery. In Proceedings of the 8th International WWW Conference, pp. 545-562, Toronto, Canada, May 1999. [4] Giles, C. L. and Councill, I. G. Who gets acknowledged: Measuring scientific contributions through automatic acknowledgement indexing. In Proceedings of the National Academy of Sciences 101(51) pp. 17599-17604, Dec. 21, 2004. [5] Najork, M. and Wiener, J. L. Breadth-First Search Crawling Yields High-Quality Pages. In Proceedings of the 10th International World Wide Web Conference, pp. 114-118, 2001. [6] Page, L., Brin, S., Motwani, R., and Winograd, T. The pagerank citation ranking: Bringing order to the web. Technical report, Stanford University Database Group, 1998. Available at http://dbpubs.stanford.edu: 8090/pub/1999-66 [7] Menczer, F., Pant, G., Ruiz, M., and Srinivasan, P. Evaluating Topic-Driven Web Crawlers.' In Proceedings of the 2001 Annual Conference of the Association of Computing Machinery, Special Interest Group in Information Retrieval, 241-249. New Orleans, September 2001. [8] Haveliwala, T. H. Topic-Sensitive PageRank. In Proceedings of the 11th International World Wide Web Conference, pp. 517-526. Honolulu, Hawaii, USA. May 2002. [9] Mukherjea, S. WTMS: a system for collecting and analyzing topic-specific Web information. Computer Networks 33(1-6): 457-471, 2000. [10] Diligenti, M., Coetzee, F.M., Lawrence, S., Giles, C. L., and Gori, M. Focused Crawling Using Context Graphs. In Proceedings of the 26th International Conference on Very Large Data Bases, pp. 527-534, 2000. [11] Aggarwal, C. C., Al-Garawi, F., and Yu, P. S. Intelligent Crawling on the World Wide Web with Arbitary Predicates. In Proceedings of the Tenth International Conference on World Wide Web, pp. 96-105, 2001. [12] Aggarwal, C. C. On Learning Strategies for Topic Specific Web Crawling. Next Generation Data Mining Applications, January 2004. [13] Pant, G., Tsjoutsiouliklis, K., Johnson, J., and Giles, C. L. Panorama: Extending Digital Libraries with Topical Crawlers. In Proceedings of the 2004 Joint ACM/IEEE Conference on Digital Libraries, pp. 142-150, 2004. [14] Menczer, F., Pant, G., and Srinivasan, P. Topical Web Crawlers: Evaluating Adaptive Algorithms. ACM TOIT 4(4): 378-419, 2004. [15] Pant, G., Srinivasan, P., and Menczer, F. Crawling the Web. In M. Levene and A. Poulovassilis, eds.: Web Dynamics, Springer, 2004. [16] Hoff, G. and Mundhenk, M. Finding scientific papers with homepagesearch and MOPS. In Proceedings of the Nineteenth Annual International Conference of Computer Documentation, Communicating in the New Millennium, pp. 201-207. October 21-24, 2001, Santa Fe, New Mexico, USA. [17] On, B. and Lee, D. PaSE: Locating Online Copy of Scientific Documents Effectively. In Proceedings of the 7th International Conference of Asian Digital Libraries (ICADL), pp. 408-418. Shanghai, China, December 2004. [18] Shakes, J., Langheinrich, M., and Etzioni, O. Dynamic Reference Sifting: a Case Study in the Homepage Domain. In Proceedings of the Sixth International World Wide Web Conference, pp. 189-200, 1997. [19] Xi, W. and Fox, E. A. Machine Learning Approach for Homepage Finding Task. In Proceedings of the Tenth Text REtrieval Conference (TREC 2001), pp. 686-698, 2001. [20] Anh, V. N. and Moffat, A. Homepage Finding and Topic Distillation using a Common Retrieval Strategy. In Proceedings of the Eleventh Text REtrieval Conference (TREC 2002), 2002. [21] Ogilvie, P. and Callan, J. Combining Structural Information and the Use of Priors in Mixed Named-Page and Homepage Finding. In Proceedings of the Twelfth Text REtrieval Conference (TREC 2003), pp. 177-184, 2003. [22] Sundaresan, N., Yi, J., and Huang, A. W. Using Metadata to Enhance a Web Information Gathering System. In Proceedings of the Third International Workshop on the Web and Databases (WebDB 2000), pp. 11-16, 2000. [23] Flesca, S., Furfaro, F., and Greco, S. Weighted Path Queries on Web Data. In Proceedings of the Fourth International Workshop on the Web and Databases (WebDB 2001), pp. 7-12, 2001. [24] Ruiz, A., Lpez-de-Teruel, P. E., and Garrido, M. C. Probabilistic Inference from Arbitrary Uncertainty using Mixtures of Factorized Generalized Gaussians. Journal of Artificial Intelligence Research (JAIR), Volume 9, pp. 167-217, 1998. [25] Russell, G., Neumller, M., and Connor, R. C. H. TypEx: A Type Based Approach to XML Stream Querying. In Proceedings of the Sixth International Workshop on the Web and Databases (WebDB 2003), pp. 55-60, 2003. 310

Digital libraries;CiteSeer;focused crawler;DBLP;harvesting;ACM

22

A Two-Phase Sampling Technique for Information Extraction from Hidden Web Databases

Hidden Web databases maintain a collection of specialised documents, which are dynamically generated in response to users' queries. However, the documents are generated by Web page templates, which contain information that is irrelevant to queries. This paper presents a Two-Phase Sampling (2PS) technique that detects templates and extracts query-related information from the sampled documents of a database. In the first phase, 2PS queries databases with terms contained in their search interface pages and the subsequently sampled documents. This process retrieves a required number of documents. In the second phase, 2PS detects Web page templates in the sampled documents in order to extract information relevant to queries. We test 2PS on a number of real-world Hidden Web databases. Experimental results demonstrate that 2PS effectively eliminates irrelevant information contained in Web page templates and generates terms and frequencies with improved accuracy.

INTRODUCTION An increasing number of databases on the Web maintain a collection of documents such as archives, user manuals or news articles. These databases dynamically generate documents in response to users' queries and are referred to as Hidden Web databases [5]. As the number of databases proliferates, it has become prohibitive for specialised search services (such as search.com) to evaluate databases individually in order to answer users' queries. Current techniques such as database selection and categorisation have been employed to enhance the effectiveness of information retrieval from databases [2, 5, 10, 11, 15]. In the domain of the Hidden Web, knowledge about the contents of databases is often unavailable. Existing approaches such as in [2, 10, 15] acquire knowledge through sampling documents from databases. For instance, query-based sampling [2] queries databases with terms that are randomly selected from those contained in the sampled documents. The techniques in [10, 15] sample databases with terms obtained from Web logs to retrieve additional topic terms. A major issue associated with existing techniques is that they also extract information irrelevant to queries. That is, information extracted is often found in Web page templates, which contain navigation panels, search interfaces and advertisements. Consequently, the accuracy of terms and frequencies generated from sampled documents has been reduced. In addition, approximate string matching techniques are adopted by [13] to extract information from Web pages, but this approach is limited to textual contents only. Alternatively, the approaches proposed in [3, 4] analyse Web pages in tree-like structures. However, such an approach requires Web pages with well-conformed HTML tag trees. Furthermore, [3] discovers dynamically generated objects from Web pages, which are clustered into groups of similar structured pages based on a set of pre-defined templates, such as exception page templates and result page templates. In this paper, we propose a sampling and extraction technique, which is referred to as Two-Phase Sampling (2PS). 2PS aims to extract information relevant to queries in order to acquire information contents of underlying databases. Our technique is applied in two phases. First, it randomly selects a term from those found in the search interface pages of a database to initiate the process of sampling documents. Subsequently, 2PS queries the database with terms randomly selected from those contained in the sampled documents. Second, 2PS detects Web page templates and extracts query-related information from which terms and frequencies are generated to summarise the database contents. Our approach utilises information contained in search interface pages of a database to initiate the sampling process. This differs from current sampling techniques such as query-based sampling, which performs an initial query with a frequently used term. Furthermore, 2PS extracts terms that are relevant to queries thus generating statistics (i.e., terms and frequencies) that represent database contents with improved accuracy. By contrast, the approaches in [2, 10, 15] extract all terms from sampled documents, including those contained in Web page templates. Consequently, information that is irrelevant to queries is also extracted. 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. WIDM'04, November 1213, 2004, Washington, DC, USA. Copyright 2004 ACM 1-58113-978-0/04/0011...$5.00. 1 Figure 1. The Two-Phase Sampling (2PS) technique. 2PS is implemented as a prototype system and tested on a number of real-world Hidden Web databases, which contain computer manuals, healthcare archives and news articles. Experimental results show that our technique effectively detects Web page templates and generates terms and frequencies (from sampled documents) that are relevant to the queries. The remainder of the paper is organised as follows. Section 2 introduces current approaches to the discovery of information contents of Hidden Web databases. Related work on the information extraction from Web pages or dynamically generated documents is also discussed. Section 3 describes the proposed 2PS technique. Section 4 presents experimental results. Section 5 concludes the paper. RELATED WORK A major area of current research into the information retrieval of Hidden Web databases focuses on the automatic discovery of information contents of databases, in order to facilitate their selection or categorisation. For instance, the technique proposed in [6] analyses the hyperlink structures of databases in order to facilitate the search for databases that are similar in content. The approach adopted by [10, 15] examines the textual contents of search interface pages maintained by data sources to gather information about database contents. A different approach is to retrieve actual documents to acquire such information. However, in the domain of Hidden Web databases, it is difficult to obtain all documents from a database. Therefore, a number of research studies [2, 10, 15] obtain information by retrieving a set of documents through sampling. For instance, query-based sampling [2] queries databases with terms that are randomly selected from those contained in the sampled documents. The techniques in [10, 15] sample databases with terms extracted from Web logs to obtain additional topic terms. These techniques generate terms and frequencies from sampled documents, which are referred to as Language Models [2], Textual Models [10, 15] or Centroids [11]. A key issue associated with the aforementioned sampling techniques is that they extract information that is often irrelevant to queries, since information contained in Web page templates such as navigation panels, search interfaces and advertisements is also extracted. For example, a language model generated from the sampled documents of the Combined Health Information Database (CHID) contains terms (such as `author' and `format') with high frequencies. These terms are not relevant to queries but are used for descriptive purposes. Consequently, the accuracy of terms and frequencies generated from sampled documents has been reduced. The use of additional stop-word lists has been considered in [2] to eliminate irrelevant terms - but it is maintained that such a technique can be difficult to apply in practice. Existing techniques in information extraction from Web pages are of varying degrees of complexity. For instance, approximate string matching techniques are adopted by [13] to extract texts that are different. This approach is limited to finding textual similarities and differences. The approaches proposed in [3, 4] analyse textual contents and tag structures in order to extract data from Web pages. However, such an approach requires Web pages that are produced with well-conformed HTML tag-trees. Computation is also needed to convert and analyse Web pages in a tree-like structure. Moreover, [3] identifies Web page templates based on a number of pre-defined templates, such as exception page templates and result page templates. Our technique examines Web documents based on textual contents and the neighbouring tag structures rather than analysing their contents in a tree-like structure. We also detect information contained in different templates through which documents are generated. Therefore, it is not restricted to a pre-defined set of page templates. Furthermore, we focus on databases that contain documents such as archives and new articles. A distinct characteristic of documents found in such a domain is that the content of a document is often accompanied by other information for supplementary or navigation purposes. The proposed 2PS technique detects and eliminates information contained in templates in order to extract the content of a document. This differs from the approaches in [1, 4], which attempt to extract a set of data from Web pages presented in a particular pattern. For example, the Web pages of a bookstore Web site contain information about authors followed by their associated list of publications. However, in the domain of document databases, information contained in dynamically generated Web pages is often presented in a structured fashion but irrelevant to queries. Other research studies [9, 8, 12] are specifically associated with the extraction of data from query forms in order to further the retrieval of information from the underlying databases. TWO-PHASE SAMPLING This section presents the proposed technique for extracting information from Hidden Web document databases in two phases, which we refer to as Two-Phase Sampling (2PS). Figure 1 depicts the process of sampling a database and extracting query-related 2 information from the sampled documents. In phase one, 2PS obtains randomly sampled documents. In phase two, it detects Web page templates. This extracts information relevant to the queries and then generates terms and frequencies to summarise the database content. The two phases are detailed in section 3.1 and 3.2. 3.1 Phase One: Document Sampling In the first phase we initiate the process of sampling documents from a database with a randomly selected term from those contained in the search interface pages of the database. This retrieves top N documents where N represents the number of documents that are the most relevant to the query. A subsequent query term is then randomly selected from terms extracted from the sampled documents. This process is repeated until a required number of documents are sampled. The sampled documents are stored locally for further analysis. Figure 2 illustrates the algorithm that obtains a number of randomly sampled documents. t q denotes a term extracted from the search interface pages of a database, D. qt p represents a query term selected from a collection of terms, Q, qt p Q, 1 p m; where m is the distinct number of terms extracted from the search interface pages and the documents that have been sampled. R represents the set of documents randomly sampled from D. t r is a term extracted from d i . d i represents a sampled document from D, d i D, 1 i n, where n is the number of document to sample. Figure 2. The algorithm for sampling documents from a database. 2PS differs from query-based sampling in terms of selecting an initial query. The latter selects an initial term from a list of frequently used terms. 2PS initiates the sampling process with a term randomly selected from those contained in the search interface pages of the database. This utilises a source of information that is closely related to its content. Moreover, 2PS analyses the sampled documents in the second phase in order to extract query-related information. By contrast, query-based sampling does not analyse their contents to determine whether terms are relevant to queries. 3.2 Phase Two: Document Content Extraction and Summarisation The documents sampled from the first phase are further analysed in order to extract information relevant to the queries. This is then followed by the generation of terms and frequencies to represent the content of the underlying database. This phase is carried out through the following processes. 3.2.1 Generate Document Content Representations The content of each sampled document is converted into a list of text and tag segments. Tag segments include start tags, end tags and single tags specified in HyperText Markup Language (HTML). Text segments are text that resides between two tag segments. The document content is then represented by text segments and their neighbouring tag segments, which we refer to as Text with Neighbouring Adjacent Tag Segments (TNATS). The neighbouring adjacent tag segments of a text segment are defined as the list of tag segments that are located immediately before and after the text segment until another text segment is reached. The neighbouring tag segments of a text segment describe how the text segment is structured and its relation to the nearest text segments. Assume that a document contains n segments, a text segment, txs, is defined as: txs = (tx i , tg-lst j , tg-lst k ), where tx i is the textual content of the i th text segment, 1 i n; tg-lst j represents p tag segments located before tx i and tg-lst k represents q tag segments located after tx i until another text segment is reached. tg-lst j = (tg 1 , ..., tg p ), 1 j p and tg-lst k = (tg 1 , ..., tg q ), 1 k q. Algorithm SampleDocument Extract t q from search interface pages of D, Q = t q For i = 1 to n Randomly select qt p from Q If (qt p has not been selected previously) Execute the query with qt p on D j = 0 While j &lt;= N If (d i R) Retrieve d i from D Extract t r from d i , R = d i Q = t r Increase j by 1 End if End while End if End for Figure 3. A template-generated document from CHID. Figure 3 shows a template-generated document retrieved from the CHID database. The source code for this document is given in Figure 4. For example, text segment, `1. Equipos Mas Seguros: Si Te Inyectas Drogas.', can be identified by the text (i.e., `1. Equipos Mas Seguros: Si Te Inyectas Drogas.') and its neighbouring tag segments. These include the list of tags located before the text (i.e., &lt;/TITLE&gt;, &lt;/HEAD&gt;, &lt;BODY&gt;, &lt;HR&gt;, &lt;H3&gt;, &lt;B&gt; and &lt;I&gt;) and the neighbouring tags located after the text (i.e., &lt;/I&gt;, &lt;/B&gt;, &lt;/H3&gt;, &lt;I&gt; and &lt;B&gt;). Thus, this segment is then represented as (`1. Equipos Mas Seguros: Si Te Inyectas Drogas.', (&lt;/TITLE&gt;, &lt;/HEAD&gt;, &lt;BODY&gt;, &lt;HR&gt;, &lt;H3&gt;, &lt;B&gt; ,&lt;I&gt;), (&lt;/I&gt;, &lt;/B&gt;, &lt;/H3&gt;, &lt;I&gt;, &lt;B&gt;)). Figure 5 shows the content 3 representation of the CHID document (given in Figure 3) generated based on TNATS. Given a sampled document, d, with n text segments, the content of d is then represented as: Content(d) = {txs 1 , ..., txs n }, where txs i represents a text segment, 1 i n. Figure 4. The source code for the CHID document. Figure 5. The content representation of the CHID document using TNATS. 3.2.2 Detect Templates In the domain of Hidden Web databases, documents are often presented to users through one or more templates. Templates are typically employed in order to describe document contents or to assist users in navigation. For example, information contained in the document (as shown in Figure 3) can be classified into the two following categories: (i) Template-Generated Information. This includes information such as navigation panels, search interfaces and advertisements. In addition, information may be given to describe the content of a document. Such information is irrelevant to a user's query. For example, navigation links (such as `Next Doc' and `Last Doc') and headings (such `Subfile' and `Format') are found in the document. (ii) Query-Related Information. This information is retrieved in response to a user's query, i.e., `1. Equipos Mas Seguros: Si Te Inyectas Drogas. ...'. The 2PS technique detects Web page templates employed by databases to generate documents in order to extract information that is relevant to queries. Figure 6 describes the algorithm that detects information contained in Web page templates from n sampled documents. d i represents a sampled document from the database D, d i , D, 1 i n. Content(d i ) denotes the content representation of d i . ... &lt;HTML&gt;&lt;HEAD&gt;&lt;TITLE&gt;CHID Document &lt;/TITLE&gt;&lt;/HEAD&gt; &lt;BODY&gt; &lt;HR&gt;&lt;H3&gt;&lt;B&gt;&lt;I&gt; 1. Equipos Mas Seguros: Si Te Inyectas Drogas. &lt;/I&gt;&lt;/B&gt;&lt;/H3&gt; &lt;I&gt;&lt;B&gt;Subfile: &lt;/B&gt;&lt;/I&gt; AIDS Education&lt;BR&gt; &lt;I&gt;&lt;B&gt;Format (FM): &lt;/B&gt;&lt;/I&gt; 08 - Brochure. &lt;BR&gt; ... Algorithm DetectTemplate For i = 1 to n If T = If S = S = d i Else if S While l &lt;= s AND T = Compare (Content(d i ),Content(d l )) If Content(d i ) Content(d l ) wpt k = Content(d i ) Content(d l ), Store wpt k , T = wpt k Delete (Content(d i ) Content(d l )) from Content(d i ), Content(d l ) G k = d i , G k = d l Delete d l from S End if End while If T = S = d i End if End if Else if T While k &lt;= r AND d i G k Compare (Content(wpt k ), Content(d i )) If Content(wpt k ) Content(d i ) Delete (Content(wpt k ) Content(d i )) from Content(d i ) G k = d i End if End while If S AND d i G k While l &lt;= s AND d i G k Compare (Content(d i ),Content(d l )) If Content(d i ) Content(d l ) wpt k = Content(d i ) Content(d l ) Store wpt k , T = wpt k Delete (Content(d i ) Content(d l )) from Content(d i ), Content(d l ) G k = d i , G k = d l Delete d l from S End if End while End if If d i G k S = d i End if End if End for ... `CHID Document', (&lt;HTML&gt;, &lt;HEAD&gt;, &lt;TITLE&gt;), (&lt;/TITLE&gt;, &lt;/HEAD&gt;, &lt;BODY&gt;, &lt;HR&gt;, &lt;H3&gt;, &lt;B&gt;, &lt;I&gt;); `1. Equipos Mas Seguros: Si Te Inyectas Drogas.', (&lt;/TITLE&gt;, &lt;/HEAD&gt;, &lt;BODY&gt;, &lt;HR&gt;, &lt;H3&gt;, &lt;B&gt;, &lt;I&gt;), (&lt;/I&gt;, &lt;/B&gt;, &lt;/H3&gt;, &lt;I&gt;, &lt;B&gt;); `Subfile:', (&lt;/I&gt;, &lt;/B&gt;, &lt;/H3&gt;, &lt;I&gt;, &lt;B&gt;), (&lt;/B&gt;, &lt;/I&gt; ); `AIDS Education', (&lt;/B&gt;, &lt;/I&gt;) , ( &lt;BR&gt;, &lt;I&gt;, &lt;B&gt;); `Format (FM):', (&lt;BR&gt;, &lt;I&gt;, &lt;B&gt;), (&lt;/B&gt;, &lt;/I&gt;); ... Figure 6. The algorithm for detecting and eliminating the information contained in Web page templates. 4 Similar to the representation for the contents of sampled documents, the content of a Web page template, wpt, is represented as Content(wpt) = {txs 1 , ..., txs q }, where q is the number of text segments, txs j , 1 j q. T represents a set of templates detected. T = {wpt 1 , ..., wpt r }, where r is the distinct number of templates, wpt k , 1 k r. G k represents a group of documents generated from wpt k . Furthermore, S represents the sampled documents from which no templates have yet been detected. Thus, S = {d 1 , ..., d s }, where s is the number of temporarily stored document, d l , 1 l s. The process of detecting templates is executed until all sampled documents are analysed. This results in the identification of one or more templates. For each template, two or more documents are assigned to a group associated with the template from which the documents are generated. Each document contains text segments that are not found in their respective template. These text segments are partially related to their queries. In addition to a set of templates, the content representations of zero or more documents in which no matched patterns are found are stored. 3.2.3 Extract Query-Related Information This process analyses a group of documents associated with each template from which documents are generated. It further identifies any repeated patterns from the remaining text segments of the documents in order to extract query-related information. We compute cosine similarity [14] given in (1) to determine the similarities between the text segments of different documents that are associated the template where the documents are generated. The textual content of each text segment is represented as a vector of terms with weights. The weight of a term is obtained by its occurrence in the segment. where txs i and txs j represent two text segments in a document; tw ik is the weight of term k in txs i , and tw jk is the weight of term k in txs j . This is only applied to text segments with identical adjacent tag segments. Two segments are considered to be similar if their similarity exceeds a threshold value. The threshold value is determined experimentally. The algorithm that extracts information relevant to queries is illustrated in Figure 7. d a and d b represent the sampled documents from the database, D, d a , d b G k , where G k denotes a group of documents associated with the template, wpt k , from which the documents are generated. tx m represents the textual content of a text segment, txs m , contained in d i , d i G k . tx n represents the textual content of a text segment, txs n , contained in d l , d l S. S represents the sampled documents from which no templates are detected. The results of the above algorithm extract text segments with different tag structures. It also extracts text segments that have identical adjacent tag structures but are significantly different in their textual contents. Figure 8 shows the information extracted from the document content (given in Figure 4) as a result of eliminating information contained in the Web page template. 3.2.4 Generate Content Summary Frequencies are computed for the terms extracted from randomly sampled documents. These summarise the information content of a database, which we refer to as Content Summary. Algorithm ExtractQueryInfo For each (d a G k ) For each (d b G k ), d a d b Compare (Content(d a ),Content(d b )) If Content(d a ) Content(d b ) Delete (Content(d a ) Content(d b )) from Content(d a ), Content(d b ) End if End for End for For each (d i G k ) Extract tx m of txs m from Content(d i ) End for For each (d l S) Extract tx n of txs n from Content(d l ) End for Figure 7. The algorithm for extracting query-related information from template-generated documents. 1. Equipos Mas Seguros: Si Te Inyectas Drogas. AIDS Education ... = = = = t k jk t k t k ik jk ik j i tw tw tw tw txs txs COSINE 1 2 1 1 2 ) ( ) ( ) ( ) , ( . (1) Figure 8. The query-related information extracted from the CHID document. Previous experiments in [2] demonstrate that a number of randomly sampled documents (i.e., 300 documents) sufficiently represent the information content of a database. In the domain of Hidden Web databases, the inverse document frequency (idf), used in traditional information retrieval, is not applicable, since the total number of documents in a database is often unknown. Therefore, document frequency (df), collection term frequency (ctf) and average term frequency (avg_tf) initially used in [2] are applied in this paper. We consider the following frequencies to compute the content summary of a Hidden Web database. Document frequency (df): the number of documents in the collection of documents sampled that contain term t, where d is the document and f is the frequency Collection term frequency (ctf): the occurrence of a term in the collection of documents sampled, where c is the collection, t is the term and f is the frequency Average term frequency (avg_tf): the average frequency of a term obtained from dividing collection term frequency by document frequency (i.e., avg_tf = ctf / df) 5 Table 1. 3 Hidden Web databases used in the experiments Database URL Subject Content Template Help Site <A href="http://www.help-site.com/">www.help-site.com Computer manuals Homogeneous Multiple templates <A href="http://www.chid.nih.gov/">CHID www.chid.nih.gov Healthcare articles Homogeneous Single template Wired News <A href="http://www.wired.com/">www.wired.com General news articles Heterogeneous Single template The content summary of a document database is defined as follows. Assume that a Hidden Web database, D, is sampled with N documents. Each sampled document, d, is represented as a vector of terms and their associated weights [14]. Thus d = (w 1 , ..., w m ), where w i is the weight of term t i , and m is the number of distinct terms in d D, 1 i m. Each w i is computed using term frequency metric, avg_tf (i. e., w i = ctf i /df i ). The content summary is then denoted as CS(D), which is generated from the vectors of sampled documents. Assume that n is the number of distinct terms in all sampled documents. CS(D) is, therefore, expressed as a vector of terms: CS(D)= {w 1 , ..., w n }, where w i is computed by adding the weights of t i in the documents sampled from D and dividing the sum by the number of sampled documents that contain t i , 1 i n. EXPERIMENTAL RESULTS This section reports on a number of experiments conducted to assess the effectiveness of the 2PS technique in terms of: (i) detecting Web page templates, and (ii) extracting relevant information from the documents of a Hidden Web databases through sampling. The experimental results are compared with those from query-based sampling (abbreviated as QS). We compare 2PS with QS as it is a well-established technique and has also been widely adopted by other relevant studies [5, 10, 11, 15]. Experiments are carried out on three real-world Hidden Web document databases including Help Site, CHID and Wired News, which provide information about user manuals, healthcare archives and news articles, respectively. Table 1 summarises these databases in terms of their subjects, contents and templates employed. For instance, Help Site and CHID contain documents relating to subjects on computing and healthcare, respectively. Their information contents are homogeneous in nature. By contrast, Wired News contains articles that relate to different subjects of interest. Where the number of templates is concerned, CHID and Wired News generate documents from one Web page template. Help Site maintains a collection of documents produced by other information sources. Subsequently, different Web page templates are found in Help Site sampled documents. The experiment conducted using QS initiates the first query to a database with a frequently used term to obtain a set of sampled documents. Subsequent query terms are randomly selected from those contained in the sampled documents. It extracts terms (including terms contained in Web page templates) and updates the frequencies after each document is sampled. By contrast, 2PS initiates the sampling process with a term contained in the search interface pages of a database. In addition, 2PS analyses the sampled documents in the second phase in order to extract query-related information, from which terms and frequencies are generated. Experimental results in [2] conclude that QS obtains approximately 80% of terms from a database, when 300 documents are sampled and top 4 documents are retrieved for each query. These two parameters are used to obtain results for our experiments in which terms and frequencies are generated for QS and 2PS after 300 documents have been sampled. The results generated from QS provide the baseline for the experiments. Three sets of samples are obtained for each database and 300 documents are retrieved for each sample. First, we manually examine each set of sampled documents to obtain the number of Web page templates used to generate the documents. This is then compared with the number of templates detected by 2PS. The detection of Web page templates from the sampled documents is important as this determines whether irrelevant information is effectively eliminated. Next, we compare the number of relevant terms (from top 50 terms) retrieved using 2PS with the number obtained by QS. Terms are ranked according to their ctf frequencies to determine their relevancy to the queries. This frequency represents the occurrences of a term contained in the sampled documents. Ctf frequencies are used to demonstrate the effectiveness of extracting query-related information from sampled documents since the terms extracted from Web page templates are often ranked with high ctf frequencies. Table 2. The number of templates employed by databases and the number detected by 2PS Number of templates Databases Employed Detected Sample 1 17 15 Sample 2 17 16 Help Site Sample 3 19 17 Sample 1 1 1 Sample 2 1 1 CHID Sample 3 1 1 Sample 1 1 1 Sample 2 1 1 Wired News Sample 3 1 1 Experimental results for QS and 2PS are summarised as follows. Firstly, Table 2 gives the number of Web page templates employed by the databases and the number detected by 2PS. It shows that 2PS effectively identifies the number of templates found in the sampled documents. However, a small number of templates are not detected from Help Site. For instance, 2PS does not detect two of the templates from the first set of sampled documents, since the two templates are very similar in terms of content and structure. 6 Table 3 summarises the number of relevant terms (from top 50 terms ranked according to their ctf frequencies) obtained for the three databases. These terms are retrieved using 2PS and QS. We determine the relevancy of a term by examining whether the term is found in Web page templates. Table 3 gives the number of retrieved terms that do not appear in Web page templates. The results show that 2PS obtains more relevant terms. For instance, in the first set of documents sampled from CHID using 2PS, the number of relevant terms retrieved is 47. By comparison, the number of terms obtained for QS is 20. The results generated from CHID and Wired News demonstrate that 2PS retrieves more relevant terms, as a large number of terms contained in the templates have been successfully eliminated from the top 50 terms. However, the elimination of template terms is less noticeable for Help Site. Our observation is that template terms attain high frequencies since the CHID and Wired News databases generate documents using a single Web page template. By comparison, a larger number of Web page templates are found in the documents sampled from Help Site. As a result, terms contained in the templates do not attain high frequencies as those found in the templates employed by CHID and Wired News. Table 4 and 5 show the results of the top 50 terms ranked according to their ctf frequencies retrieved from the first set of sampled documents of the CHID database. Table 4 shows the top 50 terms retrieved for QS whereby terms contained in Web page templates are not excluded. As a result, a number of terms (such as `author', `language' and `format') have attained much higher frequencies. By contrast, Table 5 lists the top 50 terms retrieved using 2PS. Our technique eliminates terms (such as `author' and `format') and obtains terms (such as `treatment', `disease' and `immunodeficiency') in the higher rank. Table 3. The number of relevant terms retrieved (from top 50 terms) according to ctf frequencies Number of relevant terms Databases QS 2PS Sample 1 46 48 Sample 2 47 48 Help Site Sample 3 46 48 Sample 1 20 47 Sample 2 19 47 CHID Sample 3 20 47 Sample 1 14 42 Sample 2 10 43 Wired News Sample 3 11 39 CONCLUSION This paper presents a sampling and extraction technique, 2PS, which utilises information that is contained in the search interface pages and documents of a database in the sampling process. This technique extracts information relevant to queries from the sampled documents in order to generate terms and frequencies with improved accuracy. Experimental results demonstrate that our technique effectively eliminates information contained in Web page templates, thus attaining terms and frequencies that are of a higher degree of relevancy. This can also enhance the effectiveness of categorisation in which such statistics are used to represent the information contents of underlying databases. We obtain promising results by applying 2PS in the experiments on three databases that differ in nature. However, experiments on a larger number of Hidden Web databases are required in order to further assess the effectiveness of the proposed technique. Table 4. Top 50 terms and frequencies ranked according to ctf generated from CHID when QS is applied Rank Term Rank Term Rank Term 1 hiv 18 document 35 lg 2 aids 19 disease 36 ve 3 information 20 published 37 yr 4 health 21 physical 38 ac 5 prevention 22 subfile 39 corporate 6 education 23 audience 40 mj 7 tb 24 update 41 description 8 accession 25 verification 42 www 9 number 26 major 43 cn 10 author 27 pamphlet 44 pd 11 persons 28 chid 45 english 12 language 29 human 46 national 13 sheet 30 date 47 public 14 format 31 abstract 48 immunodeficiency 15 treatment 32 code 49 virus 16 descriptors 33 ab 50 org 17 availability 34 fm 7 Table 5. Top 50 terms and frequencies ranked according to ctf generated from CHID when 2PS is applied Rank Term Rank Term Rank Term 1 hiv 18 education 35 testing 2 aids 19 virus 36 programs 3 information 20 org 37 services 4 health 21 notes 38 clinical 5 prevention 22 nt 39 people 6 tb 23 cdc 40 hepatitis 7 persons 24 service 41 community 8 sheet 25 box 42 world 9 treatment 26 research 43 listed 10 disease 27 department 44 professionals 11 human 28 positive 45 training 12 pamphlet 29 tuberculosis 46 diseases 13 www 30 control 47 accession 14 http 31 drug 48 network 15 national 32 discusses 49 general 16 public 33 ill 50 std 17 immunodeficiency 34 organizations REFERENCES [1] Arasu, A. and Garcia-Molina, H. Extracting Structured Data from Web Pages. In Proceedings of the 2003 ACM SIGMOD International Conference on Management, 2003, 337-348. [2] Callan, J. and Connell, M. Query-Based Sampling of Text Databases. ACM Transactions on Information Systems (TOIS), Vol. 19, No. 2, 2001, 97-130. [3] Caverlee, J., Buttler, D. and Liu, L. Discovering Objects in Dynamically-Generated Web Pages. Technical report, Georgia Institute of Technology, 2003. [4] Crescenzi, V., Mecca, G. and Merialdo, P. ROADRUNNER: Towards Automatic Data Extraction from Large Web Sites, In Proceedings of the 27th International Conference on Very Large Data Bases (VLDB), 2001, 109-118. [5] Gravano, L., Ipeirotis, P. G. and Sahami, M. QProber: A System for Automatic Classification of Hidden-Web Databases. ACM Transactions on Information Systems (TOIS), Vol. 21, No. 1, 2003. [6] He, M. and Drobnik, O. Clustering Specialised Web-databases by Exploiting Hyperlinks. In Proceedings of the Second Asian Digital Library Conference, 1999. [7] Hedley, Y.L., Younas, M., James, A. and Sanderson M. Query-Related Data Extraction of Hidden Web Documents. In Proceedings of the 27th Annual International ACM SIGIR Conference, 2004, 558-559. [8] Lage, J. P., da Silva, A. S., Golgher, P. B. and Laender, A. H. F. Automatic Generation of Agents for Collecting Hidden Web Pages for Data Extraction. Data & Knowledge Engineering, Vol. 49, No. 2, 2004, 177-196. [9] Liddle, S.W., Yau, S.H. and Embley, D. W. On the Automatic Extraction of Data from the Hidden Web. In Proceedings of the 20th International Conference on Conceptual Modeling, (ER) Workshops, 2001, 212-226. [10] Lin, K.I. and Chen, H. Automatic Information Discovery from the Invisible Web. International Conference on Information Technology: Coding and Computing (ITCC), 2002, 332-337. [11] Meng, W., Wang, W., Sun, H. and Yu, C. Concept Hierarchy Based Text Database Categorization. International Journal on Knowledge and Information Systems, Vol. 4, No. 2, 2002, 132-150. [12] Raghavan, S. and Garcia-Molina, H. Crawling the Hidden Web. In Proceedings of the 27th International Conference on Very Large Databases (VLDB), 2001, 129-138. [13] Rahardjo, B. and Yap, R. Automatic Information Extraction from Web Pages, In Proceedings of the 24th Annual International ACM SIGIR Conference, 2001, 430-431. [14] Salton, G. and McGill, M. Introduction to Modern Information Retrieval. New York, McCraw-Hill, 1983. [15] Sugiura, A. and Etzioni, O. Query Routing for Web Search Engines: Architecture and Experiments. In Proceedings of the 9th International World Wide Web Conference: The Web: The Next Generation, 2000, 417-430. 8

Hidden Web Databases;search interface pages;Information Extraction;hypertext markup langauges;hidden web databases;2-phase sampling technique;neighbouring adjacent tag segments;string matching techniques;information extraction;web page templates;Document Sampling;query-based sampling;irrelavant information extraction

23

A Unified Approach for Improving QoS and Provider Revenue in 3G Mobile Networks

In this paper, we introduce a unified approach for the adaptive control of 3G mobile networks in order to improve both quality of service (QoS) for mobile subscribers and to increase revenue for service providers. The introduced approach constantly monitors QoS measures as packet loss probability and the current number of active mobile users during operation of the network. Based on the values of the QoS measures just observed, the system parameters of the admission controller and packet scheduler are controlled by the adaptive performance management entity. Considering UMTS, we present performance curves showing that handover failure probability is improved by more than one order of magnitude. Moreover, the packet loss probability can be effectively regulated to a predefined level and provider revenue is significantly increased for all pricing policies.

Introduction The third generation (3G) of mobile networks is expected to complete the worldwide globalization process of mobile communication. Since different parts of the worlds emphasize different issues, the global term 3G has regional synonyms : In the US and Japan, 3G often carries the name International Mobile Telephony 2000 (IMT2000). In Europe, 3G has become Universal Mobile Telecommunications System (UMTS) following the ETSI perspective. The European industrial players have created the 3rd Generation Partnership Project (3GPP) [1] for the standardization of UMTS. 3G mobile networks provide the foundation for new services with high-rate data not provided by current second generation systems [26]. While the standardization of 3G is still ongoing the discussion of technical issues beyond 3G has already started [23,28]. Recently, Aretz et al. reported a vision for the future of wireless communication systems beyond 3G that consists of a combination of several optimized access systems on a common IP-based medium access and core network platform [5]. Charging and pricing are essential issues for network operations of 3G mobile networks. A primary target of differen-tiated pricing of Internet services is the prevention of system overload and an optimal resource usage according to different daytimes and different traffic intensities [12]. Among the proposed pricing proposals, flat-rate pricing [11] is the most common mode of payment today for bandwidth services. Flat-rate pricing is popular because of its minimal accounting overhead. A flat-rate encourages usage but does not offer any motivation for users to adjust their demand. Dynamic pricing models that take the state of the network into account in the price determination have been proposed as being more Corresponding author. responsive. Usage-based pricing regulates usage by imposing a fee based on the amount of data actually sent, whereas congestion-sensitive pricing uses a fee based on the current state of congestion in the network. Thus, a unified approach considering both dynamic pricing and controlling quality of service (i.e., performance management) provides an effective tool for the operation of 3G mobile networks. However, in previous work [8,13,19,21,25] the improvement of Quality of Service (QoS) in 3G mobile networks and the optimization of mobile service provider revenue has been considered sepa-rately . The Quality of Service (QoS) concept and architecture for UMTS networks specified in [2] provides means for sharing radio resources among different groups of users according to their individual QoS demands. Furthermore, the concept of UMTS management and control functions such as admission controller and resource manager is roughly outlined . Das et al. proposed a framework for QoS provisioning for multimedia services in 3G wireless access networks [8]. They developed an integrated framework by combining various approaches for call admission control, channel reservation , bandwidth degradation, and bandwidth compaction. In [19], we introduced a framework for the adaptive control of UMTS networks, which utilizes online monitoring of QoS measures (e.g., handover failure and call blocking probabilities ) in order to adjust system parameters of the admission controller and the packet scheduler. The presented approach is based on a lookup table called the Performance Management Information Base (P-MIB). Entries of the P-MIB have to be determined using extensive off-line simulation experiments to determine optimal parameter configuration for the considered scenarios. Given the entries of the P-MIB, we showed how to improve QoS for mobile users by periodi-cally adjusting system parameters. The practical applicability of this approach is limited if the P-MIB comprises many en-210 C. LINDEMANN ET AL. tries (i.e., many scenarios have to be considered) because of the high computational effort for determining these entries by simulation. This paper introduces a unified approach for the adaptive performance management for 3G mobile networks. As the main result of the paper, the introduced approach is based on a mathematical framework for the proposed update schemes rather than a lookup table. As a consequence, the adaptive control mechanism can be adjusted in an intuitive way and optimal system parameter configuration can efficiently be determined . We effectively utilize adaptive performance management for improving not only QoS for mobile users but also increase revenue earned by service providers. As in [19], controlled system parameters comprise queueing weights for packet scheduling, a threshold value of the access queue for admission of non real-time traffic, and a portion of the overall available bandwidth reserved for handover calls. Beyond [19], we propose a scheme for adjusting the queueing weights for both improving QoS for higher priority users that suffer from a high population of users with lower priority and for increasing the revenue earned by the service provider. For the analysis of the update strategy of the queuing weights, we consider a usage-based and a usage-/throughput-based pricing policy according to [11,12,21]. Furthermore, we introduce a hybrid pricing policy combining the notion of flat-rate and a usage-based pricing according to current policies of GSM networks. Performance curves derived by simulation evidently illustrate the gain of the unified approach for adaptive performance management. In fact, for UMTS networks, simulation results show that handover failure probability can be improved by more than one order of magnitude. Moreover, packet loss probability can be effectively regulated to a predefined level and the provider revenue is significantly increased for all considered pricing policies. The paper is organized as follows. Section 2 introduces the unified approach for adaptive performance management and describes its embedding in the system architecture of 3G mobile networks. Section 3 introduces strategies for controlling the parameters of an admission controller in order to improve QoS. Section 4 describes the parameter control of a packet scheduler for the combined improvement of both QoS and provider revenue. In section 5, we present simulation results that illustrate the benefit of employing the proposed approach for adaptive performance management. Finally, concluding remarks are given. Adaptive performance management for 3G mobile networks This section introduces the unified approach for regularly adjusting system parameters to changing traffic load, packet arrival pattern or population of users, etc. We consider a cellular mobile network in which a different transceiver station serves each cell. The purpose of the transceiver station is the modu-Figure 1. System architecture for adaptive performance management. lation of carrier frequencies and demodulation of signals. Furthermore , a base station controller (BSC) is considered that is responsible for a cluster of cells, i.e., several transceiver stations . The BSC manages the radio resources, i.e., schedules data packets, and controls handovers inside the cell cluster as well as handovers towards and from neighboring cell clusters. To improve QoS for mobile users as well as to increase revenue earned by service providers, an entity for Adaptive Performance Management (APM) is included in a BSC. Furthermore , a BSC has to be extended by an online performance monitoring component that derives QoS measures in a certain time window (e.g., handover failure probabilities of mobile users or packet loss probabilities). These QoS measures form a system pattern that is submitted in fixed time intervals (i.e., a control period) to the APM entity, which subsequently updates corresponding system parameters (i.e., parameters of traffic controlling components like the admission controller and packet scheduler). Thus, the proposed approach closes the loop between network operation and network control. Figure 1 shows the system architecture for performance management embedded in a BSC. 2.1.1. Online performance monitoring System parameters of a BSC can be effectively updated by monitoring QoS measures, which are immediately affected by these parameters. A current value for a QoS measure is determined online based on a set of relevant events corresponding to this QoS measure (e.g., packet arrivals are relevant events for computing packet loss probabilities). The online monitoring of QoS measures is done by a sliding window technique as introduced in [19]. The width of the sliding window over time depends on the number of relevant events that are occurred according to a QoS measure. Upon arrival of a new relevant event the sliding window moves in time. At the end of a control period the QoS measures are derived for each sliding window (e.g., packet loss probability can be derived from number of lost packets divided by number of all packet arrivals in the sliding window). These QoS measures and the number of events occurred in the last control period form the system pattern that is transferred to the adaptive performance management entity (see figure 1). IMPROVING QoS AND PROVIDER REVENUE IN 3G MOBILE NETWORKS 211 Note that an accurate online monitoring of QoS measures requires a specific width for the sliding window. A certain number of events representing the history of the QoS measure have to be considered to get an expressive measure. On the other hand considering a big sliding window prevents the APM entity from fast reaction on changing traffic conditions. A bigger sliding window contains more history and, thus, more events have to be collected to cause a significant change in the online monitored QoS measure. This tradeoff between accurate online monitoring and fast reaction of the APM to changing traffic conditions has to be studied carefully in several experiments to get the optimal width of the sliding window for each QoS measure. 2.1.2. Adaptive performance management Whenever a system pattern S = {(P 1 , n 1 ), . . . , (P m , n m ) }, consisting of online monitored QoS measures P 1 , . . . , P m and the numbers of relevant events n 1 , . . . , n m occurred in the last control period is transmitted to the APM an update of the system parameters can be performed. In general , an update of a system parameter is made according to a function f depending on a subset of the QoS measures P 1 , . . . , P m and the previous value ( old) of the system parameter . Let P ( 1) , . . . , P (k) , k m , be the QoS measures corresponding to system parameter , then the update is made if a certain minimum number n( ) of relevant events occurred in the last control period. That is: ( new) = f P ( 1) , . . . , P (k) , ( old) , if min {n ( 1) , . . . , n (k) } n( ). (1) We classify update functions in relative functions, that perform a parameter update relative to the old parameter value and absolute functions that set the new parameter value independent of the old value, i.e., f is independent of ( old) in (1). With relative update functions strong fluctuations of the corresponding system parameter in one update step can be avoided. In section 3, we study a special class of relative update functions in order to set the parameters of an admission controller. Furthermore, we develop in section 4 an absolute update function for adjusting the weights of a weighted fair queueing packet scheduler. 2.2. Economics and pricing policies in 3G mobile networks There are multiple requirements, which should be fulfilled for any viable pricing mechanism in multi-service class data communication networks [12]. A primary target of differen-tiated pricing of Internet services is the prevention of system overload and an optimal resource usage according to different daytimes and different traffic intensities. Furthermore, the pricing scheme should be implemented in a completely de-centralized manner and there should be multiple priorities in order to take into account the different QoS required by different applications and users. In general, pricing policies can be partitioned into usage-based (pay-as-you-go) pricing, flat-rate (all-you-can-eat) pricing, and dynamic pricing. In usage-based pricing policies a user is charged according to a connection time or traffic volume. Whereas connection based calls (e.g., in GSM) are charged by connection time, packet-switched services (e.g., in UMTS) are charging the transferred data volume. Dynamic pricing models take into account the state of the mobile radio network for determining the current price of a service. Congestion-sensitive pricing as a particular dynamic pricing model has been shown to be more responsive. MacKie-Mason and Varian introduced the concept of congestion-sensitive pricing in their smart market scheme [21]. Under this model, the actual price for each packet is determined based on the current state of network congestion. In [25], Rao and Petersen discussed the optimal pricing of priority services. Analogously to the smart market approach, Gupta et al. presented a pricing scheme that uses priorities on the packet-level [13]. They proposed to differentiate Internet traffic according to delay and loss requirements. For the analysis of the update strategy of the queuing weights, we consider in section 4 a usage-based and a usage-/ throughput-based pricing policy according to [11,12,21]. Furthermore , we introduce a hybrid pricing policy combining the notion of flat-rate and a usage-based pricing according to current policies of GSM networks. Strategies for improving Quality of Service The proposed approach distinguishes three different types of services: circuit-switched services, packet-switched real-time services (RT), and packet-switched non real-time services (NRT). Typically, circuit-switched services are voice calls from a GSM mobile station. As proposed by 3GPP, RT services belong to the conversational and streaming classes and NRT services fall into the interactive and background classes [2]. The bandwidth available in a cell must be shared by calls of these different service classes and the different service requirements have to be met. Before a mobile session begins, the user needs to specify its traffic characteristics and desired performance requirements by a QoS profile. Then, an admission controller decides to accept or reject the users request based on the QoS profile and the current network state as, e.g., given by queueing length. The purpose of the admission controller is to guarantee the QoS requirements of the user who requested admission while not violating the QoS profiles of already admitted users. The call admission criteria will be different for each service class. The QoS profile for RT sessions specifies a guaranteed bandwidth to be provided for the application in order to meet its QoS requirements. If the network cannot satisfy the desired bandwidth, the corresponding admission request is rejected. Data packets arriving at the BSC are queued until they are scheduled to be transmitted over the radio link. For NRT sessions , we consider an admission controller taking into account free buffer space in the NRT queue [8]. In order to prevent buffer overflow once a call is admitted, the current queueing length is set against certain buffer availability threshold 212 C. LINDEMANN ET AL. of the capacity, denoted by . The admission criteria for voice and RT handovers are the same as for new voice calls and RT sessions except that additional handover bandwidth can be utilized. The analysis of several admission control schemes for cellular systems presented in [24] showed that the simple reservation scheme (i.e., reserving bandwidth for handover calls) performs remarkably well. For simple cellular networks, the optimal amount of bandwidth reserved for handover calls can be determined by analytical models [14]. In the model presented here, we denote with b h the portion of the overall bandwidth that is exclusively reserved for handover calls from neighboring cells. The considered admission controller does not prioritize NRT handovers over new NRT sessions. Further details of the admission controller are given in [19]. 3.2. Adjusting the admission controller for QoS improvement In this section, we show how to utilize equation (1) for setting the parameters and b h of the admission controller in order to reduce packet loss probability and handover failure probability . For updating the system parameters, we split the general function introduced in section 2.1 into separate functions each depending only on one QoS measure. Let P 1 , . . . , P k be the QoS measures corresponding to a system parameter . Then, equation (1) can be simplified to ( new) = f 1 (P 1 ) + + f k (P k ) k ( old) , L ( new) R. (2) The interpretation of (2) is the following. Each update function f i describes the influence that the QoS measure P i should have on the system parameter . Subsequently, the overall update is performed by computing the arithmetic mean of the functions f i multiplied with the old value of the system parameter. Note that the value ( new) must be truncated at a certain lower bound L and an upper bound R in order to guarantee that the computation of ( new) results in a valid value of the system parameter. As basic update function we consider a logarithmic linear function of the form: f i (P i ) = m i log P i + b i . (3) The reason for this choice is that we want to consider QoS measures like loss probabilities and failure/blocking probabilities , which are in the range of 10 -5 to 1. Therefore, a logarithmic shape is more suitable. In previous work [19], we have studied update schemes of system parameters of an admission controller and a packet scheduler based on a lookup table. In order to determine the optimal entries of this lookup table extensive off-line simulation experiments have been conducted. Applying regression statistics to the entries of this lookup table shows that these entries are well represented by functions with logarithmic shape. Thus, besides the motivation of the update functions given here, their choice is to a large extend originated from regression statistics conducted in earlier work. The strength of the influence of f i on ( new) can be adjusted with the gradient m i . The parameter b i can be determined by the following interpretation: suppose the desired level of the QoS measure P i is i (e.g., the desired packet loss probability is 0.001). That is, if the online measured value of P i is i the system parameter should not be changed in the update step from the point of view of measure P i . Therefore, we chose f i ( i ) = 1 and from this relation we get b i = 1 - m i log i . Inserting in equation (3) results in the final form of the update function: f i (P i ) = m i log P i i + 1. (4) For ease of notation, we abbreviate the QoS measures handover failure probability and new call/session blocking probability corresponding to voice calls and RT sessions by HFP and CBP, respectively. The probability of a packet loss due to buffer overflow in the NRT queue is abbreviated by PLP. The update strategy according to equations (2)(4) is justified by its intuitive understanding and the performance results presented in section 5. The suitability of update functions other than (2)(4), is subject for further study and out of the scope of this paper. 3.2.1. Update of non real-time queue threshold Recall that a system parameter update is performed each time a system pattern arrives at the APM entity and the minimum number of relevant events corresponding to this system parameter is reached. Determining the update for the system parameter , i.e., determining ( new) , is performed corresponding to the old value ( old) and the actually observed QoS measure PLP. That is: ( new) = f (PLP) ( old) , 0.001 ( new) 1. (5) The truncation of ( new) at the lower bound guaranties that the value does not accumulate near zero for long periods of low traffic load. The minimum number of relevant events required for an update of is counted in data volume rather than in packet arrivals (in the experiments this number is 5 MB). The setting of the gradient m of the corresponding update function is derived from a couple of experiments for different values of the gradient. We found m = -0.02 to be suitable. Choosing a suitable value for the gradient is a similar tradeoff as explained for the sliding window size. A large gradient results in a fast update of the system parameter in a few number of update steps, but also introduces higher fluctuations of the system parameter over time. We demonstrate the speed of the parameter adjustment in an experiment in section 5. Furthermore , several experiments for different desired loss values are presented. 3.2.2. Update of fraction of bandwidth reserved for handover The update for the system parameter b h , i.e., determining b ( new) h , is performed based on the old value and the actually observed QoS measures HFP and CBP. That is: b ( new) h = f 1 ( HFP) + f 2 ( CBP) 2 b ( old) h , 0.001 b ( new) h R. (6) IMPROVING QoS AND PROVIDER REVENUE IN 3G MOBILE NETWORKS 213 The value b ( new) h is truncated at a lower bound of 0.1% and a certain upper bound R which is a fraction of the overall bandwidth available (in the experiments we fix R = 0.7). The truncation at the lower bound is for the same reason as explained above. In fact, for computing b ( new) h two QoS measures corresponding to the actually observed HFP and CBP are taken into account. A high HFP should increase b ( new) h but this obviously also increases the CBP because less bandwidth is available for new voice calls and RT sessions. Therefore , the HFP and the CBP influence the handover bandwidth b ( new) h . In fact, m 1 = -m 2 holds in the update functions f 1 and f 2 . From a couple of experiments for different gradients, we found m 1 = 0.08 to be suitable. A common assumption in cellular networks is to prioritize handover calls over new calls. Therefore, the desired handover failure level 1 should be smaller than the desired call blocking level 2 . According to these values the handover bandwidth is slightly increased, if HFP is equal to CBP. With the presented strategy the parameters of the update functions can be chosen in an intuitive way and optimal parameter configuration can efficiently be determined. This is the major advantage over the approach based on a Performance Management Information Base introduced in [19] which requires extensive off-line simulation experiments. Strategies for improving both QoS and provider revenue At a BSC responsible for a cluster of cells, data packets from various connections arrive and are queued until bandwidth for transmission is available. In order to distinguish different priorities for NRT traffic corresponding to the traffic handling priority defined by 3GPP [2], scheduling algorithms like Weighted Round Robin (WRR), Weighted Fair Queueing (WFQ [9]) or Class Based Queueing (CBQ [10]) have to be implemented. An overview of queueing issues for guaranteed performance services can be found in [27]. In WFQ, the weights control the amount of traffic a source may deliver relative to other active sources during some period of time. From the scheduling algorithm's point of view, a source is considered to be active, if it has data queued in the NRT queue. Let B be the overall bandwidth available for NRT sessions at time t. For an active source i with weight w i , the bandwidth B i that is allocated to this transfer at time t is given by B i = w i j w j B. (7) In (7) the sum is taken over all active NRT sources j . A class based version of WFQ serves packets of each priority class according to the weights rather than every active source. 4.2. Adjusting the packet scheduler for QoS and revenue improvement This section utilizes the proposed approach for the adaptive control of the weights of a weighted fair queueing packet scheduler in order to improve QoS as well as to increase the revenue. The strategy for adjusting the weights combined with the introduction of several pricing policies constitutes a further contribution of the paper. Recall that the revenue earned by a mobile service provider is determined by the monthly payment of mobile users as well as by the additional usage-based pricing after the monthly amount of data volume is consumed. Note, that the monthly subscription rate is only relevant for monthly revenue calculations. In this section, we consider the revenue improvement in a certain small time period regardless the monthly subscription rates. In section 5, we briefly discuss monthly revenue calculation. Let P denote the number of different priority classes, i.e., weights of the weighted fair queueing scheduler. Define by b i (t) the transferred data volume in time t of users of priority i and by r i (t) the payment of users of priority i at time t, i.e., the user pays for the transferred data volume. We distinguish a pure usage-based and a usage-/throughput-based pricing policy: (a) A user of priority i has a fixed payment p i per kbit during his session, i.e., r i (t) = p i . (b) The payment of a user of priority i consists of a fixed part p i that is increased proportional to the additional throughput i (t) he received due to the update of the queueing weights, i.e., r i (t) = p i i (t) . According to the proposed data volume based pricing with respect to different priority classes the revenue function (t) is given by (t) = P i =1 r i (t)b i (t). (8) The revenue function of equation (8) is utilized in section 5 for evaluating the strategies for revenue improvement presented below. 4.2.1. Update of WFQ weights Recall that packets of NRT users arriving at the BSC are first queued until they are scheduled for transfer by a weighted fair queueing discipline. Let w i w i +1 , i = 1, . . . , P - 1, be the basic weights of the WFQ scheduler. The update of the queueing weights, i.e., determining w ( new) i is made according to an absolute update function depending on the basic weights w i and the current number of NRT sessions belonging to priority i. Therefore, every system pattern that is transmitted from the online monitoring component to the adaptive performance management entity contains the current number of active NRT sessions with priority i in the cell. For ease of notation, the number of active non real-time sessions with priority i is abbreviated by NRT i . The idea behind the strategy for revenue improvement is to shift the overall utilization of bandwidth for NRT traffic 214 C. LINDEMANN ET AL. towards higher priority users, which pay more for the transferred data volume. Note that the update strategy should be conservative in a way that the transfer of packets of low priority is not simply blocked if packets of higher priorities are arriving, i.e., priority queueing. Assuming that the majority of users will buy a cheaper low priority service class, priority queueing will leave most users unsatisfied. Therefore, the update strategy also considers the QoS aspect. The update strategy concerning the queueing weights is developed according to the following premises: (i) If the number of active NRT users in the cell is the same for each priority class, i.e., NRT i = NRT j , i = j, the weights w ( new) i should be set according to the basic weights w i for i = 1, . . . , P . (ii) Priority classes with low population of users compared to other classes should be prioritized, i.e., the corresponding weights should be increased. (iii) The relative ordering of the weights should be preserved in a strong way, i.e., w ( new) i (w i /w i +1 ) w ( new) i +1 for i = 1, . . . , P - 1. Premise (i) constitutes the key of the update strategy. If all priority classes have the same population of users the scheduling should work as in the case without adaptive control of the weights. The rationale behind premise (ii) is to prioritize users that are consuming less bandwidth (relative to their weights) than users belonging to other classes, i.e., users of low population should be made more independent from the influence of user classes with higher population. This premise constitutes the basic idea for QoS improvement and is demon-strated by the following example that considers two priority classes, i.e., a high and low priority class. In WFQ the available bandwidth is shared among all active users according to their weights. That is, if the minority are high priority users, the overall bandwidth consumed by these users will suffer from a strong influence of low priority users that hold the majority. Therefore, increasing the weights for high priority users will result in a higher QoS for this user class. Updating the weights according to this strategy will result in a scheduling algorithm somewhere between a WFQ and a class based queueing scheduler. In fact, the benefit of both is utilized: the fair sharing of the bandwidth of WFQ and the higher bandwidth guarantees for each priority class provided by a class based queueing scheduler. Preserving the relative ordering of the weights (i.e., premise (iii)) guarantees that QoS for higher priority users and, therefore, the provider revenue can only be improved due to the adaptive control of the weights. If the intention of the update strategy is not primary on improving provider revenue the weights can be also set in a weak relation, i.e., w ( new) i w ( new) i +1 . This might be useful to increase QoS for users of low population independent of their priority class. With the following algorithm the computation of the weights w ( new) 1 , . . . , w ( new) P can be performed iteratively in P -1 minimum calculations. The iteration is given by w ( new) 1 = w 1 (NRT 1 ) , (9) w ( new) i = min w i w i -1 w ( new) i -1 , w i (NRT i ) , i = 2, . . . , P. (10) In order to smooth the influence of the number of NRT users on the queueing weights, an exponent 0 is considered (e.g., = 1/2). It is easy to show that premises (i), (ii) and (iii) hold for the weights set according to equations (9) and (10). The iteration starts with setting w ( new) 1 according to NRT 1 and continues up to w ( new) P . Note that this is only one possibility to set the new weights. Any other starting position for the iteration is possible and results in a slightly different update of the weights. Nevertheless, the algorithms work in a similar way, and therefore, we consider only the iteration of (9) and (10). If currently no users of priority i are in the cell, i.e., NRT i = 0, the algorithm skips the setting of the corresponding weight w ( new) i and the next iteration step i + 1 is related to step i - 1. Subsequently, these weights were set to zero. For other scheduling disciplines like weighted round robin or a class based queueing corresponding update strategies can be derived in a similar way. 4.2.2. Considering advanced pricing policies In pricing policy (b) introduced above, users have to pay an additional fee depending on the throughput improvement due to the update of the queueing weights. This concept of pricing indicates strong similarities to the congestion-sensitive pricing of the smart market scheme [21], where the actual price for each packet is determined based on the current state of network congestion. Similarly, in our throughput-based pricing policy the throughput of users is determined by their willingness-to-pay additional costs (according to their choice of priority class) for transmission of packets in a congested network. The additional payment is justified because the throughput for users of higher priority will be maintained, even if more and more users of lower priority attend the cell, i.e., the network is currently congested. We describe the relative throughput increase of priority class i with the function i (t) = ( new) i = w P THR i w i THR P . (11) In equation (11), THR i is the current throughput of class i derived from the corresponding sliding window and 0 1 is a scaling exponent (e.g., = 1/4) that has to be adjusted by the service provider for appropriate revenue dimensioning . In order to guarantee that revenue will be only improved, i (t) has to be truncated, i.e., i (t) 1. Next, we adjust the weights according to an advanced pricing policy that adopts ideas, which have been successful in existing GSM networks. In GSM networks, the pricing of a provided service is as follows: the proposed service is offered based on a monthly payment for a dedicated amount IMPROVING QoS AND PROVIDER REVENUE IN 3G MOBILE NETWORKS 215 of call time. If a user has consumed this amount of time before the end of the month, he has to pay for any further use of this service based on a time-dependent accounting. This idea can be generalized and extended towards packet-switched services in 3G networks. Analogously, a user has to pay a monthly charge for a dedicated amount of data volume , which can be transferred without further pricing. After using up this monthly amount of data, the user has to pay for the desired services according to the transferred data volume (byte-based). Moreover, analogous to GSM networks a user can utilize "unused" data volume, i.e., the unused fraction of the prepaid monthly amount of data volume, in subsequent months. If the monthly amount of data is unrestricted, this pricing would become a flat-rate pricing and if there is no monthly payment, the pricing follows a usage-based policy. Thus, our pricing policy constitutes a hybrid approach of flat-rate and usage-based pricing. The update of the queueing weights can now be extended in a way that users consuming their monthly amount of data are served with a lower priority than users currently paying for their data transfer. Therefore, we introduce a new weight w corresponding to the not paying users. The weight w must be sorted in the weights w 1 , . . . , w P and the iterative update algorithm (9)(10) can be applied to the P + 1 weights as described above. In order to distinguish not paying users with different priorities these users are served by the WFQ scheduler with weights w 1 , . . . , w P relative to w . That is, WFQ is applied to 2 P weights, i.e., w 1 , . . . , w P and (w /w) w 1 , . . . , (w /w) w P with w = w 1 + + w P . 4.3. Implementation issues As outlined in section 2.1, the controlled system parameters for QoS and revenue improvement, i.e., , b h , w i , i , constitute an integral component of the proposed extension to a BSC. The adjustment of system parameters is only based on implicit information that is directly measured by the online monitoring component. Therefore, no additional signaling with other BSCs is necessary for updating system parameters. The online monitored QoS measures, i.e., PLP, HFP, CBP, NRT i , and THR i , can easily be derived and stored within the BSC (see figure 1). The PLP can directly be determined by counting the number of IP packets, which are lost due to buffer overflow in the NRT queue. HFP as well as the CBP is determined by the non-admitted handover calls and new calls in the admission controller, respectively. Admission, termination , and handover of NRT calls enable the profiling of NRT i , the number of non real-time sessions with priority i. Moreover , the packet scheduler allows the throughput computation of NRT users according to their individual priorities. Furthermore , no time consuming signaling is needed to transfer the system pattern inside the BSC because the online performance monitoring component and the performance management entity both reside in the BSC. The question arises how call charging can be accomplished for the considered pricing policies in 3G mobile networks. For pricing policy (a), i.e., a fixed payment per kbit, call charging can easily be processed by the subscription management component of the operation subsystem (OSS) by means of the call charging mechanism using the home location register (HLR) [1]. Similarly, the hybrid pricing scheme can be realized except that the remaining amount of prepaid data volume has to be stored in the HLR for charging the transferred data volume. Utilizing these existing charging mechanisms, no additionally signaling overhead arises for charging data services. The throughput-based pricing policy (pricing policy (b)) just slightly changes the situation and can easily be implemented within the BSC using a local copy of the user's HLR charging data fields. This local data minimizes signaling overhead of individual user charging. According to the transferred data volume and current throughput of the user's bandwidth class, this local charging profile is continuously updated. Handovers with changing BSC of response induce the transfer of this local charging profile to the new BSC of response. Subsequently, these local data have to be updated in the HLR for individual user accounting after termination of the call. Note, that this transfer of local charging profiles can naturally be embedded in the OSS functionality. Evaluation of the adaptive performance management strategies For traffic modeling of RT applications we utilize the approach proposed in [18], where variable bit rate video traffic is modeled in terms of time-discrete M/G/ input processes. This model is based on measured video streams and efficiently captures the correlation structure of the considered video traffic applying the time-discrete M/G/ input process. The generated traffic is transformed utilizing a hybrid Gamma/Pareto numerical transformation in order to capture the marginal distribution of the measured traffic stream. Subsequently, the synthetically generated traffic is broken down to IP packets of a maximum size of 1500 bytes, which are uniformly distributed within a given frame-duration of the MPEG video sequence comprising of 1/30 s. Note that this traffic model does not propose information for modeling RT session durations. Therefore, we assume session durations to be exponentially distributed (see section 5.2). Recent recommendations for modeling NRT traffic and analytical traffic models for 3G mobile networks are proposed in [15,16], respectively. The traffic model is based on real measurements conducted at an Internet service provider dial-in link, which comprises comparable characteristics of future mobile networks [17], i.e., different access speeds, influence of the user behavior due to different tariff limits, as well as asymmetric up- and downlink traffic. Based on these measurements a NRT traffic model is conducted, applying the idea of the single user traffic model, which describes traffic characteristics on session-level, connection-level, i.e., application-level , and packet-level, respectively. The key insight of this modeling approach lies in an appropriate scaling procedure of 216 C. LINDEMANN ET AL. Table 1 Characteristics for different UMTS session types. Circuit switched Streaming real time (RT) Interactive non real time (NRT) voice service Audio Video high priority normal priority low priority Portion of arriving requests 25% 12% 3% 6% 18% 36% Session duration 120 s 180 s determined by session volume distribution Session dwell time 60 s 120 s 120 s the measured trace data towards typical bandwidth classes of 3G mobile networks, i.e., 64 kbps, 144 kbps, and 384 kbps. In this context, a bandwidth class denotes the maximum bandwidth capability of future handheld devices. We refer to [15] for details of the NRT traffic model, especially for the parameterization of the traffic characteristics. 5.2. The simulation environment In order to evaluate the proposed approach for adaptive control , we developed a simulation environment for a UMTS access network, i.e., a UMTS Terrestrial Radio Access Network (UTRAN [3]). The simulator considers a cell cluster comprising of seven hexagonal cells with corresponding transceiver stations (i.e., Node B elements), that are managed by a base station controller (i.e., a Radio Network Controller, RNC). We assume that a mobile user requests a new session in a cell according to a Poisson process. When a mobile user starts a new session, the session is classified as voice, RT, or NRT session, i.e., with the session the user utilizes voice, RT, or NRT services mutually exclusive. RT sessions consist of streaming downlink traffic corresponding to the UMTS streaming class specified by 3GPP [2] and NRT sessions consist of elastic traffic and correspond to the UMTS interactive class or background class, respectively. For the year 2010 an amount of about 50% voice calls is anticipated [26]. We assume that one half of the voice calls are served over the frequency spectrum for traditional GSM services (i.e., 890915 and 935960 MHz) and the second half is served over the new frequency spectrum allocated for UMTS. Nevertheless, the simulator considers only the new frequency spectrum. Therefore , we assume that 25% of the call requests are voice calls whereas RT and NRT sessions constitute 15% and 60% of the overall arriving requests (see table 1). Subsequently, we have to specify the QoS profile for RT and NRT sessions. For RT sessions the simulator considers two QoS profiles, i.e., a low bandwidth profile comprising of a guaranteed bit rate of 64 kbps corresponding to streaming audio and a high bandwidth profile comprising of a guaranteed bit rate of 192 kbps corresponding to streaming video. According to the RT traffic model presented in section 5.1, we assume that 80% of the RT sessions utilize the low bandwidth profile whereas the remaining 20% utilize the high bandwidth profile. Following the single user traffic model, NRT sessions are partitioned according to different bandwidth classes as follows: 60% for 64 kbps, 30% for 144 kbps, and 10% for 384 kbps, comprising of different priorities (see table 1), respectively. The amount of time that a mobile user with an ongoing session remains within the cell is called dwell time. If the session is still active after the dwell time, a handover toward an adjacent cell takes place. The call/session duration is defined as the amount of time that the call will be active, assuming it completes without being forced to terminate due to handover failure. We assume the duration of voice calls and RT sessions to be exponentially distributed. As proposed in [6], the dwell time is modeled by a lognormal distribution. All corresponding mean values are shown in table 1. A NRT session remains active until a specific data volume drawn according to a bandwidth-dependent lognormal distribution is transferred. To distinguish between NRT traffic classes, the UMTS simulator implements a WFQ scheduler with three packet priorities : 1 (high), 2 (normal), and 3 (low) with weights w 1 = 4, w 2 = 2, and w 3 = 1. These priorities correspond to the traffic handling priority specified by 3GPP. To model the user behavior in the cell, the simulator considers the handover flow of active mobile users from adjacent cells. The iterative procedure introduced in [4] is employed for balancing the incoming and outgoing handover rates. The iteration is based on the assumption that the incoming handover rate of a user class at step i + 1 is equal to the corresponding outgoing handover rate computed at step i. 5.2.1. UMTS system model assumptions The simulator exactly mimics UMTS system behavior on the IP level. The focus is not on studying link level dynamics . Therefore, we assume a reliable link layer as provided by the automatic repeat request (ARQ) mechanism of the Radio Link Control (RLC) protocol. As shown in [22] for the General Packet Radio Service (GPRS), the ARQ mechanism is fast enough to recover from packet losses before reliable protocols on higher layers (e.g., TCP) recognize these losses due to timer expiration. Thus, a reliable link level can be assumed when considering higher layer protocol actions (see, e.g., [20]). To accurately model the UMTS radio access network, the simulator represents the functionality of one radio network controller and seven Node B transceiver stations, one for each of the considered cells. Since in the end-to-end path, the wireless link is typically the bottleneck, and given the anticipated traffic asymmetry, the simulator focuses on resource contention in the downlink (i.e., the path RNC Node B MS) of the radio interface. The simulator considers the UTRAN access scheme based on Wideband-Code Division Multiple Access (W-CDMA) in Frequency Division Duplex mode (FDD) proposed by 3GPP [1]. In FDD downlink, a division of the radio frequencies into IMPROVING QoS AND PROVIDER REVENUE IN 3G MOBILE NETWORKS 217 four physical code channels with data rates of 1,920 kbps each up to 512 physical code channels with 15 kbps data rates each is possible. Therefore, the overall bandwidth that is available in one cell is 7,680 kbps. For the channel coding, we assume a convolution-coding scheme with coding factor 2. In the experiments without adaptive control the handover bandwidth portion b h is 5% and the NRT queue threshold is set to 95%. The simulation environment was implemented using the simulation library CSIM [7]. In a presimulation run the handover flow is balanced, for each cell at the boundary of the seven-cell cluster. All simulation results are derived with confidence level of 95% using the batch means method. The execution of a single simulation run requires about 4060 min of CPU time (depending on the call arrival rate) on a dual processor Sun Sparc Enterprise with one GByte main memory. 5.2.2. Implementation of the hybrid pricing policy in the simulator According to the hybrid pricing policy as introduced in section 4.2.2, the user's overall remaining amount of prepaid data volume d out of the user's monthly data volume D is determined at the beginning of a session. Moreover, the remaining amount of data volume of previous months r is determined . For simulation study purposes, this is accomplished by choosing the random value d uniformly out of the interval [0, kD]. kD captures the monthly amount of data a user typically transfers, i.e., a user typically transfers a multiple k of the data volume D that is available for a fixed monthly payment . The random value r, is sampled according to a uniform distribution out of the interval [0, 0.1D], where 0.1D measures the maximum amount of "unused" data volume of previous months. If d exceeds D + r the user has no remaining prepaid data volume, including the data volume of the current and the previous months. Otherwise, there is a remaining amount of prepaid data volume D + r - d for the considered user and additional pricing arises only, if the transferred data volume of the user session exceeds D + r - d. Thus, during the user session the remaining data volume has to be updated according to the actually transferred data. In the simulation studies we utilize the proposed hybrid-pricing scheme with a prepaid monthly data volume of 150 MB. According to the different priority classes 1, 2, and 3, the volume-based pricing for transferred data exceeding the prepaid monthly data volume comprises of 20, 15, and 10 cost-units per MB, respectively. Considering the changing traffic loads according to the daytime, this approach can be refined, by the notion of different pricing for daily periods of time. For the parameterization of the typically monthly transferred data volume, we assume k = 2. Note that the parameterization of the pricing scheme is chosen for demonstration purposes only. Due to the high flexibility of the hybrid pricing scheme, it can be easily extended towards multiple , concurrent pricing schemes comprising of, e.g., different monthly amounts of prepaid data volumes, different payments for the individual priority classes, or a pure usage-based pricing as well as pure flat-rate pricing. (a) (b) Figure 2. Impact of adaptive performance management on non real-time traffic. 5.3. Performance results Using simulation experiments, we illustrate the benefit of the proposed unified approach for adaptive performance management of UMTS systems. In particular, we show the improvement of QoS measures and the increase in revenue earned by service providers. The presented curves plot the mean values of the confidence intervals for the considered QoS measures. In almost all figures, the overall call/session arrival rate of new mobile users is varied to study the cell under increasing load conditions. For ease of notation, results with and without adaptive performance management (APM) are abbreviated by APM on and APM off, respectively. In a first experiment, we investigate the effect of adaptive control on the threshold for the buffer size of the NRT queue denoted by . Figure 2 shows the NRT packet loss probability (a) and the average number of NRT users in the cell (b) for the UMTS system with and without adaptive control . Furthermore, the figures distinguish between different desired loss levels as introduced in section 3.2. We observe that the APM achieves a substantially decrease in packet loss probability. Moreover, the packet loss probability can be kept below a constant level for increasing arrival rates of mobile users. Note, that this level slightly differs from the desired level of the QoS measure. This is due to the fact that the update function only decreases the NRT threshold if the online 218 C. LINDEMANN ET AL. Figure 3. Number of packet losses for a half day window of a weekly usage pattern. measured packet loss probability is greater than . Therefore , the packet loss probability is in steady state also slightly greater than . Nevertheless, figure 2 shows that the resulting packet loss probability can be adjusted quite well. For very low arrival rates, the packet loss probability is increased compared to the case without adaptive control. This is because the packet loss probability is below the desired level and is adjusted towards 100%. Figure 2(b) shows the average number of NRT users admitted in the cell. For all curves, the number of NRT users in the cell first increases up to about 70 users for an arrival rate of 1.0 arrivals per second. For higher arrival rates the admission controller decides to reject requests depending on the choice of the NRT threshold. In the case without APM the number of NRT users approaches 100 whereas in the cases with adaptive control less users are admitted in the cell because the threshold parameter is decreased (e.g., about 80 users for = 0.001). For high arrival rates a slightly decrease of the average number of NRT users can be observed. This is due to the fact that with increasing arrival rate the competition between voice, RT and NRT traffic decreases the bandwidth capacity available for NRT traffic. Therefore, less NRT users are admitted. In the experiment presented in figure 3, we study the absolute number of packet losses observed in one hour for a transient scenario, i.e., the arrival rate of new calls is changing every hour according to a half day window of a weekly usage pattern [15]. The purpose of this experiment is to show that the adaptive performance management is fast enough to react on changing traffic conditions, i.e., to effectively adjust the NRT threshold in order to reduce packet losses. The bars shown in figure 3 correspond to the number of packet losses for experiments with and without adaptive control. Furthermore , the figure distinguishes between a desired loss level of 0.01 and 0.001, respectively. The new call arrival rates considered in one hour are depicted above the bars. We conclude from figure 3 that for a real-life pattern of changing arrival rates the packet losses can be effectively controlled by the APM. This justifies the choice of the gradient m = -0.02 in the update function for the NRT threshold. (a) (b) Figure 4. Impact of adaptive performance management on handover traffic. Next, we study the effect of the APM on the handover traffic. Figure 4 shows the handover failure probability (a) and the new call blocking probability (b) for the UMTS system with and without APM. Similar to figure 2, we distinguish between different desired levels for the handover failure probability. The desired level for new call blocking is fixed to 0.1. Note, that for controlling the handover bandwidth the desired level can be used only to adjust the degree of prioritization of handover failure over new call blocking . Distinct from the packet loss probability, it cannot be expected to keep the handover failure probability at a constant level for increasing traffic load. That is for two reasons : (1) the handover bandwidth is adjusted according to two QoS measures that have a contrary influence and (2) the increase of the handover bandwidth must be limited by a certain portion of the overall available bandwidth (see section 3.2). If this limit is reached handover failures occur more frequently for further increasing call arrival rate. These two effects can be observed in the curves of figure 4. Nevertheless , the handover failure probability is improved more than one order of magnitude for call arrival rates between 0.75 and 1.25 call requests per second and a desired loss level = 0.001. When studying the blocking probability of new voice calls and RT sessions (see figure 4(b)), we surly observe a higher blocking probability of new calls in the case with adaptive control and high arrival rate. In IMPROVING QoS AND PROVIDER REVENUE IN 3G MOBILE NETWORKS 219 Figure 5. Improving QoS for high priority non real-time users. Figure 6. Effect of adjusting WFQ weights on bandwidth utilization of NRT traffic. fact, almost all call requests are blocked if system load is high. In a next experiment, we study the impact of NRT users on QoS by the adaptive control of the queueing weights as introduced in section 4.2. Figure 5 plots the average throughput per user for each priority class of NRT traffic. As shown in table 1, we assume 10% NRT users with high priority, 30% with normal priority, and 60% with low priority. Recall, that higher priority service is more expensive and, hence, more users choose low priority service. If the overall load in the cell is very low (i.e., less than 0.3 call arrivals per second) each NRT user receives the maximal throughput independent of the priority class. However, when the cell load is further increased (arrival rates of more than 0.5 arrivals per second), throughput for users of all priority classes decreases. The intention of adaptively controlling the queueing weights is to reduce heavy throughput degradation of high priority users in this case. The performance increase of high priority users and the decrease of low priority users are shown in figure 5. Figure 6 plots the bandwidth portion utilized for each priority class of NRT traffic. For low arrival rate (i.e., less than 0.5 call arrivals per second) NRT users with low priority utilize the greatest portion of the NRT bandwidth because most NRT users have priority low. When the cell load is increased (arrival rates of more than 0.5 arrivals per second), the band (a) (b) Figure 7. Revenue improvement for usage-based pricing policy. width will be utilized more and more by high priority users. The adaptive control of the WFQ weights decides to intensify this effect because users belonging to priority high suffer from the high population of low priority users. Figures 5 and 6 are derived from simulation runs with = 1/2 (see section 4.2). In the following experiments, we study the impact of controlling the queueing weights on the revenue function (see equation (8)) for the three proposed pricing policies, i.e., usage-based, usage-/throughput-based, and the hybrid pricing policy. From the revenue function the average (steady state) provider revenue in the considered cell can be derived. Recall that the available bandwidth for NRT traffic is variable for different call arrival rates. Therefore, we consider the revenue earned by the provider in one hour per available bandwidth unit, i.e., per available kbit, for NRT traffic. Figure 7 shows the provider revenue for the usage-based pricing policy (i.e., = 0) and different values of the exponent . As discussed in section 4.2, the best revenue improvement will be achieved with priority queueing. From the curves we conclude that the update strategy increases the revenue in one cell successfully for the considered traffic assumptions. Recall that the revenue improvement stems from a shift in bandwidth utilization towards higher priority users (see figure 6) if the population of high priority users is low compared to users of lower priority. Figure 7(b) shows the revenue improvement for different user populations. In the experiment the percentage of high 220 C. LINDEMANN ET AL. Figure 8. Revenue improvement for usage-/throughput-based pricing policy. Figure 9. Revenue improvement for hybrid pricing policy. priority users among the arriving user requests is varied. The remaining users are assumed to be low priority users. Normal priority users are not considered in this experiment (i.e., 0% normal priority users). This figure shows how the adaptive control of the queueing weights works. As expected, for a low percentage of high priority users the corresponding weight is increased. Therefore, QoS for high priority users and the provider revenue is also increased. For more than 50% high priority users the revenue is the same as in the case without adaptive control. No further revenue improvement is allowed because degradation of QoS for low priority users would be unacceptable. Considering a weak relation among the weights as introduced in section 4.2 would decrease the revenue compared to the case without adaptive control for more than 50% high priority users. This might be useful to increase QoS for users of low population independent of their priority class. Figure 8 shows the revenue improvement for the usage-/ throughput-based pricing policy and scaling exponents = 1/4 and = 1/16. In the last experiment we studied the revenue improvement for the hybrid pricing policy (see figure 9). We assume that half of the arriving users start their session in non-paying mode (i.e., k = 2). The curves distinguish between weights w = 1 and w = 2 for the non-paying users. Furthermore, the revenue for the case with and without adaptive control is compared. The curves are derived from simulations with = 0 and = 1/2. From the revenue curves of figures 79 the average monthly revenue can be computed considering a daily/weekly usage-pattern and different splits of call arrival rates of users requesting different services (i.e., voice, RT, NRT with different priorities). Comparing the monthly revenue for the pricing policies used in figures 79 with the monthly revenue for the hybrid pricing policy a provider can determine values such as the monthly free data volume and monthly payment per user. Conclusions We introduced a unified approach based on a mathematical framework for the adaptive performance management of 3G mobile networks. Opposed to previous work [8,13,19,21,25], the improvement of quality of service (QoS) and the optimization of mobile service provider revenue was considered in an integrated way. The unified approach aims at improving both QoS for mobile subscribers and increasing revenue earned by service providers. System parameters controlled by adaptive performance management constitute the portion of bandwidth reserved for handovers, the buffer threshold of the queue for non real-time traffic, and the weights of a weighted fair queueing packet scheduler. Using the UMTS traffic model of [15] and a simulator on the IP level for the UMTS system, we presented performance curves for various QoS measures to illustrate the benefit of the unified approach for adaptive performance management. We introduced update functions that effectively control the packet loss probability and the handover failure probability. Considering usage-based, usage-/throughput-based, and hybrid pricing policies, we showed that the provider revenue in one cell can be significantly increased by the adaptive control of the queueing weights. Throughout the paper, we considered the services and QoS profiles standardized for UMTS. Thus, the proposed approach for adaptive control is tailored to UMTS networks. However , by considering other services and QoS profiles, the basic ideas underlying the unified approach for adaptive performance management can also be applied for the adaptive control of other kinds of multi-service IP networks. References [1] 3GPP, http://www.3gpp.org [2] 3GPP, QoS concept and architecture, Technical Specification TS 23.107 (September 2001). [3] 3GPP, UTRAN overall description, Technical Specification TS 25.401 (September 2001). [4] M. Ajmone Marsan, S. Marano, C. Mastroianni and M. Meo, Performance analysis of cellular mobile communication networks supporting multimedia services, Mobile Networks and Applications 5 (2000) 167 177. [5] K. Aretz, M. Haardt, W. Konhuser and W. Mohr, The future of wireless communications beyond the third generation, Computer Networks 37 (2001) 8392. [6] F. Barcel and J. Jordn, Channel holding time distribution in public cellular telephony, in: Proceedings of the 16th International Teletraffic Congress, Edinburgh, Scotland (1999) pp. 107116. IMPROVING QoS AND PROVIDER REVENUE IN 3G MOBILE NETWORKS 221 [7] CSIM18 The Simulation Engine, http://www.mesquite.com [8] S.K. Das, R. Jayaram, N.K. Kakani and S.K. Sen, A call admission and control scheme for Quality-of-Service provisioning in next generation wireless networks, Wireless Networks 6 (2000) 1730. [9] A. Demers, S. Keshav and S. Shenker, Analysis and simulation of a fair queueing algorithm, in: Proceedings of the International Symposium on Communications Architectures and Protocols (SIGCOMM), Austin, TX (1989) pp. 112. [10] S. Floyd and V. Jacobson, Link-sharing and resource management models for packet networks, IEEE/ACM Transactions on Networking 3 (1995) 365386. [11] X. Geng and A.B. Whinston, Profiting from value-added wireless services , IEEE Computer 34 (August 2001) 8789. [12] A. Gupta, D.O. Stahl and A.B. Whinston, The economics of network management, Communications of the ACM 42 (1999) 5763. [13] A. Gupta, D.O. Stahl and A.B. Whinston, Priority pricing of integrated services networks, in: Internet Economics, eds. L. McKnight and J. Bailey (MIT Press, 1995) pp. 323378. [14] G. Haring, R. Marie and K.S. Trivedi, Loss formulas and their application to optimization for cellular networks, IEEE Transactions on Vehicular Technology 50 (2001) 664673. [15] A. Klemm, C. Lindemann and M. Lohmann, Traffic modeling and characterization for UMTS networks, in: Proceedings of GLOBECOM 2001, San Antonio, TX (November 2001) pp. 17411746. [16] A. Klemm, C. Lindemann and M. Lohmann, Traffic modeling of IP networks using the batch Markovian arrival process, in: Proceedings of Tools 2002, London, Great Britain (April 2002) pp. 92110. [17] J. Kilpi and I. Norros, Call level traffic analysis of a large ISP, in: Proceedings of the 13th ITC Specialist Seminar on Measurement and Modeling of IP Traffic, Monterey, CA (2000) pp. 6.16.9. [18] M. Krunz and A. Makowski, A source model for VBR video traffic based on M/G/ input processes, in: Proceedings of the 17th Conference on Computer Communications (IEEE INFOCOM), San Francisco, CA (1998) pp. 14411449. [19] C. Lindemann, M. Lohmann and A. Thmmler, Adaptive performance management for UMTS networks, Computer Networks 38 (2002) 477 496. [20] R. Ludwig, A. Konrad and A.D. Joseph, Optimizing the end-to-end performance of reliable flows over wireless links, in: Proceedings of the 5th Conference on Mobile Computing and Networking (ACM MobiCom), Seattle, WA (1999) pp. 113119. [21] J.K. MacKie-Mason and H.R. Varian, Pricing the Internet, in: Public Access to the Internet, eds. B. Kahin and J. Keller (MIT Press, 1995) pp. 269314. [22] M. Meyer, TCP performance over GPRS, in: Proceedings of the First Wireless Communications and Networking Conference (IEEE WCNC), New Orleans, MS (1999) pp. 12481252. [23] Mobile Wireless Internet Forum (MWIF), OpenRAN architecture in 3rd generation mobile systems, Technical report MTR-007 (September 2001) http://www.mwif.org [24] J.M. Peha and A. Sutivong, Admission control algorithms for cellular systems, Wireless Networks 7 (2001) 117125. [25] S. Rao and E.R. Petersen, Optimal pricing of priority services, Operations Research 46 (1998) 4656. [26] UMTS-Forum, UMTS/IMT-2000 Spectrum, Report No. 6 (1999). [27] H. Zhang, Service disciplines for guaranteed performance service in packet-switched networks, Proceedings of the IEEE 83 (1995) 1374 1396. [28] Wireless World Research Forum (WWRF), http://www. wireless-world-research.org Christoph Lindemann is an Associate Professor in the Department of Computer Science at the University of Dortmund and leads the Computer Systems and Performance Evaluation group. From 1994 to 1997, he was a Senior Research Scientist at the GMD Institute for Computer Architecture and Software Technology (GMD FIRST) in Berlin. In the summer 1993 and during the academic year 1994/1995, he was a Visiting Scientist at the IBM Almaden Research Center, San Jose, CA. Christoph Lindemann is a Senior Member of the IEEE. He is author of the monograph Performance Modelling with Deterministic and Stochastic Petri Nets (Wiley, 1998). Moreover , he co-authored the survey text Performance Evaluation Origins and Directions (Springer-Verlag, 2000). He served on the program committees of various well-known international conferences. His current research interests include mobile computing, communication networks, Internet search technology , and performance evaluation. E-mail: cl@cs.uni-dortmund.de WWW: http://www4.cs.uni-dortmund.de/ Lindemann/ Marco Lohmann received the degree Diplom-Infor-matiker (M.S. in computer science) with honors from the University of Dortmund in March 2000. Presently, he is a Ph.D. student in the Computer Systems and Performance Evaluation group at the University of Dortmund. He is a student member of the IEEE and the ACM. His research interests include mobile computing, Internet search technology, and stochastic modeling. E-mail: ml@ls4.cs.uni-dortmund.de Axel Thmmler received the degree Diplom-Infor-matiker (M.S. in computer science) from the University of Dortmund in April 1998. Presently, he is a Ph.D. student in the Computer Systems and Performance Evaluation group at the University of Dortmund . His research interests include mobile computing , communication networks, and performance evaluation. E-mail: at@ls4.cs.uni-dortmund.de

QoS;packet loss probability;Quality of Service in mobile systems;provider revenue;performance evaluation of next generation mobile systems;packet scheduler;adaptive performance management;admission control in mobile system;pricing policy;admission control;3G mobile networks;pricing and revenue optimization

24

A WEIGHTED RANKING ALGORITHM FOR FACET-BASED COMPONENT RETRIEVAL SYSTEM

Facet-based component retrieval techniques have been proved to be an effective way for retrieving. These Techniques are widely adopted by component library systems, but they usually simply list out all the retrieval results without any kind of ranking. In our work, we focus on the problem that how to determine the ranks of the components retrieved by user. Factors which can influence the ranking are extracted and identified through the analysis of ER-Diagram of facet-based component library system. In this paper, a mathematical model of weighted ranking algorithm is proposed and the timing of ranks calculation is discussed. Experiment results show that this algorithm greatly improves the efficiency of component retrieval system.

Motivations A high efficiency retrieval system for software component library is important for the reuse of software components. The point of high efficiency is not that the time performance in one matching or retrieving process which can be measured by how many seconds or how many milliseconds elapsed, but that the efficiency to make the component consumers be able to find what they need as soon as possible, even though the former is the basis of the latter. No matter accuracy matching or fuzzy matching, our component retrieval system usually simply lists out all the retrieval results without any kind of ranking, or at least without a systematic ranking. Users have to view the detail information of all the retrieval results one by one to find out which is the best to fit their requirements, or else they have to adjust their query conditions to retrieve again. If there are a large number of components retrieved from the component library, it could be a tough and torturous experience to find a proper component. However, it's a fact that there's a matching degree between the query conditions and retrieval results. The matching degree is just the similarity and relevancy between the query condition and its retrieval results. Only when we rank the retrieval results by the matching degree as the Web search engines can component consumers easily find what they need. They only have to compare the first several retrieval results but not all of them. According to the discussion above, it's clear that a formula to calculate the matching degree and its corresponding ranking algorithm, which can greatly improve the retrieval efficiency for software component library, are needed. In this paper, we propose a weighted ranking algorithm for facet-based component retrieval system. This algorithm has been implemented in a software component library, called DLCL, and greatly improves the efficiency of the retrieval system. Introduction to Retrieval Methods for Component Library 2.1 Existing Retrieval Methods for Component Library Retrieval of software components is a core technique of component library. Today there are lots of retrieval methods for software component library. The main are as follows [1, 2]: (1) Specification matching method; (2) AI Based method; (3) Information science method; (4) Hypertext browsing method. As to the four methods, each has its own features and there's no a general formula to calculate the matching degree. For example, specification matching method uses formal specifications to describe the behavior of software components and relies on theorem proving to determine match and mismatch. AI Based method relies on the use of AI planning techniques to automatically search software components in component library. So we have to use different calculating strategies to calculate the matching degree of each retrieval method. Among the retrieval methods discussed above, information science method is widely used in practice. Information science method usually comprises several different retrieval methods which are attribute-value, enumerated, faceted, and keyword method. Of the four methods, facet-based component retrieval method has been proved to be an effective way for retrieving and has been widely adopted by component library systems. In the following section, we'll discuss the facet-based retrieval method. 505-049 274 2.2 Facet-based Retrieval Method A component classification is a set of {facet, facet term} pairs, also called descriptors [3]. Reusable software components (RSC) are classified by assigning appropriate facet terms for all applicable facets. The objective in classifying the RSC is to make it possible for all users who might use the RSC to retrieve it as a result of their requests. Faceted classification scheme is an effective way for classifying the components and widely adopted by component library systems. Correspondingly, there are several retrieving algorithms for faceted classification scheme. Some systems use the traditional database query techniques in facet-based retrieval. Wang YF proposed a tree matching algorithm in his PH.D dissertation [4]. This algorithm maps the component facets into a facet tree and maps the query conditions into a query tree. The matching algorithm deals with the match of the facet tree and query tree and calculates the matching cost. This algorithm bases on the tree matching theories, such as tree embedding, tree inclusion, and tree containment. These three tree matching methods are becoming more and more elastic in order to improve the retrieving recall while maintaining the precision to a certain extent. Matching cost of the tree matching will be calculated to measure the approximate degree between the facet trees of the components and the query tree. The data structure of a tree is represented by a three-tuple: T= (V, E, root (T)), V represents a limited set of vectors, root (T) represents the root of the tree, E represents the set of edges. Weighted Ranking Algorithm for Facet-based Component Retrieval System There's no a general formula to calculate the matching degree due to the different feature of each retrieval method. Facet-based retrieval method has been widely adopted by existing component library systems, such as REBOOT, Proteus, Asset Library, and JBCL [5]. It has been proved to be an effective method to the retrieval of component library system. And therefore, it makes great sense to propose a component ranking algorithm for facet-based retrieval system. 3.1 ER-Diagram of Software Component Library The extraction and identification of the influential factors which are used to calculate the matching degree is the first step to establish a mathematical model. To Analyze the ER-Diagram of software component library is an effective way to extract the factors. An ER-diagram of facet-based component library was given below: Producer Provide Consume r Component Reuse Facet Term Describe Feedback Summary Feedback 1 n 1 1 1 1 n n n n m m Relate n m Include Describe Fig. 1. ER-Diagram of Component Library Entities list: Component: component is the basic and primary entity in component library. Besides the attributes, there are facet-term pairs and information summary to describe a component. User Feedback: an opinion, a comment or a score provided by users after they have used a component. Component Summary: an information summary to describe a component which enables users to know well the component quickly. Facet: facet and its terms are used to classify and represent the components. 3.2 Factors of Weighted Ranking Algorithm As to a facet-based component library system, facet is the most important method to classify and represent the components. Correspondingly, facet-based retrieval methods, such as facet tree matching method, are important for the component retrieval system. The matching degree between facet tree and query tree is of much importance for ranking. However, matching degree of facet is not the only factor which is able to influence the ranking. Retrieval system of component library usually has two search modes: simple query and complex query. Simple query just simply uses the traditional database query method to match the Attribute-Valued pairs. In contrast, complex query is a much more effective way which combines several query methods together to match different kinds of component information. And therefore, 275 we should take other factors into account besides the facet for ranking the retrieval results. According to the analysis of ER-Diagram above, we can extract some other factors which are able to influence the ranking of component retrieval results while using the complex query. Attributes of component, such as component name, can be used to match the keywords in the query conditions. The matching degree of Attribute-Valued pairs should be an influential factor for ranking. Summary of component can also be used to match the query conditions. Query conditions usually consist of several keywords. The density, prominence, and position of keywords within the component summary will influence the ranking of components. The keyword density is just the number of occurrences of the keywords within the component summary divided by the total number of words. Keyword prominence is related to the location of keywords in the summary. For example, keywords placed at the beginning of the summary maybe carry more weight than those towards the end of it. User feedback of a component is very useful for other users who want to use it to evaluate the quality and other features of the component. They can acquire much more objective description and useful information about the component besides the component attributes and summary. How many times the component information has been visited and how many times the component has been downloaded for reusing should also be taken into account as the factors to calculate the matching degree for ranking. They reflect the popularity and reusability of the component from another aspect. 3.3 Mathematical Model Retrieval results consist of a collection of components matching the query conditions: Definition 1: Components (C 1 , C 2 , ......, C i , ......, Cn); (n N, n1) It makes no sense to discuss the circumstance of empty retrieval results, since that we are going to discuss the ranking of component lists. Accordingly, each component has a rank value: Definition 2: Ranks (R 1 , R 2 , ......, R i , ......, Rn); (n N, n1) The query condition consists of a collection of keywords: Definition 3: Keywords (K 1 , K 2 , ......, K i , ......, K n0 ); (n 0 N, n 0 1) Each component is described by a set of Attribute-Valued pairs: Definition 4: Attributes (A 1 , A 2 , ......, A i , ......, A n1 ); (n 1 N, n 1 1) Besides with Attribute-Valued pairs, components are also classified and represented by a set of facets and their terms: Definition 5: Facets (F 1 , F 2 , ......, F i , ......, F n2 ); (n 2 N, n 2 1) Summary of component information differs from Attribute-Valued pairs. It provides a comprehensive description of a component in context. Definition 6: Summary (S); User feedback includes all the comments and feedback to a specific component: Definition 7: User Feedback (U 1 , U 2 , ......, U i , ......, U n3 ); (n 3 N, n 3 1) User feedback must be analyzed and evaluated to a relative number. We use E to represent the Evaluation number of user feedback. Definition 8: E = Evaluate (User Feedback). Definition 9: Visited times of a component: Visited times (V); Definition 10: Downloaded times of a component: Downloaded times (D); We have listed out all the influential factors above, which constitute a six-tuple: Factors (A, F, S, U, V, D); Their influential weights differ from each other according to their feature and importance: Definition 11: Weights (W A , W F , W S , W U , W V , W D ); (0W A , W F , W S , W U , W V , W D 1, W A +W F +W S +W U +W V +W D = 1) W A represents the weight of Attributes; W F represents the weight of Facets; W S , W U , W V , and W D also represent the weight of corresponding factor discussed above. There are several functions for calculating the matching degree of some factors. The core calculating formula of each function relies on its corresponding matching algorithm. Functions Formulas Summary F S (Keywords, Summary) = no i S Ki match 1 ) , ( Facets F F (Keywords, Facets) = = no i n j Fj Ki match 1 2 1 ) , ( Attributes F A (Keywords, Attributes) = = no i n j Aj Ki match 1 1 1 ) , ( Match function of component summary uses the content-based similarity measurement algorithm of the search engine techniques. A Best-First algorithm was proposed by Cho [6]. This algorithm uses a vector space model to calculate the similarity between the keywords and the content. Its formula is given as following: = p k q k kq kp p q k kp kq W W W W p q sim 2 2 | ) , ( The variable q represents the collection of keywords, p represents the content, and W kp represents the importance of k to a specific topic. In our mathematical model, variable q represents the query condition, and the variable p represents the summary of the component information. 276 Facet-based retrieving method usually adopts the facet tree matching. And therefore, its match function calculates the matching degree between the facet tree of component and the query tree. A formula to calculate the matching cost of tree containment matching was given by Xu [7]: Q=(V,E,root(Q)), D=(W,F,root(D)) are two unordered label tree, TCostM(Q, D) represents the tree containment matching cost from tree Q to tree D. ( ) { } D Q f f = : | min D) TCostM(Q, ( ) ( ) ( ) ( + + = ) ( ) ( ) ( ) ( ) ( ) ( )) ( ( ) ( f Range f spectrum w f domain V v f domain v v label v label v f label v label f ) If f is a tree containment matching from tree Q to tree D, and (f) = TCostM(Q, D), then f is the tree containment matching which obtains the minimum matching cost from tree Q to tree D. This definition could be also applied to the containment matching between tree and forest or between forest and forest. As to the match function of component attributes, we just use the traditional database query methods to deal with it. E, V, and D are three ranking factors without any relation to query keywords. Even though they are numbers, we could not use them directly for ranking. Functions should be provided to transform them. Functions Evaluation of Feedback F E (E) Visited Times F V (V) Downloaded Times F D (D) According to the discussion above, we finally draw out a very simple formula to calculate the rank for each component: Rank = F A W A + F F W F + F S W S + F E W E + F V W V + F D W D We can use matrix operation to represent the calculation of rank value for each component in the retrieval results. There are n components and 6 influential factors. F 6n W 16 = R 1n : = n i D V E S F A Dn Vn En Sn Fn An Di Vi Ei Si Fi Ai D V E S F A D V E S F A R R R R W W W W W W F F F F F F F F F F F F F F F F F F F F F F F F M M M M M M M M M M M M M M 2 1 2 2 2 2 2 2 1 1 1 1 1 1 We specify the weights with experiential values at the very beginning, and then use the data mining technology to analyze the user logs to dynamically and iteratively adjust those values. 3.4 Timing of Ranks Calculation It's time for us to determine when to deal with the calculation of ranks, since that we have designed how to deal with it. It's just the matter of the timing of rank calculation. There are two probable times for calculation: calculating after the retrieving process has been finished; calculating during the process of retrieving. Both of the probable times have their own advantages and disadvantages. If we calculate the rank after the retrieving process has been finished, we can deal with the retrieving and ranking separately. It will be much easier for us to design and maintain the system, since the retrieving and ranking process are independent. However, it costs much more time and space to calculate the rank. It needs a lot of memory space to store a large number of retrieval results temporally before they are ranked, and costs much more time to manage the transmission of data between storage devices and CPU. On the contrary, if we calculate the rank during the process of retrieving, we have to combine the retrieving with the ranking process completely or partly. Undoubtedly, it will be hard for us to implement and maintain the system, but it can greatly improve the time and the space performances. According to the discussion above, we can choose a proper time to calculate the ranks. Which solution we should choose depends on the requirements of the system. The solution which calculates the ranks during the retrieving process should be adopted if the time and the space performances are rigorously required. Implementation Our component library system, named DLCL, is implemented with J2EE platform. The mathematical model and its algorithm we discussed above have been implemented in this system with Java. Java is an object-oriented language. Its implementation consists of several core interfaces and classes. The core interfaces, classes and the relationship between them are demonstrated in the UML class diagram: 277 Fig. 2. Class Diagram of Ranking Module We calculate the ranks during the retrieving process to improve the time and space performances, and therefore, we have to combine the ranking module partly with the retrieving module. In order to lower the coupled degree between these two modules, a callback mechanism was adopted. We define an interface Rank, which consists of only one method to calculate the ranks. RankComponentImpl is a class implementing the interface Rank to calculate the ranks of those components retrieved by users. The method of its concrete object can be executed by the searcher during the retrieving process. Class Component encapsulates those methods which provide the component information. ComponentMatchingDegree is a class providing those methods for calculating the matching degree between the query keywords and component. Each influential factor has its own strategy for calculation. There's also a class Weight, which provides the methods to get the influential weight of each factor. Experiment and its Results In order to verify the efficiency of the component retrieving system which adoptes our weighted ranking algorithm, we design an experiment and carry out this experiment in our component library system, named DLCL. There are more than 1000 components in this system. Retrieving system of DLCL splits the retrieval results into several pages if there are too many components retrieved, and lists out 10 components per page. The experiment separates the users into two groups, group 1 and group 2. Each group consists of 10 persons. All the users know about the knowledge of component reuse to a certain extent. Both groups use the facet-based component retrieving method to retrieve the components. Retrieval results of group 1 are listed out without any ranking, however, those of group 2 are ranked by our weighted ranking algorithm. There are several aspects to measure the efficiency of each group: how many pages they turned; how many times they had to adjust the query condition; and the most important, how many time was elapsed during the whole retrieving process. The experimental results are given in the following table: Group 1 Group 2 Average Turned Pages (pages) 2.7 1.4 Average Adjusted Times (times) 2.3 1.1 Average Time elapsed (minutes) 26.6 9.5 The experimental results obviously shows that the efficiency of group 2 is greatly higher than group 1. By applying the weighted ranking algorithm into the retrieving system of DLCL, users needn't turn too many pages to view and compare the component information or to adjust the query condition to improve the query precision. Only to view the first page of retrieval results will be enough most of the time. And therefore, it greatly saves the time and retrieval costs. Related Works The idea of Component Rank comes from computing fair impact factors of published papers [8]. Google is a web search engine. Its method can be considered as an HTML extension of the method proposed for counting impact of publications, called influence weight in [8]. Google computes the ranks (called PageRanks) for HTML documents in the Internet [9, 10]. In reference [11], the authors present the Component Rank model for ranking software components, and show a system for computing Component Rank. In this model, a collection of software components is represented as a weighted directed graph whose nodes correspond to the components and edges correspond to the usage relations. Similar components are clustered into one node so that effect of simply duplicated nodes is removed. The nodes in the graph are ranked by their weights which are defined as the elements of the eigenvector of an adjacent matrix for the directed graph. A major distinction of Component Rank model in [11] from PageRank and the influence weight in [9, 10] is that Component Rank model explores similarity between components before the weight computation. In this paper, we also propose a weighted ranking algorithm for component retrieval system. This weighted ranking algorithm uses different calculating strategies according to the feature of facet-based retrieval methods. While in [11], the authors employed only statical use relations. Conclusion In this paper, a mathematical model of weighted ranking algorithm is proposed and the timing of ranks calculation is discussed. We have applied this ranking algorithm into our component library system, named DLCL. The experiment we carried out shows that this algorithm greatly improves the efficiency of component retrieving system, saving the time and retrieval costs for component reusing. Acknowledgement 278 This research is partially supported by the National High Technology Development 863 Program under Grant No. 2004AA116010. References [1] Frakes WB, Pole TP, An empirical study of representation methods for reusable software components, IEEE Transactions on Software Engineering, 1994, l20(8), pp617-630 [2] H. Mili, R. Rada, W. Wang, K. Strickland, C. Boldyreff, L. Olsen, J. Witt, J. Heger, W. Scherr, and P. Elzer, Practitioner and SoftClass: A Comparative Study of Two Software Reuse Research Projects, J. Systems and Software, 1994, 27(5) [3] NEC Software Engineering Laboratory, NATO Standard for Management of a Reusable Software Component Library, NATO Communications and Information Systems Agency, 1991 [4] Wang YF. Research on retrieving reusable components classified in faceted scheme [Ph.D. Thesis]. Shanghai: Fudan University, 2002. [5] Chang JC, et al. Representation and Retrieval of Reusable Software Components [J].Computer Science, 1999, 26(5):41-48. [6] Cho J, Garcia-Molina H, Page L. Efficient Crawling Through URL Ordering [J]. Computer Networks, 1998, 30(1~7):161-172 [7] Xu RZ, et al. Research on Matching Algorithm for XML-Based Software Component Query, Journal of Software, 2003, 14(7):1195-1202. [8] G. Pinski and F. Narin. "Citation Influence for Journal Aggregates of Scientific Publications: Theory, with Application to the Literature of Physics". Information Processing and Management, 12(5):297.312, 1976. [9] L. Page, S. Brin, R. Motwani, and T. Winograd. "The PageRank Citation Ranking: Bringing Order to the Web". Technical Report of Stanford Digital Library Technologies Project, 1998. "http://www-db .stanford.edu/.backrub/ pageranksub.ps". [10] J. Kleinberg. "Authoritative Sources in a Hyperlinked Environment". Journal of the ACM, 46(5):604.632, 1999. [11] Katsuro Inoue, Reishi Yokomori, Hikaru Fujiwara, Tetsuo Yamamoto, Makoto Matsushita, Shinji Kusumoto: Component Rank: Relative Significance Rank for Software Component Search. ICSE 2003: 14-24 279

retrieval system;facet;component rank;component retrieval;and component library;ranking algorithm;Weighted ranking algorithm;matching degree;facet-based component retrieval;component library

25

Accelerated Focused Crawling through Online Relevance Feedback

The organization of HTML into a tag tree structure, which is rendered by browsers as roughly rectangular regions with embedded text and HREF links, greatly helps surfers locate and click on links that best satisfy their information need. Can an automatic program emulate this human behavior and thereby learn to predict the relevance of an unseen HREF target page w.r.t. an information need, based on information limited to the HREF source page? Such a capability would be of great interest in focused crawling and resource discovery, because it can fine-tune the priority of unvisited URLs in the crawl frontier, and reduce the number of irrelevant pages which are fetched and discarded. We show that there is indeed a great deal of usable information on a HREF source page about the relevance of the target page. This information, encoded suitably, can be exploited by a supervised apprentice which takes online lessons from a traditional focused crawler by observing a carefully designed set of features and events associated with the crawler. Once the apprentice gets a sufficient number of examples, the crawler starts consulting it to better prioritize URLs in the crawl frontier. Experiments on a dozen topics using a 482-topic taxonomy from the Open Directory (Dmoz) show that online relevance feedback can reduce false positives by 30% to 90%.

Introduction Keyword search and clicking on links are the dominant modes of accessing hypertext on the Web. Support for keyword search through crawlers and search engines is very mature, but the surfing paradigm is not modeled or assisted (Note: The HTML version of this paper is best viewed using Microsoft Internet Explorer. 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.) Contact author, email soumen@cse.iitb.ac.in Copyright is held by the author/owner(s). WWW2002, May 711, 2002, Honolulu, Hawaii, USA. ACM 1-58113-449-5/02/0005 Baseline learner Dmoz topic taxonomy Class models consisting of term stats Frontier URLS priority queue Crawler Pick best Newly fetched page u Submit page for classification If Pr(c*|u) is large enough then enqueue all outlinks v of u with priority Pr(c*|u) Crawl database Seed URLs Figure 1: A basic focused crawler controlled by one topic classifier/learner. as well. Support for surfing is limited to the basic interface provided by Web browsers, except for a few notable research prototypes. While surfing, the user typically has a topic-specific information need, and explores out from a few known relevant starting points in the Web graph (which may be query responses) to seek new pages relevant to the chosen topic/s. While deciding for or against clicking on a specific link (u, v), humans use a variety of clues on the source page u to estimate the worth of the (unseen) target page v, including the tag tree structure of u, text embedded in various regions of that tag tree, and whether the link is relative or remote. "Every click on a link is a leap of faith" <A href="25.html#12">[19], but humans are very good at discriminating between links based on these clues. Making an educated guess about the worth of clicking on a link (u, v) without knowledge of the target v is central to the surfing activity. Automatic programs which can learn this capability would be valuable for a number of applications which can be broadly characterized as personalized, topic-specific information foragers. Large-scale, topic-specific information gatherers are called focused crawlers <A href="25.html#12">[1, 9, 14, 28, 30]. In contrast to giant, all-purpose crawlers which must process large portions of the Web in a centralized manner, a distributed federation of focused crawlers can cover specialized topics in more depth and keep the crawl more fresh, because there is less to cover for each crawler. In its simplest form, a focused crawler consists of a supervised topic classifier (also called a `learner') controlling the priority of the unvisited frontier of a crawler (see Figure <A href="25.html#1">1). The classifier is trained a priori on document samples embedded in a topic taxonomy such as Yahoo! or Dmoz. It thereby learns to label new documents as belonging to topics in the given taxonomy <A href="25.html#12">[2, 5, 21]. The goal of the focused crawler is to start from nodes relevant to a focus topic c in the Web graph and explore links to selectively collect pages about c , while avoiding fetching pages not about c . Suppose the crawler has collected a page u and 148 encountered in u an unvisited link to v. A simple crawler (which we call the baseline) will use the relevance of u to topic c (which, in a Bayesian setting, we can denote Pr(c |u)) as the estimated relevance of the unvisited page v. This reflects our belief that pages across a hyperlink are more similar than two randomly chosen pages on the Web, or, in other words, topics appear clustered in the Web graph <A href="25.html#12">[11, 23]. Node v will be added to the crawler's priority queue with priority Pr(c |u). This is essentially a "best-first" crawling strategy. When v comes to the head of the queue and is actually fetched, we can verify if the gamble paid off, by evaluating Pr(c |v). The fraction of relevant pages collected is called the harvest rate. If V is the set of nodes collected, the harvest rate is defined as (1/ |V |) v V Pr(c |v). Alternatively, we can measure the loss rate, which is one minus the harvest rate, i.e., the (expected) fraction of fetched pages that must be thrown away. Since the effort on relevant pages is well-spent, reduction in loss rate is the primary goal and the most appropriate figure of merit. For focused crawling applications to succeed, the "leap of faith" from u to v must pay off frequently. In other words, if Pr(c |v) is often much less than the preliminary estimate Pr(c |u), a great deal of network traffic and CPU cycles are being wasted eliminating bad pages. Experience with random walks on the Web show that as one walks away from a fixed page u 0 relevant to topic c 0 , the relevance of successive nodes u 1 , u 2 , . . . to c 0 drops dramatically within a few hops <A href="25.html#12">[9, 23]. This means that only a fraction of outlinks from a page is typically worth following. The average out-degree of the Web graph is about 7 <A href="25.html#12">[29]. Therefore, a large number of page fetches may result in disappointment, especially if we wish to push the utility of focused crawling to topic communities which are not very densely linked. Even w.r.t. topics that are not very narrow, the number of distracting outlinks emerging from even fairly relevant pages has grown substantially since the early days of Web authoring <A href="25.html#12">[4]. Template-based authoring, dynamic page generation from semi-structured databases, ad links, navigation panels, and Web rings contribute many irrelevant links which reduce the harvest rate of focused crawlers. Topic-based link discrimination will also reduce these problems. 1.1 Our contribution: Leaping with more faith In this paper we address the following questions: How much information about the topic of the HREF target is available and/or latent in the HREF source page, its tag-tree structure, and its text? Can these sources be exploited for accelerating a focused crawler? Our basic idea is to use two classifiers. Earlier, the regular baseline classifier was used to assign priorities to unvisited frontier nodes. This no longer remains its function. The role of assigning priorities to unvisited URLs in the crawl frontier is now assigned to a new learner called the apprentice, and the priority of v is specific to the features associated with the (u, v) link which leads to it <A href="25.html#2">1 <A href="25.html#2">. The features used by the apprentice are derived from the Document Object Model or 1 If many u's link to a single v, it is easiest to freeze the priority of v when the first-visited u linking to v is assessed, but combinations of scores are also possible. Baseline learner (Critic) Dmoz topic taxonomy Class models consisting of term stats Frontier URLS priority queue Crawler Pick best Newly fetched page u Submit page for classification If Pr(c*|u) is large enough... An instance (u,v) for the apprentice u v Pr(c*|v) Pr(c|u) for all classes c Crawl database Apprentice learner Class models + Online training ... submit (u,v) to the apprentice Apprentice assigns more accurate priority to node v Figure 2: The apprentice is continually presented with training cases (u, v) with suitable features. The apprentice is interposed where new outlinks (u, v) are registered with the priority queue, and helps assign the unvisited node v a better estimate of its relevance. DOM (<A href="http://www.w3.org/DOM/">http://www.w3.org/DOM/) of u. Meanwhile, the role of the baseline classifier becomes one of generating training instances for the apprentice, as shown in Figure <A href="25.html#2">2. We may therefore regard the baseline learner as a critic or a trainer, which provides feedback to the apprentice so that it can improve "on the job." The critic-apprentice paradigm is related to reinforcement learning and AI programs that learn to play games <A href="25.html#12">[26, 1.2]. We argue that this division of labor is natural and effective. The baseline learner can be regarded as a user specification for what kind of content is desired. Although we limit ourselves to a generative statistical model for this specification, this can be an arbitrary black-box predicate. For rich and meaningful distinction between Web communities and topics, the baseline learner needs to be fairly sophisticated, perhaps leveraging off human annotations on the Web (such as topic directories). In contrast, the apprentice specializes in how to locate pages to satisfy the baseline learner. Its feature space is more limited, so that it can train fast and adapt nimbly to changing fortunes at following links during a crawl. In Mitchell's words <A href="25.html#12">[27], the baseline learner recognizes "global regularity" while the apprentice helps the crawler adapt to "local regularity." This marked asymmetry between the classifiers distinguishes our approach from Blum and Mitchell's co-training technique <A href="25.html#12">[3], in which two learners train each other by selecting unlabeled instances. Using a dozen topics from a topic taxonomy derived from the Open Directory, we compare our enhanced crawler with the baseline crawler. The number of pages that are thrown away (because they are irrelevant), called the loss rate, is cut down by 3090%. We also demonstrate that the fine-grained tag-tree model, together with our synthesis and encoding of features for the apprentice, are superior to simpler alternatives. 1.2 Related work Optimizing the priority of unvisited URLs on the crawl frontier for specific crawling goals is not new. FishSearch by De Bra et al. <A href="25.html#12">[12, 13] and SharkSearch by Hersovici et al. <A href="25.html#12">[16] were some of the earliest systems for localized searches in the Web graph for pages with specified keywords. 149 In another early paper, Cho et al. <A href="25.html#12">[10] experimented with a variety of strategies for prioritizing how to fetch unvisited URLs. They used the anchor text as a bag of words to guide link expansion to crawl for pages matching a specified keyword query, which led to some extent of differentiation among out-links, but no trainer-apprentice combination was involved. No notion of supervised topics had emerged at that point, and simple properties like the in-degree or the presence of specified keywords in pages were used to guide the crawler. Topical locality on the Web has been studied for a few years. Davison made early measurements on a 100000-node Web subgraph <A href="25.html#12">[11] collected by the DiscoWeb system. Using the standard notion of vector space TFIDF similarity <A href="25.html#12">[31], he found that the endpoints of a hyperlink are much more similar to each other than two random pages, and that HREFs close together on a page link to documents which are more similar than targets which are far apart. Menczer has made similar observations <A href="25.html#12">[23]. The HyperClass hypertext classifier also uses such locality patterns for better semi-supervised learning of topics <A href="25.html#12">[7], as does IBM's Automatic Resource Compilation (ARC) and Clever topic distillation systems <A href="25.html#12">[6, 8]. Two important advances have been made beyond the baseline best-first focused crawler: the use of context graphs by Diligenti et al. <A href="25.html#12">[14] and the use of reinforcement learning by Rennie and McCallum <A href="25.html#12">[30]. Both techniques trained a learner with features collected from paths leading up to relevant nodes rather than relevant nodes alone. Such paths may be collected by following backlinks. Diligenti et al. used a classifier (learner) that regressed from the text of u to the estimated link distance from u to some relevant page w, rather than the relevance of u or an outlink (u, v), as was the case with the baseline crawler. This lets their system continue expanding u even if the reward for following a link is not immediate, but several links away. However, they do favor links whose payoffs are closest. Our work is specifically useful in conjunction with the use of context graphs: when the context graph learner predicts that a goal is several links away, it is crucial to offer additional guidance to the crawler based on local structure in pages, because the fan-out at that radius could be enormous. Rennie and McCallum <A href="25.html#12">[30] also collected paths leading to relevant nodes, but they trained a slightly different classifier, for which: An instance was a single HREF link like (u, v). The features were terms from the title and headers (&lt;h1&gt;...&lt;/h1&gt; etc.) of u, together with the text in and `near' the anchor (u, v). Directories and pathnames were also used. (We do not know the precise definition of `near', or how these features were encoded and combined.) The prediction was a discretized estimate of the number of relevant nodes reachable by following (u, v), where the reward from goals distant from v was geometrically discounted by some factor &lt; 1/2 per hop. Rennie and McCallum obtained impressive harvests of research papers from four Computer Science department sites, and of pages about officers and directors from 26 company Websites. Lexical proximity and contextual features have been used extensively in natural language processing for disambiguating word sense <A href="25.html#12">[15]. Compared to plain text, DOM trees and hyperlinks give us a richer set of potential features. Aggarwal et al. have proposed an "intelligent crawling" framework <A href="25.html#12">[1] in which only one classifier is used, but similar to our system, that classifier trains as the crawl progresses. They do not use our apprentice-critic approach, and do not exploit features derived from tag-trees to guide the crawler. The "intelligent agents" literature has brought forth several systems for resource discovery and assistance to browsing <A href="25.html#12">[19]. They range between client- and site-level tools. Letizia <A href="25.html#12">[18], Powerscout, and WebWatcher <A href="25.html#12">[17] are such systems. Menczer and Belew proposed InfoSpiders <A href="25.html#12">[24], a collection of autonomous goal-driven crawlers without global control or state, in the style of genetic algorithms. A recent extensive study <A href="25.html#12">[25] comparing several topic-driven crawlers including the best-first crawler and InfoSpiders found the best-first approach to show the highest harvest rate (which our new system outperforms). In all the systems mentioned above, improving the chances of a successful "leap of faith" will clearly reduce the overheads of fetching, filtering, and analyzing pages. Furthermore, whereas we use an automatic first-generation focused crawler to generate the input to train the apprentice, one can envisage specially instrumented browsers being used to monitor users as they seek out information. We distinguish our work from prior art in the following important ways: Two classifiers: We use two classifiers. The first one is used to obtain `enriched' training data for the second one. (A breadth-first or random crawl would have a negligible fraction of positive instances.) The apprentice is a simplified reinforcement learner. It improves the harvest rate, thereby `enriching' the data collected and labeled by the first learner in turn. No manual path collection: Our two-classifier framework essentially eliminates the manual effort needed to create reinforcement paths or context graphs. The input needed to start off a focused crawl is just a pre-trained topic taxonomy (easily available from the Web) and a few focus topics. Online training: Our apprentice trains continually, acquiring ever-larger vocabularies and improving its accuracy as the crawl progresses. This property holds also for the "intelligent crawler" proposed by Aggarwal et al., but they have a single learner, whose drift is controlled by precise relevance predicates provided by the user. No manual feature tuning: Rather than tune ad-hoc notions of proximity between text and hyperlinks, we encode the features of link (u, v) using the DOM-tree of u, and automatically learn a robust definition of `nearness' of a textual feature to (u, v). In contrast, Aggarwal et al use many tuned constants combining the strength of text-and link-based predictors, and Rennie et al. use domain knowledge to select the paths to goal nodes and the word bags that are submitted to their learner. 150 Methodology and algorithms We first review the baseline focused crawler and then describe how the enhanced crawler is set up using the apprentice-critic mechanism. 2.1 The baseline focused crawler The baseline focused crawler has been described in detail elsewhere <A href="25.html#12">[9, 14], and has been sketched in Figure <A href="25.html#1">1. Here we review its design and operation briefly. There are two inputs to the baseline crawler. A topic taxonomy or hierarchy with example URLs for each topic. One or a few topics in the taxonomy marked as the topic(s) of focus. Although we will generally use the terms `taxonomy' and `hierarchy', a topic tree is not essential; all we really need is a two-way classifier where the classes have the connotations of being `relevant' or `irrelevant' to the topic(s) of focus. A topic hierarchy is proposed purely to reduce the tedium of defining new focused crawls. With a two-class classifier, the crawl administrator has to seed positive and negative examples for each crawl. Using a taxonomy, she composes the `irrelevant' class as the union of all classes that are not relevant. Thanks to extensive hierarchies like Dmoz in the public domain, it should be quite easy to seed topic-based crawls in this way. The baseline crawler maintains a priority queue on the estimated relevance of nodes v which have not been visited, and keeps removing the highest priority node and visiting it, expanding its outlinks and checking them into the priority queue with the relevance score of v in turn. Despite its extreme simplicity, the best-first crawler has been found to have very high harvest rates in extensive evaluations <A href="25.html#12">[25]. Why do we need negative examples and negative classes at all? Instead of using class probabilities, we could maintain a priority queue on, say, the TFIDF cosine similarity between u and the centroid of the seed pages (acting as an estimate for the corresponding similarity between v and the centroid, until v has been fetched). Experience has shown <A href="25.html#12">[32] that characterizing a negative class is quite important to prevent the centroid of the crawled documents from drifting away indefinitely from the desired topic profile. In this paper, the baseline crawler also has the implicit job of gathering instances of successful and unsuccessful "leaps of faith" to submit to the apprentice, discussed next. 2.2 The basic structure of the apprentice learner In estimating the worth of traversing the HREF (u, v), we will limit our attention to u alone. The page u is modeled as a tag tree (also called the Document Object Model or DOM). In principle, any feature from u, even font color and site membership may be perfect predictors of the relevance of v. The total number of potentially predictive features will be quite staggering, so we need to simplify the feature space and massage it into a form suited to conventional learning algorithms. Also note that we specifically study properties of u and not larger contexts such as paths leading to u, meaning that our method may become even more robust and useful in conjunction with context graphs or reinforcement along paths. Initially, the apprentice has no training data, and passes judgment on (u, v) links according to some fixed prior obtained from a baseline crawl run ahead of time (e.g., see the statistics in <A href="25.html#7">3.3). Ideally, we would like to train the apprentice continuously, but to reduce overheads, we declare a batch size between a few hundred and a few thousand pages. After every batch of pages is collected, we check if any page u fetched before the current batch links to some page v in the batch. If such a (u, v) is found, we extract suitable features for (u, v) as described later in this section, and add (u, v), Pr(c |v) as another instance of the training data for the apprentice. Many apprentices, certainly the simple naive Bayes and linear perceptrons that we have studied, need not start learning from scratch; they can accept the additional training data with a small additional computational cost. 2.2.1 Preprocessing the DOM tree First, we parse u and form the DOM tree for u. Sadly, much of the HTML available on the Web violates any HTML standards that permit context-free parsing, but a variety of repair heuristics (see, e.g., HTML Tidy, available at <A href="http://www.w3.org/People/Raggett/tidy/">http://www.w3.org/People/Raggett/tidy/) let us generate reasonable DOM trees from bad HTML. a HREF TEXT font TEXT li li li ul li TEXT TEXT em TEXT tt TEXT TEXT @0 @0 @1 @2 @3 @-1 @-2 Figure 3: Numbering of DOM leaves used to derive offset attributes for textual tokens. `@' means "is at offset". Second, we number all leaf nodes consecutively from left to right. For uniformity, we assign numbers even to those DOM leaves which have no text associated with them. The specific &lt;a href...&gt; which links to v is actually an internal node a v , which is the root of the subtree containing the anchor text of the link (u, v). There may be other element tags such as &lt;em&gt; or &lt;b&gt; in the subtree rooted at a v . Let the leaf or leaves in this subtree be numbered (a v ) through r(a v ) (a v ). We regard the textual tokens available from any of these leaves as being at DOM offset zero w.r.t. the (u, v) link. Text tokens from a leaf numbered , to the left of (a v ), are at negative DOM offset - (a v ). Likewise, text from a leaf numbered to the right of r(a v ) are at positive DOM offset - r(a v ). See Figure <A href="25.html#4">3 for an example. 2.2.2 Features derived from the DOM and text tokens Many related projects mentioned in <A href="25.html#2">1.2 use a linear notion of proximity between a HREF and textual tokens. In the ARC system, there is a crude cut-off distance measured 151 in bytes to the left and right of the anchor. In the Clever system, distance is measured in tokens, and the importance attached to a token decays with the distance. In reinforcement learning and intelligent predicate-based crawling, the exact specification of neighborhood text is not known to us. In all cases, some ad-hoc tuning appears to be involved. We claim (and show in <A href="25.html#7">3.4) that the relation between the relevance of the target v of a HREF (u, v) and the proximity of terms to (u, v) can be learnt automatically. The results are better than ad-hoc tuning of cut-off distances, provided the DOM offset information is encoded as features suitable for the apprentice. One obvious idea is to extend the Clever model: a page is a linear sequence of tokens. If a token t is distant x from the HREF (u, v) in question, we encode it as a feature t, x . Such features will not be useful because there are too many possible values of x, making the t, x space too sparse to learn well. (How many HREFS will be exactly five tokens from the term `basketball' ?) Clearly, we need to bucket x into a small number of ranges. Rather than tune arbitrary bucket boundaries by hand, we argue that DOM offsets are a natural bucketing scheme provided by the page author. Using the node numbering scheme described above, each token t on page u can be annotated w.r.t. the link (u, v) (for simplicity assume there is only one such link) as t, d , where d is the DOM offset calculated above. This is the main set of features used by the apprentice. We shall see that the apprentice can learn to limit |d| to less than d max = 5 in most cases, which reduces its vocabulary and saves time. A variety of other feature encodings suggest themselves. We are experimenting with some in ongoing work ( <A href="25.html#11">4), but decided against some others. For example, we do not expect gains from encoding specific HTML tag names owing to the diversity of authoring styles. Authors use &lt;div&gt;, &lt;span&gt;, &lt;layer&gt; and nested tables for layout control in non-standard ways; these are best deflated to a nameless DOM node representation. Similar comments apply to HREF collections embedded in &lt;ul&gt;, &lt;ol&gt;, &lt;td&gt; and &lt;dd&gt;. Font and lower/upper case information is useful for search engines, but would make features even sparser for the apprentice. Our representation also flattens two-dimensional tables to their "row-major" representation. The features we ignore are definitely crucial for other applications, such as information extraction. We did not see any cases where this sloppiness led to a large loss rate. We would be surprised to see tables where relevant links occurred in the third column and irrelevant links in the fifth, or pages where they are rendered systematically in different fonts and colors, but are not otherwise demarcated by the DOM structure. 2.2.3 Non-textual features Limiting d may lead us to miss features of u that may be useful at the whole-page level. One approach would be to use "d = " for all d larger in magnitude than some threshold. But this would make our apprentice as bulky and slow to train as the baseline learner. Instead, we use the baseline learner to abstract u for the apprentice. Specifically, we use a naive Bayes baseline learner to classify u, and use the vector of class probabilities returned as features for the apprentice. These features can help the apprentice discover patterns such as "Pages about /Recreation/Boating/Sailing often link to pages about /Sports/Canoe_and_Kayaking." This also covers for the baseline classifier confusing between classes with related vocabulary, achieving an effect similar to context graphs. Another kind of feature can be derived from co-citation. If v 1 has been fetched and found to be relevant and HREFS (u, v 1 ) and (u, v 2 ) are close to each other, v 2 is likely to be relevant. Just like textual tokens were encoded as t, d pairs, we can represent co-citation features as , d , where is a suitable representation of relevance. Many other features can be derived from the DOM tree and added to our feature pool. We discuss some options in <A href="25.html#11">4. In our experience so far, we have found the t, d features to be most useful. For simplicity, we will limit our subsequent discussion to t, d features only. 2.3 Choices of learning algorithms for the apprentice Our feature set is thus an interesting mix of categorical, ordered and continuous features: Term tokens t, d have a categorical component t and a discrete ordered component d (which we may like to smooth somewhat). Term counts are discrete but can be normalized to constant document length, resulting in continuous attribute values. Class names are discrete and may be regarded as synthetic terms. The probabilities are continuous. The output we desire is an estimate of Pr(c |v), given all the observations about u and the neighborhood of (u, v) that we have discussed. Neural networks are a natural choice to accommodate these requirements. We first experimented with a simple linear perceptron, training it with the delta rule (gradient descent) <A href="25.html#12">[26]. Even for a linear perceptron, convergence was surprisingly slow, and after convergence, the error rate was rather high. It is likely that local optima were responsible, because stability was generally poor, and got worse if we tried to add hidden layers or sigmoids. In any case, convergence was too slow for use as an online learner. All this was unfortunate, because the direct regression output from a neural network would be convenient, and we were hoping to implement a Kohonen layer for smoothing d. In contrast, a naive Bayes (NB) classifier worked very well. A NB learner is given a set of training documents, each labeled with one of a finite set of classes/topic. A document or Web page u is modeled as a multiset or bag of words, { , n(u, ) } where is a feature which occurs n(u, ) times in u. In ordinary text classification (such as our baseline learner) the features are usually single words. For our apprentice learner, a feature is a t, d pair. NB classifiers can predict from a discrete set of classes, but our prediction is a continuous (probability) score. To bridge this gap, We used a simple two-bucket (low/high relevance) special case of Torgo and Gama's technique of using classifiers for discrete labels for continuous regression <A href="25.html#12">[33], using "equally probable intervals" as far as possible. 152 Torgo and Gama recommend using a measure of centrality, such as the median, of each interval as the predicted value of that class. Rennie and McCallum <A href="25.html#12">[30] corroborate that 23 bins are adequate. As will be clear from our experiments, the medians of our `low' and `high' classes are very close to zero and one respectively (see Figure <A href="25.html#7">5). Therefore, we simply take the probability of the `high' class as the prediction from our naive Bayes apprentice. The prior probability of class c, denoted Pr(c) is the fraction of training documents labeled with class c. The NB model is parameterized by a set of numbers c, which is roughly the rate of occurrence of feature in class c, more exactly, c, = 1 + u V c n(u, ) |T | + u, n(u, ) , (1) where V c is the set of Web pages labeled with c and T is the entire vocabulary. The NB learner assumes independence between features, and estimates Pr(c |u) Pr(c) Pr(u |c) Pr(c) u n(u, ) c, . (2) Nigam et al. provide further details <A href="25.html#12">[22]. Experimental study Our experiments were guided by the following requirements. We wanted to cover a broad variety of topics, some `easy' and some `difficult', in terms of the harvest rate of the baseline crawler. Here is a quick preview of our results. The apprentice classifier achieves high accuracy in predicting the relevance of unseen pages given t, d features. It can determine the best value of d max to use, typically, 46. Encoding DOM offsets in features improves the accuracy of the apprentice substantially, compared to a bag of ordinary words collected from within the same DOM offset window. Compared to a baseline crawler, a crawler that is guided by an apprentice (trained offline) has a 30% to 90% lower loss rate. It finds crawl paths never expanded by the baseline crawler. Even if the apprentice-guided crawler is forced to stay within the (inferior) Web graph collected by the baseline crawler, it collects the best pages early on. The apprentice is easy to train online. As soon as it starts guiding the crawl, loss rates fall dramatically. Compared to t, d features, topic- or cocitation-based features have negligible effect on the apprentice. To run so many experiments, we needed three highly optimized and robust modules: a crawler, a HTML-to-DOM converter, and a classifier. We started with the w3c-libwww crawling library from <A href="http://www.w3c.org/Library/">http://www.w3c.org/Library/, but replaced it with our own crawler because we could effectively overlap DNS lookup, HTTP access, and disk access using a select over all socket/file descriptors, and prevent memory leaks visible in w3c-libwww. With three caching DNS servers, we could achieve over 90% utilization of a 2Mbps dedicated ISP connection. We used the HTML parser libxml2 library to extract the DOM from HTML, but this library has memory leaks, and does not always handle poorly written HTML well. We had some stability problems with HTML Tidy <A href="http://www.w3.org/People/Raggett/tidy/">(http://www. w3.org/People/Raggett/tidy/), the well-known HTML cleaner which is very robust to bad HTML. At present we are using libxml2 and are rolling our own HTML parser and cleaner for future work. We intend to make our crawler and HTML parser code available in the public domain for research use. For both the baseline and apprentice classifier we used the public domain BOW toolkit and the Rainbow naive Bayes classifier created by McCallum and others <A href="25.html#12">[20]. Bow and Rainbow are very fast C implementations which let us classify pages in real time as they were being crawled. 3.1 Design of the topic taxonomy We downloaded from the Open Directory <A href="http://dmoz.org/">(http://dmoz. org/) an RDF file with over 271954 topics arranged in a tree hierarchy with depth at least 6, containing a total of about 1697266 sample URLs. The distribution of samples over topics was quite non-uniform. Interpreting the tree as an is-a hierarchy meant that internal nodes inherited all examples from descendants, but they also had their own examples. Since the set of topics was very large and many topics had scarce training data, we pruned the Dmoz tree to a manageable frontier by following these steps: 1. Initially we placed example URLs in both internal and leaf nodes, as given by Dmoz. 2. We fixed a minimum per-class training set size of k = 300 documents. 3. We iteratively performed the following step as long as possible: we found a leaf node with less than k example URLs, moved all its examples to its parent, and deleted the leaf. 4. To each internal node c, we attached a leaf subdirectory called Other. Examples associated directly with c were moved to this Other subdirectory. 5. Some topics were populated out of proportion, either at the beginning or through the above process. We made the class priors more balanced by sampling down the large classes so that each class had at most 300 examples. The resulting taxonomy had 482 leaf nodes and a total of 144859 sample URLs. Out of these we could successfully fetch about 120000 URLs. At this point we discarded the tree structure and considered only the leaf topics. Training time for the baseline classifier was about about two hours on a 729MHz Pentium III with 256kB cache and 512MB RAM. This was very fast, given that 1.4GB of HTML text had to be processed through Rainbow. The complete listing of topics can be obtained from the authors. 3.2 Choice of topics Depending on the focus topic and prioritization strategy, focused crawlers may achieve diverse harvest rates. Our 153 early prototype <A href="25.html#12">[9] yielded harvest rates typically between 0.25 and 0.6. Rennie and McCallum <A href="25.html#12">[30] reported recall and not harvest rates. Diligenti et al. <A href="25.html#12">[14] focused on very specific topics where the harvest rate was very low, 46%. Obviously, the maximum gains shown by a new idea in focused crawling can be sensitive to the baseline harvest rate. To avoid showing our new system in an unduly positive or negative light, we picked a set of topics which were fairly diverse, and appeared to be neither too broad to be useful (e.g., /Arts, /Science) nor too narrow for the baseline crawler to be a reasonable adversary. We list our topics in Figure <A href="25.html#7">4. We chose the topics without prior estimates of how well our new system would work, and froze the list of topics. All topics that we experimented with showed visible improvements, and none of them showed deteriorated performance. 3.3 Baseline crawl results We will skip the results of breadth-first or random crawling in our commentary, because it is known from earlier work on focused crawling that our baseline crawls are already far better than breadth-first or random crawls. Figure <A href="25.html#7">5 shows, for most of the topics listed above, the distribution of page relevance after running the baseline crawler to collect roughly 15000 to 25000 pages per topic. The baseline crawler used a standard naive Bayes classifier on the ordinary term space of whole pages. We see that the relevance distribution is bimodal, with most pages being very relevant or not at all. This is partly, but only partly, a result of using a multinomial naive Bayes model. The naive Bayes classifier assumes term independence and multiplies together many (small) term probabilities, with the result that the winning class usually beats all others by a large margin in probability. But it is also true that many outlinks lead to pages with completely irrelevant topics. Figure <A href="25.html#7">5 gives a clear indication of how much improvement we can expect for each topic from our new algorithm. 3.4 DOM window size and feature selection A key concern for us was how to limit the maximum window width so that the total number of synthesized t, d features remains much smaller than the training data for the baseline classifier, enabling the apprentice to be trained or upgraded in a very short time. At the same time, we did not want to lose out on medium- to long-range dependencies between significant tokens on a page and the topic of HREF targets in the vicinity. We eventually settled for a maximum DOM window size of 5. We made this choice through the following experiments. The easiest initial approach was an end-to-end cross-validation of the apprentice for various topics while increasing d max . We observed an initial increase in the validation accuracy when the DOM window size was increased beyond 0. However, the early increase leveled off or even reversed after the DOM window size was increased beyond 5. The graphs in Figure <A href="25.html#8">6 display these results. We see that in the Chess category, though the validation accuracy increases monotonically, the gains are less pronounced after d max exceeds 5. For the AI category, accuracy fell beyond d max = 4. Topic #Good #Bad /Arts/Music/Styles/Classical/Composers 24000 13000 /Arts/Performing_Arts/Dance/Folk_Dancing 7410 8300 /Business/Industries.../Livestock/Horses... 17000 7600 /Computers/Artificial_Intelligence 7701 14309 /Computers/Software/Operating_Systems/Linux 17500 9300 /Games/Board_Games/C/Chess 17000 4600 /Health/Conditions_and_Diseases/Cancer 14700 5300 /Home/Recipes/Soups_and_Stews 20000 3600 /Recreation/Outdoors/Fishing/Fly_Fishing 12000 13300 /Recreation/Outdoors/Speleology 6717 14890 /Science/Astronomy 14961 5332 /Science/Earth_Sciences/Meteorology 19205 8705 /Sports/Basketball 26700 2588 /Sports/Canoe_and_Kayaking 12000 12700 /Sports/Hockey/Ice_Hockey 17500 17900 Figure 4: We chose a variety of topics which were neither too broad nor too narrow, so that the baseline crawler was a reasonable adversary. #Good (#Bad) show the approximate number of pages collected by the baseline crawler which have relevance above (below) 0.5, which indicates the relative difficulty of the crawling task. 0 0 . 2 0 . 4 0 . 6 0 . 8 1 AI Astronomy Basketball Cancer Chess Composers FlyFishing FolkDance Horses IceHockey Kayaking Linux Meteorology Soups Tobacco 10 100 1000 10000 100000 Expected #pages Relevance probability Figure 5: All of the baseline classifiers have harvest rates between 0.25 and 0.6, and all show strongly bimodal relevance score distribution: most of the pages fetched are very relevant or not at all. It is important to notice that the improvement in accuracy is almost entirely because with increasing number of available features, the apprentice can reject negative (low relevance) instances more accurately, although the accuracy for positive instances decreases slightly. Rejecting unpromising outlinks is critical to the success of the enhanced crawler. Therefore we would rather lose a little accuracy for positive instances rather than do poorly on the negative instances. We therefore chose d max to be either 4 or 5 for all the experiments. We verified that adding offset information to text tokens was better than simply using plain text near the link <A href="25.html#12">[8]. One sample result is shown in Figure <A href="25.html#8">7. The apprentice accuracy decreases with d max if only text is used, whereas it increases if offset information is provided. This highlights 154 Chess 65 70 75 80 85 90 95 100 0 2 4 6 8 d_max % A c c u r a c y Negative Positive Average AI 65 70 75 80 85 90 0 2 4 6 8 d_max % A c c u r a c y Negative Positive Average Figure 6: There is visible improvement in the accuracy of the apprentice if d max is made larger, up to about 5 7 depending on topic. The effect is more pronounced on the the ability to correctly reject negative (low relevance) outlink instances. `Average' is the microaverage over all test instances for the apprentice, not the arithmetic mean of `Positive' and `Negative'. AI 76 78 80 82 84 86 0 1 2 3 4 5 6 7 8 d_max % A c c u r a c y Text Offset Figure 7: Encoding DOM offset information with textual features boosts the accuracy of the apprentice substantially. the importance of designing proper features. To corroborate the useful ranges of d max above, we compared the value of average mutual information gain for terms found at various distances from the target HREF. The experiments revealed that the information gain of terms found further away from the target HREF was generally lower than those that were found closer, but this reduction was not monotonic. For instance, the average information Chess 0.00002 0.00004 0.00006 0.00008 0.0001 0.00012 0.00014 0.00016 0.00018 0.0002 -8 -6 -4 -2 0 2 4 6 8 d I n f o g a i n d_max=8 d_max=5 d_max=4 d_max=3 AI 4.00E-05 5.00E-05 6.00E-05 7.00E-05 8.00E-05 9.00E-05 1.00E-04 -8 -6 -4 -2 0 2 4 6 8 d I n f o G a i n d_max=8 d_max=5 d_max=4 d_max=3 Figure 8: Information gain variation plotted against distance from the target HREF for various DOM window sizes. We observe that the information gain is insensitive to d max . gain at d = -2 was higher than that at d = -1; see Figure <A href="25.html#8">8. For each DOM window size, we observe that the information gain varies in a sawtooth fashion; this intriguing observation is explained shortly. The average information gain settled to an almost constant value after distance of 5 from the target URL. We were initially concerned that to keep the computation cost manageable, we would need some cap on d max even while measuring information gain, but luckily, the variation of information gain is insensitive to d max , as Figure <A href="25.html#8">8 shows. These observations made our final choice of d max easy. In a bid to explain the occurrence of the unexpected saw-tooth form in Figure <A href="25.html#8">8 we measured the rate t,d at which term t occurred at offset d, relative to the total count of all terms occurring at offset d. (They are roughly the multinomial naive Bayes term probability parameters.) For fixed values of d, we calculated the sum of values of terms found at those offsets from the target HREF. Figure <A href="25.html#9">9(a) shows the plot of these sums to the distance(d) for various categories. The values showed a general decrease as the distances from the target HREF increased, but this decrease, like that of information gain, was not monotonic. The values of the terms at odd numbered distances from the target HREF were found to be lower than those of the terms present at the even positions. For instance, the sum of values of terms occurring at distance -2 were higher than that of terms at position -1. This observation was explained by observing the HTML tags that are present at various distances from the target HREF. We observed that tags located at odd d are mostly non-text tags, thanks to authoring idioms such as &lt;li&gt;&lt;a...&gt;&lt;li&gt;&lt;a...&gt; and &lt;a...&gt;&lt;br&gt;&lt;a...&gt;&lt;br&gt; etc. A plot of the frequency of HTML tags against the distance from the HREF at which 155 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 -5 -4 -3 -2 -1 0 1 2 3 4 5 d T h e t a _ { t , d } AI Chess Horses Cancer IceHockey Linux Bball+ Bball-Tags at various DOM offsets 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 -5 -4 -3 -2 -1 0 1 2 3 4 5 d N u m b e r o f o c c u r r e n c e s font td img b br p tr li comment div table center i span hr Figure 9: Variation of (a) relative term frequencies and (b) frequencies of HTML tags plotted against d. they were found is shown in Figure <A href="25.html#9">9(b). (The &lt;a...&gt; tag obviously has the highest frequency and has been removed for clarity.) These were important DOM idioms, spanning many diverse Web sites and authoring styles, that we did not anticipate ahead of time. Learning to recognize these idioms was valuable for boosting the harvest of the enhanced crawler. Yet, it would be unreasonable for the user-supplied baseline black-box predicate or learner to capture crawling strategies at such a low level. This is the ideal job of the apprentice. The apprentice took only 310 minutes to train on its (u, v) instances from scratch, despite a simple implementation that wrote a small file to disk for each instance of the apprentice. Contrast this with several hours taken by the baseline learner to learn general term distribution for topics. 3.5 Crawling with the apprentice trained off-line In this section we subject the apprentice to a "field test" as part of the crawler, as shown in Figure <A href="25.html#2">2. To do this we follow these steps: 1. Fix a topic and start the baseline crawler from all example URLs available from the given topic. 2. Run the baseline crawler until roughly 2000025000 pages have been fetched. 3. For all pages (u, v) such that both u and v have been fetched by the baseline crawler, prepare an instance from (u, v) and add to the training set of the apprentice. 4. Train the apprentice. Set a suitable value for d max . 0 2000 4000 6000 0 2000 4000 6000 8000 10000 Expected #pages lost #Pages fetched Folk Dancing Baseline Apprentice 0 4000 8000 0 4000 8000 12000 16000 20000 Expected #pages lost #Pages fetched Ice Hockey Baseline Apprentice Figure 10: Guidance from the apprentice significantly reduces the loss rate of the focused crawler. 5. Start the enhanced crawler from the same set of pages that the baseline crawler had started from. 6. Run the enhanced crawler to fetch about the same number of pages as the baseline crawler. 7. Compare the loss rates of the two crawlers. Unlike with the reinforcement learner studied by Rennie and McCallum, we have no predetermined universe of URLs which constitute the relevant set; our crawler must go forth into the open Web and collect relevant pages from an unspecified number of sites. Therefore, measuring recall w.r.t. the baseline is not very meaningful (although we do report such numbers, for completeness, in <A href="25.html#10">3.6). Instead, we measure the loss (the number of pages fetched which had to be thrown away owing to poor relevance) at various epochs in the crawl, where time is measured as the number of pages fetched (to elide fluctuating network delay and bandwidth). At epoch n, if the pages fetched are v 1 , . . . , v n , then the total expected loss is (1/n) i (1 - Pr(c |v i )). Figure <A href="25.html#9">10 shows the loss plotted against the number of pages crawled for two topics: Folk dancing and Ice hockey. The behavior for Folk dancing is typical; Ice hockey is one of the best examples. In both cases, the loss goes up substantially faster with each crawled page for the baseline crawler than for the enhanced crawler. The reduction of loss for these topics are 40% and 90% respectively; typically, this number is between 30% and 60%. In other words, for most 156 topics, the apprentice reduces the number of useless pages fetched by one-third to two-thirds. In a sense, comparing loss rates is the most meaningful evaluation in our setting, because the network cost of fetching relevant pages has to be paid anyway, and can be regarded as a fixed cost. Diligenti et al. show significant improvements in harvest rate, but for their topics, the loss rate for both the baseline crawler as well as the context-focused crawler were much higher than ours. 3.6 URL overlap and recall The reader may feel that the apprentice crawler has an unfair advantage because it is first trained on DOM-derived features from the same set of pages that it has to crawl again. We claim that the set of pages visited by the baseline crawler and the (off-line trained) enhanced crawler have small overlap, and the superior results for the crawler guided by the apprentice are in large part because of generalizable learning. This can be seen from the examples in Figure <A href="25.html#10">11. Baseline Apprentice Intersect Basketball 27220 26280 2431 FolkDance 14011 8168 2199 IceHockey 34121 22496 1657 FlyFishing 19252 14319 6834 Basketball 49% 47% 4% Baseline Apprentice Intersect FolkDance 57% 34% 9% Baseline Apprentice Intersect IceHockey 58% 39% 3% Baseline Apprentice Intersect FlyFishing 48% 35% 17% Baseline Apprentice Intersect Figure 11: The apprentice-guided crawler follows paths which are quite different from the baseline crawler because of its superior priority estimation technique. As a result there is little overlap between the URLs harvested by these two crawlers. Given that the overlap between the baseline and the enhanced crawlers is small, which is `better' ? As per the verdict of the baseline classifier, clearly the enhanced crawler is better. Even so, we report the loss rate of a different version of the enhanced crawler which is restricted to visiting only those pages which were visited by the baseline learner. We call this crawler the recall crawler. This means that in the end, both crawlers have collected exactly the same set of pages, and therefore have the same total loss. The test then is how long can the enhanced learner prevent the loss from approaching the baseline loss. These experiments are a rough analog of the `recall' experiments done by Rennie and McCallum. We note that for these recall experiments, the apprentice does get the benefit of not having to generalize, so the gap between baseline loss and recall loss could be optimistic. Figure <A href="25.html#10">12 compares the expected total loss of the baseline crawler, the recall crawler, and the apprentice-guided crawler (which is free to wander outside the baseline collection) plotted against the number of pages fetched, for a few topics. As expected, the recall crawler has loss generally 0 1000 0 1000 2000 3000 4000 5000 6000 Expected #pages lost #Pages fetched Ice Hockey Baseline Recall Apprentice 0 5000 10000 0 5000 10000 15000 20000 Expected #pages lost #Pages fetched Kayaking Baseline Recall Apprentice Figure 12: Recall for a crawler using the apprentice but limited to the set of pages crawled earlier by the baseline crawler. somewhere between the loss of the baseline and the enhanced crawler. 3.7 Effect of training the apprentice online Next we observe the effect of a mid-flight correction when the apprentice is trained some way into a baseline and switched into the circuit. The precise steps were: 1. Run the baseline crawler for the first n page fetches, then stop it. 2. Prepare instances and train the apprentice. 3. Re-evaluate the priorities of all unvisited pages v in the frontier table using the apprentice. 4. Switch in the apprentice and resume an enhanced crawl. We report our experience with "Folk Dancing." The baseline crawl was stopped after 5200 pages were fetched. Re-evaluating the priority of frontier nodes led to radical changes in their individual ranks as well as the priority distributions. As shown in Figure <A href="25.html#11">13(a), the baseline learner is overly optimistic about the yield it expects from the frontier, whereas the apprentice already abandons a large fraction of frontier outlinks, and is less optimistic about 157 (a) Folk Dancing 0 2000 4000 6000 8000 10000 12000 0 0-.2 .2-.4 .4-.6 .6-.8 .8-1 Estimated relevance of outlinks F r e q u e n c y Baseline Apprentice (b) Folk Dancing 2100 2200 2300 2400 2500 2600 2700 4500 5500 #Pages crawled E x p e c t e d l o s s ( # p a g e s ) Train apprentice Collect instances for apprentice Apprentice guides crawl Figure 13: The effect of online training of the apprentice. (a) The apprentice makes sweeping changes in the estimated promise of unvisited nodes in the crawl frontier. (b) Resuming the crawl under the guidance of the apprentice immediately shows significant reduction in the loss accumulation rate. the others, which appears more accurate from the Bayesian perspective. Figure <A href="25.html#11">13(b) shows the effect of resuming an enhanced crawl guided by the trained apprentice. The new (u, v) instances are all guaranteed to be unknown to the apprentice now. It is clear that the apprentice's prioritization immediately starts reducing the loss rate. Figure <A href="25.html#11">14 shows an even more impressive example. There are additional mild gains from retraining the apprentice at later points. It may be possible to show a more gradual online learning effect by retraining the classifier at a finer interval, e.g., every 100 page fetches, similar to Aggarwal et al. In our context, however, losing a thousand pages at the outset because of the baseline crawler's limitation is not a disaster, so we need not bother. 3.8 Effect of other features We experimented with two other kinds of feature, which we call topic and cocitation features. Our limiting d max to 5 may deprive the apprentice of important features in the source page u which are far from the link (u, v). One indirect way to reveal such features to the apprentice is to classify u, and to add the names of some of the top-scoring classes for u to the instance (u, v). <A href="25.html#5">2.2.3 explains why this may help. This modification resulted in a 1% increase in the accuracy of the apprentice. A further increase of 1% was observed if we added all Classical Composers 600 800 1000 1200 1400 1600 1800 2000 3000 4000 5000 6000 7000 8000 #Pages fetched C u m u l a t i v e e x p e c t e d l o s s ( # p a g e s ) Collect instances for apprentice Train apprentice Apprentice guides crawl Figure 14: Another example of training the apprentice online followed by starting to use it for crawl guidance. Before guidance, loss accumulation rate is over 30%, after, it drops to only 6%. prefixes of the class name. For example, the full name for the Linux category is /Computers/Software/Operating_ Systems/Linux. We added all of the following to the feature set of the source page: /, /Computers, /Computers/ Software, /Computers/Software/Operating_Systems and /Computers/Software/Operating_Systems/Linux. We also noted that various class names and some of their prefixes appeared amongst the best discriminants of the positive and negative classes. Cocitation features for the link (u, v) are constructed by looking for other links (u, w) within a DOM distance of d max such that w has already been fetched, so that Pr(c |w) is known. We discretize Pr(c |w) to two values high and low as in <A href="25.html#5">2.3, and encode the feature as low, d or high, d . The use of cocitation features did not improve the accuracy of the apprentice to any appreciable extent. For both kinds of features, we estimated that random variations in crawling behavior (because of fluctuating network load and tie-breaking frontier scores) may prevent us from measuring an actual benefit to crawling under realistic operating conditions. We note that these ideas may be useful in other settings. Conclusion We have presented a simple enhancement to a focused crawler that helps assign better priorities to the unvisited URLs in the crawl frontier. This leads to a higher rate of fetching pages relevant to the focus topic and fewer false positives which must be discarded after spending network, CPU and storage resources processing them. There is no need to manually train the system with paths leading to relevant pages. The key idea is an apprentice learner which can accurately predict the worth of fetching a page using DOM features on pages that link to it. We show that the DOM features we use are superior to simpler alternatives. Using topics from Dmoz, we show that our new system can cut down the fraction of false positives by 3090%. We are exploring several directions in ongoing work. We wish to revisit continuous regression techniques for the apprentice, as well as more extensive features derived from the DOM. For example, we can associate with a token t the length of the DOM path from the text node containing t to 158 the HREF to v, or the depth of their least common ancestor in the DOM tree. We cannot use these in lieu of DOM offset, because regions which are far apart lexically may be close to each other along a DOM path. t, , d features will be more numerous and sparser than t, d features, and could be harder to learn. The introduction of large numbers of strongly dependent features may even reduce the accuracy of the apprentice. Finally, we wish to implement some form of active learning where only those instances (u, v) with the largest | Pr(c |u) - Pr(c |v)| are chosen as training instances for the apprentice. Acknowledgments: Thanks to the referees for suggesting that we present Figure <A href="25.html#8">7. References [1] C. C. Aggarwal, F. Al-Garawi, and P. S. Yu. Intelligent crawling on the World Wide Web with arbitrary predicates. In WWW2001, Hong Kong, May 2001. ACM. Online at <A href="http://www10.org/cdrom/papers/110/">http: <A href="http://www10.org/cdrom/papers/110/">//www10.org/cdrom/papers/110/. [2] C. Apte, F. Damerau, and S. M. Weiss. Automated learning of decision rules for text categorization. ACM Transactions on Information Systems, 1994. IBM Research Report RC18879. [3] A. Blum and T. M. Mitchell. Combining labeled and unlabeled data with co-training. In Computational Learning Theory, pages 92100, 1998. [4] S. Chakrabarti. Integrating the document object model with hyperlinks for enhanced topic distillation and information extraction. In WWW 10, Hong Kong, May 2001. Online at <A href="http://www10.org/cdrom/papers/489">http://www10.org/cdrom/papers/489. [5] S. Chakrabarti, B. Dom, R. Agrawal, and P. Raghavan. Scalable feature selection, classification and signature generation for organizing large text databases into hierarchical topic taxonomies. VLDB Journal, Aug. 1998. Online at <A href="http://www.cs.berkeley.edu/~soumen/VLDB54_3.PDF">http: <A href="http://www.cs.berkeley.edu/~soumen/VLDB54_3.PDF">//www.cs.berkeley.edu/~soumen/VLDB54_3.PDF. [6] 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 Worldwide web conference (WWW7), 1998. Online at <A href="http://www7.scu.edu.au/programme/fullpapers/1898/com1898.html">http://www7. scu.edu.au/programme/fullpapers/1898/com1898.html. [7] S. Chakrabarti, B. Dom, and P. Indyk. Enhanced hypertext categorization using hyperlinks. In SIGMOD Conference. ACM, 1998. Online at <A href="http://www.cs.berkeley.edu/~soumen/sigmod98.ps">http://www.cs.berkeley.edu/~soumen/sigmod98. ps. [8] S. Chakrabarti, B. E. Dom, D. A. Gibson, R. Kumar, P. Raghavan, S. Rajagopalan, and A. Tomkins. Topic distillation and spectral filtering. Artificial Intelligence Review, 13(5 6):409435, 1999. [9] 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 <A href="http://www8.org/w8-papers/5a-search-query/crawling/index.html">http://www8.org/ <A href="http://www8.org/w8-papers/5a-search-query/crawling/index.html">w8-papers/5a-search-query/crawling/index.html. [10] J. Cho, H. Garcia-Molina, and L. Page. Efficient crawling through URL ordering. In 7th World Wide Web Conference, Brisbane, Australia, Apr. 1998. Online at <A href="http://www7.scu.edu.au/programme/fullpapers/1919/com1919.htm">http://www7.scu.edu. au/programme/fullpapers/1919/com1919.htm. [11] B. D. Davison. Topical locality in the Web. In Proceedings of the 23rd Annual International Conference on Research and Development in Information Retrieval (SIGIR 2000), pages 272279, Athens, Greece, July 2000. ACM. Online at <A href="http://www.cs.rutgers.edu/~davison/pubs/2000/sigir/">http:// www.cs.rutgers.edu/~davison/pubs/2000/sigir/. [12] P. M. E. De Bra and R. D. J. Post. Information retrieval in the world-wide web: Making client-based searching feasible. In Proceedings of the First International World Wide Web Conference, Geneva, Switzerland, 1994. Online at <A href="http://www1.cern.ch/PapersWWW94/reinpost.ps">http://www1. cern.ch/PapersWWW94/reinpost.ps. [13] P. M. E. De Bra and R. D. J. Post. Searching for arbitrary information in the WWW: The fish search for Mosaic. In Second World Wide Web Conference '94: Mosaic and the Web, Chicago, Oct. 1994. Online at <A href="http://archive.ncsa.uiuc.edu/SDG/IT94/Proceedings/Searching/debra/article.html">http://archive.ncsa.uiuc.edu/ <A href="http://archive.ncsa.uiuc.edu/SDG/IT94/Proceedings/Searching/debra/article.html">SDG/IT94/Proceedings/Searching/debra/article.html and <A href="http://citeseer.nj.nec.com/172936.html">http: //citeseer.nj.nec.com/172936.html. [14] M. Diligenti, F. Coetzee, S. Lawrence, C. L. Giles, and M. Gori. Focused crawling using context graphs. In A. E. Abbadi, M. L. Brodie, S. Chakravarthy, U. Dayal, N. Kamel, G. Schlageter, and K.-Y. Whang, editors, VLDB 2000, Proceedings of 26th International Conference on Very Large Data Bases, September 10-14, 2000, Cairo, Egypt, pages 527534. Morgan Kaufmann, 2000. Online at <A href="http://www.neci.nec.com/~lawrence/papers/focus-vldb00/focus-vldb00.pdf">http://www.neci.nec.com/~lawrence/ papers/focus-vldb00/focus-vldb00.pdf. [15] W. A. Gale, K. W. Church, and D. Yarowsky. A method for disambiguating word senses in a large corpus. Computer and the Humanities, 26:415439, 1993. [16] M. Hersovici, M. Jacovi, Y. S. Maarek, D. Pelleg, M. Shtalhaim, and S. Ur. The shark-search algorithm--an application: Tailored Web site mapping. In WWW7, 1998. Online at <A href="http://www7.scu.edu.au/programme/fullpapers/1849/com1849.htm">http://www7.scu. edu.au/programme/fullpapers/1849/com1849.htm. [17] T. Joachims, D. Freitag, and T. Mitchell. WebWatcher: A tour guide for the web. In IJCAI, Aug. 1997. Online at <A href="http://www.cs.cmu.edu/~webwatcher/ijcai97.ps">http://www. cs.cmu.edu/~webwatcher/ijcai97.ps. [18] H. Leiberman. Letizia: An agent that assists Web browsing. In International Joint Conference on Artificial Intelligence (IJCAI ), Montreal, Aug. 1995. See Website at <A href="http://lieber.www.media.mit.edu/people/lieber/Lieberary/Letizia/Letizia.html">http://lieber.www. media.mit.edu/people/lieber/Lieberary/Letizia/Letizia.html. [19] H. Leiberman, C. Fry, and L. Weitzman. Exploring the Web with reconnaissance agents. CACM, 44(8):6975, Aug. 2001. <A href="http://www.acm.org/cacm">http://www.acm.org/cacm. [20] A. McCallum. Bow: A toolkit for statistical language modeling, text retrieval, classification and clustering. Software available from <A href="http://www.cs.cmu.edu/~mccallum/bow/">http://www.cs.cmu.edu/~mccallum/bow/, 1998. [21] A. McCallum and K. Nigam. A comparison of event models for naive Bayes text classification. In AAAI/ICML-98 Workshop on Learning for Text Categorization, pages 4148. AAAI Press, 1998. Online at <A href="http://www.cs.cmu.edu/~knigam/">http://www.cs.cmu.edu/~knigam/. [22] A. McCallum and K. Nigam. A comparison of event models for naive Bayes text classification. In AAAI/ICML-98 Workshop on Learning for Text Categorization, pages 4148. AAAI Press, 1998. Also technical report WS-98-05, CMU; online at <A href="http://www.cs.cmu.edu/~knigam/papers/multinomial-aaaiws98.pdf">http: <A href="http://www.cs.cmu.edu/~knigam/papers/multinomial-aaaiws98.pdf">//www.cs.cmu.edu/~knigam/papers/multinomial-aaaiws98.pdf. [23] F. Menczer. Links tell us about lexical and semantic Web content. Technical Report Computer Science Abstract CS.IR/0108004, arXiv.org, Aug. 2001. Online at <A href="http://arxiv.org/abs/cs.IR/0108004">http://arxiv. org/abs/cs.IR/0108004. [24] F. Menczer and R. K. Belew. Adaptive retrieval agents: Internalizing local context and scaling up to the Web. Machine Learning, 39(2/3):203242, 2000. Longer version available as Technical Report CS98-579, <A href="http://dollar.biz.uiowa.edu/~fil/Papers/MLJ.ps">http://dollar.biz.uiowa.edu/~fil/ Papers/MLJ.ps, University of California, San Diego. [25] F. Menczer, G. Pant, M. Ruiz, and P. Srinivasan. Evaluating topic-driven Web crawlers. In SIGIR, New Orleans, Sept. 2001. ACM. Online at <A href="http://dollar.biz.uiowa.edu/~fil/Papers/sigir-01.pdf">http://dollar.biz.uiowa.edu/~fil/Papers/ <A href="http://dollar.biz.uiowa.edu/~fil/Papers/sigir-01.pdf">sigir-01.pdf. [26] T. Mitchell. Machine Learning. McGraw Hill, 1997. [27] T. Mitchell. Mining the Web. In SIGIR 2001, Sept. 2001. Invited talk. [28] S. Mukherjea. WTMS: a system for collecting and analyzing topic-specific Web information. WWW9/Computer Networks, 33(16):457471, 2000. Online at <A href="http://www9.org/w9cdrom/293/293.html">http://www9.org/w9cdrom/293/ 293.html. [29] S. RaviKumar, P. Raghavan, S. Rajagopalan, D. Sivakumar, A. Tomkins, and E. Upfal. Stochastic models for the Web graph. In FOCS, volume 41, pages 5765. IEEE, nov 2000. Online at <A href="http://www.cs.brown.edu/people/eli/papers/focs00.ps">http://www.cs.brown.edu/people/eli/papers/focs00.ps. [30] J. Rennie and A. McCallum. Using reinforcement learning to spider the web efficiently. In ICML, 1999. Online at <A href="http://www.cs.cmu.edu/~mccallum/papers/rlspider-icml99s.ps.gz">http:// www.cs.cmu.edu/~mccallum/papers/rlspider-icml99s.ps.gz. [31] G. Salton and M. J. McGill. Introduction to Modern Information Retrieval. McGraw-Hill, 1983. [32] M. Subramanyam, G. V. R. Phanindra, M. Tiwari, and M. Jain. Focused crawling using TFIDF centroid. Hypertext Retrieval and Mining (CS610) class project, Apr. 2001. Details available from manyam@cs.utexas.edu. [33] L. Torgo and J. Gama. Regression by classification. In D. Borges and C. Kaestner, editors, Brasilian AI Symposium, volume 1159 of Lecture Notes in Artificial Intelligence, Curitiba, Brazil, 1996. Springer-Verlag. Online at <A href="http://www.ncc.up.pt/~ltorgo/Papers/list_pub.html">http://www.ncc.up.pt/~ltorgo/ Papers/list_pub.html. 159

focused crawlers;Reinforcement learning;URLs;Focused crawling;taxonomy;DOM;HREF link;classifiers;Document object model

26

Accelerating 3D Convolution using Graphics Hardware

Many volume filtering operations used for image enhancement, data processing or feature detection can be written in terms of three-dimensional convolutions. It is not possible to yield interactive frame rates on todays hardware when applying such convolutions on volume data using software filter routines. As modern graphics workstations have the ability to render two-dimensional convoluted images to the frame buffer, this feature can be used to accelerate the process significantly. This way generic 3D convolution can be added as a powerful tool in interactive volume visualization toolkits.

Introduction Direct volume rendering is a very important technique for visualizing three dimensional data. Several fundamental different methods have been introduced [2, 4, 5, 6, 8, 12]. Hardware-based volume texturing [9, 14] is one of the most prominent variants for interactive visualization due to the high frame rates that can be achieved with this technique. The basic principle of texture based volume rendering is depicted in Figure 1. According to the sampling theorem a 3D view of the volume is generated by drawing an adequate number of equidistant, semi-transparent polygons parallel to the image plane with respect to the current viewing direction ("volume slicing"). Filtering on the other hand is a major part of the visualization pipeline. It is broadly used for improving images, reducing noise, and enhancing detail structure. Volume rendering can benefit from filter operations, as low pass filters reduce the noise e.g. of sam-pled medical volume images and high pass filters can be used for edge extraction, visualizing prominent data features. Multiscale approaches as [13] regularly use disjunct filtering and downsampling steps and can benefit from any speedups of the filtering process. Filters can be classified as linear or non-linear. Discrete linear filters can be written as convolutions with filter kernels that completely specify the filtering operation. Non-linear filters, as for instance morphological operators, were recently used for volume analysis and visualization [7]. Segmentation and classification depend heavily on filtering operations as well. Bro-Nielson [1] already thought about using convolution hardware for accelerating the registration process. For texture based volume rendering the data set has to be loaded into special texture memory, which can be addressed by the graphics pipe very fast. The loading process itself is relatively slow, taking several seconds for big data sets even on the fastest available graphics workstations. As the data set has to be reloaded after a filter operation has been performed in software, interactive filtering will benefit a lot from convolution algorithms that directly work on the texture hardware. Additionally, we will show in the following that computing the convolution with graphics hardware is much faster than software solutions. 3D Convolution The general three-dimensional discrete convolution can be written as ~ f (x, y, z) = i 1 ,i 2 ,i 3 k(i 1 , i 2 , i 3 ) f(x + i 1 , y + i 2 , z + i 3 ) with f being the input data function and k being the filter kernel, resulting in the convoluted data ~ f . In the following examination we will assume without loss of generality that k(i 1 , i 2 , i 3 ) is given for 0 i 1 , i 2 , i 3 &lt; n and vanishes outside this interval. Also, we will assume that the input data function vanishes for (x, y, z) outside the interval [0, m) 3 . In the special case of k(i 1 , i 2 , i 3 ) = k 1 (i 1 ) k 2 (i 2 ) k 3 (i 3 ) the 0-7803-5897-X/99/$10.00 Copyright 1999 IEEE 471 0 0.1 0.2 0.3 0.4 0.5 -3 -2 -1 0 1 2 3 0.1 0.2 0.3 0 1 2 3 4 5 6 7 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 -3 -2 -1 0 1 2 3 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0 1 2 3 4 5 6 7 Figure 2: The Gau filter function and its second derivative Figure 3: Example image; original, filtered with Gau, and filtered with second derivative kernel is called separable. In this case the number of operations necessary for the convolution can be reduced down to O(m 3 n), from O(m 3 n 3 ) in the non-separable case: ~ f 1 (x, y, z) = i 1 k 1 (i 1 ) f(x + i 1 , y, z) (1) ~ f 2 (x, y, z) = i 2 k 2 (i 2 ) ~ f 1 (x, y + i 2 , z) (2) ~ f (x, y, z) = i 3 k 3 (i 3 ) ~ f 2 (x, y, z + i 3 ) (3) Of course special care has to be taken near the boundaries of the input data function, as convolution routines are generally written on a very low language level for speed purposes. Figure 2 shows two well known convolution filters, the Gau filter and its second derivative, both in their continuous and discrete forms. They can be used for noise reduction and edge detection, respectively. An example image that was filtered with these kernels can be seen in Figure 3. Hardware Acceleration In order to accelerate the convolution process, special purpose hardware can be used. On systems that have built-in Digital Signal Processors (DSPs), for example for multimedia acceleration, a spe-cialized convolution subroutine could be downloaded to the signal 2D Convolution Figure 4: The first pass of the hardware filtering algorithm 0 1 000000000000000000000000000000 000000000000000000000000000000 000000000000000000000000000000 000000000000000000000000000000 111111111111111111111111111111 111111111111111111111111111111 111111111111111111111111111111 111111111111111111111111111111 Used texture coordinates 0 00 00 11 11 the data value 15/16 Texel inside a texel Exact position of 1/16 3/4 7/8 1 Texture coordinates 5/8 1/8 1/4 3/8 1/2 00 00 11 11 00 00 11 11 00 00 11 11 00 00 11 11 00 00 11 11 Figure 5: Texture coordinates used for exact texel hits processor. On the other hand, most times these DSPs are not well documented or the run-time system can not be modified by the user. In general they are not faster than the main processor as well. Additionally , there exists a wide range of different DSP systems, all of which are incompatible to each other. The approach we have implemented in our system is to combine a 2D and a 1D convolution kernel in order to calculate three-dimensional separable convolutions. Several vendors of the graphics API OpenGL -- as for example Silicon Graphics Inc. [10] -have included extensions for one- and two-dimensional filtering. In contrast to most implementations that emulate these extensions only in software, the SGI graphics pipes MXE and BasicReality calculate the convolutions on-board, boosting performance by an order of magnitude already for reasonably sized filters. The CRM graphics system of the O 2 is capable of rendering convolutions in hardware as well, but it does not support volume textures, which are crucial for the algorithm as well. Recall that the volume data is already stored in texture memory for visualization using texture hardware. Now (1) and (2) are combined to one 2D convolution that is to be applied to every plane of the volume data perpendicular to the z-axis. Therefore, plane by plane is drawn by rendering textured triangle strips (two triangles per strip) into the frame buffer as it is sketched in Figure 4. The texture coordinates assigned to the vertices of the triangle strips are specified in such way that no interpolation of the texture is necessary (see Figure 5). In order to increase the potential speedup and to avoid rounding problems, nearest neighbor interpolation is activated during the rendering process. Each plane is then read back with the OpenGL routine glCopyTexSubImage3DEXT , which replaces one plane of the active volume texture orthogonal to the z-axis by data that is read directly from the frame buffer. While transfering the data to the texture memory, the separable 2D convolution filter is activated using glSeparableFilter2DEXT . After this first pass, the volume texture contains data filtered along the x- and y-axes . 472 1D Convolution Figure 6: The second pass Framebuffer Operations Memory Texture Pixel Data Textures Pixel Transfer Modes Primitives Per-Fragment Rasterization Engine Geometry Geometric Scale, Bias Clamping Storage Mode Convolution Post-Convolution Scale, Bias Pixel Figure 7: The OpenGL graphics pipeline Applying the convolution to the third axis is more complicated and depicted in Figure 6. In this second pass planes are rendered perpendicular to one of the other axes. Assume that the y-axis has been chosen. As glCopyTexSubImage3DEXT can not write planes orthogonal to any other axis than the z-axis to the texture memory, they have to be transfered to a second volume texture. OpenGL's texture objects are used by calling glBindTexture for switching between them, which implies only a very small overhead. While copying the data from the frame buffer to the texture memory, a one-dimensional convolution filter is activated. As we are dealing with two-dimensional image data, we specify a 2D convolution filter with glConvolutionFilter2DEXT , using a filter kernel that is exactly one pixel wide. After this second pass the convoluted volume data has been mirrored at the plane y - z = 0. For texture based volume renderers this does not impose any restrictions, as they only have to swap coordinates during texture coordinate calculations. When data order is crucial for the application, the algorithm of pass two can be used for both passes, thus restoring the data order in the second pass. However, the textured planes have to be drawn two times perpendicular to the y-axis. Cache misses are much more likely in this case in respect to planes rendered orthogonal to the z-axis. This can increase the convolution times on big volumes by up to 50% even on fast graphics hardware. Figure 7 depicts the relevant part of the OpenGL pipeline. It reveals , that pixel fragments read from the frame buffer are clamped to [0, 1) before they can be written back to the frame buffer or into the texture memory. Filter kernels with negative coefficients can compute negative intermediary values during the two-dimensional convolution pass, which will not contribute to the final 1D convolution . Additionally, no negative results can be stored in the output volume. These values are especially needed when the filter kernel is not symmetrical. The strategy for avoiding these effects depends on the particular application. For edge detection the absolute maxima of the filtered volume data are of interest. In this case, calculating the absolute value can be performed in hardware as well, further reducing necessary computations on the CPU. In most other cases post-convolution scaling and bias can be used to map the expected results to the interval [0, 1) just before the clamping takes place. OpenGL provides the GL POST CONVOLUTION c SCALE EXT and GL POST CONVOLUTION c BIAS EXT parameters, which are applied to pixel color values after convolution and before clamping as depicted in Figure 7. Results The used data sets are presented on the color plate. Figures 8 to 12 show slices of a head data set of size 128 3 . Figure 8 reveals the unfiltered data set, whereas Figure 9 and 10 present slices of the software and hardware low pass filtered volume data, respectively. Here, a Gau filter kernel of size 5 3 has been used. Almost no differences can be detected. Figure 11 and 12 depict the results for high pass filtering using the second derivative of the Gau filter , again computed in software and in hardware. The hardware convoluted volume displays noticeable artifacts that occur due to the already mentioned clamping step in the OpenGL pipeline. By using post-convolution scaling and bias the artifacts disappear completely . The Figures 13 to 18 picture another data set that has been used for testing the implementation. They have been visualized with the hardware based volume rendering toolkit TiVOR [11], again with the first picture being rendered with the original data set. While the noise reduction effect of the Gau filter is rather bothering in Figure 14 by smearing volume details, it has remarkable positive effects on ISO surface generation (compare Figures 17 and 18). Please remark that the ISO surfaces are rendered in real-time using a hardware accelerated volume rendering approach described in [14]. Noise interferes with high pass filters, which can be seen in Figure 15. Using a high pass filter on the already low pass filtered data set reveals by far more and better separable details (see Figure 16) compared to the directly filtered volume. We have tested the speed our implementation against a well tuned software convolution filter. Unsurprisingly, the software convolution is almost completely memory bound. Even extremely fast workstations as the Octane are limited by the main memory band-473 Filter size 2 3 3 3 5 3 7 3 head 0.33/0.72 0.33/1.02 0.33/1.56 0.48/2.0 angio 2.5/6.0 2.5/8.7 2.5/14.7 3.7/21.3 Data set created by computer tomography, 128 3 Data set created by MR angiography, 256 3 All times were measured on a Silicon Graphics Onyx2 equipped with a BasicReality graphics pipe. The system has two R10000/195 MHz processors and 640 MB main memory. Table 1: Convolution times in seconds using hardware / software width, as today's caches are far too small for the values needed for convolution along the z-axis. High end machines as the Onyx2 perform huge 3D convolutions three times faster than the Octane, even when equipped with slower CPUs. Standard PCs cannot cope with the memory bandwidth of the Onyx2 system, and multipro-cessor options will not accelerate the process because it is not CPU bound. Table 1 shows convolution times for different data sets and filter sizes, using software and hardware convolution. All times have been measured on an Onyx2 equipped with a BasicReality graphics pipe. The maximum filter size supported by the graphics system is 7 2 . Therefor, the maximum 3D convolution that can be performed in hardware on this system is 7 3 . Noteworthy is the fact that the BasicReality graphics system is optimized for filter kernels of size 5 2 . Convolutions with smaller kernels need exactly the same computation time. Filters of size 6 2 and 7 2 share their timing results as well. The x and y coordinates of the volume are swapped during the hardware based convolution process, which is a side effect of the presented 3D convolution algorithm. Conclusion As several of today's graphics workstation vendors have added two-dimensional convolution to their OpenGL pipeline, using this capability for accelerating 3D convolution is an almost straightforward approach. We have determined that by using the implemented algorithm three-dimensional convolution can be performed even on big data sets with nearly interactive rates. First promising approaches of accelerating wavelet decomposition and reconstruction have been investigated as well [3]. As all intermediary data is transfered to the frame buffer, clamping can swallow negative values that result from the two-dimensional convolution as well as final negative results. Thus this approach is currently most useful for symmetrical filter kernels. By using post-convolution scaling and bias extensions these problems can be easily overcome. Non-separable convolutions are not possible right now with this algorithm . However, by applying several two-dimensional filter kernels and blending convoluted images in the frame buffer the use of non-separable 3D kernels will be a possibility for the future as well. References [1] M Bro-Nielson. Medical Image Registration and Surgery Simulation. PhD thesis, IMM, Technical University of Denmark, 1996. [2] R. A. Crawfis and N. L. Max. Texture Splats for 3D Scalar and Vector Field Visualization. In G. M. Nielson and Bergeron D., editors, Visualization 93, pages 261265, Los Alamitos, CA, 1993. IEEE Computer Society, IEEE Computer Society Press. [3] M. Hopf and T. Ertl. Hardware Based Wavelet Transformations. In Erlangen Workshop '99 on Vision, Modeling and Visualization, Erlangen , November 1999. IEEE. accepted for publication, November 17 19. [4] A. Kaufman. Volume Visualization. IEEE Computer Society Press, 1991. [5] P. Lacroute and M. Levoy. Fast Volume Rendering Using a Shear-Warp Factorization of the Viewing Transformation. In Computer Graphics Proceedings, Annual Conference Series, pages 451457, Los Angeles, California, July 1994. ACM SIGGRAPH, Addison-Wesley Publishing Company, Inc. [6] L. Lippert, M. H. Gross, and C. Kurmann. Compression Domain Volume Rendering for Distributed Environments. In D. Fellner and L. Szirmay-Kalos, editors, EUROGRAPHICS '97, volume 14, pages C95C107. Eurographics Association, Blackwell Publishers, 1997. [7] C. L urig and T. Ertl. Hierarchical Volume Analysis and Visualization Based on Morphological Operators. In Proc. IEEE Visualization '98, pages 335341, 1998. [8] P. Schroder and G. Stoll. Data Parallel Volume Rendering as Line Drawing. In 1992 Workshop on Volume Visualization. ACM SIGGRAPH , October 1992. [9] Silicon Graphics Inc., Mountain View, California. Volume Rendering using RE2 Technology, 1994. [10] Silicon Graphics Inc., Mountain View, California. OpenGL on Silicon Graphics Systems, 1996. [11] O. Sommer, A. Dietz, R. Westermann, and T. Ertl. An Interactive Visualization and Navigation Tool for Medical Volume Data. In N. M. Thalmann and V. Skala, editors, WSCG '98, The Sixth International Conference in Central Europe on Computer Graphics and Visualization '98, volume II, pages 362371, Plzen, Czech Republic, February 1998. University of West Bohemia Press. [12] T. Totsuka and M. Levoy. Frequency Domain Volume Rendering. Computer Graphics, 27(4):27178, August 1993. [13] R. Westermann and T. Ertl. A Multiscale Approach to Integrated Volume Segmentation and Rendering. In Computer Graphics Forum 16(3) (Proc. EUROGRAPHICS '97), number 3, pages 117129. Blackwell, 1997. [14] R. Westermann and T. Ertl. Efficiently Using Graphics Hardware in Volume Rendering Applications. In Computer Graphics Proceedings, Annual Conference Series, pages 169177, Orlando, Florida, 1998. ACM SIGGRAPH. 474 Figure 8: The unfiltered head data set Figure 9: Head, low pass filtered in software Figure 10: Head, low pass filtered in hardware Figure 11: Head, high pass filtered in software Figure 12: Head, high pass filtered in hardware Figure 13: The original angiography data set Figure 14: The Gau filtered data set Figure 15: Data, after direct filtering with Gau' second derivative Figure 16: First low pass, then high pass filtered data Figure 17: ISO surfaces on the original angiography data set Figure 18: ISO surfaces on the Gau filtered data set

3D convolution;Convolution;visualization;filtering;Volume Visualization;Hardware Acceleration;volume rendering

27

Agent Technology for Personalized Information Filtering: The PIA-System

As today the amount of accessible information is overwhelming, the intelligent and personalized filtering of available information is a main challenge. Additionally, there is a growing need for the seamless mobile and multi-modal system usage throughout the whole day to meet the requirements of the modern society ("anytime, anywhere, anyhow"). A personal information agent that is delivering the right information at the right time by accessing, filtering and presenting information in a situation-aware matter is needed. Applying Agent-technology is promising, because the inherent capabilities of agents like autonomy, pro- and reactiveness offer an adequate approach. We developed an agent-based personal information system called PIA for collecting, filtering, and integrating information at a common point, offering access to the information by WWW, e-mail, SMS, MMS, and J2ME clients. Push and pull techniques are combined allowing the user to search explicitly for specific information on the one hand and to be informed automatically about relevant information divided in pre-, work and recreation slots on the other hand. In the core of the PIA system advanced filtering techniques are deployed through multiple filtering agent communities for content-based and collaborative filtering. Information-extracting agents are constantly gathering new relevant information from a variety of selected sources (internet, files, databases, web-services etc.). A personal agent for each user is managing the individual information provisioning, tailored to the needs of this specific user, knowing the profile, the current situation and learning from feedback.

Introduction Nowadays, desired information often remains unfound, because it is hidden in a huge amount of unnecessary and irrelevant data. On the Internet there are well maintained search engines that are highly useful if you want to do full-text keyword-search [1], but they are not able to support you in a personalized way and typically do not offer any "push-services" or in other words no information will be sent to you when you are not active. Also, as they normally do not adapt themselves to mobile devices, they cannot be used throughout a whole day because you are not sitting in front of a standard browser all the time and when you return, these systems will treat you in the very same way as if you have never been there before (no personalization, no learning). Users who are not familiar with domain-specific keywords won't be able to do successful research, because no support is offered. Predefined or auto-generated keywords for the search-domains are needed to fill that gap. As information demands are continuously increasing today and the gathering of information is time-consuming, there is a growing need for a personalized support. Labor-saving information is needed to increase productivity at work and also there is an increasing aspiration for a personalized offer of general information, specific domain knowledge, entertainment, shopping, fitness, lifestyle and health information. Existing commercial "personalized" systems are far away from that functionality, as they usually do not offer much more than allowing to choose the kind of the layout or collecting some of the offered information channels (and the information within is not personalized). To overcome that situation you need a personal information agent (PIA) who "knows" the way of your thinking. and can really support you throughout the whole day by accessing, 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. SAC'05, March 13-17, 2005, Santa Fe, New Mexico, USA. Copyright 2005 ACM 1-58113-964-0/05/0003...$5.00. 54 2005 ACM Symposium on Applied Computing filtering and presenting information to you in a situation-aware matter (figure 1). Some systems exist (FAB [2], Amalthaea [3], WAIR [4], P-Tango [5], TripMatcher [6], PIAgent [7], Letizia [8], Let's Browse [9], Newt [10], WebWatcher [11], PEA [12], BAZAR [13]) that implement advanced algorithmic technology, but did not become widely accepted, maybe because of real world requirements like availability, scalability and adaptation to current and future standards and mobile devices. In this paper we present an agent-based approach for the efficient, seamless and tailored provisioning of relevant information on the basis of end-users' daily routine. As we assume the reader will be familiar with agent-technology (see [14], [15] for a good introduction), we will concentrate on the practical usage and the real-world advantages of agent-technology . We describe the design and architecture in the following section and afterwards depict the system in section three. Design of PIA The Personal Information Agent To meet the discussed requirements and to support the user in that matter, we designed a multi-agent system composed of four classes of agents: many information extracting agents, agents that implement different filtering strategies, agents for providing different kinds of presentation and one personal agent for each user. Logically, all this can be seen as a classical three tier application (figure 2). Concerning the information extraction, general search engines on the one hand but also domain-specific portals on the other hand have to be integrated. Additional information sources (Databases, Files, Mailinglists etc.) should also be integrated easily at run-time. Several agents realize different filtering strategies (content-based and collaborative filtering [16], [5]) that have to be combined in an intelligent matter. Also agents for providing information actively via SMS, MMS, Fax, e-mail (push-services) are needed. A Multi-access service platform has to manage the presentation of the results tailored to the used device and the current situation. The personal agent should constantly improve the knowledge about "his" user by learning from the given feedback, which is also taken for collaborative filtering, as information that has been rated as highly interesting might be useful for a user with a similar profile as well. As users usually are not very keen on giving explicit feedback (ratings), implicit feedback like the fact that the user stored an article can also be taken into account [18]. A "keywordassistant" should support the user to be able to define queries even if he is not familiar with a certain domain. Keywords predefined by experts should be offered and also the possibility to point to a "basis-paper" serving as an example. The keywordassistant will extract automatically the most important keywords of that paper and will provide them for searching. The whole system is designed to be highly scalable, easy to modify, to adapt and to improve and therefore an agent-based approach that allows to integrate or to remove agents even at run-time is a smart choice. The different filtering techniques are needed to provide accurate results, because the weakness of individual techniques should be compensated by the strengths of others. Documents should be logically clustered by their domains to allow fast access, and for each document several "models" [19] will be built, all including stemming and stop-word elimination, but some tailored for very efficient retrieval at run-time and others to support advanced filtering algorithms for a high accuracy. PIA Demand for information is increasing continuously Information gathering is time consuming Users have different devices Facts Portals Databases General Information News Weather Information to increase productivity / Labor-saving information Domain knowledge Recreation Entertainment Shopping Health / Fitness / Lifestyle Technologies End Devices Software Networks www / Internet Not structured content Structured Dedicated content Structured content Profile Demand for information Figure 1: Demand for a personal information agent 55 Figure 2: The PIA System seen as a three tier application If the system notices that the content-based filtering is not able to offer sufficient results, additional information should be offered by collaborative filtering, i.e. information that was rated as interesting by a user with a similar profile will be presented. With the "push-services", the user can decide to get new integrated relevant information immediately and on a mobile device, but for users who do not want to get new information immediately, a personalized newsletter also has to be offered. This newsletter is collecting new relevant information to be conveniently delivered by e-mails, allowing users to stay informed even if they are not actively using the system for some time. Deployment and Evaluation We implemented the system using Java and standard open source database and web-technology and based on the JIAC IV 1 agent framework [20]. JIAC IV is FIPA 2000 compliant [21], that is it is conforming to the latest standards. It consists of a communication infrastructure as well as services for administrating and organizing agents (Agent Management Service, AMS and Directory Facilitator, DF). The JIAC IV framework provides a variety of management and security functions, management services including configuration, fault management and event logging, security aspects including authorization, authentication and mechanisms for measuring and ensuring trust and therefore has been a good choice to be used from the outset to the development of a real world application. Within JIAC IV, agents are arranged on platforms, allowing the arrangement of agents that belong together with the control of at least one "manager". A lot of visual tools are offered to deal with administration aspects. Figure 3 shows a platform, in this case with agents for the building of different models specialized for different retrieval algorithms. The prototypical system is currently running on Sun-Fire-880 , Sun-Fire-480R and Sun Fire V65x, whereas the main filtering computation, database- and web-server and information-extraction is placed on different machines for performance reasons. 1 JIAC IV is funded by T-Systems Nova GmbH Figure 3: Several Agents are building different models specialized for different retrieval algorithms. 56 New content is stored, validated and consolidated in a central relational database (update-driven). Information can be accessed by WWW, e-mail, SMS, MMS, and J2ME Clients, where the system adapts the presentation accordingly, using the CC/PP (Preferences Profile) with a tailored layout for a mobile phone and a PDA (see section 3.5). The personalized newsletter and the push-services are sent via e-mail, SMS or MMS. The user can use self-defined keywords for a request for information or choose a category and therefore the system will use a list of keywords predefined by experts and updated smoothly by learning from collaborative filtering. A combination of both is also possible. The keyword assistant is able to extract the most import keywords of a given article using the term frequency inverse document frequency (TFIDF)-algorithm [22]. 3.2 Gathering new Information New information is constantly inserted in the system by information extraction agents, e.g. web-reader agents or agents that are searching specified databases or directories. Additional agents for the collection of new content can easily be integrated even at runtime, as all that is necessary for a new agent is to register himself at the system, store the extracted information at a defined database table and inform the modeling-manager agent about the insertion. As a file reader-agent is constantly observing a special directory, manual insertion of documents can be done simply by drag-and-drop and an e-mail and upload-interface also exists. Source can also be integrated by Web services. New Readers can be created using a easy-to-handle tool and another tool is enabling to conveniently observe the extraction-agents, as this is the interface to the outside that might become critical if for example the data-format of a source is changed. 3.3 Pre processing for efficient retrieval The first step of pre processing information for efficient retrieval is the use of distinct tables in the global database for different domains like e.g. news, agent-related papers, etc. Depending on the filtering request, tables with no chance of being relevant can therefore be omitted. The next step is the building of several models for each document. Stemming and stop-word elimination is implemented in every model but different models are built by computing a term importance either based only on local frequencies, or based on term frequency inverse document frequency (TFIDF) approach. Furthermore number of words that should be included in models is different which makes models either more accurate or more efficient. Created models are indexed either on document or word level, which facilitate their efficient retrieval. The manager agent is assigning the appropriate modeling agents to start building their models but might decide (or the human system administrator can tell him) at runtime to delay latest time-consuming modeling activity for a while if system load is critical at a moment. This feature is important for a real-world application, as overloading has been a main reason for the unusability of advanced academic systems. 3.4 Filtering technology As the quality of results to a particular filtering request might heavily depend on the information domain (news, scientific papers, conference calls), different filtering communities are implemented. For each domain, there is at least one community which contains agents being tailored to do specific filtering and managing tasks in an efficient way. Instead of having only filtering agents, each and every community has also one so called manager agent that is mainly responsible for doing coordination of filtering tasks. The coordination is based on quality, CPU, DB and memory fitness values, which are the measures being associated to each filtering agent [23]. These measures respectively illustrate filtering agent successfulness in the past, its efficiency in using available CPU and DB resources, and the amount of memory being required for filtering. A higher CPU, DB or memory fitness value means that filtering agent needs a particular resource in a smaller extent for performing a filtering task. This further means that an insufficiency of a particular resource has a smaller influence on filtering agents with a higher particular fitness value. The introduced different fitness values together with the knowledge about the current system runtime performance can make coordination be situation aware (see also [23]) in the way that when a particular resource is highly loaded a priority in coordination should be given to filtering agents for which a particular resource has a minor importance. This situation aware coordination provides a way to balance response time and filtering accuracy, which is needed in overcoming the problem of finding a perfect filtering result after few hours or even few days of an expensive filtering. Instead of assigning filtering task to the agent with the best combination of fitness values in the current runtime situation, manager is going to employ a proportional selection principle [24] where the probability for the agent to be chosen to do actual filtering is proportional to the mentioned combination of its fitness values. By not always relying only on the most promising agents, but also sometimes offering a job to other agents, manager gives a chance to each and every agent to improve its fitness values. While the adaptation of quality fitness value can be accomplished after the user feedback became available, the other fitness values can be changed immediately after the filtering was finished through the response time analyses. The adaptation scheme has a decreasing learning rate that prevents already learnt fitness values of being destroyed, which further means that proven agents pay smaller penalties for bad jobs than the novice ones [17]. In the case where the received filtering task cannot be successfully locally accomplished usually because of belonging to unsupported information domain, manager agent has to cooperate with other communities. While coordination takes place inside each community between manager and filtering agents, cooperation occurs between communities among manager agents. Cooperation is based on either finding a community which supports given domain or in splitting received task on sub-tasks where for each sub-task a community with good support exists. Figure 4 presents a high level overview of the filtering framework being composed of three different filtering communities (FC), where each community has one filter manager agent (M) and different number of specialized filtering agents (F). There are two different databases (DB) with information from different domains, and they are accessed at least by one community. On the figure cooperation is illustrated as a circle with arrows which connect manager agents. 57 Figure 4: Filtering framework with three different communities 3.5 Presentation As one of the main design principles has been to support the user throughout the whole day, the PIA system provides several different access methods and adapts its interfaces to the used device (Figure 5). To fulfill these requirements an agent platform (&quot;Multi Access Service Platform&quot;) was developed that optimizes the graphical user interface for the access by Desktop PCs, PDAs and smart phones. If the user wants to use the PIA system, the request is received by the Multi Access Service Platform (MAP). The MAP delegates the request to an agent, providing the logic for this service. In the PIA system the requests are forwarded either to login agent or the personal agent. The chosen agent performs the service specific actions and sends the MAP an abstract description of the formular that should be presented to the user. For this purpose the XML based Abstract Interaction Description Language (AIDL) has been defined. Based on the abstract description and the features of the used device the MAP generates an optimized interface presented to the user. The conversion is implemented as a XSLT transformation in which the optimal XSLT style sheet is selected based on the CC/PP information about the user's device. This approach simplifies the creation of optimized user interfaces for different devices. The abstract interface description can be easily transformed into HTML, PDA optimized HTML or WML. If the user wants to have a voice interface, a style sheet for converting the abstract user interface description into VoiceXML has to be added to the MAP. Additional changes at the PIA system are not needed. Beside of the features provided by MAP the design of the user interface must create an easy to use system even on devices with a tiny screen and without a keyboard. That is why the PIA interface provides additional navigation elements on complex forms and minimizes the use of text input fields. New results matching a defined request are presented first as a list of short blocks containing only title, abstract and some meta-information (as this is familiar to every user from well-known search-engines ). This information is also well readable on PDAs or even mobile phones. Important articles can be stored in a repository. This allows the user to choose the articles on his PDA he wants to read later at his desktop PC. Depending on the personal options specified by the user, old information found for a specific query may be deleted automatically step by step after a given time, that is, there is always up to date information that is presented to the user (we call this &quot;smart mode&quot;). This is for example convenient for getting personalized filtering news. The other option is to keep that information unlimited (&quot;global mode&quot;) for a query for e.g. basic scientific papers. For the &quot;push-services&quot; (that is, the system is becoming active and sending the user information without an explicit request), the user specifies his working time (e.g. 9 am to 5 pm). This divides the day in a pre-, work, and a recreation slot, where the PIA system assumes different demands of information. For each slot an adequate delivering technology can be chosen (e-mail, sms, mms, fax or Voice). If you decide to subscribe to the personalized newsletter, new relevant information for you will be collected and sent by e-mail or fax once a day with a similar layout and structure for convenient reading if you have not seen it already by other pull- or push services. Therefore you can also stay informed without having to log into the system and if you do not want to get all new information immediately. Figure 5: Information accessed by browser or tailored for presentation on a PDA or a mobile phone 58 Conclusion and future work The implemented system has an acceptable runtime performance and shows that it is a good choice to develop a personal information system using agent-technology based on a solid agent-framework like JIAC IV. Currently, PIA system supports more than 120 different web sources, grabs daily around 3.000 new semi-structured and unstructured documents, has almost 500.000 already pre-processed articles, and actively helps about fifty scientists related to our laboratory in their information retrieval activities. Their feedback and evaluation is a valuable input for the further improvement of PIA. In the near future we plan to increase the number of users to thousands, and therefore we plan to work on the further optimization of the filtering algorithms to be able to simultaneously respond to multiple filtering requests. Also, we think about integrating additional services for the user that provide information tailored to his geographical position (GPS), a natural speech interface and innovative ways to motivate the user to give precise explicit feedback, as the learning ability of the system is depending on that information. REFERENCES [1] Brin, S.; Page, L.: The anatomy of a large-scale hyper textual (Web) search engine, Proc. 7th International World Wide Web Conference on Computer Networks, 30(1-7), pp. 107-117, 1998. [2] Balabanovic, M.; Yoav, S.: FAB: Content-Based Collaborative Recommendation, Communication of the ACM, Volume 40, Number 3, pp. 66-72, 1997. [3] Moukas, A.: Amalthaea: Information Discovery and Filtering using a Multi agent Evolving Ecosystem, Practical Application of Intelligent Agents & Multi-Agent Technology, London 1998. [4] Zhang, B.; Seo, Y.: Personalized Web-Document Filtering Using Reinforcement Learning, Applied Artificial Intelligence, Volume 15 Number 7, pp. 665-685, 2001. [5] Claypool, M.; Gokhale, A.; Miranda, T.; Murnikov, P.; Netes, D.; Sartin, N.: Combining Content-Based and Collaborative Filters in an Online Newspaper, ACM SIGIR Workshop on Recommender Systems, Berkeley, CA, August 19, 1999. [6] Delgado, J.; Davidson, R.: Knowledge bases and user profiling in travel and hospitality recommender systems, In Proceedings of the ENTER 2002 Conference, Innsbruck, Austria, January 22-25 2002, Springer Verlag, pp. 1-16. [7] Kuropka, D.; Serries, T.: Personal Information Agent, Informatik Jahrestagung 2001, pp. 940-946. [8] Lieberman, H.: Letizia: An Agent That Assists Web Browsing, International Joint Conference on Artificial Intelligence, Montreal, August 1995. [9] Lieberman, H.; Van Dyke, N,; Vivacqua, A.: Let's Browse: A Collaborative Browsing Agent, Knowledge-Based Systems, 12(8), 1999, pp. 427431. [10] Sheth, B.: A Learning Approach to Personalized Information Filtering, M.S. Thesis, MIT- EECS dept, USA, 1994. [11] Joachims, T.; Freitag, D.; Mitchell, T.: WebWatcher: A Tour Guide for the World Wide Web, In IJCAI (1), 1997, pp. 770-777 . [12] Winiwarter, W.: PEA - A Personal Email Assistant with Evolutionary Adaptation, International Journal of Information Technology, Vol. 5, No. 1, 1999. [13] Thomas, C.; Fischer, G.: Using agents to improve the usability and usefulness of the world wide web. In Proceedings of the Fifth International Conference on User Modelling, pages 5--12, 1996. [14] Jennings, N; Wooldridge, M: Agent-oriented software engineering, Handbook of Agent Technology (ed. J. Bradshaw). AAAI/MIT Press, 2000. [15] Sesseler, R.; Albayrak, S.: Agent-based Marketplaces for Electronic Commerce, International Conference on Artificial Intelligence, IC-AI 2001. [16] Resnick, P.; Neophytos, J.; Suchak, M.; Bergstrom, P.; Riedl, J.: GroupLens: An open architecture for collaborative filtering of net news, Proceedings ACM Conference on Computer-Supported Cooperative Work, pp. 175-186, 1994. [17] Albayrak, S.; Milosevic, D.: Self Improving Coordination in Multi Agent Filtering Framework, IEEE/WIC/ACM International Joint Conference on Intelligent Agent technology (IAT 04) and Web Intelligence (WI 04), Beijing, China, September 2004., (to appear). [18] Nichols, D.: Implicit Rating and Filtering, Proc. Fifth DELOS Workshop on Filtering and Collaborative Filtering, Budapest, Hungary, 10-12 November, ERCIM, pp. 31-36, 1997. [19] Tauritz, D.: Adaptive Information Filtering: concepts and algorithms, Ph.D. dissertation, Leiden University, 2002. [20] Fricke, S.; Bsufka, K.; Keiser, J.; Schmidt, T.; Sesseler, R.; Albayrak, S.: Agent-based Telematic Services and Telecom Applications, Communications of the ACM, April. 2001. [21] Foundation for Intelligent Physical Agents, www.fipa.org, 2004. [22] Jing, L.; Huang, H.; Shi, H.: Improved Feature Selection Approach TFIDF in Text Mining, Proc. 1 st Internat. Conference on Machine Learning and Cybernetics, Beijing, 2002. [23] Albayrak, S; Milosevic D.: Situation-aware Coordination in Multi Agent Filtering Framework, The 19th International Symposium on Computer and Information Sciences (ISCIS 04), Antalya, Turkey, October 2004., (to appear). [24] Zbigniew, M.; Fogel, D.: How to Solve It: Modern Heuristics, Springer-Verlag New York, Inc., New York, NY, 2000. 59

Adaptation and Learning;filtering;Recommendation systems;Agent-based deployed applications;Evolution;Intelligent and personalized filtering;Agents and complex systems;personal information agent;agent technology;Ubiquitous access

28

An Adaptive Information Retrieval System based on Associative Networks

In this paper we present a multilingual information retrieval system that provides access to Tourism information by exploiting the intuitiveness of natural language. In particular, we describe the knowledge representation model underlying the information retrieval system. This knowledge representation approach is based on associative networks and allows the definition of semantic relationships between domain-intrinsic information items. The network structure is used to define weighted associations between information items and augments the system with a fuzzy search strategy. This particular search strategy is performed by a constrained spreading activation algorithm that implements information retrieval on associative networks. Strictly speaking, we take the relatedness of terms into account and show, how this fuzzy search strategy yields beneficial results and, moreover, determines highly associated matches to users' queries. Thus, the combination of the associative network and the constrained spreading activation approach constitutes a search algorithm that evaluates the relatedness of terms and, therefore, provides a means for implicit query expansion.

Introduction Providing easy and intuitive access to information still remains a challenge in the area of information system research and development. Moreover, as Van Rijsbergen (1979) points out, the amount of available information is increasing rapidly and offering accurate and speedy access to this information is becoming ever more difficult. This quote, although about 20 years old, is still valid nowadays if you consider the amount of information offered on the Internet. But how to address these problems? How to overcome the limitations associated with conventional search interfaces ? Furthermore, users of information retrieval systems are often computer illiterate and not familiar with the required logic for formulating appropriate queries, e.g. the burdens associated with Boolean Copyright c 2004, Australian Computer Society, Inc. This paper appeared at First Asia-Pacific Conference on Conceptual Modelling (APCCM 2004), Dunedin, New Zealand. Conferences in Research and Practice in Information Technology, Vol. 31. Sven Hartmann and John Roddick, Ed. Reproduction for academic, not-for profit purposes permitted provided this text is included. logic. This goes hand in hand with the urge to understand what users really want to know from information retrieval systems. Standard information retrieval interfaces consist of check boxes, predefined option sets or selection lists forcing users to express her or his needs in a very restricted manner. Therefore, an approach leaving the means of expression in users' hands, narrows the gap between users' needs and interfaces used to express these needs. An approach addressing this particular problem is to allow query formulation in natural language. Natural language interfaces offer easy and intuitive access to information sources and users can express their information needs in their own words. Hence, we present a multilingual information retrieval system allowing for query formulation in natural language. To reduce word sense ambiguities the system operates on a restricted domain. In particular, the system provides access to tourism information, like accommodations and their amenities throughout Austria. However, the core element of the information retrieval system remains the underlying knowledge representation model. In order to provide a knowledge representation model allowing to define relations among information items, an approach based on a network structure, namely an associative network, is used. More precisely, this associative network incorporates a means for knowledge representation allowing for the definition of semantic relationships of domain-intrinsic information. Processing the network and, therefore, result determination is accomplished by a technique refereed to as spreading activation. Some nodes of the network act as sources of activation and, subsequently, activation is propagated to adjacent nodes via weighted links. These newly activated nodes, in turn, transmit activation to associated nodes, and so on. We introduce a knowledge representation approach based on an associative network consisting of three layers. Moreover, a constrained spreading activation algorithm implements a processing technique that operates on the network. Due to the network structure of the knowledge representation model and the processing technique, implicit query expansion enriches the result set with additional matches. Hence, a fuzzy search strategy is implemented. The remainder of the paper is organized as follows . In Section <A href="28.html#2">2 we review the architecture of the information retrieval system that acts as a basis for the redeveloped approach presented herein. Moreover , Section <A href="28.html#4">3 gives an overview about associative networks and we present an algorithm for processing such networks, i.e. spreading activation. In Section <A href="28.html#5">4 we describe our approach based on associative net-27 works and finally, some conclusions are given in Section <A href="28.html#9">5. AD.M.IN A Natural Language Information Retrieval System Crestani (1997) points out that information retrieval is a science that aims to store and allow fast access to a large amount of data. In contrast to conventional database systems, an information retrieval system does not provide an exact answer to a query but tries to produce a ranking that reflects the intention of the user. More precisely, documents are ranked according to statistical similarities based on the occurrence frequency of terms in queries and documents. The occurrence frequency of a term provides an indicator of the significance of this term. Moreover, in order to get a measure for determining the significance of a sentence, the position of terms within a sentence is taken into account and evaluated. For comprehensive reports about information retrieval see Salton & McGill (1983), Salton (1989) and Baeza-Yates & Ribeiro-Neto (1999). In order to adapt information retrieval systems to the multilingual demands of users, great efforts have been made in the field of multilingual information retrieval . Hull & Grafenstette (1996) subsume several attempts to define multilingual information retrieval, where Harman (1995) formulates the most concise one: "multilingual information retrieval is information retrieval in any language other than English". Multilingual information retrieval systems have to be augmented by mechanisms for query or document translation to support query formulation in multiple languages. Information retrieval is such an inexact discipline that it is not clear whether or not query translation is necessary or even optimal for identifying relevant documents and, therefore, to determine appropriate matches to the user query. Strictly speaking , the process of translating documents or queries represents one of the main barriers in multilingual information retrieval. Due to the shortness of user queries, query translation introduces ambiguities that are hard to overcome . Contrarily, resolving ambiguity in document translation is easier to handle because of the quantity of text available. Nevertheless, state-of-the-art machine translation systems provide only an insufficient means for translating documents. Therefore, resolving ambiguities associated with translations remains a crucial task in the field of multilingual information retrieval. Ballesteros & Croft (1998), for instance, present a technique based on co-occurrence statistics from unlinked text corpora which can be used to reduce the ambiguity associated with translations . Furthermore, a quite straightforward approach in reducing ambiguities is to restrict the domain a multilingual information retrieval system operates on. Xu, Netter & Stenzhorn (2000) describe an information retrieval system that aims at providing uniform multilingual access to heterogeneous data sources on the web. The MIETTA (Multilingual Tourist Information on the World Wide Web) system has been applied to the tourism domain containing information about three European regions, namely Saarland, Turku, and Rome. The languages supported are English, Finnish, French, German, and Italian. Since some of the tourism information about the regions were available in only one language, machine translation was used to deal with these web documents. Due to the restricted domain, automatic translation should suffice to understand the basic meaning of the translated document without having knowledge of the source language. Users can query the system in various ways, such as free text queries, form-based queries, or browsing through the concept hierarchy employed in the system. MIETTA makes it transparent to the users whether they search in a database or a free-form document collection. 2.1 The Architecture of the Original System The software architecture of the natural language information retrieval system is designed as a pipeline structure. Hence, successively activated pipeline elements apply transformations on natural language queries that are posed via arbitrary client devices, such as, for instance, web browsers, PDAs or mobile phones. Due to the flexibility of this approach, different pipeline layouts can be used to implement different processing strategies. Figure <A href="28.html#3">1 depicts the layout of the software architecture and illustrates the way of interaction of the pipeline elements. In a first step, the natural language query is evaluated by an automatic language identification module to determine the language of the query. Next, the system corrects typographic errors and misspellings to improve retrieval performance. Before adding grammar rules and semantic information to the query terms, a converter transforms numerals to their nu-meric equivalents. Depending on the rules assigned to the query terms, a mapping process associates these terms with SQL fragments that represent the query in a formal way. Due to the fact that the system uses a relational database as backend this mapping process is crucial. In a next step the SQL fragments are combined according to the modifiers (e.g. "and", "or", "near", "not") identified in the query and a single SQL statement that reflects the intention of the query is obtained. Then the system determines the appropriate result and generates an XML representation for further processing. Finally, the XML result set is adapted to fit the needs of the client device. The remainder of this section gives a brief outline of the system. 2.1.1 The Knowledge Base A major objective of the Ad.M.In.(Adaptive Multilingual Interfaces) system was to separate the program logic from domain dependent data. In particular , language, domain and device dependent portions are placed in the knowledge base. Thus, the knowledge base represents the backbone of the system and consists of a relational database and a set of ontologies . The database stores information about domain entities, as, for instance, amenities of accommodations . The ontologies store synonyms, define semantic relations and grammar rules. Basically, the knowledge base consists of separate XML files, whereas the synonym ontology is used to associate terms having the same semantic meaning, i.e. describes linguistic relationships like synonymy. The synonym ontology is based on a flat structure, allowing to define synonymy. Taking a look at the tourism domain, "playground" represents a concept possessing several semantic equivalents, as, for instance , "court". Unfortunately, the synonym ontology provides no means to associate concepts. Consider, for example, the three concepts "sauna", "steam bath" and "vegetarian kitchen". Straightforward, someone might derive a stronger degree of relatedness between the concepts "sauna" and "steam bath" as between "sauna" and "vegetarian kitchen". The second component of the knowledge base stores a set of grammar rules. More precisely, a lightweight grammar describes how certain concepts 28 Figure 1: Software Architecture may be modified by prepositions, adverbial or adjectival structures that are also specified in the synonym ontology. For a more detailed description we refer to Berger (2001). 2.1.2 Language Identification To identify the language of a query, an n-gram-based text classification approach (cf. Cavnar & Trenkle (1994)) is used. An n-gram is an n-character slice of a longer character string. As an example, for n = 3, the trigrams of the string "language" are: { la, lan, ang, ngu, gua, uag, age, ge }. Dealing with multiple words in a string, the blank character is usu-ally replaced by an underscore " " and is also taken into account for the construction of an n-gram document representation. This language classification approach using n-grams requires sample texts for each language to build statistical models, i.e. n-gram frequency profiles, of the languages. We used various tourism-related texts, e.g. hotel descriptions and holiday package descriptions, as well as news articles both in English and German language. The n-grams, with n ranging from 1...5, of these sample texts were an-alyzed and sorted in descending order according to their frequency, separately for each language. These sorted histograms are the n-gram frequency profiles for a given language. For a comprehensive description see Berger, Dittenbach & Merkl (2003). 2.1.3 Error Correction To improve retrieval performance, potential orthographic errors have to be considered in the web-based interface. After identifying the language, a spell-checking module is used to determine the correctness of query terms. The efficiency of the spell checking process improves during the runtime of the system by learning from previous queries. The spell checker uses the metaphone algorithm (cf. Philips (1990)) to transform the words into their soundalikes. Because this algorithm has originally been developed for the English language, the rule set defining the mapping of words to the phonetic code has to be adapted for other languages. In addition to the base dictionary of the spell checker, domain-dependent words and proper names like names of cities, regions or states, have to be added to the dictionary. For every misspelled term of the query, a list of potentially correct words is returned. First, the misspelled word is mapped to its metaphone equivalent, then the words in the dictionary , whose metaphone translations have at most an edit distance (cf. Levenshtein (1966)) of two, are added to the list of suggested words. The suggestions are ranked according to the mean of first, the edit distance between the misspelled word and the suggested word, and second, the edit distance between the misspelled word's metaphone and the suggested word's. The smaller this value is for a suggestion, the more likely it is to be the correct substitution from the orthographic or phonetic point of view. However, this ranking does not take domain-specific information into account. Because of this deficiency, correctly spelled words in queries are stored and their respective number of occurrences is counted. The words in the suggestion list for a misspelled query term are looked up in this repository and the suggested word having the highest number of occurrences is chosen as the replacement of the erroneous original query term. In case of two or more words having the same number of occurrences the word that is ranked first is selected. If the query term is not present in the repository up to this moment, it is replaced by the first suggestion, i.e. the word being phonetically or orthographically closest. Therefore, suggested words that are very similar to the misspelled word, yet make no sense in the context of the application domain, might be rejected as replacements. Consequently, the word correction process described above is improved by dynamic adaptation to past knowledge. Another important issue in interpreting the natural language query is to detect terms consisting of multiple words. Proper names like "Bad Kleinkirch-heim" or nouns like "parking garage" have to be treated as one element of the query. Therefore, all multi-word denominations known to the system are stored in an efficient data structure allowing to identify such cases. More precisely, regular expressions are used to describe rules applied during the identification process. 2.1.4 SQL Mapping With the underlying relational database management system PostgreSQL, the natural language query has to be transformed into a SQL statement to retrieve 29 the requested information. As mentioned above the knowledge base describes parameterized SQL fragments that are used to build a single SQL statement representing the natural language query. The query terms are tagged with class information, i.e. the relevant concepts of the domain (e.g. "hotel" as a type of accommodation or "sauna" as a facility provided by a hotel), numerals or modifying terms like "not", "at least", "close to" or "in". If none of the classes specified in the ontology can be applied, the database tables containing proper names have to be searched. If a noun is found in one of these tables, it is tagged with the respective table's name, such that "Tyrol" will be marked as a federal state. In the next step, this class information is used by the grammar to select the appropriate SQL fragments. Finally, the SQL fragments have to be combined to a single SQL statement reflecting the natural language query of the user. The operators combining the SQL fragments are again chosen according to the definitions in the grammar. Associative Networks Quillian (1968) introduced the basic principle of a semantic network and it played, since then, a central role in knowledge representation. The building blocks of semantic networks are, first, nodes that express knowledge in terms of concepts, second, concept properties, and third,the hierarchical sub-super class relationship between these concepts. Each concept in a semantic network represents a semantic entity. Associations between concepts describe the hierarchical relationship between these semantic entities via is-a or instance-of links. The higher a concept moves up in the hierarchy along is-a relations, the more abstract is its semantic meaning . Properties are attached to concepts and, therefore , properties are also represented by concepts and linked to nodes via labeled associations. Furthermore, a property that is linked to a high-level concept is inherited by all descendants of the concept. Hence, it is assumed that the property applies to all subsequent nodes. An example of a semantic network is depicted in Figure <A href="28.html#5">2. Semantic networks initially emerged in cognitive psychology and the term itself has been used in the field of knowledge representation in a far more general sense than described above. In particular, the term semantic network has been commonly used to refer to a conceptual approach known as associative network. An associative network defines a generic network which consists of nodes representing information items (semantic entities) and associations between nodes, that express, not necessarily defined or labeled, relations among nodes. Links between particular nodes might be weighted to determine the strength of connectivity. 3.1 Spreading Activation A commonly used technique, which implements information retrieval on semantic or associative networks, is often referred to as spreading activation. The spreading activation processing paradigm is tight-knit with the supposed mode of operation of human memory . It was introduced to the field of artificial intelligence to obtain a means of processing semantic or associative networks. The algorithm, which underlies the spreading activation (SA) paradigm, is based on a quite simple approach and operates on a data structure that reflects the relationships between information items. Thus, nodes model real world entities and links between these nodes define the relatedness of entities. Furthermore, links might possess , first, a specific direction, second, a label and, third, a weight that reflects the degree of association. This conceptual approach allows for the definition of a more general, a more generic network than the basic structure of a semantic network demands. Nevertheless , it could be used to model a semantic network as well as a more generic one, for instance an associative network. The idea, underlying spreading activation, is to propagate activation starting from source nodes via weighted links over the network. More precisely, the process of propagating activation from one node to adjacent nodes is called a pulse. The SA algorithm is based on an iterative approach that is divided into two steps: first, one or more pulses are triggered and, second, a termination check determines if the process has to continue or to halt. Furthermore, a single pulse consists of a pre-adjustment phase, the spreading process and a post-adjustment phase. The optional pre- and post-adjustment phases might incorporate a means of activation decay, or to avoid reactivation from previous pulses. Strictly speaking, these two phases are used to gain more control over the network. The spreading phase implements propagation of activation over the network. Spreading activation works according to the formula: I j (p) = k i (O i (p - 1) w ij ) (1) Each node j determines the total input I j at pulse p of all linked nodes. Therefore, the output O i (p - 1) at the previous pulse p - 1 of node i is multiplied with the associated weight w ij of the link connecting node i to node j and the grand total for all k connected nodes is calculated. Inputs or weights can be expressed by binary values (0/1), inhibitory or reinforcing values (-1/+1), or real values defining the strength of the connection between nodes. More precisely, the first two options are used in the application of semantic networks, the latter one is commonly used for associative networks. This is due to the fact that the type of association does not necessarily have some exact semantic meaning. The weight rather describes the relationship between nodes. Furthermore, the output value of a node has to be determined. In most cases, no distinction is made between the input value and the activation level of a node, i.e. the input value of a node and its activation level are equal. Before firing the activation to adjacent nodes a function calculates the output depending on the activation level of the node: O i = f (I i ) (2) Various functions can be used to determine the output value of a node, for instance the sigmoid function , or a linear activation function, but most commonly used is the threshold function. The threshold function determines, if a node is considered to be active or not, i.e. the activation level of each node is compared to the threshold value. If the activation level exceeds the threshold, the state of the node is set to active. Subsequent to the calculation of the activation state, the output value is propagated to adjacent nodes. Normally, the same output value is sent to all adjacent nodes. The process described above is repeated, pulse after pulse, and activation spreads through the network and activates more and more nodes until a termination condition is met. Finally , the SA process halts and a final activation state is obtained. Depending on the application's task the 30 accomodation hotel animal farm pension pig sheep facility steam bath sauna hot is_a is_a is_a is_a is_a is_a is_a has offers is is Figure 2: A semantic network example of tourism-related terms activation levels are evaluated and interpreted accordingly . 3.2 Taming Spreading Activation Unfortunately, the basic approach of spreading activation entails some major drawbacks. Strictly speaking , without appropriate control, activation might be propagated all over the network. Furthermore, the semantics of labeled associations are not incorporated in SA and it is quite difficult to integrate an inference mechanism based on the semantics of associations . To overcome these undesired side-effects the integration of constraints helps to tame the spreading process (cf. Crestani (1997)). Some constraints commonly used are described as follows. Fan-out constraint: Nodes with a broad semantic meaning possess a vast number of links to adjacent nodes. This circumstance implies that such nodes activate large areas of the network. Therefore, activation should diminish at nodes with a high degree of connectivity to avoid this unwanted effect. Distance constraint: The basic idea underlying this constraint is, that activation ceases when it reaches nodes far away from the activation source. Thus, the term far corresponds to the number of links over which activation was spread, i.e. the greater the distance between two nodes, the weaker is their semantic relationship. According to the distance of two nodes their relation can be classified. Directly connected nodes share a first order relation. Two nodes connected via an intermediate node are associated by a second order relation, and so on. Activation constraint: Threshold values are assigned to nodes (it is not necessary to apply the same value to all nodes) and are interpreted by the threshold function. Moreover, threshold values can be adapted during the spreading process in relation to the total amount of activity in the network. Path constraint: Usually, activation spreads over the network using all available links. The integration of preferred paths allows to direct activation according to application-dependent rules. Another enhancement of the spreading activation model is the integration of a feedback process. The activation level of some nodes or the entire network is evaluated by, for instance, another process or by a user. More precisely, a user checks the activation level of some nodes and adapts them according to her or his needs. Subsequently, activation spreads depending on the user refinement. Additionally, users may indicate preferred paths for spreading activation and, therefore, are able to adapt the spreading process to their own needs. Recommendation via Spreading Activation One of the first information retrieval systems using constrained spreading activation was GRANT. Kjeldsen & Cohen (1987) developed a system that handles information about research proposals and potential funding agencies. GRANT's domain knowledge is stored in a highly associated semantic network. The search process is carried out by constrained spreading activation over the network. In particular, the system extensively uses path constraints in the form of path endorsement. GRANT can be considered as an inference system applying repeatedly the same inference schema: IF x AND R(x, y) y (3) R(x, y) represents a path between two nodes x and y. This inference rule can be interpreted as follows: "if a founding agency is interested in topic x and there is a relation between topic x and topic y then the founding agency might be interested in the related topic y." Croft, Lucia, Crigean & Willet (1989) developed an information retrieval system initially intended to study the possibility of retrieving documents by plausible inference. In order to implement plausible inference constrained spreading activation was chosen accidently. The I 3 R system acts as a search intermediary (cf. Croft & Thompson (1987)). To accomplish this task the system uses domain knowledge to refine user queries, determines the appropriate search strategy, assists the user in evaluating the output and reformulating the query. In its initial version, the domain knowledge was represented using a tree structure of concepts. The design was later refined to meet the requirements of a semantic network. Belew (1989) investigated the use of connectionist techniques in an information retrieval system called Adaptive Information Retrieval (AIR). The system handles information about scientific publications , like the publication title and the author. AIR uses a weighted graph as knowledge representation paradigm. For each document, author and keyword (keywords are words found in publication titles) a node is created and associations between nodes are constructed from an initial representation of documents and attributes. A user's query causes initial activity to be placed on some nodes of the network. 31 This activity is propagated to other nodes until certain conditions are met. Nodes with the highest level of activation represent the answer to the query by the AIR system. Furthermore, users are allowed to assign a degree of relevance to the results (++, +, -, --). This causes new links to be created and the adaptation of weights between existing links. Moreover, feedback is averaged across the judgments of many users. A mentionable aspect of the AIR system is that no provision is made for the traditional Boolean operators like AND and OR. Rather, AIR emulates these logical operations because "the point is that the difference between AND and OR is a matter of degree". This insight goes back to Von Neumann (as pointed out by Belew (1989)). A system based on a combination of an ostensive approach with the associative retrieval approach is described in Crestani & Lee (2000). In the WebSCSA (Web Searching by Constrained Spreading Activation ) approach a query does not consist of keywords. Instead, the system is based on an ostensive approach and assumes that the user has already identified relevant Web pages that act as a basis for the following retrieval process. Subsequently, relevant pages are parsed for links and they are followed to search for other relevant associated pages. The user does not ex-plicitly refine the query. More precisely, users point to a number of relevant pages to initiate a query and the WebSCSA system combines the content of these pages to build a search profile. In contrast to conventional search engines WebSCSA does not make use of extensive indices during the search process. Strictly speaking , it retrieves relevant information only by navigating the Web at the time the user searches for information . The navigation is processed and controlled by means of a constrained spreading activation model. In order to unleash the power of WebSCSA the system should be used when users already have a point to start for her or his search. Pragmatically speaking, the intention of WebSCSA is to enhance conventional search engines, use these as starting points and not to compete with them. Hartmann & Strothotte (2002) focus on a spreading activation approach to automatically find associations between text passages and multimedia material like illustrations, animations, sounds, and videos. Moreover, a media-independent formal representation of the underlying knowledge is used to automatically adapt illustrations to the contents of small text segments . The system contains a hierarchical representation of basic anatomic concepts such as bones, muscles , articulations, tendons, as well as their parts and regions. Network structures provide a flexible model for adaptation and integration of additional information items. Nevertheless, Crestani (1997) points out that "... the problem of building a network which effec-tively represents the useful relations (in terms of the IRs aims) has always been the critical point of many of the attempts to use SA in IR. These networks are very difficult to build, to maintain and keep up to date. Their construction requires in depth application domain knowledge that only experts in the application domain can provide." Dittenbach, Merkl & Berger (2003) present an approach based on neural networks for organizing words of a specific domain according to their semantic relations . A two-dimensional map is used to display semantically similar words in spatially regions. This representation can support the construction and enrichment of information stored in the associative network . 4.1 The Redeveloped System Architecture To overcome the limitations of the knowledge base of the original system, the development of an alternative approach to model domain knowledge was necessary . Basically, the unassociated, non-hierarchic knowledge representation model inhibits the power of the system. Strictly speaking, the original system failed to retrieve results on a fuzzy basis, i.e. the results determined by the system provide exact matches only, without respect to first, interactions users made during past sessions, second, personal preferences of users, third, semantic relations of domain intrinsic information , and fourth, locational interdependencies. In order to adapt the system architecture accordingly , an approach based on associative networks was developed. This associative network replaces the flat synonym ontology used in the original system. Moreover , both the grammar rules and the SQL fragments have been removed from the knowledge base. More precisely, the functionality and logic is now covered by newly developed pipeline elements or implic-itly resolved by the associative network. In analogy to the original pipeline, the first three processing steps are accomplished. Next, a newly implemented initializationelement associates concepts extracted from the query with nodes of the associative network. These nodes act as activation sources. Subsequently , the newly designed spreading element implements the process of activation propagation. Finally , the new evaluationelement analyzes the activation level of the associative network determined during the spreading phase and produces a ranking according to this activation level. 4.1.1 The Knowledge Representation Model Basically, the knowledge base of the information retrieval system is composed of two major parts: first, a relational database that stores information about domain entities and, second, a data structure based on an associative network that models the relationships among terms relevant to the domain. Each domain entity is described by a freely definable set of attributes . To provide a flexible and extensible means for specifying entity attributes, these attributes are organized as &lt;name, value&gt; pairs. An example from the tourism domain is depicted in Table <A href="28.html#6">1. Hotel Wellnesshof category 4 facility sauna facility solarium facility ... Table 1: &lt;name,value&gt;-pair example for entity "Hotel Wellnesshof " The associative network consists of a set of nodes and each node represents an information item. Moreover , each node is member of one of three logical layers defined as follows: Abstraction layer: One objective of the rede-velopment of the knowledge base was to integrate information items with abstract semantic meaning . More precisely, in contrast to the knowledge base used in the original system which only supported modeling of entity attributes, the new approach allows the integration of a broader set of terms, e.g. terms like "wellness" or "summer activities" that virtually combine several information items. 32 Conceptual layer: The second layer is used to associate entity attributes according to their semantic relationship. Thus, each entity attribute has a representation at the conceptual layer. Furthermore , the strengths of the relationships between information items are expressed by a real value associated with the link. Entity layer: Finally, the entity layer associates entities with information items (entity attributes ) of the conceptual layer, e.g. an entity possessing the attribute "sauna" is associated with the saunanode of the conceptual layer. The building blocks of the network are concepts. A concept represents an information item possessing several semantically equivalent terms, i.e. synonyms, in different languages. Each concept possesses one of three different roles: Concrete concepts are used to represent information items at the conceptual layer. More precisely, concrete concepts refer to entity attributes . Concepts with an abstract role refer to terms at the abstraction layer. Finally, the modifier role is used to categorize concepts that alter the processing rules for abstract or concrete concepts. A modifier like, for instance, "not" allows the exclusion of concepts by negation of the assigned initialization value. Moreover, concepts provide, depending on their role, a method for expressing relationships among them. The connectedTo relation defines a bidirec-tional weighted link between two concrete concepts, e.g. the concrete concept "sauna" is linked to "steam bath". The second relation used to link information items is the parentOf association. It is used to express the sub-super class relationship between abstract concepts or concrete and abstract concepts. A set of concepts representing a particular domain is described in a single XML file and acts as input source for the information retrieval system. During initialization, the application parses the XML file, in-stantiates all concepts, generates a list of synonyms pointing at corresponding concepts, associates concepts according to their relations and, finally, links the entities to concrete concepts. Currently, the associative network consists of about 2,200 concepts, 10,000 links and more than 13,000 entities. The concept network includes terms that describe the tourism domain as well as towns, cities and federal states throughout Austria. To get a better picture of the interdependencies associated with the layers introduced above see Figure <A href="28.html#8">3. Each layer holds a specific set of concepts. Abstract concepts associate concepts at the same or at the conceptual layer. Concepts at the conceptual layer define links between entity attributes and associate these attributes with entities at the entity layer. Finally, entities are placed at the lowest layer, the entity layer. Concepts at the entity layer are not associated with items at the same layer. Consider, for example, the abstract concept "indoor sports" and the concept "sauna" as concepts from which activation originates from. First, activation is propagated between the abstraction layer to the conceptual layer via the dashed line from "indoor sports" to "table tennis" . We shall note, that dashed lines indicate links between concepts of different layers. Thus, "sauna" and "table tennis" act as source concepts and, moreover , activation is spread through the network along links at the conceptual layer. Activation received by concepts at the conceptual layer is propagated to the entities at the entity layer stimulating, in this particular case, the entities "Hotel Stams", "Hotel Thaya" as well as "Wachauerhof ". Moreover, a fraction of activation is propagated to adjacent concept nodes at the conceptual layer, i.e. "solarium", "whirlpool" as well as "tennis", and to entities, i.e. "Hotel Wiental" and "Forellenhof ", respectively. 4.1.2 Processing the Associative Network Due to the flexibility and adaptivity of the original system, the integration of the redesigned parts has been accomplished with relatively little effort. In particular , the existing knowledge base has been replaced by the associative network and additional pipeline elements to implement spreading activation have been incorporated. Figure <A href="28.html#9">4 depicts the redeveloped knowledge base on which the processing algorithm operates. The conceptual layer stores concrete concepts and the weighted links among them. Associating abstract concepts with concrete concepts is done at the abstraction layer. Each entity has a unique identifier that is equivalent to the entity identifier stored in the relational database. Furthermore, entities are connected to concepts at the conceptual layer. More precisely, an entity is connected to all attributes it possesses. As an example consider the entity "Hotel Stams" as depicted in Figure <A href="28.html#9">4. This hotel offers a "sauna", a "steam bath" and a "solarium" and is, therefore, linked to the corresponding concepts at the conceptual layer. First, a user's query, received by the information retrieval system, is decomposed into single terms. After applying an error correction mechanism and a phrase detection algorithm to the query, terms found in the synonym lexicon are linked to their corresponding concept at the abstraction or conceptual layer. These concepts act as activation sources and, subsequently , the activation process is initiated and activation spreads according to the algorithm outlined below. At the beginning, the role of each concept is evaluated . Depending on its role, different initialization strategies are applied: Modifier role: In case of the "not" modifier, the initialization value of the subsequent concept is multiplied with a negative number. Due to the fact that the "and" and "or" modifiers are im-plicitly resolved by the associative network, they receive no special treatment. More precisely, if, for instance, somebody is searching for an accommodation with a sauna or solarium, those accommodations offering both facilities will be ranked higher than others, providing only one of the desired facilities. Furthermore, the "near" modifier reflecting geographic dependencies, is automatically resolved by associating cities or towns within a circumference of 15km. Depending on the distance, the weights are adapted accordingly , i.e. the closer they are together, the higher is the weight of the link in the associative network . Abstract role: If a source concept is abstract, the set of source concepts is expanded by resolving the parentOf relation between parent and child concepts. This process is repeated until all abstract concepts are resolved, i.e. the set of source concepts contains members of the conceptual layer only. The initial activation value is propagated to all child concepts, with respect to the weighted links. 33 Figure 3: Network layer interdependencies Concrete role: The initial activation level of concrete concepts is set to initialization value defined in the XML source file. The spreading activation process takes place at the conceptual layer, i.e. the connectedTo relations between adjacent concepts are used to propagate activation through the network. After the initialization phase has completed, the iterative spreading process is activated. During a single iteration one pulse is performed, i.e. the number of iterations equals the number of pulses. Starting from the set of source concepts determined during initialization , in the current implementation activation is spread to adjacent nodes according to the following formula: O i (p) = 0 if I i (p) &lt; , F i p+1 I i (p) else, with F i = (1 C i C T ) (4) The output, O i (p), sent from node i at pulse p, is calculated as the fraction of F i , which limits the propagation according to the degree of connectivity of node i (i.e. fan-out constraint, cf. Section <A href="28.html#5">3.2), and p + 1, expressing the diminishing semantic relationship according to the distance of node i to activation source nodes (i.e. distance constraint, cf. Section <A href="28.html#5">3.2). Moreover, F i is calculated by dividing the number of concepts C i directly connected to node i by the total number of nodes C T building the associative network. Note, represents a threshold value. Simultaneous to calculating the output value for all connected nodes, the activation level I i (p) of node i is added to all associated entities. More precisely, each entity connected to node i receives the same value and adds it to an internal variable representing the total activation of the entity. As an example, if the concept node "sauna" is activated, the activation potential is propagated to the entities "Hotel Stams" and "Hotel Thaya" (cf. Figure <A href="28.html#9">4). Next, all newly activated nodes are used in the subsequent iteration as activation sources and the spreading process continues until the maximum number of iterations is reached. After the spreading process has terminated, the system inspects all entities and ranks them according to their activation. Figure <A href="28.html#10">5 depicts the results determined for the example query Ich und meine Kinder m ochten in einem Hotel in Kitzb uhel Urlaub machen. Es sollte ein Dampfbad hab<A href="28.html#8">en. <A href="28.html#8">1 In this particular case, the entities "Schwarzer Adler Kitzb uhel" and "Hotel Schloss Lebenberg Kitzb uhel" located in "Kitzb uhel" are suggested to be the best matching answers to the query. Moreover, the result set includes matches that are closely related to the user's query. Thus, depending on the relations stored in the associative network, entities offering related concepts are activated accordingly. More precisely , not only the attributes "hotel", "steam bath" and "kids" are taken into account, but also all other related entity attributes (e.g. "sauna", "whirlpool", "solarium", etc.) have some influence on the ranking position. Furthermore, accommodations in cities in the vicinity of "Kitzb uhel" providing the same or even better offers are also included in the result set. Thus, the associative network provides a means for exact information retrieval and incorporates a fuzzy search strategy that determines closely related matches to the user's query. 1 Me and my kids would like to spend our holidays in a hotel in Kitzbhel. It should have a steam bath. 34 Figure 4: Knowledge base architecture Conclusion A natural language system based on an approach described in Berger (2001) and Berger, Dittenbach, Merkl & Winiwarter (2001) has been reviewed in this paper and, furthermore, provided the basis for the research presented herein. The reviewed system offers multilingual access to information on a restricted domain . In this particular case the system operates on the tourism domain. Moreover, users of the search interface are encouraged to formulate queries in natural language, i.e. they are able to express their intentions in their own words. We developed a knowledge representation model that facilitates the definition of semantic relations between information items exemplified by terms of the tourism domain. In particular, an associative network based on a three layered structure was introduced . First, the abstraction layer allows modelling of terms with a subjective or broader semantic meaning , second, the conceptual layer is used to define relations via weighted links between terms, and, finally , the entity layer provides a means to associate elements stored in a relational database with information items in the associative network. Moreover, a constrained spreading activation algorithm implements a processing technique operating on the network . Generally, the combination of the associative nature of the knowledge representation model and the constrained spreading activation approach constitutes a search algorithm that evaluates the relatedness of terms and, therefore, provides a means for implicit query expansion. The flexible method of defining relationships between terms unleashes the ability to determine highly associated results as well as results that are predefined due to personal preferences. Moreover, especially designed associative networks can be used to model scenarios , as, for instance, a winter holiday scenario that favors accommodations offering winter sports activities by adapting the weights of links accordingly. One important task for further enhancement is the possibility to express the relevance of query terms. Users should be able to assign a degree of significance to terms. Consider, for example, a user searching for an accommodation with several amenities in the capital city of Austria. Moreover, the user is a vegetarian. Therefore, a means for expressing the importance of vegetarian kitchen is needed. In order to accomplish this requirement, the system might be extended to understand words that emphasis terms, e.g. in analogy to modifiers like "and", "or", "near", etc. the word "important" is handled like a modifier and influences the activation level of the following query term. Additionally, an interface providing a graphical instrument to express relevance by means of a slide controller might be considered. Furthermore, an associative network might act as a kind of short term memory. More precisely, during a user session a particular network is used to store the activation level determined during past user interactions . A user, for instance, is searching for a hotel in Vienna. Thus, the associative network stores the activation level for further processing. Next, the user might restrict the results to accommodations offering a sauna. This spreading process is carried out using the associative network determined during the previous interaction. References Baeza-Yates, R. A. & Ribeiro-Neto, B. (1999), Modern Information Retrieval, Addison-Wesley, Reading, MA. Ballesteros, L. & Croft, W. B. (1998), Resolving ambiguity for cross-language retrieval, in `Re-search and Development in Information Retrieval' , pp. 6471. Belew, R. K. (1989), Adaptive information retrieval: Using a connectionist representation to retrieve and learn about documents, in N. J. Nicholas J. Belkin & C. J. Van Rijsbergen, eds, `Proceedings of the 12th International Conference on Research and Development in Information Retrieval (SIGIR'89)', ACM, pp. 1120. Berger, H. (2001), Adaptive multilingual interfaces, Master's thesis, Vienna University of Technology . Berger, H., Dittenbach, M. & Merkl, D. (2003), Querying tourism information systems in natural language, in `Proceedings of the 2nd International Conference on Information System Technology and its Applications (ISTA 2003)', Kharkiv, Ukraine. Berger, H., Dittenbach, M., Merkl, D. & Winiwarter, W. (2001), Providing multilingual natural language access to tourism information, in W. Winiwarter , S. Bressan & I. K. Ibrahim, eds, `Proceedings of the 3rd International Conference on 35 Figure 5: Weighted result set determined by constrained spreading activation Information Integration and Web-based Applications and Services (IIWAS 2001)', Austrian Computer Society, Linz, Austria, pp. 269276. Cavnar, W. B. & Trenkle, J. M. (1994), N-gram-based text categorization, in `International Symposium on Document Analysis and Information Retrieval', Las Vegas, NV. Crestani, F. (1997), `Application of spreading activation techniques in information retrieval', Artificial Intelligence Review 11(6), 453582. Crestani, F. & Lee, P. L. (2000), `Searching the web by constrained spreading activation', Information Processing and Management 36(4), 585 605. Croft, W., Lucia, T., Crigean, J. & Willet, P. (1989), `Retrieving documents by plausible inference: an experimental study', Information Processing & Management 25(6), 599614. Croft, W. & Thompson, R. H. (1987), `I 3 R: A New Approach to the Design of Document Retrieval Systems', Journal of the American Society for Information Science 38(6), 389404. Dittenbach, M., Merkl, D. & Berger, H. (2003), Using a connectionist approach for enhancing domain ontologies: Self-organizing word category maps revisited, in `Proceedings of the 5th International Conference on Data Warehousing and Knowledge Discovery - (DaWaK 2003)'. Accepted for publication. Harman, D. K. (1995), Overview of the 3rd Text Retrieval Conference (TREC-3), in D. K. Harman , ed., `Proceedings of the 3rd Text Retrieval Conference (TREC-3)', NIST Special Publication 500225, pp. 119. Hartmann, K. & Strothotte, T. (2002), A spreading activation approach to text illustration, in `Proceedings of the 2nd International Symposium on Smart graphics', ACM Press, pp. 3946. Hull, D. A. & Grafenstette, G. (1996), Querying across languages: A dictionary-based approach to multilingual information retrieval, in `Proceedings of ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 1996)', pp. 4957. Kjeldsen, R. & Cohen, P. (1987), `The evolution and performance of the GRANT system', IEEE Expert pp. 7379. Levenshtein, V. I. (1966), `Binary codes capable of correcting deletions, insertions and reversals', Soviet Physics Doklady 10(8), 707710. Philips, L. (1990), `Hanging on the metaphone', Computer Language Magazine 7(12). Quillian, M. R. (1968), Semantic memory, in M. Min-sky , ed., `Semantic Information Processing', MIT Press, pp. 227270. Salton, G. (1989), Automatic Text Processing: The Transformation, Analysis, and Retrieval of Information by Computer, Addison-Wesley, Reading , MA. Salton, G. & McGill, M. J. (1983), Introduction to Modern Information Retrieval, McGraw-Hill, New York. Van Rijsbergen, C. J. (1979), Information Retrieval, Department of Computer Science, University of Glasgow. Xu, F., Netter, K. & Stenzhorn, H. (2000), Mietta a framework for uniform and multilingual access to structured database and web information, in `Proceedings of the 5th International Workshop on Information Retrieval with Asian languages'. 36

natural language information retrieval;constrained spreading activation;query expansion;spreading activation;multilingual information retrieval system;knowledge representation model;associative networks;knowledge representation;natural language query

29

An Analytical Model Based on G/M/1 with Self-Similar Input to Provide End-to-End QoS in 3G Networks

The dramatic increase in demand for wireless Internet access has lead to the introduction of new wireless architectures and systems including 3G, Wi-Fi and WiMAX. 3G systems such as UMTS and CDMA2000 are leaning towards an all-IP architecture for transporting IP multimedia services, mainly due to its scalability and promising capability of inter-working heterogeneous wireless access networks. During the last ten years, substantial work has been done to understand the nature of wired IP traffic and it has been proven that IP traffic exhibits self-similar properties and burstiness over a large range of time scales. Recently, because of the large deployment of new wireless architectures, researchers have focused their attention towards understanding the nature of traffic carried by different wireless architecture and early studies have shown that wireless data traffic also exhibits strong long-range dependency. Thus, the classical tele-traffic theory based on a simple Markovian process cannot be used to evaluate the performance of wireless networks. Unfortunately, the area of understanding and modeling of different kinds of wireless traffic is still immature which constitutes a problem since it is crucial to guarantee tight bound QoS parameters to heterogeneous end users of the mobile Internet. In this paper, we make several contributions to the accurate modeling of wireless IP traffic by presenting a novel analytical model that takes into account four different classes of self-similar traffic. The model consists of four queues and is based on a G/M/1 queueing system. We analyze it on the basis of priority with no preemption and find exact packet delays. To date, no closed form expressions have been presented for G/M/1 with priority.

INTRODUCTION During the past decade, researchers have made significant efforts to understand the nature of Internet traffic and it has been proven that Internet traffic exhibits self-similar properties. The first study, which stimulated research on self-similar traffic, was based on measurements of Ethernet traffic at Bellcore [1]. Subsequently, the self-similar feature has been discovered in many other types of Internet traffic including studies on Transmission Control Protocol (TCP) [2, 3], WWW traffic [4], VBR video [5] and Signaling System No 7 [6]. Deeper studies into the characteristics of Internet traffic has discovered and investigated various properties such as self-similarity [7], long-range dependence [8] and scaling behavior at small time-scale [9]. The references [10, 11] provide two extensive bibliographies on self-similarity and long-range dependence research covering both theoretical and applied papers on the subject. Concurrently, over the past few years, we have witnessed a growing popularity of Third Generation Systems (3G), which have been designed to provide high-speed data services and multimedia applications over mobile personal communication networks. The Universal Mobile Telecommunication System (UMTS) is the predominant global standard for 3G developed by Third Generation Partnership Project (3GPP) [12]. The UMTS architecture is shown in Fig. 1. It consists of two service domains, a Circuit Switched (CS) service domain and a Packet Switched (PS) service domain, which is of interest in this paper. In the PS service domain, a UMTS network connects to a public data network (PDN) through Serving GPRS Support node (SGSN) and Gateway GPRS support node (GGSN). 3GPP has defined four different QoS classes for UMTS; (1) Conversational (2) Interactive (3) Streaming and (4) Background, conversational being the most delay-sensitive and background the least delay sensitive class [12]. With the increasing demand of Internet connectivity and the flexibility and wide deployment of IP technologies, there has emerged a paradigm shift towards IP-based solutions for wireless networking [13]. Several Wireless IP architectures have been proposed [17-23] based on three main IP QoS models, IntServ [14], DiffServ [15] and MPLS [16]. 3GPP has also recently introduced a new domain called IP Multimedia Subsystem (IMS) for UMTS. The main objective of IMS is to deliver innovative 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. MobiWAC'06, October 2, 2006, Torremolinos, Malaga, Spain. Copyright 2006 ACM 1-59593-488-X/06/0010...$5.00. 180 Fig. 1: A Simplified UMTS Network Architecture and cost-effective services such as IP telephony, media streaming and multiparty gaming by providing IP connectivity to every mobile device [24]. In the light of this, researchers have recently focused on understanding the nature of wireless IP traffic and early studies have shown that wireless data traffic also exhibits self-similarity and long-range dependency [25-28]. Much of the current understanding of wireless IP traffic modeling is based on the simplistic Poisson model, which can yield misleading results and hence poor wireless network planning. Since the properties and behavior of self-similar traffic is very different from traditional Poisson or Markovian traffic, several issues need to be addressed in modeling wireless IP traffic to provide end-to-end QoS to a variety of heterogeneous applications. We begin by giving an overview of related work on wired and wireless IP traffic modeling along with a comparison of our model with previous work. RELATED WORK In this section, we first discuss the related work which has been done in the area of performance evaluation of wired IP and Wireless IP networks under self-similar input and then we compare our model with the previous ones. 2.1 Previous Work on IP Traffic Modeling There has been much work done on Internet traffic modeling based on queueing theory in the presence of self-similar traffic [29-34]. In [33], a Matrix Geometric (analytical) method is used to compute numerical results for a two class DiffServ link being fed by a Markovian Modulated Poisson Process (MMPP) input. A weakness of this model is that MMPP may require an estimation of a large number of parameters. An OPNET based simulation approach was adopted in [34] to see the impact of self-similarity on the performance evaluation of DiffServ networks. As a result, an idea of expected queue length was given in relation to the Hurst parameter and server utilization. It is difficult to offer guaranteed QoS parameters on the basis of such analysis. The major weakness of the majority of available queueing based results is that only the FIFO queueing discipline has been considered for serving the incoming traffic and thus differential treatment to different kinds of traffic can not be provided. In addition, the previous results are asymptotic. We also refer the readers to [35-39] for an overview of previous work that has been carried out to evaluate the performance of IP networks. The major drawback of the existing work is that, the queueing models considered are not able to capture the self-similar characteristics of Internet traffic. Furthermore, it is important to note that most of the previous work is focused on the analysis of one type of traffic only without discussing its affect on the performance of other kinds of network traffic. 2.2 Previous Work on Wireless IP Traffic Modeling Few studies have focused on wireless traffic modeling and here we discuss the most relevant work. As shown in Fig. 1, the principle of allocation of data flows between end users and GGSN leads to increasing load on the network elements when moving closer to the GGSN. Hence, GGSN is the node most exposed to self-similar influence in UMTS [40]. The influence of self-similar input on GGSN performance in the UMTS Release 5 IM-subsystem has been analyzed on the basis of a FBM/D/1/W queueing system (FBM-Fractional Brownian Motion) in [40]. In this work, different probabilistic parameters of GGSN such as average queue length and average service rate were also found. The work in [41] presents modeling and a simulation study of the Telus Mobility (a commercial service provider) Cellular Digital Packet Data (CDPD) network. The collected results on average queueing delay and buffer overflow probability indicated that genuine traffic traces 181 produce longer queues as compared to traditional Poisson based traffic models. To get an overview of the analysis done in wireless IP traffic modeling with self-similar input, we refer the readers to [42-45]. These studies are merely based on characterization of wireless traffic. To provide differential treatment to multiple traffic classes with different QoS demands, there is a need to accurately determine end-to-end QoS parameters such as delay, jitter, throughput, packet loss, availability and per-flow sequence preservation. 2.3 Comparison of our Model with Prior Work To overcome the limitations of the previous work in traffic modeling (wired and wireless IP traffic), we present a realistic and novel analytical model by considering four different classes of traffic that exhibit long-range dependence and self-similarity. Our model implements four queues based on a G/M/1 queueing system and we analyze it on the basis of priority with no preemption. The traffic model considered is parsimonious with few parameters and has been studied in [46]. The model is furthermore similar to on/off processes, in particular to its variation N-Burst model studied in [47] where packets are incorporated. However, only a single type of traffic is considered in [47]. We present a novel analytical approach and make the following contributions to Wireless IP traffic modeling. Interarrival Time Calculations: We calculate the packet interarrival time distributions for the particular self-similar traffic model [46] for the first time in this paper. The distribution of cross interarrival time between different types of packets is derived on the basis of single packet results. Packet Delays for Multiple Self-Similar Traffic Classes: We consider a G/M/1 queueing system which takes into account four different classes of self-similar input traffic denoted by SS/M/1 and analyze it on the basis of non preemptive priority and find exact packet delays. To date, no closed form expressions have been presented for G/M/1 with priority. Embedded Markov Chain Formulation: We also formulate the embedded Markov chain of G/M/1 by considering all possible states and derive the corresponding transition probabilities. The rest of the paper is organized as follows. Section 3 and 4 are devoted to explaining the self-similar traffic model with multiple classes and the calculation of interarrival times respectively. Section 5 explains the procedure of formulating the embedded Markov Chain along with the derivation of packet delays. The applications of the model are discussed in section 6. Finally, conclusion and future work is given in Section 7. TRAFFIC MODEL The traffic model considered here [46] belongs to a particular class of self-similar traffic models also called telecom process in [48], recently. The model captures the dynamics of packet generation while accounting for the scaling properties of the traffic in telecommunication networks. Such models, also called infinite source models, are similar to on/off processes with heavy tailed on and/or off times. What is more, our model abstracts the packet arrival process in particular and facilitates queueing analysis by the approaches developed in the sequel. In the framework of a Poisson point process, the model represents an infinite number of potential sources. The traffic is found by aggregating the number of packets generated by such sources. Each source initiates a session with a heavy-tailed distribution, in particular a Pareto distribution whose density is given by , 1 ) ( = r b r g r &gt; b where is related to the Hurst parameter by 2 / ) 3 ( = H . The sessions arrive according to a Poisson process with rate . The packets arrive according to a Poisson process with rate , locally, over each session. For each class, the traffic Y (t) measured as the total number of packets injected in [0, t] is found by ) ) ( ( ) ( = t S i i i i S t R U t Y where denote the local Poisson process, the duration and the arrival time of session i, respectively. Hence, Y(t) corresponds to the sum of packets generated by all sessions initiated in [0,t] until the session expires if that happens before t, and until t if is does not. The stationary version of this model based on an infinite past is considered in calculations below. The packet sizes are assumed to be fixed because each queue corresponds to a certain type of application where the packets have fixed size or at least fixed service time distribution. i i i S R U , , The traffic model Y is long-range dependent and almost second-order self-similar; the auto covariance function of its increments is that of fractional Gaussian noise. Three different heavy traffic limits are possible depending on the rate of increase in the traffic parameters [46, 48]. Two of these limits are well known self-similar processes, fractional Brownian motion and Levy process, which do not account for packet dynamics in particular. INTERARRIVAL TIMES Packet interarrival time distributions for the particular self-similar traffic model are calculated for the first time in this paper. We consider a single type of packet first. The distributions of cross interarrival time between different types of packets are derived on the basis of single packet results. 4.1 Interarrival Times for a Single Class Although the packet arrival process itself is long-range dependent and shows self-similarity, the number of alive sessions at a period of time, say of length t, has a stationary distribution and is Poisson distributed. The alive sessions at any time can be further split into independent components as those session that last longer than t and those that expire before t. Such results are well known [49, pg.273] and will be used to derive the interarrival time distribution of the packets. Given that there is a packet arrival at an instant in time, we aim to find the distribution of the time until next arrival denoted by T. We will find ) (t F = , for . When the event is considered, the information that there is a packet arrival is equivalent to the information that there is at least one session alive at the given instant. This follows from the assumption that local packet generation process is Poisson over each session. The probability that next interarrival is greater than t } { t T P &gt; 0 t } { t T &gt; 182 on a particular session is the same as the probability that the remaining time until next arrival is greater than t due to the memoryless property of exponential distribution. That is, = &gt; } { t T P P {Next packet interarrival is greater than t | there is a packet arrival} = 1 P {Next packet interarrival is greater than t, there is at least one alive session} where is the probability that there is at least one alive session, in other words the utilization of an system. The event that next packet interarrival is greater than t can be split as follows: / / G M The active sessions that expire before t do not incur any new arrivals. The active sessions that expire after t do not incur any new arrivals No new session arrivals in t or at least one session arrival with no packet arrival in t. We find the probability that all three events occur at the same time by using the independence of a Poisson point process over disjoint sets. The result is } { t T P &gt; = )] 1 ( exp[ { 1 ) ( ) ( t B v A v e t e e t t ) 1 ] / ) 1 )( ( exp[ ) 1 ] ) ( (exp[ t e B v e A v t t t t ) 1 ] e ) ( )](exp[ 1 ( exp[ e e ) ( ) ( + t t t B A A e t t t } ) 1 ] / ) e 1 )( ( )](exp[ 1 ( exp[ e e ) ( ) ( + t B e t t t t B A t t where = ) ( t A t dy y g t y ) ( ) ( (1) + = t t t G t dy y yg B 0 ) ( ) ( ) ( (2) and ) ] duration session [ E exp( 1 = ] ) 1 /( exp[ 1 = b because the steady state number in the system in / / G M queue is Poisson distributed with mean E[Session duration] [50], and and b are the parameters of the session duration with complementary distribution function G and density 1 ) ( = r b r g b r &gt; which is Pareto. 4.2 Interarrival Times for Multiple Classes Here we explain the detailed procedure to find out the Interarrival times for two classes, the Interarrival times for more than two classes can be found in a similar way. Let denote the interarrival time between a class i packet that comes first and a class j packet that follows, The analysis, which can be extended to , provides a method for other self-similar models as well provided that the distribution of interarrivals are available. ij T . 2 , 1 , = j i 3 , j i i T For the consecutive packet 1 arrival time , we have 11 T , { } { 1 11 t T P t T P &gt; = &gt; no arrivals of class 2 in } 1 T = t T ds s f P ) ( } s in 2 class of arrivals no { 1 = t T ds s f s F ) ( ) ( 1 0 2 where ) ( ) ( 0 2 2 2 ) ( t t A v B v e e t F = . ] ) ( exp[ )] 1 ( exp[ 2 2 2 2 t t t e A v e t (3) ] / ) 1 )( ( exp[ 2 2 2 t e B v t t ) ( 2 t A and ) ( 2 t B are defined analogously as in (1) and (2), and we used the independence of class 1 and 2 packet inputs. Here, 0 2 F is found through similar arguments used for P {T&gt;t} in the last subsection, without assuming that there is an alive session of type 2 . As a result, by differentiation we find ) ( ) ( ) ( 0 2 1 11 t F t f t f T T = Now consider the interarrival time T 12 occurring between a class 1 packet followed by a class 2 packet. For T 12 , we get = } { 12 t T P t ds s P s f 0 0 2 } arrived packet class1 a | in 1 class of arrivals no { ) ( = t T ds s F s f 0 0 2 ) ( ) ( 1 where is the density function corresponding to the event that there is an arrival of class 2 packet at time s, and we used independence of class 1 and 2 packet streams. As a matter of fact, can be obtained by taking the derivative of the complementary distribution function ) ( 0 2 s f ) ( 0 2 s f 0 2 F given in (3). As a result, we get ) ( ) ( ) ( 1 12 0 2 t F t f t f T T = 183 Similarly, it can be shown that ) ( ) ( ) ( 0 1 2 22 t F t f t f T T = , ) ( ) ( ) ( 2 21 0 1 t F t f t f T T = QUEUEING MODEL We consider a model of four queues based on G/M/1 by considering four different classes of self-similar input traffic denoted by SS/M/1, and analyze it on the basis of priority with no preemption. Let the service time distribution have rate 1 , 2 3 , and 4 for type 1, type 2, type 3 and type 4 packets, respectively, and let type 1 packets have the highest priority and type 4 packets have the lowest priority. 5.1 With Four Classes 1 / / M SS The usual embedded Markov chain [51] formulation of is based on the observation of the queueing system at the time of arrival instants, right before an arrival. At such instants, the number in the system is the number of packets that arriving packet sees in the queue plus packets in service, if any, excluding the arriving packet itself. We specify the states and the transition probability matrix P of the Markov chain with the self-similar model for four types of traffic. 1 / / M G Let denote the embedded Markov chain at the time of arrival instants. As the service is based on priority, the type of packet in service is important at each arrival instant of a given type of packet to determine the queueing time. Therefore, we define the state space as: } 0 : { n X n } , , , }, , , , , { }, , , , { : ) , , , , , {( 4 3 2 1 4 3 2 1 4 3 2 1 4 3 2 1 + = Z i i i i I s s s s s a a a a a s a i i i i S (4) where are labels to denote the type of arrival, are labels to denote the type of packet in service, are the number of packets in each queue including a possible packet in service, I denotes the idle state in which no packet is in service or queued and is the set of nonnegative integers. Some of the states in the state space S given in (4) have zero probability. For example, is impossible. The particular notation in (4) for S is chosen for simplicity, although the impossible states could be excluded from S. Each possible state, the reachable states from each and the corresponding transition probabilities will be calculated. 4 3 2 1 , , , a a a a 4 3 2 1 , , , s s s s 4 3 2 1 , , , i i i i + Z ) , , , , 0 , ( 2 1 4 3 1 s a i i i 5.2 States of the Embedded Markov Chain The states of the Markov chain and the possible transitions with respective probabilities can be enumerated by considering each case. We will only analyze the states with non-empty queues in this paper. 5.2.1 States with ) , , , , , ( 4 3 2 1 s a i i i i 0 , , , 4 3 2 1 i i i i We can divide the states and transitions into 256 groups. Because (a, s) can occur 4x4=16 different ways, and the next state (p, q) can be composed similarly in 16 different ways as } , , , { , 4 3 2 1 a a a a p a and . We will analyze only the first one in detail; the others follow similarly. } , , , { , 4 3 2 1 s s s s q s 5.2.2 Transition from ) , , , , , ( ) , , , , , ( 2 2 4 3 2 1 1 1 4 3 2 1 s a j j j j s a i i i i This is the case where a transition occurs from an arrival of type 1 to an arrival of type 2 such that the first arrival has seen a type 1 packet in service, packets of type 1 (equivalently, total of queue 1 and the packet in service) and packets of type 2 (in this case only queue 2), packets of type 3 and packets of type 4 in the system. The transition occurs to packets of type 1, packets of type 2, with a type 2 packet in service, packets of type 3 and packets of type 4 in the system. Due to priority scheduling, an arrival of type 2 can see a type 2 packet in service in the next state only if all type 1 packets including the one that arrived in the previous state are exhausted during the interarrival time. That is why can take only the value 0 and exactly 1 i 2 i 3 i 4 i 1 j 2 j 3 j 4 j 1 j 1 1 + i packets of type 1 are served. In contrast, the number of packets served from queue 2, say k, can be anywhere between 0 and 1 2 i as at least one type 2 packet is in the system, one being in service, when a new arrival occurs. The transition probabilities are )} , , , , , ( | ) , , , , , 0 ( { 1 1 4 3 2 1 2 2 4 3 2 1 s a i i i i X s a i i k i X P n n = = + 1 { 1 + = i P served from type 1, k served from type 2 and a type 2 packet remains in service during } 12 T where we use the fact that the remaining service time of a type 1 packet in service has the same exponential distribution Exp( 1 ), due to the memory-less property of a Markovian service. Therefore, for 1 , , 0 2 = i k K )} , , , , , ( | ) , , , , , 0 ( { 1 1 4 3 2 1 2 2 4 3 2 1 s a i i i i X s a i i k i X P n n = = + + + = 0 0 ) ( ) ( ) ( 12 2 1 1 1 2 t x t T S S S dt dx ds t f x f s f k i where : sum of l independent service times of type m packets, m=1, 2, l m S + Z l . Note that has an Erlang distribution with parameters l m S ) , ( m l as each service time has an exponential distribution, and the sum being the sum of several exponentially distributed random variables has a hypoexponential distribution. The density functions of all these distributions can easily be evaluated numerically. Similarly, we can enumerate all 256 cases. The results for first 64 cases are given in Table 1 in the Appendix. 2 1 2 1 l l S S + 184 5.3 Limiting Distribution and Waiting Times Steady state distribution as seen by an arrival can be found by solving = P using the transition matrix P of the Markov chain analyzed above. In practice, the queue capacity is limited in a router. So, the steady state distribution exists. To the best of our knowledge, no previous analytical expressions are available for the waiting time of a G/M/1 queue with priority. Our analysis relies on the limiting distribution of the state of the queue at the arrival instances, which can be computed using the analysis given above for our self-similar traffic model. In general, the following analysis is valid for any G/M/1 queueing system where the limiting distribution at the arrival instances can be computed. The expected waiting time for the highest priority queue can be found as + + + = = = = = = = = = ) , , , , , ( ) 1 ( ) , , , , , ( ] [ 2 1 4 3 2 1 1 0 1 0 0 2 1 1 1 1 4 3 2 1 1 1 0 0 0 1 1 1 1 1 2 2 3 3 4 4 1 1 2 2 3 3 4 4 s a j j j j j s a j j j j j W E J j J j J j J j J j J j J j J j ) , , , , , ( ) 1 ( ) , , , , , ( ) 1 ( 4 1 4 3 2 1 1 0 0 0 1 4 1 1 3 1 4 3 2 1 3 1 0 0 1 0 1 1 1 1 2 2 3 3 4 4 1 1 2 2 3 3 4 4 s a j j j j j s a j j j j j J j J j J j J j J j J j J j J j = = = = = = = = + + + where and are the respective capacities of each queue. This follows clearly from the fact that an arriving packet of higher priority will wait until all packets of the same priority as well as the packet in service are served. Depending on the type of the packet in service, we have the constituent expressions in the sum. , , 2 1 J J 3 J 4 J On the other hand, we obtain the expected waiting time for the low priority queues by analyzing the events that constitute this delay. The amount of work in the system at any time is defined as the (random) sum of all service times that will be required by the packets in the system at that instant. The waiting time of a type 2 packet (which is 2 nd highest priority queue) can be written as .... 3 2 1 2 + + + = Z Z Z W (5) where Z 1 is the amount of work seen by the arriving packet in queue 1 and queue 2 (i.e, higher priority and equal priority), Z 2 is the amount of work associated with higher priority (i.e.type 1) packets arriving during Z 1 , Z 3 is the amount of work associated with type 1 packets arriving during Z 2 , and so on. As illustrated in Fig.2, the waiting time of an arriving packet of type 2 is indeed given by the total workload building in front of it. The arrows in the figure denote the arrival times of type 1 packets, and all the oblique lines have 45 degrees angle with the time axis. In this figure the waiting time is 4 3 2 1 2 Z Z Z Z W + + + = for example. Let M j denote the number of type j arrivals over Z i , j=1,2,....Then L + + + = 2 1 1 1 1 2 M M S S Z W where denotes the random sum of M j M S 1 j independent service times of type 1 packets. Then, L + + + = ] [ ] [ ] [ ] [ ] [ [ 2 1 1 1 1 2 ] M E S E M E S E Z E W E since the service times and the arrival process are independent. For a stationary packet arrival process, we get ] [ ] [ ]] | [ [ ] [ 1 1 j j j j j Z E c Z c E Z M E E M E = = = due to mentioned independence, where is a constant particular to the arrival process. That is, expectation of the number of arrivals in any period of time is proportional to the length of that period because of stationarity in time and linearity of expectation. In our stationary self-similar traffic input process, c 0 1 &gt; c 1 is the expected number of arrivals per unit time which can be called the arrival rate, given by the product of the arrival rate of session arrivals, the arrival rate of packets over a session, and the expected session length [46]. Explicitly, . Hence, the expected waiting time reduces to ) 1 /( 1 = b c L + + + = ] [ ] [ ] [ ] [ ] [ [ 2 1 1 1 1 1 1 2 ] Z E c S E Z E c S E Z E W E ] [ ] [ ) ] [ ] [ ( ] [ 2 1 1 1 2 1 1 1 1 W E c Z E Z E Z E c Z E + = + + + = L In view of (5), therefore we get from ] [ 2 W E + + + + = = = = = = = = = ) , , , , , ( ) ( ) , , , , , ( ) ( ] [ 2 2 4 3 2 1 2 2 0 1 1 0 0 1 1 1 2 4 3 2 1 2 2 1 1 0 0 0 1 1 2 1 1 2 2 3 3 4 4 1 1 2 2 3 3 4 4 s a j j j j j j s a j j j j j j W E J j J j J j J j J j J j J j J j 1 2 1 4 2 4 3 2 1 4 0 1 0 0 1 2 2 1 1 3 2 4 3 2 1 3 0 1 0 1 0 2 2 1 1 ] [ ) , , , , , ( ) 1 ( ) , , , , , ( ) 1 ( 1 1 2 2 3 3 4 4 1 1 2 2 3 3 4 4 W E c s a j j j j j j s a j j j j j j J j J j J j J j J j J j J j J j + + + + + + = = = = = = = = Z 4 Z 3 Z 2 Z 1 time Work (time) Fig.2 Waiting time of a type 2 packet in terms of Z j 's. 185 Similarly, we can directly write down the expected waiting time for a packet of type 3 (3 rd priority queue) and type 4 (lowest priority queue). The expected waiting time for a packet of type 3 can be found from: + + + + + + = = = = = = = = = ) , , , , , ( ) ( ) , , , , , ( ) ( ] [ 2 3 4 3 2 1 3 3 0 1 1 0 0 2 2 1 1 1 3 4 3 2 1 3 3 2 2 1 0 1 0 0 1 1 3 1 1 2 2 3 3 4 4 1 1 2 2 3 3 4 4 s a j j j j j j j s a j j j j j j j W E J j J j J j J j J j J j J j J j ] [ ) ( ) , , , , , ( ) 1 ( ) , , , , , ( ) ( 3 2 2 1 1 4 3 4 3 2 1 4 0 0 1 0 1 3 3 2 2 1 1 3 3 4 3 2 1 3 3 2 2 0 0 1 1 0 1 1 1 1 2 2 3 3 4 4 1 1 2 2 3 3 4 4 W E c c s a j j j j j j j s a j j j j j j j J j J j J j J j J j J j J j J j + + + + + + + + = = = = = = = = and can be determined from ] [ 4 W E + + + + + + + + = = = = = = = = = ) , , , , , ( ) ( ) , , , , , ( ) ( ] [ 2 4 4 3 2 1 4 4 3 3 0 1 0 1 0 2 2 1 1 1 4 4 3 2 1 4 4 3 3 2 2 1 0 0 1 0 1 1 4 1 1 2 2 3 3 4 4 1 1 2 2 3 3 4 4 s a j j j j j j j j s a j j j j j j j j W E J j J j J j J j J j J j J j J j + + + + + + + + = = = = = = = = ) , , , , , ( ) ( ) , , , , , ( ) ( 4 4 4 3 2 1 4 4 0 0 0 1 1 3 3 2 2 1 1 3 4 4 3 2 1 4 4 0 0 1 1 0 3 3 2 2 1 1 1 1 2 2 3 3 4 4 1 1 2 2 3 3 4 4 s a j j j j j j j j s a j j j j j j j j J j J j J j J j J j J j J j J j ] [ ) ( 4 3 3 2 2 1 1 W E c c c + + APPLICATIONS OF THE MODEL Here we give an overview of the prime application of the model. 3G systems such as UMTS and CDMA2000 are leaning towards an all-IP network architecture for transporting IP multimedia services [52]. An all-IP DiffServ platform is the currently most promising architecture to interwork the heterogeneous wireless access networks and the Internet to provide broadband access, seamless global roaming and QoS guarantees for various IP multimedia services [53]. To transport UMTS services through IP networks without loosing end-to-end QoS provisioning, a consistent and efficient QoS mapping between UMTS services and IP QoS classes is required. According to 3GPP, UMTS-to-IP QoS mapping is performed by a translation function in the GGSN router that classifies each UMTS packet flow and maps it to a suitable IP QoS class [52]. In order to make accurate mappings and to ensure guaranteed QoS parameters to the end user of mobile Internet, it is essential to being able to accurately model the end-to -end behavior of different classes of wireless IP traffic (conversational, interactive, streaming and background) passing through a DiffServ domain. Several queueing tools have been developed that can be implemented in IP routers within different QoS domains including Priority Queueing (PQ), Custom Queueing (CQ), Weighted Fair Queueing (WFQ), Class Based Weighted Fair Queueing (CBWFQ) and Low-Latency Queueing (LLQ) [54]. Our model is directly applicable to the problem of determining the end-to-end queueing behavior of IP traffic through both Wired and wireless IP domains, but modeling accuracy is more crucial in resource constrained environments such as wireless networks. For example, our model is directly able to analyze the behavior of four different QoS classes of UMTS traffic passing through a DiffServ domain, in which routers are implemented with priority queueing. Thus, the model enables tighter bounds on actual behavior so that over-provisioning can be minimized. It also enables translations of traffic behavior between different kinds of QoS domains so that it is possible to map reservations made in different domains to provide session continuity. CONCLUSION AND FUTURE WORK In this paper, we have presented a novel analytical model based on G/M/1 queueing system for accurate modeling of wireless IP traffic behavior under the assumption of four different classes of self-similar traffic. We have analyzed it on the basis of non-preemptive priority and explicit expressions of expected waiting time for the corresponding classes have been derived. The model represents an important step towards the overall aim of finding realistic (under self-similar traffic assumptions) end-to-end QoS behavior (in terms of QoS parameters such as delay, jitter and throughput) of multiple traffic classes passing through heterogeneous wireless IP domains (IntServ, DiffServ and MPLS). At the moment, we are working on the numerical analysis to find solutions to the suggested imbedded Markov Chain in order to find exact QoS parameter bounds for a given system. Our future work will focus to analyze the performance of different QoS domains implemented with different queueing disciplines. REFERENCES [1] W. Leland, M. Taqqu, W. Willinger and D. Wilson, "On the self-similar nature of Ethernet traffic (extended version)", IEEE/ACM Transactions on Networking, vol. 2. no. 1, pp. 1-15, Feb. 1994. [2] V. Paxon, "Empirically derived analytical models of wide-area TCP connections", IEEE/ACM Transactions on Networking, vol. 2, pp. 316-336, Aug. 1994 [3] V. Paxon and S. Floyd , "Wide-area traffic: the failure of Poisson modeling", in Proc. ACM SIGCOMM 94, London, U.K., Aug. 1994, pp. 257-268 [4] M. Crovella and A. Bestavros, "Explaining World Wide Web Traffic Self-Similarity", Tech. Rep. TR-95-015, Boston University, CS Dept, Boston, MA 02215, Aug. 1995 [5] M. W. Garrett and W. Willinger, "Analysis, Modeling and Generation of Self-Similar VBR Video Traffic", ACM Computer Communication Review, vol. 24, Oct. 1994, SIGCOMM 94 Symposium [6] W. Willinger et al, "Statistical analysis of CCSN/SS7 traffic data from working CCS subnetworks", IEEE. Journal on Selected Areas of Communication, vol. 12, no. 3, pp. 544-551, Apr. 1994 [7] M. Crovella and A. Bestavros, "Self-Similarity in World Wide Web Traffic: Evidence and Possible Causes", in ACM Sigmetrics, May 1996 [8] J.C Bolot and M. Grossglauser, "On the Relevance of Long-Range Dependence in Network Traffic", Computer Communication Review, vol. 26, no. 4, pp. 15-24, October 1996. 186 [9] Z. L. Zhang, V. Ribeiro, S. Moon and C. Diot, "Small-Time Scaling behavior of internet backbone traffic: An Empirical Study", in IEEE INFOCOM, march 2003 [10] M. S. Taqqu, "Self-Similar processes". In S. Kotz and N. Johnson, editors, Encyclopedia of Statistical Sciences, vol. 8, pp. 352-357. Wiley, New York, 1988. [11] W. Willinger, M.S Taqqu and A. Erramilli, "A bibliographical guide to self-similar traffic and performance modeling for modern high speed networks", In F. P. Kelly, S. Zachary and I. Ziedins, editors, Stochastic Networks: Theory and Applications, pp. 339-366 , Claredon Press, Oxford, 1996 [12] H. Holma and A. Taskala, "WCDMA for UMTS, Radio Access for Third Generation Mobile Communications, 2 nd Edition", John Wiley & Sons, Ltd. 2002, pp. 1-5 [13] J. Yang and I. Kriaras, "Migration to all-IP based UMTS networks, " IEEE 1 st International Conference on 3G Mobile Communication Technologies , 27-29 March, 2000, pp. 19-23 [14] W. Stallings, "Integrated Services Architecture: The Next-Generation Internet", International Journal of Network Management, 9, 1999, pp. 38-43 [15] S. Blake et al., "An Architecture for Differentiated Services", IETF RFC 2475, Dec. 1998 [16] Rosen E. et al., "Multiprotocol Label Switching (MPLS) Architecture", RFC 3031, Jan. 2001 [17] K. Venken, J. De Vriendt and D. De Vleeschauwer, "Designing a DiffServ-capable IP-backbone for the UTRAN", IEEE 2 nd International Conference on 3G Mobile Communication Technologies, 26-28 March 2001, pp. 47-52 [18] S. Maniatis, C. Grecas and I. Venieris, "End-to-End QoS Issues Over Next Generation Mobile Internet", IEEE Symposium on Communication and Vehicular Technology, 2000, SVCT-2000, 19 Oct, 2000, pp. 150-154 [19] P. Newman, Netillion Inc. "In Search of the All-IP Mobile Network", IEEE Communication Magazine, vol. 42, issue 12, Dec. 2004, pp. S3-S8 [20] G. Araniti, F. Calabro, A. Iera, A. Molinaro and S. Pulitano, "Differentiated Services QoS Issues in Next Generation Radio Access Network: a New Management Policy for Expedited Forwarding Per-Hop Behavior", IEEE Vehicular Technology Conference, VTC 2004-Fall, vol. 4, 26-29 Sept. 2004, pp. 2693-2697 [21] S. Uskela, "All IP Architectures for Cellular Networks", 2 nd International Conference on 3G Mobile Communication Technologies, 26-28 March 2001, pp. 180-185 [22] Jeong-Hyun Park, "Wireless Internet Access for Mobile Subscribers Based on GPRS/UMTS Network" IEEE Communication Magazine, vol. 40, issue 4, April 2002, pp. 38-39 [23] K. Daniel Wong and Vijay K. Varma, "Supporting Real-Time IP Multimedia Services in UMTS", IEEE Communication Magazine, vol. 41, issue 11, Nov. 2003, pp. 148-155 [24] 3GPP, "Universal Mobile Telecommunication System (UMTS); QoS Concepts and Architecture", TS23.107V6, March 2004 [25] R. Chakravorty, J. Cartwright and I. Pratt, "Practical Experience with TCP over GPRS", in IEEE GlobeCom, Nov. 2002 [26] D. Schwab and R. Bunt, "Characterizing the use of a Campus Wireless Network", in IEEE INFOCOM, March 2004 [27] X. Meng, S. Wong, Y. Yuan and S.Lu, "Characterizing Flows in Large Wireless Data Networks", in ACM Mobicom, Sep 2004 [28] A. Balachandran, G. M. Voelker, P. Bahl and P. Venkat Rangan, "Characterizing user behavior and network performance in a public Wireless LAN", Sigmetrics Performance Evaluation. Review, vol. 30. no. 1, 2002, pp. 195-205 [29] A. Adas and A. Mukherjee, "On Resource Management and QoS guarantees for long-range dependant traffic", In Proc IEEE INFOCOM, 1995, pp. 779-787 [30] M. Parulekar and A. Makowski, "Tail Probabilities for a Multiplexer with self-similar input", In proc IEEE INFOCOM, 1996, pp. 1452-1459 [31] I. Norros, "A Storage Model with self-similar input", Queueing System, 16, 1994, pp. 387-396 [32] B. Tsybakov and N. D. Georganas, "Self-Similar traffic and upper bounds to buffer overflow in ATM queue", Performance Evaluation, 36, 1998, pp. 57-80 [33] M. Zukerman et al, "Analytical Performance Evaluation of a Two Class DiffServ link", IEEE ICS, 25-28 Nov. 2002, vol. 1, pp. 373-377 [34] J. M. Chung, Z. Quan, "Impact of Self-Similarity on Performance Evaluation in DiffServ Networks", IEEE MWSCAS, 4-7 Aug. 2002, vol. 2, pp. 326-329 [35] C. F. Chou et al, "Low Latency and efficient packet scheduling for streaming applications", IEEE ICC, 20-24 June, 2004, vol. 4, pp. 1963-1967 [36] A. Kos and B. Klepec, "Performance of VoIP applications in a simple Differentiated Services network architecture", IEEE EUROCON, 4-7 July, 2001, vol. 1, pp. 214-217 [37] J. M. Chung and H. M. Soo, "Analysis of non preemptive priority queueing of MPLS networks with Bulk arrivals", IEEE MWSCAS, 4-7 Aug. 2002, vol. 3. pp. 81-84 [38] Salil S. Kanhere and Harish Sethu, "Fair, Efficient and Low-Latency Packet Scheduling using Nested Deficit Round Robin", Proceedings of the IEEE Workshop on High Performance Switching and Routing (HSPR), May 2001 [39] N. F. MIR and A. Chien, "Simulation of Voice over MPLS communication networks", IEEE ICCS, 25-28 Nov. 2002, vol. 1, pp. 389-393 [40] A. Krendzel, Y. Koucheryavy, J. Harju and S. Lopatin, "Traffic and QoS management in Wireless Multimedia Networks" COST 290:: Wi-QoST, Working group N3 http://www.cost290.org [41] M. Jiang, M. Nikolic, S. Hardy and L. Trajkovic, "Impact of Self-Similarity on Wireless Data Network Performance", IEEE ICC, 2001, vol. 2, pp. 477-481 [42] J. Ridoux, A. Nucci and D. Veitch, "Characterization of Wireless Traffic based on Semi-Experiments", Technical Report-LIP6, December 2005 [43] Z. Sahinoglu and S. Tekinay, "On Multimedia Networks: Self-Similar Traffic and Network Performance", IEEE Communication Magazine, vol. 37, issue 1, Jan. 1999, pp. 48-52 [44] I. Norros, "On the use of Fractional Brownian Motion in theory of connectionless networks", IEEE Journal on Selected Areas in Communications, vol. 13. no. 6, August 1995, pp. 953-962 [45] P. Benko, G. Malicsko and A. Veres, "A Large-scale, passive analysis of end-to-end TCP Performances over GPRS", in IEEE INFOCOM, March 2004 [46] M. Caglar, "A Long-Range Dependant Workload Model for Packet Data Traffic", Mathematics of Operations Research, 29, 2004, pp. 92-105 187 [47] H. P. Schwefel, L. Lipsky, "Impact of aggregated self-similar ON/OFF traffic on delay in stationary queueing models (extended version)", Performance Evaluation, 43, 2001, pp. 203-221 [51] E. Cinlar, "Introduction to Stochastic Processes", 1975, pp. 178 [52] R. Ben Ali, Y Lemieux and S. Pierre, "UMTS-to-IP QoS Mapping for Voice and Video Telephony Services, IEEE Network , vol. 19, issue 2, March/April 2005, pp. 26-32 [48] I. Kaj, "Limiting fractal random processes in heavy-tailed systems", In Fractals in Engineering, New Trends in Theory and Applications, Eds.J. Levy-Lehel, E. Lutton, Springer-Verlag London, 2005, pp. 199-218 [53] Y. Cheng, H, Jiang, W, Zhuang, Z. Niu and C. Lin, "Efficient Resource Allocation for China's 3G/4G Wireless Networks, IEEE Communication Magazine, vol. 43, issue 1, Jan 2005, pp. 76-83 [49] S.M. Ross, "Introduction to Probability Models" Academic Press, 1997. [54] W. Odom and M. J. Cavanaugh, "IP Telephony Self-Study Cisco DQoS Exam Certification Guide", Cisco Press, 2004, pp. 3-314 [50] K.S. Trivedi, Probability and statistics with reliability, queueing, and computer science applications Wiley, New York, 2002. APPENDIX Table: 1 The States of the Markov Chain and Transition Probabilities Initial State Reachable States ( 4 , 3 , 2 , 1 = m ) Transition Probability ) , , , , , 1 ( 1 4 3 2 1 s a i i i k i m + , 1 , , 0 i k K = 0 0 ) ( ) ( ) ( 1 1 1 t x t T S S dt dx ds t f x f s f m k ) , , , , , 0 ( 2 4 3 2 s a i i k i m , 1 , , 0 2 = i k K + + 0 0 ) ( ) ( ) ( 1 2 1 1 1 2 t x t T S S S dt dx ds t f x f s f m k i 1 ,....... 1 , 0 ) , , , , 0 , 0 ( 3 3 4 3 = i k s a i k i m dsdxdt t f x f s f m k i i T S S S t x t S ) ( ) ( ) ( 1 3 2 2 1 1 1 3 0 0 + + + ) , , , , , ( 1 1 4 3 2 1 s a i i i i 1 ..... ,......... 1 , 0 ) , , , 0 , 0 , 0 ( 4 4 4 = i k s a k i m + + + + 0 0 ) ( ) ( ) ( 1 4 3 3 2 2 1 1 1 4 t x t T S S S S S dsdxdt t f x f s f m k i i i ) , , , , 1 , 1 ( 1 4 3 2 1 s a i i i k i m + , 1 , , 0 i k K = + 0 0 ) ( ) ( ) ( 1 1 2 1 1 t x t T S S S dt dx ds t f x f s f m k ) , , , , , 1 ( 2 4 3 2 1 s a i i i i m + 0 ) ( ) ( 1 2 t T S dt ds t f s f m ) , , , , , 0 ( 2 4 3 2 s a i i k i m K , 3 , 2 2 = i and 1 , , 1 2 = i k K + + 0 0 ) ( ) ( ) ( 1 2 1 1 1 2 t x t T S S S dt dx ds t f x f s f m k i ) , , , , , ( 2 1 4 3 2 1 s a i i i i 1 ... ,......... 1 , 0 ) , , , , 0 , 0 ( 3 3 4 3 = i k s a i k i m + + + 0 0 ) ( ) ( ) ( 1 3 2 2 1 1 1 3 t x t T S S S S dsdxdt t f x f s f m k i i 188 1 ... ,......... 2 , 1 , 0 ) , , , 0 , 0 , 0 ( 4 4 4 = i k s a k i m + + + + 0 0 ) ( ) ( ) ( 1 4 3 3 2 2 1 1 1 4 t x t T S S S S S dsdxdt t f x f s f m k i i i 1 1 4 3 2 1 ... ,......... 1 , 0 ) , , , 1 , , 1 ( i k s a i i i k i m = + + 0 0 ) ( ) ( ) ( 1 1 3 1 1 t x t T S S S dsdxdt t f x f s f m k 1 ... ,......... 1 , 0 ) , , , 1 , , 0 ( 2 2 4 3 2 = i k s a i i k i m + + + 0 0 ) ( ) ( ) ( 1 1 3 2 1 1 1 2 t x t T S S S S dsdxdt t f x f s f m k i ) , , , , , 1 ( 3 4 3 2 1 s a i i i i m + 0 ) ( ) ( 1 3 t T S dt ds t f s f m ) , , , , 0 , 0 ( 3 4 3 s a i k i m ,...... 3 , 2 3 = i 1 , , 1 3 = i k K + + + 0 0 ) ( ) ( ) ( 1 3 2 2 1 1 1 3 t x t T S S S S dt dx ds t f x f s f m k i i ) , , , , , ( 3 1 4 3 2 1 s a i i i i 1 ..... ,......... 1 , 0 ) , , , 0 , 0 , 0 ( 4 4 4 = i k s a k i m + + + + 0 0 ) ( ) ( ) ( 1 4 3 3 2 2 1 1 1 4 t x t T S S S S S dsdxdt t f x f s f m k i i i 1 1 4 3 2 1 .. ,......... 1 , 0 ) , , 1 , , , 1 ( i k s a i i i k i m = + + 0 0 ) ( ) ( ) ( 1 1 4 1 1 t x t T S S S dsdxdt t f x f s f m k 1 .. ,......... 1 , 0 ) , , 1 , , , 0 ( 2 2 4 3 2 = i k s a i i k i m + + + 0 0 ) ( ) ( ) ( 1 1 4 2 1 1 1 2 t x t T S S S S dsdxdt t f x f s f m k i 1 .... ,......... 1 , 0 ) , , 1 , , 0 , 0 ( 3 3 4 3 = i k s a i k i m + + + + 0 0 ) ( ) ( ) ( 1 1 4 3 2 2 1 1 1 3 t x t T S S S S S dsdxdt t f x f s f m k i i ) , , , , , 1 ( 4 4 3 2 1 s a i i i i m + 0 ) ( ) ( 1 4 t T S dsdt t f s f m ) , , , , , ( 4 1 4 3 2 1 s a i i i i .. ,......... 3 , 2 ) , , , 0 , 0 , 0 ( 4 4 4 = i s a k i m 1 ,... 2 , 1 4 = i k + + + + 0 0 ) ( ) ( ) ( 1 4 3 3 2 2 1 1 1 4 t x t T S S S S S dsdxdt t f x f s f m k i i i 189

QoS;3G networks;traffic modelling;3G;Self-Similar;GGSN;self-similar traffic;wireless IP traffic;UMTS;queuing model

3

A Computational Approach to Reflective Meta-Reasoning about Languages with Bindings

We present a foundation for a computational meta-theory of languages with bindings implemented in a computer-aided formal reasoning environment. Our theory provides the ability to reason abstractly about operators, languages, open-ended languages, classes of languages, etc. The theory is based on the ideas of higher-order abstract syntax, with an appropriate induction principle parameterized over the language (i.e. a set of operators) being used. In our approach , both the bound and free variables are treated uniformly and this uniform treatment extends naturally to variable-length bindings . The implementation is reflective, namely there is a natural mapping between the meta-language of the theorem-prover and the object language of our theory. The object language substitution operation is mapped to the meta-language substitution and does not need to be defined recursively. Our approach does not require designing a custom type theory; in this paper we describe the implementation of this foundational theory within a general-purpose type theory. This work is fully implemented in the MetaPRL theorem prover, using the pre-existing NuPRL-like Martin-Lof-style computational type theory. Based on this implementation, we lay out an outline for a framework for programming language experimentation and exploration as well as a general reflective reasoning framework. This paper also includes a short survey of the existing approaches to syntactic reflection.

Introduction 1.1 Reflection Very generally, reflection is the ability of a system to be "self-aware" in some way. More specifically, by reflection we mean the property of a computational or formal system to be able to access and internalize some of its own properties. There are many areas of computer science where reflection plays or should play a major role. When exploring properties of programming languages (and other languages) one often realizes that languages have at least two kinds of properties -- semantic properties that have to do with the meaning of what the language's constructs express and syntactic properties of the language itself. Suppose for example that we are exploring some language that contains arithmetic operations. And in particular, in this language one can write polynomials like x 2 + 2x + 1. In this case the number of roots of a polynomial is a semantic property since it has to do with the valuation of the polynomial. On the other hand, the degree of a polynomial could be considered an example of a syntactic property since the most natural way to define it is as a property of the expression that represents that polynomial. Of course, syntactic properties often have semantic consequences, which is what makes them especially important. In this example, the number of roots of a polynomial is bounded by its degree. Another area where reflection plays an important role is run-time code generation -- in most cases, a language that supports run-time code generation is essentially reflective, as it is capable of manipulating its own syntax. In order to reason about run-time code generation and to express its semantics and properties, it is natural to use a reasoning system that is reflective as well. There are many different flavors of reflection. The syntactic reflection we have seen in the examples above, which is the ability of a system to internalize its own syntax, is just one of these many flavors. Another very important kind of reflection is logical reflection, which is the ability of a reasoning system or logic to internalize and reason about its own logical properties. A good example of a logical reflection is reasoning about knowledge -since the result of reasoning about knowledge is knowledge itself, the logic of knowledge is naturally reflective <A href="3.html#10">[Art04]. In most cases it is natural for reflection to be iterated. In the case of syntactic reflection we might care not only about the syntax of our language, but also about the syntax used for expressing the syntax, the syntax for expressing the syntax for expressing the syntax and so forth. In the case of the logic of knowledge it is natural to have iterations of the form "I know that he knows that I know . . .". When a formal system is used to reason about properties of programming languages, iterated reflection magnifies the power of the 2 system, making it more natural to reason not just about individual languages, but also about classes of languages, language schemas, and so on. More generally, reflection adds a lot of additional power to a formal reasoning system <A href="3.html#10">[GS89, Art99]. In particular, it is well-known <A href="3.html#10">[God36, <A href="3.html#11">Mos52, <A href="3.html#10">EM71, <A href="3.html#11">Par71] that reflection allows a super-exponential reduction in the size of certain proofs. In addition , reflection could be a very useful mechanism for implementing proof search algorithms <A href="3.html#9">[ACU93, <A href="3.html#10">GWZ00, CFW04]. See also <A href="3.html#10">[Har95] for a survey of reflection in theorem proving. 1.2 Uniform Reflection Framework For each of the examples in the previous section there are many ad-hoc ways of achieving the specific benefits of a specific flavor of reflection. This work aims at creating a unifying reflective framework that would allow achieving most of these benefits in a uniform manner, without having to reinvent and re-implement the basic reflective methodology every time. We believe that such a framework will increase the power of the formal reasoning tools, and it may also become an invaluable tool for exploring the properties of novel programming languages, for analyzing run-time code generation, and for formalizing logics of knowledge. This paper establishes a foundation for the development of this framework -- a new approach to reflective meta-reasoning about languages with bindings. We present a theory of syntax that: in a natural way provides both a higher-order abstract syntax (HOAS) approach to bindings and a de Bruijn-style approach to bindings, with easy and natural translation between the two; provides a uniform HOAS-style approach to both bound and free variables that extends naturally to variable-length "vectors" of binders; permits meta-reasoning about languages -- in particular, the operators, languages, open-ended languages, classes of languages etc. are all first-class objects that can be reasoned about both abstractly and concretely; comes with a natural induction principle for syntax that can be parameterized by the language being used; provides a natural mapping between the object syntax and meta-syntax that is free of exotic terms, and allows mapping the object-level substitution operation directly to the meta-level one (i.e. -reduction); is fully derived in a pre-existing type theory in a theorem prover; is designed to serve as a foundation for a general reflective reasoning framework in a theorem prover; is designed to serve as a foundation for a programming language experimentation framework. The paper is structured as follows. Our work inherits a large number of ideas from previous efforts and we start in Section <A href="3.html#2">2 with a brief survey of existing techniques for formal reasoning about syntax. Next in Section <A href="3.html#4">3 we outline our approach to reasoning about syntax and in Section <A href="3.html#5">4 we present a formal account of our theory based on a Martin-Lof style computational type theory <A href="3.html#10">[CAB <A href="3.html#10">+ <A href="3.html#10">86, HAB <A href="3.html#10">+ <A href="3.html#10">] and the implementation of that account in the MetaPRL theorem prover <A href="3.html#11">[Hic97, Hic99, Hic01, HNC <A href="3.html#11">+ <A href="3.html#11">03, <A href="3.html#11">HNK <A href="3.html#11">+ <A href="3.html#11">, <A href="3.html#10">HAB <A href="3.html#10">+ <A href="3.html#10">]. Then in Section <A href="3.html#8">5 we outline our plan for building a uniform reflection framework based on the syntactic reflection. Finally, in Section <A href="3.html#9">6 we resume the discussion of related work that was started in Section <A href="3.html#2">2. 1.3 Notation and Terminology We believe that our approach to reasoning about syntax is fairly general and does not rely on any special features of the theorem prover we use. However, since we implement this theory in MetaPRL, we introduce some basic knowledge about MetaPRL terms. A MetaPRL term consists of: 1. An operator name (like "sum"), which is a unique name indicating the logic and component of a term; 2. A list of parameters representing constant values; and 3. A set of subterms with possible variable bindings. We use the following syntax to describe terms, based on the NuPRL definition <A href="3.html#9">[ACHA90]: opname operator name [ p 1 ; ; p n ] parameters {v 1 .t 1 ; ; v m .t m } subt er ms In addition, MetaPRL has a meta-syntax somewhat similar to the higher-order abstract syntax presented in Pfenning and Elliott <A href="3.html#11">[PE88]. MetaPRL uses the second-order variables in the style of Huet and Lang <A href="3.html#11">[HL78] to describe term schemas. For example, x.V [x], where V is a second-order variable of arity 1, is a schema that stands for an arbitrary term whose top-level operator is . This meta-syntax requires that every time a binding occurrence is explicitly specified in a schema, all corresponding bound occurrences have to be specified as well. This requirement makes it very easy to specify free variable restrictions -- for example, x.V , where V is a second-order meta-variable of arity 0, is a schema that stands for an arbitrary term whose top-level operator is and whose body does not have any free occurrences of the variable bound by that . In particular, the schema x.V matches the term y.1, but not the term x.x. In addition, this meta-language allows specifying certain term transformations, including implicit substitution specifications. For example, a beta reduction transformation may be specified using the following schema: (x.V 1 [x]) V 2 V 1 [V 2 ] Here the substitution of V 2 for x in V 1 is specified implicitly. Throughout this paper we will use this second-order notation to denote arbitrary terms -- namely, unless stated otherwise, when we write "x.t [x]" we mean an arbitrary term of this form, not a term containing a concrete second-order variable named "t". As in LF <A href="3.html#11">[HHP93] we assume that object level variables (i.e. the variables of the language whose syntax we are expressing) are directly mapped to meta-theory variables (i.e. the variable of the language that we use to express the syntax). Similarly, we assume that the object-level binding structure is mapped to the meta-level binding structure. In other words, the object-level notion of the "binding/bound occurrence" is a subset of that in the metalanguage . We also consider -equal terms -- both on the object level and on the meta-level -- to be identical and we assume that substitution avoids capture by renaming. The sequent schema language we use <A href="3.html#11">[NH02] contains a number of more advanced features in addition to those outlined here. However, for the purposes of this presentation, the basic features outlined above are sufficient. Previous Models of Reflection In 1931 Godel used reflection to prove his famous incompleteness theorem <A href="3.html#10">[God31]. To express arithmetic in arithmetic itself, he assigned a unique number (a Godel number) to each arithmetic 3 formula. A Godel number of a formula is essentially a numeric code of a string of symbols used to represent that formula. A modern version of the Godel's approach was used by Aitken et al. <A href="3.html#9">[ACHA90, AC92, ACU93, <A href="3.html#10">Con94] to implement reflection in the NuPRL theorem prover <A href="3.html#10">[CAB <A href="3.html#10">+ <A href="3.html#10">86, <A href="3.html#9">ACE <A href="3.html#9">+ <A href="3.html#9">00]. A large part of this effort was essentially a reimplementation of the core of the NuPRL prover inside NuPRL's logical theory. In Godel's approach and its variations (including Aitken's one), a general mechanism that could be used for formalizing one logical theory in another is applied to formalizing a logical theory in itself. This can be very convenient for reasoning about reflection, but for our purposes it turns out to be extremely impractical. First, when formalizing a theory in itself using generic means, the identity between the theory being formalized and the one in which the formalization happens becomes very obfuscated, which makes it almost impossible to relate the reflected theory back to the original one. Second, when one has a theorem proving system that already implements the logical theory in question, creating a completely new implementation of this logical theory inside itself is a very tedious redundant effort. Another practical disadvantage of the Godel numbers approach is that it tends to blow up the size of the formulas; and iterated reflection would cause the blow-up to be iterated as well, making it exponential or worse. A much more practical approach is being used in some programming languages, such as Lisp and Scheme. There, the common solution is for the implementation to expose its internal syntax representation to user-level code by the quote constructor (where quote (t) prevents the evaluation of the expression t). The problems outlined above are solved instantly by this approach: there is no blow-up, there is no repetition of structure definitions, there is even no need for verifying that the reflected part is equivalent to the original implementation since they are identical. Most Scheme implementations take this even further: the eval function is the internal function for evaluating a Scheme expression, which is exposed to the user-level; Smith <A href="3.html#11">[Smi84] showed how this approach can achieve an infinite tower of processors. A similar language with the quotation and antiquotation operators was introduced in <A href="3.html#10">[GMO03]. This approach, however, violates the congruence property with respect to computation: if two terms are computationally equal then one can be substituted for the other in any context. For instance, although 2 2 is equal to 4, the expressions "2*2" and "4" are syntactically different, thus we can not substitute 2*2 by 4 in the expression quote(2*2). The congruence property is essential in many logical reasoning systems, including the NuPRL system mentioned above and the MetaPRL system <A href="3.html#11">[HNC <A href="3.html#11">+ <A href="3.html#11">03, HNK <A href="3.html#11">+ <A href="3.html#11">, <A href="3.html#10">HAB <A href="3.html#10">+ <A href="3.html#10">] that our group uses. A possible way to expose the internal syntax without violating the congruence property is to use the so-called "quoted" or "shifted" operators <A href="3.html#9">[AA99, <A href="3.html#10">Bar01, Bar05] rather than quoting the whole expression at once. For any operator op in the original language , we add the quoted operator (denoted as op) to represent a term built with the operator op. For example, if the original language contains the constant "0" (which, presumably, represents the number 0), then in the reflected language, 0 would stand for the term that denotes the expression "0". Generally, the quoted operator has the same arity as the original operator, but it is defined on syntactic terms rather than on semantic objects. For instance, while is a binary operator on numbers, is a binary operator on terms. Namely, if t 1 and t 2 are syntactic terms that stand for expressions e 1 and e 2 respectively, then t 1 t 2 is a new syntactic term that stands for the expression e 1 e 2 . Thus, the quotation of the expression 12 would be 1 2. In general, the well-formedness (typing) rule for a quoted operator is the following: t 1 Term . . . t n Term op{t 1 ; . . . ; t n } Term (1) where Term is a type of terms. Note that quotations can be iterated arbitrarily many times, allowing us to quote quoted terms. For instance, 1 stands for the term that denotes the term that denotes the numeral 1. Problems arise when quoting expressions that contain binding variables. For example, what is the quotation of x.x? There are several possible ways of answering this question. A commonly used approach <A href="3.html#11">[PE88, <A href="3.html#10">DH94, DFH95, <A href="3.html#9">ACM02, ACM03] in logical frameworks such as Elf <A href="3.html#11">[Pfe89], LF <A href="3.html#11">[HHP93], and Isabelle <A href="3.html#11">[PN90, Pau94] is to construct an object logic with a concrete operator that has a type like (Term Term) Term or (Var Term) Term. In this approach, the quoted x.x might look like (x.x) and the quoted x.1 might look like (x.1). Note that in these examples the quoted terms have to make use of both the syntactic (i.e. quoted) operator and the semantic operator . Exotic Terms. Naive implementations of the above approach suffer from the well-known problem of exotic terms <A href="3.html#10">[DH95, DFH95]. The issue is that in general we can not allow applying the operator to an arbitrary function that maps terms to terms (or variables to terms) and expect the result of such an application to be a "proper" reflected term. Consider for example the following term: (x. if x = 1 then 1 else 2) It is relatively easy to see that it is not a real syntactic term and can not be obtained by quoting an actual term. (For comparison, consider (x. if x = 1 then 1 else 2), which is a quotation of x. if x = 1 then 1 else 2). How can one ensure that e denotes a "real" term and not an "exotic" one? That is, is it equal to a result of quoting an actual term of the object language? One possibility is to require e to be a substitution function; in other words it has to be equal to an expression of the form x.t [x] where t is composed entirely of term constructors (i.e. quoted operators) and x, while using destructors (such as case analysis, the if operator used in the example above, etc) is prohibited. There are a number of approaches to enforcing the above restriction . One of them is the usage of logical frameworks with restricted function spaces <A href="3.html#11">[PE88, HHP93], where -terms may only contain constructors. Another is to first formalize the larger type that does include exotic terms and then to define recursively a predicate describing the "validity" or "well-formedness" of a term <A href="3.html#10">[DH94, DFH95] thus removing the exotic terms from consideration. Yet another approach is to create a specialized type theory that combines the idea of restricted function spaces with a modal type operator <A href="3.html#10">[DPS97, DL99, DL01]. There the case analysis is disallowed on objects of "pure" type T , but is allowed on objects of a special type T . This allows expressing both the restricted function space "T 1 T 2 " and the unrestricted one "( T 1 ) T 2 " within a single type theory. Another way of regarding the problem of exotic terms is that it is caused by the attempt to give a semantic definition to a primarily syntactic property. A more syntax-oriented approach was used by Barzilay et al. <A href="3.html#10">[BA02, BAC03, Bar05]. In Barzilay's approach, the quoted version of an operator that introduces a binding has the same shape (i.e. the number of subterms and the binding structure) as the original one and the variables (both the binding and the 4 bound occurrences) are unaffected by the quotation. For instance, the quotation of x.x is just x.x. The advantages of this approach include: This approach is simple and clear. Quoted terms have the same structure as original ones, inheriting a lot of properties of the object syntax. In all the above approaches, the -equivalence relation for quoted terms is inherited "for free". For example, x.x and y.y are automatically considered to be the same term. Substitution is also easy: we do not need to re-implement the substitution that renames binding variables to avoid the capture of free variables; we can use the substitution of the original language instead. To prune exotic terms, Barzilay says that x.t [x] is a valid term when x.t [x] is a substitution function. He demonstrates that it is possible to formalize this notion in a purely syntactical fashion. In this setting, the general well-formedness rule for quoted terms with bindings is the following: is subst k {x 1 , , x k .t[x]} is subst l {z 1 , , z l .s[z]} op{x 1 , , x k .t[x]; ; z 1 , , z l .s[z]} Term (2) where is subst n {x 1 , , x n .t[x]} is the proposition that t is a substitution function over variables x 1 , , x n (in other words, it is a syntactic version of the Valid predicate of <A href="3.html#10">[DH94, DFH95]). This proposition is defined syntactically by the following two rules: is subst n {x 1 , , x n . x i } and is subst n+k {x 1 , , x n , y 1 , , y k .t[x; y]} . . . is subst n+l {x 1 , , x n , z 1 , , z l .s[x; z]}} is subst n {x 1 x n .op{y 1 y k .t[x; y]; ; z 1 z l .s[x; z]}} In this approach the is subst n {} and operators are essentially untyped (in NuPRL type theory, the computational properties of untyped terms are at the core of the semantics; types are added on top of the untyped computational system). Recursive Definition and Structural Induction Principle. A difficulty shared by both the straightforward implementations of the (Term Term) Term approach and by the Barzilay's one is the problem of recursively defining the Term type. We want to define the Term type as the smallest set satisfying rules <A href="3.html#3">(1) and <A href="3.html#4">(2). Note, however, that unlike rule <A href="3.html#3">(1), rule <A href="3.html#4">(2) is not monotonic in the sense that is subst k {x 1 , , x k .t[x]} depends non-monotonically on the Term type. For example, to say whether x.t [x] is a term, we should check whether t is a substitution function over x. It means at least that for every x in Term, t [x] should be in Term as well. Thus we need to define the whole type Term before using <A href="3.html#4">(2), which produces a logical circle. Moreover, since has type (Term Term) Term, it is hard to formulate the structural induction principle for terms built with the term constructor. Variable-Length Lists of Binders. In Barzilay's approach, for each number n, is subst n {} is considered to be a separate operator -- there is no way to quantify over n, and there is no way to express variable-length lists of binders. This issue of expressing the unbounded-length lists of binders is common to some of the other approaches as well. Meta-Reasoning. Another difficulty that is especially apparent in Barzilay's approach is that it only allows reasoning about concrete operators in concrete languages. This approach does not provide the ability to reason about operators abstractly; in particular, there is no way to state and prove meta-theorems that quantify over operators or languages, much less classes of languages. Higher-Order Abstract Syntax with Inductive Definitions Although it is possible to solve the problems outlined in the previous Section (and we will return to the discussion of some of those solutions in Section <A href="3.html#9">6), our desire is to avoid these difficulties from the start. We propose a natural model of reflection that manages to work around those difficulties. We will show how to give a simple recursive definition of terms with binding variables, which does not allow the construction of exotic terms and does allow structural induction on terms. In this Section we provide a conceptual overview of our approach ; details are given in Section <A href="3.html#5">4. 3.1 Bound Terms One of the key ideas of our approach is how we deal with terms containing free variables. We extend to free variables the principle that variable names do not really matter. In fact, we model free variables as bindings that can be arbitrarily -renamed. Namely, we will write bterm{x 1 , , x n .t[x]} for a term t over variables x 1 , , x n . For example, instead of term xy we will use the term bterm{x, y.xy} when it is considered over variables x and y and bterm{x, y, z.xy} when it is considered over variables x, y and z. Free occurrences of x i in t [x] are considered bound in bterm{x 1 , , x n .t[x]} and two -equal bterm{} expressions ("bterms") are considered to be identical. Not every bterm is necessarily well-formed. We will define the type of terms in such a way as to eliminate exotic terms. Consider for example a definition of lambda-terms. E XAMPLE 1. We can define a set of reflected lambda-terms as the smallest set such that bterm{x 1 , , x n .x i }, where 1 i n, is a lambda-term (a variable); if bterm x 1 , , x n , x n+1 .t[x] is a lambda-term, then bterm x 1 , , x n .x n+1 .t[x] is also a lambda-term (an abstraction); if bterm{x 1 , , x n .t 1 [x]} and bterm{x 1 , , x n .t 2 [x]} are lambda-terms, then bterm{x 1 ; ; x n .apply{t 1 [x]; t 2 [x]}} is also a lambda-term (an application). In a way, bterms could be understood as an explicit coding for Barzilay's substitution functions. And indeed, some of the basic definitions are quite similar. The notion of bterms is also very similar to that of local variable contexts <A href="3.html#10">[FPT99]. 3.2 Terminology Before we proceed further, we need to define some terminology. D EFINITION 1. We change the notion of subterm so that the subterms of a bterm are also bterms. For example, the immediate subterms of bterm{x , y.x y} are bterm{x , y.x } and bterm{x , y.y}; the immediate subterm of bterm{x .y.x } is bterm{x, y.x }. D EFINITION 2. We call the number of outer binders in a bterm expression its binding depth. Namely, the binding depth of the bterm bterm{x 1 , , x n .t[x]} is n. D EFINITION 3. Throughout the rest of the paper we use the notion of operator shape. The shape of an operator is a list of natural numbers each stating how many new binders the operator introduces on 5 the corresponding subterm. The length of the shape list is therefore the arity of the operator. For example, the shape of the + operator is [0; 0] and the shape of the operator is [1]. The mapping from operators to shapes is also sometimes called a binding signature of a language <A href="3.html#10">[FPT99, <A href="3.html#11">Plo90]. D EFINITION 4. Let op be an operator with shape [d 1 ; ; d N ], and let btl be a list of bterms [b 1 ; ; b M ]. We say that btl is compatible with op at depth n when, 1. N = M; 2. the binding depth of bterm b j is n + d j for each 1 j N . 3.3 Abstract Operators Expressions of the form bterm{x.op{ }} can only be used to express syntax with concrete operators. In other words, each expression of this form contains a specific constant operator op. However, we would like to reason about operators abstractly; in particular, we want to make it possible to have variables of the type "Op" that can be quantified over and used in the same manner as operator constants. In order to address this we use explicit term constructors in addition to bterm{x.op{ }} constants. The expression mk bterm{n; "op"; btl}, where "op" is some encoding of the quoted operator op, stands for a bterm with binding depth n, operator op and subterms btl. Namely, mk bterm{n; op; bterm{x 1 , , x n , y 1 .t 1 [x; y 1 ]} :: :: bterm{x 1 , , x n , y k .t k [x; y k ]} :: nil} is bterm{x 1 , , x n .op {y 1 .t 1 [x; y 1 ]; ; y k .t k [x; y k ]}}. Here, nil is the empty list and :: is the list cons operator and therefore the expression b 1 :: :: b n :: nil represents the concrete list [b 1 ; ; b n ]. Note that if we know the shape of the operator op and we know that the mk bterm expression is well-formed (or, more specifically, if we know that btl is compatible with op at depth n), then it would normally be possible to deduce the value of n (since n is the difference between the binding depth of any element of the list btl and the corresponding element of the shape(op) list). There are two reasons, however, for supplying n explicitly: When btl is empty (in other words, when the arity of op is 0), the value of n can not be deduced this way and still needs to be supplied somehow. One could consider 0-arity operators to be a special case, but this results in a significant loss of uniformity. When we do not know whether an mk bterm expression is necessarily well-formed (and as we will see it is often useful to allow this to happen), then a lot of definitions and proofs are greatly simplified when the binding depth of mk bterm expressions is explicitly specified. Using the mk bterm constructor and a few other similar constructors that will be introduced later, it becomes easy to reason abstractly about operators. Indeed, the second argument to mk bterm can now be an arbitrary expression, not just a constant. This has a cost of making certain definitions slightly more complicated. For example, the notion of "compatible with op at depth n" now becomes an important part of the theory and will need to be explicitly formalized. However, this is a small price to pay for the ability to reason abstractly about operators, which easily extends to reasoning abstractly about languages, classes of languages and so forth. 3.4 Inductively Defining the Type of Well-Formed Bterms There are two equivalent approaches to inductively defining the general type (set) of all well-formed bterms. The first one follows the same idea as in Example <A href="3.html#4">1: bterm{x 1 , , x n .x i } is a well-formed bterm for 1 i n; mk bterm{n; op; btl} is a well-formed bterm when op is a well-formed quoted operator and btl is a list of well-formed bterms that is compatible with op at some depth n. If we denote bterm{x 1 , , x l , y, z 1 , , z r .y} as var{l; r}, we can restate the base case of the above definition as "var{l; r }, where l and r are arbitrary natural numbers, is a well-formed bterm". Once we do this it becomes apparent that the above definition has a lot of similarities with de Bruijn-style indexing of variables <A href="3.html#10">[dB72]. Indeed, one might call the numbers l and r the left and right indices of the variable var{l; r }. It is possible to provide an alternate definition that is closer to pure HOAS: bnd{x .t [x]}, where t is a well-formed substitution function, is a well-formed bterm (the bnd operation increases the binding depth of t by one by adding x to the beginning of the list of t's outer binders). mk term{op; btl}, where op is a well-formed quoted operator, and btl is a list of well-formed bterms that is compatible with op at depth 0, is a well-formed bterm (of binding depth 0). Other than better capturing the idea of HOAS, the latter definition also makes it easier to express the reflective correspondence between the meta-syntax (the syntax used to express the theory of syntax, namely the one that includes the operators mk bterm, bnd, etc.) and the meta-meta-syntax (the syntax that is used to express the theory of syntax and the underlying theory, in other words, the syntax that includes the second-order notations.) Namely, provided that we define the subst{bt; t } operation to compute the result of substituting a closed term t for the first outer binder of the bterm bt, we can state that subst{bnd{x .t 1 [x]} ; t 2 } t 1 [t 2 ] (3) (where t 1 and t 2 are literal second-order variables). In other words, we can state that the substitution operator subst and the implicit second-order substitution in the "meta-meta-" language are equivalent . The downside of the alternate definition is that it requires defining the notion of "being a substitution function". 3.5 Our Approach In our work we try to combine the advantages of both approaches outlined above. In the next Section we present a theory that includes both the HOAS-style operations (bnd, mk term) and the de Bruijn-style ones (var, mk bterm). Our theory also allows deriving the equivalence <A href="3.html#5">(3). In our theory the definition of the basic syntactic operations is based on the HOAS-style operators; however, the recursive definition of the type of well-formed syntax is based on the de Bruijn-style operations. Our theory includes also support for variable-length lists of binders. Formal Implementation in a Theorem Prover In this Section we describe how the foundations of our theory are formally defined and derived in the NuPRL-style Computational Type Theory in the MetaPRL Theorem Prover. For brevity, we will present a slightly simplified version of our implementation; full details are available in the extended version of this paper <A href="3.html#11">[NKYH05, Appendix]. 4.1 Computations and Types In our work we make heavy usage of the fact that our type theory allows us to define computations without stating upfront (or even knowing) what the relevant types are. In NuPRL-style type theo-6 ries (which some even dubbed "untyped type theory"), one may define arbitrary recursive functions (even potentially nonterminating ones). Only when proving that such function belongs to a particular type, one may have to prove termination. See <A href="3.html#10">[All87a, All87b] for a semantics that justifies this approach. The formal definition of the syntax of terms consists of two parts: The definition of untyped term constructors and term operations , which includes both HOAS-style operations and de Bruijn-style operations. As it turns out, we can establish most of the reduction properties without explicitly giving types to all the operations. The definition of the type of terms. We will define the type of terms as the type that contains all terms that can be legitimately constructed by the term constructors. 4.2 HOAS Constructors At the core of our term syntax definition are two basic HOAS-style constructors: bnd{x .t [x]} is meant to represent a term with a free variable x. The intended semantics (which will not become explicit until later) is that bnd{x.t [x]} will only be considered well-formed when t is a substitution function. Internally, bnd{x.t [x]} is implemented simply as the pair 0, x.t [x] . This definition is truly internal and is used only to prove the properties of the two destructors presented below; it is never used outside of this Section (Section <A href="3.html#6">4.2). mk term{op; ts} pairs op with ts. The intended usage of this operation (which, again, will only become explicit later) is that it represents a closed term (i.e. a bterm of binding depth 0) with operator op and subterms ts. It will be considered well-formed when op is an operator and ts is a list of terms that is compatible with op at depth 0. For example, mk term{; bnd{x.x}} is x.x. Internally, mk term{op; ts} is implemented as the nested pair 1, op, ts . Again, this definition is never used outside of this Section. We also implement two destructors: subst{bt; t } is meant to represent the result of substituting term t for the first variable of the bterm bt. Internally, subst{bt; t } is defined simply as an application (bt.2) t (where bt.2 is the second element of the pair bt ). We derive the following property of this substitution operation: subst{bnd{x.t 1 [x]} ; t 2 } t 1 [t 2 ] where "" is the computational equality relation <A href="3.html#6">1 and t 1 and t 2 may be absolutely arbitrary, even ill-typed. This derivation is the only place where the internal definition of subst{bt; t} is used. Note that the above equality is exactly the "reflective property of substitution" <A href="3.html#5">(3) that was one of the design goals for our theory. weak dest {bt; bcase; op, ts.mkt case[op; ts]} is designed to provide a way to find out whether bt is a bnd{} or a mk term{op; ts} 1 In NuPRL-style type theories the computational equality relation (which is also sometimes called "squiggle equality" and is sometimes denoted as "" or "") is the finest-grained equality relation in the theory. When a b is true, a may be replaced with b in an arbitrary context. Examples of computational equality include beta-reduction x.a[x] b a[b], arithmetical equalities (1 + 2 3), and definitional equality (an abstraction is considered to be computationally equal to its definition). and to "extract" the op and ts in the latter case. In the rest of this paper we will use the "pretty-printed" form for weak dest -- "match bt with bnd{ } bcase | mk term{op; ts} mkt case[op; ts]". Internally, it is defined as if bt. 1 = 0 then bcase else mkt case[bt.2.1; bt.2.2]. From this internal definition we derive the following properties of weak dest: match bnd{x.t[x]} with bnd{ } bcase | mk term{op; ts} mkt case[op; ts] bcase match mk term{op; ts} with bnd{ } bcase | mk term{o; t} mkt case[o; t] mkt case[op; ts] 4.3 Vector HOAS Operations As we have mentioned at the end of Section <A href="3.html#2">2, some approaches to reasoning about syntax make it hard or even impossible to express arbitrary-length lists of binders. In our approach, we address this challenge by allowing operators where a single binding in the metalanguage stands for a list of object-level bindings. In particular, we allow representing bnd{x 1 .bnd{x 2 . bnd{x n .t[x 1 ; . . . ; x n ]} }} as vbnd{n; x .t [nth{1; x} ; . . . ; nth{n; x}]}, where "nth{i ; l}" is the "i th element of the list l" function. We define the following vector-style operations: vbnd{n; x .t [x]} represents a "telescope" of nested bnd operations . It is defined by induction <A href="3.html#6">2 on the natural number n as follows: vbnd{0; x.t [x]} := t [nil] vbnd{n + 1; x.t [x]} := bnd{v.vbnd{n; x .t [v :: x ]}} We also introduce vbnd{n; t } as a simplified notation for vbnd{n; x .t } when t does not have free occurrences of x. vsubst{bt; ts} is a "vector" substitution operation that is meant to represent the result of simultaneous substitution of the terms in the ts list for the first |ts| variables of the bterm bt (here |l| is the length of the list l). vsubst{bt; ts} is defined by induction on the list ts as follows: vsubst{bt; nil} := bt vsubst{bt; t :: ts} := vsubst{subst{bt; t } ; ts} Below are some of the derived properties of these operations: bnd{v.t [v]} vbnd{1; hd(v)} (4) m, n N. vbnd{m + n; x .t [x]} vbnd{m; y.vbnd{n; z.t [y@z]}} (5) l List. (vsubst{vbnd{|l|; v.t[v]} ; l} t[l]) (6) l List.n N. (n |l|) (vsubst{vbnd{n; v.t[v]} ; l} vbnd{n - |l|; v.bt[l@v]}) (7) n N. vbnd{n; l.vsubst{vbnd{n; v.t [v]} ; l}} vbnd{n; l.t [l]} (8) where "hd" is the list "head" operation, "@" is the list append operation, "List" is the type of arbitrary lists (the elements of a list do not have to belong to any particular type), N is the type of natural numbers, and all the variables that are not explicitly constrained to a specific type stand for arbitrary expressions. 2 Our presentation of the inductive definitions is slightly simplified by omitting some minor technical details. See <A href="3.html#11">[NKYH05, Appendix] for complete details. 7 Equivalence <A href="3.html#6">(5) allows the merging and splitting of vector bnd operations. Equivalence <A href="3.html#6">(6) is a vector variant of equivalence <A href="3.html#5">(3). Equivalence <A href="3.html#6">(8) is very similar to equivalence <A href="3.html#6">(6) applied in the vbnd{n; l. } context, except that <A href="3.html#6">(8) does not require l to be a member of any special type. 4.4 De Bruijn-style Operations Based on the HOAS constructors defined in the previous two sections , we define two de Bruijn-style constructors. var{i ; j } is defined as vbnd{i ; bnd{v.vbnd{ j ; v}}}. It is easy to see that this definition indeed corresponds to the informal bterm{x 1 , , x l , y, z 1 , , z r .y} definition given in Section <A href="3.html#5">3.4. mk bterm{n; op; ts} is meant to compute a bterm of binding depth n, with operator op, and with ts as its subterms. This operation is defined by induction on natural number n as follows: mk bterm{0; op; ts} := mk term{op; ts} mk bterm{n + 1; op; ts} := bnd{v.mk bterm{n; op; map t.subst{t ; v} ts}} Note that, if ts is a list of bnd expressions (which is the intended usage of the mk bterm operation), then the bnd{v. map t.subst{t ; v} ts } has the effect of stripping the outer bnd from each of the members of the ts list and "moving" them into a single "merged" bnd on the outside. We also define a number of de Bruijn-style destructors, i.e., operations that compute various de Bruijn-style characteristics of a bterm. Since the var and mk bterm constructors are defined in terms of the HOAS constructors, the destructors have to be defined in terms of HOAS operations as well. Because of this, these definitions are often far from straightforward. It is important to emphasize that the tricky definitions that we use here are only needed to establish the basic properties of the operations we defined. Once the basic theory is complete, we can raise the level of abstraction and no usage of this theory will ever require using any of these definitions, being aware of these definitions, or performing similar tricks again. bdepth{t } computes the binding depth of term t . It is defined recursively using the Y combinator as Y f.b.match b with bnd{ } 1 + f subst{b; mk term{0; 0}} | mk term{ ; } 0 t In effect, this recursive function strips the outer binders from a bterm one by one using substitution (note that here we can use an arbitrary mk bterm expression as a second argument for the substitution function; the arguments to mk bterm do not have to have the "correct" type) and counts the number of times it needs to do this before the outermost mk bterm is exposed. We derive the following properties of bdepth: l, r N. bdepth{var{l; r}} (l + r + 1) ; n N. bdepth{mk bterm{n; op; ts}} n . Note that the latter equivalence only requires n to have the "correct" type, while op and ts may be arbitrary. Since the bdepth operator is needed for defining the type of Term of well-formed bterms, at this point we would not have been able to express what the "correct" type for ts would be. left{t } is designed to compute the "left index" of a var expression . It is defined as Y f.b.l. match b with bnd{ } 1 + f subst{b; mk term{l; 0}} (l + 1) | mk term l ; l t 0 In effect, this recursive function substitutes mk term{0; 0} for the first binding of t , mk term{1; 0} for the second one, mk term{2; 0} for the next one and so forth. Once all the binders are stripped and a mk term{l; 0} is exposed, l is the index we were looking for. Note that here we intentionally supply mk term with an argument of a "wrong" type (N instead of Op); we could have avoided this, but then the definition would have been significantly more complicated. As expected, we derive that l, r N.(left{var{l; r}} l). right{t } computes the "right index" of a var expression. It is trivial to define in terms of the previous two operators: right{t } := bdepth{t } - left{t } - 1. get op{t ; op} is an operation such that n N. get op mk bterm{n; op; ts} ; op op , l, r N. (get op{var{i; j} ; op} op . Its definition is similar to that of left{}. subterms{t } is designed to recover the last argument of a mk bterm expression. The definition is rather technical and complicated, so we omit it; see <A href="3.html#11">[NKYH05, Appendix C] for details. The main property of the subterms operation that we derive is n N.btl List. subterms{mk bterm{n; op; btl}} map b.vbnd{n; v.vsubst{b; v}} btl The right-hand side of this equivalence is not quite the plain "btl" that one might have hoped to see here. However, when btl is a list of bterms with binding depths at least n, which is necessarily the case for any well-formed mk bterm{n; op; btl}, equivalence <A href="3.html#6">(8) would allow simplifying this right-hand side to the desired btl. 4.5 Operators For this basic theory the exact representation details for operators are not essential and we define the type of operators Op abstractly. We only require that operators have decidable equality and that there exist a function of the type Op N List that computes operators' shapes. Using this shape function and the bdepth function from Section <A href="3.html#7">4.4, it is trivial to formalize the "ts is compatible with op at depth n" predicate of Definition <A href="3.html#5">4. We denote this predicate as shape compat{n; op; ts} and define it as |shape{op}| = |btl| i 1..|btl|.bdepth{nth{btl; i}} = n + nth{shape{op}; i} 4.6 The Type of Terms In this section we will define the type of terms (i.e. well-formed bterms), Term, as the type of all terms that can be constructed by the de Bruijn constructors from Section <A href="3.html#7">4.4. That is, the Term type contains all expressions of the forms: var{i ; j } for all natural numbers i, j ; or 8 mk bterm{n; op; ts} for any natural number n, operator op, and list of terms ts that is compatible with op at depth n. The Term type is defined as a fixpoint of the following function from types to types: Iter(X ) := Image(dom(X ); x .mk(x )), where Image is a type constructor such that Image(T ; x. f [x]) is the type of all the f [t ] for t T (for it to be well-formed, T must be a well-formed type and f must not have any free variables except for x); dom(X ) is a type defined as (NN)+ n:Nop:Op{ts:X List | shape compat{n; op; ts}} ; and mk(x) (where x is presumably a member of the type dom(X )) is defined as match x with inl (i, j) var{i; j} | inr (n, op, ts) mk bterm{n; op; ts} . The fixpoint of Iter is reached by defining Term 0 := Void (an empty type) Term n+1 := Iter(Term n ) Term := n N Term n We derive the intended introduction rules for the Term type: i N j N var{i ; j } Term and n N op Op ts Term List shape compat{n; op; ts} mk bterm{n; op; ts} Term . Also, the structural induction principle is derived for the Term type. Namely, we show that to prove that some property P[t ] holds for any term t , it is sufficient to prove (Base case) P holds for all variables, that is, P[var{i ; j }] holds for all natural numbers i and j ; (Induction step) P[mk bterm{n; op; ts}] is true for any natural number n, any operator op, and any list of terms ts that is compatible with op at depth n, provided P[t ] is true for any element t of the list ts. Note that the type of "terms over n variables" (where n = 0 corresponds to closed terms) may be trivially defined using the Term type and the "subset" type constructor -- {t : Term | bdepth{t } = n}. Conclusions and Future Work In Sections <A href="3.html#4">3 and <A href="3.html#5">4 we have presented a basic theory of syntax that is fully implemented in a theorem prover. As we mentioned in the introduction, the approach is both natural and expressive, and provides a foundation for reflective reasoning about classes of languages and logics. However, we consider this theory to be only the first step towards building a user-accessible uniform reflection framework and a user-accessible uniform framework for programming language reasoning and experimentation, where tasks similar to the ones presented in the P OPL M ARK challenge <A href="3.html#9">[ABF <A href="3.html#9">+ <A href="3.html#9">05] can be performed easily and naturally. In this section we provide an outline of our plans for building such frameworks on top of the basic syntactic theory. 5.1 Higher-Level User Interface One obvious shortcoming of the theory presented in Sections <A href="3.html#4">3 and <A href="3.html#5">4 is that it provides only the basic low-level operations such as bnd, var, subterms, etc. It presents a very low-level account of syntax in a way that would often fail to abstract away the details irrelevant to the user. To address this problem we are planning to provide user interface functionality capable of mapping the high-level concepts to the low-level ones. In particular, we are going to provide an interface that would allow instantiating general theorems to specific collections of operators and specific languages. Thus, the user will be able to write something like "reflect language [x.; apply{; }]" and the system will create all the components outlined in Example <A href="3.html#4">1: It will create a definition for the type Language[x.; apply{; }] of reflected lambda-terms (where Language[l] is a general definition of a language over a list of operators l); It will state and derive the introduction rules for this type; It will state and derive the elimination rule for this type (the induction principle). Moreover, we are planning to support even more complicated language declarations, such as t := int | t t ; e := v | x : t.e[x ] | apply{e; e} that would cause the system to create mutually recursive type definitions and appropriate rules. Finally, we are also planning to support "pattern bindings" that are needed for a natural encoding of ML-like pattern matching (such as the one sketched in the P OPL M ARK challenge <A href="3.html#9">[ABF <A href="3.html#9">+ <A href="3.html#9">05]). As far as the underlying theory goes, we believe that the mechanisms very similar to the "vector bindings" presented in Section <A href="3.html#6">4.3 will be sufficient here. 5.2 "Dereferencing" Quoted Terms As in Barzilay's work, the quoted operator approach makes it easy to define the "unquoting" (or "dereferencing") operator [[]] unq . If t is a syntactic term, then [[t ]] unq is the value represented by t . By definition, [[op{t 1 ; . . . ; t n }]] unq = op{[[t 1 ]] unq ; . . . ; [[t n ]] unq }. For instance, [[2 3]] unq is 2 3 (i.e. 6). In order to define unquoting on terms with bindings, we need to introduce the "guard" operation such that [[ t ]] unq is t for an arbitrary expression t . Then [[]] unq can be defined as follows: [[op{x 1 , , x k .t[x 1 ; . . . ; x k ]; ;z 1 , , z l .s[z 1 ; . . . ; z l ]}]] unq = op{x 1 , , x k .[[t[ x 1 ; . . . ; x k ]]] unq ; ; z 1 , , z l .[[s[ z 1 ; . . . ; z l ]]] unq }. For example, [[x.2 x]] unq = x.[[2 x ]] unq = x.[[2]] unq [[ x ]] unq = x.2 x. The unquote operation establishes the identity between the original syntax and the reflected syntax, making it a "true" reflection. Note that the type theory (which ensures, in particular, that only terminating functions may be shown to belong to a function type) would keep the [[ ]] unq operation from introducing logical paradox<A href="3.html#8">es. <A href="3.html#8">3 3 This is, obviously, not a proper argument. While a proper argument can be made here, it is outside of the scope of this particular paper. 9 Also, since the notion of the quoted operators is fully open-ended , each new language added to the system will automatically get to use the [[ ]] unq operation for all its newly introduced operators . 5.3 Logical Reflection After defining syntactic reflection, it is easy to define logical reflection . If we consider the proof system open-ended, then the logical reflection is trivial -- when P is a quotation of a proposition, we can regard "[[P]] unq " as meaning " P is true". The normal modal rules for the [[]] unq modality are trivially derivable. For example modus ponens [[P Q]] unq [[P]] unq [[Q]] unq is trivially true because if we evaluate the first [[]] unq (remember, [[P Q]] unq = [[P]] unq [[Q]] unq by definition of [[]] unq ), we get an obvious tautology ([[P]] unq [[Q]] unq ) [[P]] unq [[Q]] unq . In order to consider a closed proof system (in other words, if we want to be able to do induction over derivations), we would need to define a provability predicate for that system. We are planning to provide user interface functionality that would allow users to describe a set of proof rules and the system would generate appropriate proof predicate definitions and derive appropriate rules (in a style similar to the one outlined in Section <A href="3.html#8">5.1 for the case of language descriptions). Related Work In Section <A href="3.html#2">2 we have already discussed a number of approaches that we consider ourselves inheriting from. Here we would like to revisit some of them and mention a few other related efforts. Our work has a lot in common with the HOAS implemented in Coq by Despeyroux and Hirschowitz <A href="3.html#10">[DH94]. In both cases, the more general space of terms (that include the exotic ones) is later restricted in a recursive manner. In both cases, the higher-order analogs of first-order de Bruijn operators are defined and used as a part of the "well-formedness" specification for the terms. Despeyroux and Hirschowitz use functions over infinite lists of variables to define open terms, which is similar to our vector bindings. There are a number of significant differences as well. Our approach is sufficiently syntactical, which allows eliminating all exotic terms, even those that are extensionally equal to the well-formed ones, while the more semantic approach of <A href="3.html#10">[DH94, <A href="3.html#10">DFH95] has to accept such exotic terms (their solution to this problem is to consider an object term to be represented by the whole equivalence class of extensionally equal terms); more generally while <A href="3.html#10">[DH94] states that "this problem of extensionality is recurrent all over our work", most of our lemmas establish identity and not just equality, thus avoiding most of the issues of extensional equality. In our implementation, the substitution on object terms is mapped directly to -reduction, while Despeyroux et al. <A href="3.html#10">[DFH95] have to define it recursively. In addition, we provide a uniform approach to both free and bound variables that naturally extends to variable-length "vector" bindings. While our approach is quite different from the modal -calculus one <A href="3.html#10">[DPS97, DL99, DL01], there are some similarities in the intuition behind it. Despeyroux et al. <A href="3.html#10">[DPS97] says "Intuitively, we interpret B as the type of closed objects of type B. We can iterate or distinguish cases over closed objects, since all constructors are statically known and can be provided for." The intuition behind our approach is in part based on the canonical model of the NuPRL type theory <A href="3.html#10">[All87a, All87b], where each type is mapped to an equivalence relations over the closed terms of that type. Gordon and Melham <A href="3.html#10">[GM96] define the type of -terms as a quotient of the type of terms with concrete binding variables over -equivalence. Michael Norrish <A href="3.html#11">[Nor04] builds upon this work by replacing certain variable "freshness" requirements with variable "swapping". This approach has a number of attractive properties; however, we believe that the level of abstraction provided by the HOAS-style approaches makes the HOAS style more convenient and accessible. Ambler, Crole, and Momigliano <A href="3.html#9">[ACM02] have combined the HOAS with the induction principle using an approach which in some sense is opposite to ours. Namely, they define the HOAS operators on top of the de Bruijn definition of terms using higher order pattern matching. In a later work <A href="3.html#9">[ACM03] they have de-scribed the notion of "terms-in-infinite-context" which is quite similar to our approach to vector binding. While our vector bindings presented in Section <A href="3.html#6">4.3 are finite length, the exact same approach would work for the infinite-length "vectors" as well. Acknowledgments The authors are grateful to Eli Barzilay whose ideas were an inspiration for some of the work that lead to this paper. We are also grateful for his comments on an early draft of this paper. We are grateful to the anonymous reviewers for their very thorough and fair feedback and many helpful suggestions. References [AA99] Eric Aaron and Stuart Allen. Justifying calculational logic by a conventional metalinguistic semantics. Technical Report TR99-1771, Cornell University, Ithaca, New York, September 1999. [ABF + 05] Brian E. Aydemir, Aaron Bohannon, Matthew Fairbairn, J. Nathan Foster, Benjamin C. Pierce, Peter Sewell, Dimitrios Vytiniotis, Geoffrey Washburn, Stephanie Weirich, and Steve Zdancewic. Mechanized metatheory for the masses: The POPLmark challenge. Available from <A href="http://www.cis.upenn.edu/group/proj/plclub/mmm/">http://www.cis. upenn.edu/group/proj/plclub/mmm/, 2005. [AC92] William Aitken and Robert L. Constable. Reflecting on NuPRL : Lessons 14. Technical report, Cornell University, Computer Science Department, Ithaca, NY, 1992. [ACE + 00] Stuart Allen, Robert Constable, Richard Eaton, Christoph Kreitz, and Lori Lorigo. The NuPRL open logical environment . In David McAllester, editor, Proceedings of the 17 t h International Conference on Automated Deduction, volume 1831 of Lecture Notes in Artificial Intelligence, pages 170176. Springer Verlag, 2000. [ACHA90] Stuart F. Allen, Robert L. Constable, Douglas J. Howe, and William Aitken. The semantics of reflected proof. In Proceedings of the 5 t h Symposium on Logic in Computer Science, pages 95197. IEEE Computer Society Press, June 1990. [ACM02] Simon Ambler, Roy L. Crole, and Alberto Momigliano. Combining higher order abstract syntax with tactical theorem proving and (co)induction. In TPHOLs '02: Proceedings of the 15th International Conference on Theorem Proving in Higher Order Logics, pages 1330, London, UK, 2002. Springer-Verlag. [ACM03] S. J. Ambler, R. L. Crole, and Alberto Momigliano. A definitional approach to primitive recursion over higher order abstract syntax. In Proceedings of the 2003 workshop on Mechanized reasoning about languages with variable binding, pages 111. ACM Press, 2003. [ACU93] William Aitken, Robert L. Constable, and Judith Underwood. Metalogical Frameworks II: Using reflected decision procedures . Journal of Automated Reasoning, 22(2):171221, 1993. 10 [All87a] Stuart F. Allen. A Non-type-theoretic Definition of Martin-Lof's Types. In D. Gries, editor, Proceedings of the 2 nd IEEE Symposium on Logic in Computer Science, pages 215224. IEEE Computer Society Press, June 1987. [All87b] Stuart F. Allen. A Non-Type-Theoretic Semantics for Type-Theoretic Language. PhD thesis, Cornell University, 1987. [Art99] Sergei Artemov. On explicit reflection in theorem proving and formal verification. In Ganzinger <A href="3.html#10">[Gan99], pages 267 281. [Art04] Sergei Artemov. Evidence-based common knowledge. Technical Report TR-2004018, CUNY Ph.D. Program in Computer Science Technical Reports, November 2004. [BA02] Eli Barzilay and Stuart Allen. Reflecting higher-order abstract syntax in NuPRL. In Victor A. Carre~no, Cezar A. Mu~noz, and Sophi`ene Tahar, editors, Theorem Proving in Higher Order Logics; Track B Proceedings of the 15 t h International Conference on Theorem Proving in Higher Order Logics (TPHOLs 2002), Hampton, VA, August 2002, pages 2332. National Aeronautics and Space Administration, 2002. [BAC03] Eli Barzilay, Stuart Allen, and Robert Constable. Practical reflection in NuPRL. Short paper presented at 18th Annual IEEE Symposium on Logic in Computer Science, June 22 25, Ottawa, Canada, 2003. [Bar01] Eli Barzilay. Quotation and reflection in NuPRL and Scheme. Technical Report TR2001-1832, Cornell University, Ithaca, New York, January 2001. [Bar05] Eli Barzilay. Implementing Reflection in NuPRL. PhD thesis, Cornell University, 2005. In preparation. [CAB + 86] Robert L. Constable, Stuart F. Allen, H. M. Bromley, W. R. Cleaveland, J. F. Cremer, R. W. Harper, Douglas J. Howe, T. B. Knoblock, N. P. Mendler, P. Panangaden, James T. Sasaki, and Scott F. Smith. Implementing Mathematics with the NuPRL Proof Development System. Prentice-Hall, NJ, 1986. [CFW04] Luis Crus-Filipe and Freek Weidijk. Hierarchical reflection. In Slind et al. <A href="3.html#11">[SBG04], pages 6681. [Con94] Robert L. Constable. Using reflection to explain and enhance type theory. In Helmut Schwichtenberg, editor, Proof and Computation, volume 139 of NATO Advanced Study Institute , International Summer School held in Marktoberdorf, Germany, July 20-August 1, NATO Series F, pages 65100. Springer, Berlin, 1994. [dB72] N. G. de Bruijn. Lambda calculus notation with nameless dummies, a tool for automatic formula manipulation, with application to the Church-Rosser theorem. Indagaciones Mathematische, 34:381392, 1972. This also appeared in the Proceedings of the Koninklijke Nederlandse Akademie van Wetenschappen, Amsterdam, series A, 75, No. 5. [DFH95] Joelle Despeyroux, Amy Felty, and Andre Hirschowitz. Higher-order abstract syntax in Coq. In M. Dezani-Ciancaglini and G. Plotkin, editors, Proceedings of the International Conference on Typed Lambda Calculus and its Applications, volume 902 of Lecture Notes in Computer Science, pages 124138. Springer-Verlag, April 1995. Also appears as <A href="http://www.inria.fr/rrrt/rr-2556.html">INRIA research report RR-2556. [DH94] Joelle Despeyroux and Andre Hirschowitz. Higher-order abstract syntax with induction in Coq. In LPAR '94: Proceedings of the 5th International Conference on Logic Programming and Automated Reasoning, volume 822 of Lecture Notes in Computer Science, pages 159173. Springer-Verlag, 1994. Also appears as <A href="http://www.inria.fr/rrrt/rr-2292.html">INRIA research report RR-2292. [DH95] James Davis and Daniel Huttenlocher. Shared annotations for cooperative learning. In Proceedings of the ACM Conference on Computer Supported Cooperative Learning, September 1995. [DL99] Joelle Despeyroux and Pierre Leleu. A modal lambda calculus with iteration and case constructs. In T. Altenkirch, W. Naraschewski, and B. Reus, editors, Types for Proofs and Programs: International Workshop, TYPES '98, Kloster Irsee, Germany, March 1998, volume 1657 of Lecture Notes in Computer Science, pages 4761, 1999. [DL01] Joelle Despeyroux and Pierre Leleu. Recursion over objects of functional type. Mathematical Structures in Computer Science, 11(4):555572, 2001. [DPS97] Joelle Despeyroux, Frank Pfenning, and Carsten Schurmann. Primitive recursion for higherorder abstract syntax. In R. Hindley, editor, Proceedings of the Third International Conference on Typed Lambda Calculus and Applications (TLCA'97), volume 1210 of Lecture Notes in Computer Science, pages 147163. Springer-Verlag, April 1997. An extended version is available as <A href="http://reports-archive.adm.cs.cmu.edu/anon/1996/CMU-CS-96-172.ps.gz">Technical Report CMU-CS-96 -172, Carnegie Mellon University. [EM71] Andrzej Ehrenfeucht and Jan Mycielski. Abbreviating proofs by adding new axioms. Bulletin of the American Mathematical Society, 77:366367, 1971. [F + 86] Solomon Feferman et al., editors. Kurt Godel Collected Works, volume 1. Oxford University Press, Oxford, Clarendon Press, New York, 1986. [FPT99] Marcelo Fiore, Gordon Plotkin, and Daniele Turi. Abstract syntax and variable binding. In Proceedings of 14 t h IEEE Symposium on Logic in Computer Science, pages 193+. IEEE Computer Society Press, 1999. [Gan99] Harald Ganzinger, editor. Proceedings of the 16 t h International Conference on Automated Deduction, volume 1632 of Lecture Notes in Artificial Intelligence, Berlin, July 710 1999. Trento, Italy. [GM96] A. D. Gordon and T. Melham. Five axioms of alpha-conversion . In J. von Wright, J. Grundy, and J. Harrison, editors, Theorem Proving in Higher Order Logics: 9th International Conference, Turku, Finland, August 1996: Proceedings, volume 1125 of Lecture Notes in Computer Science, pages 173190. Springer-Verlag, 1996. [GMO03] Jim Grundy, Tom Melham, and John O'Leary. A reflective functional language for hardware design and theorem proving. Technical Report PRG-RR-03-16, Oxford Univerity, Computing Laboratory, 2003. [God31] Kurt Godel. Uber formal unentscheidbare satze der principia mathematica und verwandter systeme I. Monatshefte fur Mathematik und Physik, 38:173198, 1931. English version in <A href="3.html#11">[vH67]. [God36] K. Godel. Uber die Lange von beweisen. Ergebnisse eines mathematischen Kolloquiums, 7:2324, 1936. English translation in <A href="3.html#10">[F <A href="3.html#10">+ <A href="3.html#10">86], pages 397399. [GS89] F. Giunchiglia and A. Smaill. Reflection in constructive and non-constructive automated reasoning. In H. Abramson and M. H. Rogers, editors, Meta-Programming in Logic Programming, pages 123140. MIT Press, Cambridge, Mass., 1989. [GWZ00] H. Geuvers, F. Wiedijk, and J. Zwanenburg. Equational reasoning via partial reflection. In J. Harrison and M. Aagaard, editors, Theorem Proving in Higher Order Logics: 13 t h International Conference, TPHOLs 2000, volume 1869 of Lecture Notes in Computer Science, pages 162178. Springer-Verlag, 2000. [HAB + ] Jason J. Hickey, Brian Aydemir, Yegor Bryukhov, Alexei Kopylov, Aleksey Nogin, and Xin Yu. A listing of MetaPRL theories. <A href="http://metaprl.org/theories.pdf">http://metaprl.org/theories.pdf. [Har95] J. Harrison. Metatheory and reflection in theorem proving: A survey and critique. Technical Report CRC-53, SRI International, Cambridge Computer Science Research Centre, Millers Yard, Cambridge, UK, February 1995. 11 [HHP93] Robert Harper, Furio Honsell, and Gordon Plotkin. A framework for defining logics. Journal of the Association for Computing Machinery, 40(1):143184, January 1993. A revised and expanded verion of '87 paper. [Hic97] Jason J. Hickey. NuPRL-Light: An implementation framework for higher-order logics. In William McCune, editor, Proceedings of the 14 t h International Conference on Automated Deduction, volume 1249 of Lecture Notes in Artificial Intelligence, pages 395399. Springer, July 1317 1997. An extended version of the paper can be found at <A href="http://www.cs.caltech.edu/~jyh/papers/cade14_nl/default.html">http://www.cs.caltech.edu/~jyh/papers/cade14_ nl/default.html. [Hic99] Jason J. Hickey. Fault-tolerant distributed theorem proving. In Ganzinger <A href="3.html#10">[Gan99], pages 227231. [Hic01] Jason J. Hickey. The MetaPRL Logical Programming Environment. PhD thesis, Cornell University, Ithaca, NY, January 2001. [HL78] Gerard P. Huet and Bernard Lang. Proving and applying program transformations expressed with second-order patterns. Acta Informatica, 11:3155, 1978. [HNC + 03] Jason Hickey, Aleksey Nogin, Robert L. Constable, Brian E. Aydemir, Eli Barzilay, Yegor Bryukhov, Richard Eaton, Adam Granicz, Alexei Kopylov, Christoph Kreitz, Vladimir N. Krupski, Lori Lorigo, Stephan Schmitt, Carl Witty, and Xin Yu. MetaPRL -- A modular logical environment . In David Basin and Burkhart Wolff, editors, Proceedings of the 16 t h International Conference on Theorem Proving in Higher Order Logics (TPHOLs 2003), volume 2758 of Lecture Notes in Computer Science, pages 287303. Springer-Verlag, 2003. [HNK + ] Jason J. Hickey, Aleksey Nogin, Alexei Kopylov, et al. MetaPRL home page. <A href="http://metaprl.org/">http://metaprl.org/. [Mos52] Andrzej Mostowski. Sentences undecidable in formalized arithmetic: an exposition of the theory of Kurt Godel. Amsterdam: North-Holland, 1952. [NH02] Aleksey Nogin and Jason Hickey. Sequent schema for derived rules. In Victor A. Carre~no, Cezar A. Mu~noz, and Sophi`ene Tahar, editors, Proceedings of the 15 t h International Conference on Theorem Proving in Higher Order Logics (TPHOLs 2002), volume 2410 of Lecture Notes in Computer Science, pages 281297. Springer-Verlag, 2002. [NKYH05] Aleksey Nogin, Alexei Kopylov, Xin Yu, and Jason Hickey. A computational approach to reflective meta-reasoning about languages with bindings. Technical Report CaltechCSTR :2005.003, California Institure of Technology, 2005. Available at <A href="http://resolver.caltech.edu/CaltechCSTR:2005.003">http://resolver.caltech.edu/ CaltechCSTR:2005.003. [Nor04] Michael Norrish. Recursive function definition for types with binders. In Slind et al. <A href="3.html#11">[SBG04], pages 241256. [Par71] R. Parikh. Existence and feasibility in arithmetic. The Journal of Symbolic Logic, 36:494508, 1971. [Pau94] Lawrence C. Paulson. Isabelle: A Generic Theorem Prover, volume 828 of Lecture Notes in Computer Science. Springer-Verlag , New York, 1994. [PE88] Frank Pfenning and Conal Elliott. Higher-order abstract syntax. In Proceedings of the ACM SIGPLAN '88 Conference on Programming Language Design and Implementation (PLDI), volume 23(7) of SIGPLAN Notices, pages 199208, Atlanta, Georgia, June 1988. ACM Press. [Pfe89] Frank Pfenning. Elf: a language for logic definition and verified metaprogramming. In Proceedings of the 4 t h IEEE Symposium on Logic in Computer Science, pages 313322, Asilomar Conference Center, Pacific Grove, California, June 1989. IEEE Computer Society Press. [Plo90] Gordon Plotkin. An illative theory of relations. In R. Cooper, K. Mukai, and J. Perry, editors, Situation Theory and Its Applications, Volume 1, number 22 in CSLI Lecture Notes, pages 133146. Centre for the Study of Language and Information, 1990. [PN90] L. Paulson and T. Nipkow. Isabelle tutorial and user's manual . Technical report, University of Cambridge Computing Laboratory, 1990. [SBG04] Konrad Slind, Annette Bunker, and Ganesh Gopalakrishnan, editors. Proceedings of the 17 t h International Conference on Theorem Proving in Higher Order Logics (TPHOLs 2004), volume 3223 of Lecture Notes in Computer Science. Springer-Verlag, 2004. [Sch01] Carsten Schurmann. Recursion for higher-order encodings. In L. Fribourg, editor, Computer Science Logic, Proceedings of the 10 t h Annual Conference of the EACSL, volume 2142 of Lecture Notes in Computer Science, pages 585599. Springer-Verlag, 2001. [Smi84] B.C. Smith. Reflection and semantics in Lisp. Principles of Programming Languages, pages 2335, 1984. [vH67] J. van Heijenoort, editor. From Frege to Godel: A Source Book in Mathematical Logic, 18791931. Harvard University Press, Cambridge, MA, 1967. 12

system reflection;programming language;High order abstract syntax;formal languages;Theorem prover;NuPRL;Meta-syntax;MetaPRL theorem prover;Languages with bindings;Uniform reflection framework;Higher-Order Abstract Syntax;Bruijn-style operations;HOAS-style operations;NuPRL-like Martin-Lof-style computational type theory;higher-order abstract syntax;Type Theory;formal definition and theory;computer aided reasoning;Meta-reasoning;Recursive definition;Reflective reasoning;Reflection;Languages with Bindings;Substitution;MetaPRL;Runtime code generation;Meta-language;uniform reflection framework;Theory of syntax;Programming Language Experimentation

30

An Architectural Style for High-Performance Asymmetrical Parallel Computations

Researchers with deep knowledge of scientific domains are now developing highly-adaptive and irregular (asymmetrical ) parallel computations, leading to challenges in both delivery of data for computation and mapping of processes to physical resources. Using software engineering principles, we have developed a new communications protocol and architectural style for asymmetrical parallel computations called ADaPT. Utilizing the support of architecturally-aware middleware, we show that ADaPT provides a more efficient solution in terms of message passing and load balancing than asymmetrical parallel computations using collective calls in the Message-Passing Interface (MPI) or more advanced frameworks implementing explicit load-balancing policies. Additionally , developers using ADaPT gain significant windfall from good practices in software engineering, including implementation-level support of architectural artifacts and separation of computational loci from communication protocols

INTRODUCTION In recent years, as the cost-to-performance ratio of consumer hardware has continued to decrease, computational clusters consisting of fast networks and commodity hardware have become a common sight in research laboratories. A Copyright is held by the author/owner. ICSE'06, May 2028, 2006, Shanghai, China. ACM 1-59593-085-X/06/0005. growing number of physicists, biologists, chemists, and computer scientists have developed highly-adaptive and irregular parallel applications that are characterized by computational intensity, loosely-synchronous parallelism and dynamic computation. Because the computation time of each parallel process varies significantly for this class of computation , we shall refer to them as asymmetrical parallel computations . Adaptive mesh refinements for the simulation of crack growth, combinatorial search applications used in artificial intelligence, and partial differential equation field solvers [2] are examples of asymmetrical computations. While supercomputing platforms available to us continue to increase performance, our ability to build software capable of matching theoretical limits is lacking [8]. At the same time, researchers with significant depth of knowledge in a scientific domain but with limited software experience are confounded by the interface bloat of libraries such the Message-Passing Interface (MPI), which has 12 different routines for point-to-point communications alone [5]. Would-be practitioners of high-performance computing are introduced early to the mantra of optimization. The myth that high-level concepts inherent to software engineering principles, such as "separation of concerns," result in inefficiencies at the performance level has caused these researchers to eschew best practices of traditional software development in favor of highly-optimized library routines. We contend that a sound software engineering solution to asymmetrical parallel computations provides decoupling of connectors from computational loci and reduces the complexity of development for the programmer while still providing an efficient solution both in terms of load-balancing and message-delivery. In this paper, we present such a solution . In the next section, we will discuss our motivations for creating the ADaPT protocol and architecture, including the load-balancing inefficiencies of "optimized" communications libraries when computing asymmetrical parallel computations . We will then present ADaPT, a communications protocol and associated software architecture for asymmetrical computations. Additionally, we will present analysis which shows ADaPT's ability to outperform both MPI and other load-balancing frameworks using traditional work-sharing strategies. We conclude with an overview of future research opportunities afforded by ADaPT. MOTIVATION This work has been motivated by our experience with two key classes of existing approaches: use of optimized 857 communications libraries such as MPI [4], and message-passing frameworks which implement load-balancing strategies based on work sharing. 2.1 Message-Passing Interface High-performance communications libraries such as MPI are optimized to reduce the bandwidth needed to communicate a large amount of data to subprocesses. In order to accomplish this reduction, collective calls in MPI are synchronous , causing barriers at data-distribution points in the software. When used to compute uniform parallel computations barriers are unobtrusive. In asymmetrical computations , however, an effective mapping of processes to physical resources contributes more significantly to wall-clock time to completion than efficient communications. For these computations , asynchronous communications are needed, despite increased bandwidth. To illustrate this phenomena, let us consider a mapping of a large normalized population of computation times with a high level of variance onto a significantly smaller number of physical nodes (a strategy known as overaggregation). The MPI library offers developers efficient use of bandwidth via collective scatter and gather commands. While bandwidth is conserved using these collective calls, analyses made by Gropp, et. al. and Skjellum [10, 4] suggest that most implementations of MPI's scatter are built on top of MPI's rendezvous protocol and result in a synchronous barrier at each subsequent distribution of data. Since each process has variable computation time, a number of subprocesses will remain idle until the longest process completes during each of the scatters. In [1] we have shown that the smallest contribution to overall wall-clock time to completion made by this idle time is given as n , where n is the number of subprocesses and is the mean of the computation times. In comparison, the wall-clock time saved using the collective calls to reduce bandwidth is negligible. While these collective calls only consider bandwidth optimizations , it is clear that in asymmetrical parallel computations , process load-balancing across subprocesses is a more important optimization to pursue. 2.2 Load-Balancing Frameworks Attempts to develop message-passing frameworks that can assist computational scientists in the development of asymmetrical parallel computations can be divided into two groups: static load-balancing frameworks and dynamic load-balancing frameworks. Because a priori knowledge of the computation involved in asymmetrical parallel computations is required of static load balancers, such frameworks are inapplicable to this class of problems. Unlike static load balancers, dynamic load-balancing frameworks do not require information a priori and are able to redeploy balanced distributions of data during program execution . Notable examples of parallel development frameworks which provide dynamic load-balancing are PREMA [2] and Charm++ [6]. Unfortunately, these frameworks often incur significant performance losses due to the introduction of barriers for load-balancing. Additionally, these frameworks do not provide explicit support for consistency of structure and development. A software architectural solution can provide a number of benefits in addition to load balancing. Employing a sound software engineering principle, the separation of communication from computational elements shields the developer from the need to optimize communications and provides enforcement of architectural constraints. An added benefit is that architectural components reified as explicit implementation-level artifacts allow for easy reconfiguration of software in principle. We will revisit this point below. A NOVEL PROTOCOL Two overlooked aspects of performance optimizations that must be addressed in order to provide a truly efficient solution are asynchronous load-balancing and event pattern optimization . In addition to simply providing a load-balanced distribution, asynchronous load-balancing provides a best effort redistribution of processes without introducing a barrier to computation. Event pattern optimization suggests that a protocol is capable of utilizing the predictability of future messages given analysis of past messages. During overaggregated parallel computations, a number of computations need to be distributed to each of the subprocesses over the course of the parallel computation, causing a pattern to emerge. In order to incorporate each of these optimizations into a high-performance communications protocol, we have developed ADaPT, an Adaptive Data-parallel Publication/ Subscription Transport protocol and software architecture. The thesis of ADaPT is that it is possible to exploit the sequence that emerges from sending multiple messages to each parallel process in order to reduce the overall wall clock time to completion of the computation while still making a best-effort to avoid sending messages to each subprocess to quickly for the process to buffer. 3.1 ADaPT Defined We feel that for the purposes of this paper it is most helpful to define ADaPT's protocol, architectural elements, and implementation. 3.1.1 Protocol ADaPT views each parallel process as an independent software component (Worker) residing on a physical node capable of performing computations on data. Each Worker initiates computation by subscribing to a coordination component (Master). An important distinction between ADaPT and traditional publication/subscription systems is that unlike traditional pub/sub systems, ADaPT does not duplicate messages to service multiple downstream requests. Rather, it distributes messages uniquely from a queue in a round-robin fashion. Upon receipt of a subscription, the Master publishes a message to the Worker. There is another divergence from traditional pub/sub systems at this point. The Master waits for another request from the subscribed Worker before publishing another message to that Worker. Using data from each subscribed Worker on its computation time, or , the Master tracks an average processing time, or . Because the protocol is adaptive, when a predetermined number of messages have been sent to the Workers and a has been calculated, the Master switches from this conservative phase to an aggressive phase during which it sends messages of the requested type to the process at the regular interval dictated by . The protocol exploits the emerging event pattern to reduce the overall processing time at each physical node. 858 0 5 0 100 150 200 0 5 10 15 20 Number of Workers % Overhead Computation Time Variance = 10 LBF MP I ADaPT 0 50 100 150 200 0 5 10 15 20 Number of Workers % Overhead Computation Time Variance = 50 LBF MPI ADaPT 0 50 100 150 200 0 5 10 15 20 Number of Workers % Overhead Computation Time Variance = 100 LBF MPI ADaPT 0 50 100 150 200 0 5 10 15 20 Number of Workers % Overhead Computation Time Variance = 200 LBF MPI ADaPT Figure 1: Monte Carlo simulations of overhead for asymmetrical computations exhibiting multiple variances. Similar to MPI's eager protocol, this phase of ADaPT can be too aggressive, flooding the process's buffer (datasets in high-performance computing tend to be very large causing memory limitations to surface frequently). If the number of messages in the Worker's buffer reaches a maximum, the Worker unsubscribes from the Master. After the Worker has computed each of the messages in its buffer, it re-subscribes to the Master, starting once again with the conservative phase of delivery as described above. 3.1.2 Architectural Model and Implementation We have further codified ADaPT in a software architectural style [9]. In addition to Master and Worker components , the ADaPT connector utilizes an adaptive dispatcher to deliver messages to each subscribed Worker using the ADaPT protocol. The dispatcher uses a priority-based round-robin algorithm which utilizes the calculated and attempts to saturate each Worker's computation load without flooding the Worker's buffer. This handler auto-matically switches between the conservative and aggressive phases. The key contribution of this connector is the encap-sulation of underlying protocols, allowing the developer to focus instead on the computations to be performed. Similar to the C2 software architecture [3], messages triggering computation travel downstream from one or more Masters to the ADaPT connector. Messages typed as results originating at Workers travel upstream through the ADaPT connector back to the Masters. We have implemented these architectural rules through extensions to the Prism framework [7]. Prism-MW, a middleware designed to enforce architectural rules at the level of software artifacts, is a light-weight event-based framework consisting of a core set of functionality with handles to extensible components, connectors, and event handlers. Topological rules for each architectural style are also en-forced through overloaded methods for connecting artifacts. 3.2 Performance Analysis In analyzing ADaPT's performance in comparison to load-balancing frameworks as well as synchronous scatters and gathers using MPI, it is first important to define a base metric with which to compare protocols. This metric, the "natural rate" of parallel computation, is the sum of all individual computations to be completed divided by the number of nodes in the parallel computation. In this section we will present comparisons of protocols as measured by percentage overhead (calculated as the wall-clock time for the parallel process to complete minus the natural rate, divided by the natural rate). In order to properly compare ADaPT's ability to reduce message traffic as well as to efficiently map asymmetrical computations to physical resources, we developed a Monte Carlo simulation in which a normalized population of computations was delivered to virtual processors via three different communications policies/architectures and the percentage overhead was calculated for each. All massage-passing costs were uniform across the network for each policy implemented . MPI (collective calls) - Costs of synchronous scatters and gathers using MPI were modeled using equations from [10, 4]. In this policy each worker receives a computation via a scatter and returns via a gather before scattering the next subset until all computations are completed. This process is known as a multi-part scatter [1]. Load-balancing framework - The Monte Carlo simulation of the load-balancing framework uses work-sharing methods. All events are delivered to workers before they begin processing and a barrier is periodically introduced. At 859 this barrier, the workload is redistributed evenly between all processors. In order to idealize load balancing, the cost of this calculation was treated as negligible. ADaPT implementation - Using the routing policies of ADaPT, this implementation assumes that workers are capable of buffering only two events and each worker is homogeneous . We made each of these assumptions in order to conservatively profile ADaPT's performance. In each of four simulations, a normalized population of 1000 computations was generated with a mean computation time of 100 milliseconds and a variance of 10 milliseconds 2 , 50 milliseconds 2 , 100 milliseconds 2 , and 200 milliseconds 2 , respectively. For each simulation, the aggregation of the parallel computation (i.e, the ratio of workers to computations) was varied from 1:500 to 1:5. Results of this comparison are shown in Figure 1. DISCUSSION It can be seen in these plots that while ADaPT performs uniformly at all aggregations smaller than 1:100 (i.e., &gt;=10 workers), MPI collective commands and load-balancing frameworks decrease performance as the aggregation is reduced. For load-balancing frameworks, this is due to the increased volume of messaging required in order to re-balance the load across all processors at each barrier. In the presence of load-balancing , the idle time is significantly reduced, but the cost of rerouting messages to new processors makes ADaPT the better performer especially in high variance computations. From these initial results, we feel that our implementation of ADaPT outperforms collective calls via MPI as well as load-balancing frameworks employing a full worksharing scheme for significantly varied aggregations and computation time variances. In Monte Carlo simulations, ADaPT produced a better mapping of computations to resources, reducing computational overhead to under 1% for aggregations less than 1:100. In the simulations of aggregations greater than 1:100, ADaPT does not perform as well as MPI or other load-balancing frameworks due to the increased percentage of time each worker's buffer remains empty before another event is pushed to the Worker at the calculated rate of computation. This situation seems of little consequence, however, in that data sets are seldom overagreggated to this extreme. ADaPT offers a significant decrease in overhead for event delivery in parallel computations and also outperforms established load-balancing techniques for use with asymmetrical parallel computations. Additionally, ADaPT, through its implementation in Prism, offers developers architectural artifacts at the level of implementation, clear division between the computation loci (in the form of extensible Workers ) and communications algorithms, and reduction of communications knowledge needed by the developer in order to implement asymmetrical parallel computations. 4.2 Future Work While ADaPT is clearly an applicable architectural style to high-performance computing, we make no claim as to its monopoly of the field. In future work, we hope to build a more substantial architectural framework for high-performance computing which provides more underlying protocol choices and further assists developers in code migration to new platforms including SMP and other shared-memory machines. We hope to demonstrate the ease of system design and implementation via architectures to the high-performance community without serious performance degradation, as is cur-rently the prevalent though. Further enhancements to the ADaPT protocol and architecture will include refinement of its topological constraints to encapsulate both data-parallel stages of computation and higher-level workflow stages using multiple layers of masters and workers connected between more advanced ADaPT connectors (themselves perhaps distributed across multiple physical nodes). Also, we hope to further investigate the tradeoffs associated with alternate unsubscription policies and the effects of "pumping" the parallel computation by modifying delivery rates to be faster than average computation rates. This work was supported by the NSF 0312780 grant. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect those of the National Science Foundation. The authors also wish to thank the anonymous reviewers for their helpful comments. REFERENCES [1] D. Woollard et. al. Adapt: Event-passing protocol for reducing delivery costs in scatter-gather parallel processes. In Proceeding of the Workshop on Patterns in High Performance Computing, Urbana, Illinois, May 2005. [2] K. Barker et. al. A load balancing framework for adaptive and asynchronous applications. Parallel and Distributed Systems, IEEE Transactions on, 15:183192, 2004. [3] R. Taylor et. al. A component- and message-based architectural style for gui software. IEEE Transactions on Software Engineering, June, 1996. [4] W. Gropp, E. Lusk, and A. Skjellum. Using MPI: Portable Programming with the Message Passing Interface. MIT Press, 1999. [5] S. Guyer and C. Lin. Broadway: A software architecture for scientific computing. In Proceedings of the IFIP TC2/WG2.5 Working Conference on the Architecture of Scientific Software, pages 175192, Deventer, The Netherlands, The Netherlands, 2001. Kluwer, B.V. [6] L. Kale and S. Krishnan. CHARM++: A Portable Concurrent Object Oriented System Based on C++. In A. Paepcke, editor, Proceedings of OOPSLA'93, pages 91108. ACM Press, September 1993. [7] S. Malek, M. Mikic-Rakic, and N. Medvidovic. A style-aware architectural middleware for resource-constrained, distributed systems. IEEE Transactions on Software Engineering, March, 2005. [8] D. Post and L. Votta. Computational science demands a new paradigm. Physics Today, 58(1):3541, 2005. [9] M. Shaw and D. Garlan. Software Architecture: Perspectives on an Emerging Discipline. Prentice-Hall, 1996. [10] A. Skjellum. High performance mpi: Extending the message passing interface for higher performance and higher predictability, 1998. 860

collective calls;ADaPT;MPI;software engineering;asymamtrical parallel computations;load balancing;communication protocols;high-performance computing;High-Performance Computing;Asymmetrical Parallel Computations

31

An empirical comparison of supervised machine learning techniques in bioinformatics

Research in bioinformatics is driven by the experimental data. Current biological databases are populated by vast amounts of experimental data. Machine learning has been widely applied to bioinformatics and has gained a lot of success in this research area. At present, with various learning algorithms available in the literature, researchers are facing difficulties in choosing the best method that can apply to their data. We performed an empirical study on 7 individual learning systems and 9 different combined methods on 4 different biological data sets, and provide some suggested issues to be considered when answering the following questions: (i) How does one choose which algorithm is best suitable for their data set? (ii) Are combined methods better than a single approach? (iii) How does one compare the effectiveness of a particular algorithm to <A href="31.html#1">the others?

Introduction In the post-genome era, research in bioinformatics has been overwhelmed by the experimental data. The complexity of biological data ranges from simple strings (nucleotides and amino acids sequences) to complex graphs (biochemical networks); from 1D (sequence data) to 3D (protein and RNA structures). Considering the amount and complexity of the data, it is becoming impossible for an expert to compute and compare the entries within the current databases. Thus, machine learning and artificial intelligence techniques have been widely applied in this domain to discover and mine the knowledge in the databases. Quoting from Baldi and Brunak (Baldi and Brunak, 2001) "As a result, the need for computer / statistical / machine learning techniques is today stronger rather than weaker." Shavlik et al. (Shavlik et al., 1995) described the field of molecular biology as tailor-made for machine learning approaches. This is due to the nature of machine learning approaches that performs well in domains where there is a vast amount of data but little theory this is exactly the situation in bioinformatics. Since the introduction of machine learning to this field, various algorithms and methods have been produced and applied to study different data sets. Most of these studies compare a `new' algorithm with the conventional ones, asserting the effectiveness and efficiencies of their methods in particular data sets. The variety of learning algorithms currently available for the researchers are enormous and the main problems faced by researchers are: (i) How does one choose which algorithm is best suitable for their data set? (ii) Are combined methods better than a single approach? (iii) How does one compare the effectiveness of a particular algorithm to the others? Copyright 2003, Australian Computer Society, Inc. This paper appeared at First Asia-Pacific Bioinformatics Conference, Adelaide, Australia. Conferences in Research and Practice in Information Technology, Vol. 19. Yi-Ping Phoebe Chen, Ed. Reproduction for academic, not-for profit purposes permitted provided this text is included. The objective of this study is to provide some suggestions for the community by answering the above questions. This paper is organised as follows. Section 2 presents a brief summary of machine learning. Section 3 outlines the materials and methods used in this study. Section 4 presents the results and discussion, and the final section summarises this work. Machine Learning Background A machine learning algorithm is one that can learn from experience (observed examples) with respect to some class of tasks and a performance measure. (Mitchell, 1997). Machine learning methods are suitable for molecular biology data due to the learning algorithm's ability to construct classifiers/hypotheses that can explain complex relationships in the data. The classifiers or hypotheses can then be interpreted by a domain expert who suggests some wet-lab experiments to validate or refute the hypotheses. This feedback loop between in silico and in vivo / in vitro experiments accelerates the knowledge discovery process over the biological data. This feedback is an important characteristic of machine learning in bioinformatics. Generally, there are two types of learning schemes in machine learning: supervised learning where the output has been given a priori labelled or the learner has some prior knowledge of the data; and unsupervised learning where no prior information is given to the learner regarding the data or the output. The overall tasks for the learner are to classify, characterise, and cluster the input data. Classification is the most common task in biological problem where given two different sets of examples, namely positive E + and negative E examples (E + E = ), the learner needs to construct a classifier to distinguish between the positive examples and the negative set. This classifier can then be used as the basis for classifying as yet unseen data in the future. Usually, for a supervised classification problem, the training examples are in the form of a set of tuples {( where x )} , ( ),..., , 1 1 j n n j y x y x i is the class label and y ij is the set of attributes for the instances. The task of the learning algorithm is to produce a classifier (hypothesis, function) to classify the instances into the correct class. In this study, we only consider supervised machine learning applied to classification. Materials and Methodologies We performed an empirical comparison of rule-based learning systems (Decision trees, One Rule, Decision rules), statistical learning system (Nave Bayes, Instance Based, SVM and neural networks) and ensemble methods (Stacking, Bagging and Boosting) on the data listed in Table 1 based on the accuracy, positive predicted value, specificity and sensitivity of the learning algorithms. All the learning methods used in this study were obtained from the WEKA machine learning package (http://www.cs.waikato.ac.nz/~ml/weka/). 3.2 Data set In this study we used the following data sets obtained from UCI machine learning repository (Blake and Merz, 1998). We briefly describe the biological motivation for the data sets; interested readers should refer to the cited papers for details. E.coli data set The objective of this data set is to predict the cellular localisation sites of E.coli proteins (Horton and Nakai, 1996). There are 8 different cellular sites, which are cytoplasm (cp), inner membrane without signal sequence (im), periplasm (pp), inner membrane with uncleavable signal sequence (imU), outer membrane (om), outer membrane lipoprotein (omL), inner membrane lipoprotein (imL) and inner membrane with cleavable signal sequence (imS). The attributes are signal sequence recognition methods (specifically those of McGeoch and von Heijne), the presence of charge on N-terminus of predicted lipoproteins and 3 different scoring functions on the amino acid contents whether predicted as a outer membrane or inner membrane, cleavable or uncleavable sequence signal. Yeast data set The objective is similar to the E.coli data, which is to determine the cellular localisation of the yeast proteins (Horton and Nakai, 1996). There are 10 different sites, which include: CYT (cytosolic or cytoskeletal); NUC (nuclear); MIT (mitochondrial); ME3 (membrane protein, no N-terminal signal); ME2 (membrane protein, uncleaved signal); ME1 (membrane protein, cleaved signal); EXC (extracellular); VAC (vacuolar); POX (peroxisomal) and ERL (endoplasmic reticulum lumen). The attributes are similar to the E.coli data set with the addition of nuclear localisation information. Promoter data set. The task of the classifier is to predict whether a DNA sequence from E.coli is either a promoter or not (Towell et al., 1990). The input data is a 57-nucleotide sequence (A, C, T or G). HIV data set The data set contains 362 octamer protein sequences each of which needs to be classified as an HIV protease cleavable site or uncleavable site (Cai and Chou, 1998). Data set E.coli Yeast Promoters HIV Continuous Attribute 2 0 57 8 Discrete Attribute 5 8 0 0 Classes 8 10 2 2 Data Size 336 1484 106 362 Table 1: Data sets used in this study. 3.3 Evaluation We constructed a confusion matrix (contingency table) to evaluate the classifier's performance. Table 2 shows a generic contingency table for a binary class problem. True positives (TP) denote the correct classifications of positive examples. True negatives (TN) are the correct classifications of negative examples. False positives (FP) represent the incorrect classifications of negative examples into class positive and False negatives (FN) are the positive examples incorrectly classified into class negative. Predicted Positive Negative Positive TP FN Actual Negative FP TN Table 2: A contingency table for a binary class problem. Based on the contingency table, several measurements can be carried out to evaluate the performance of the induced classifier. The most popular performance evaluation measure used in prediction or classification learning is classifier accuracy which measures the proportion of correctly classified instances; FN FP TN TP TN TP + + + Acc + = . Positive Predictive Accuracy (PPV, or the reliability of positive predictions of the induced classifier) is computed by FP TP TP PPV + = . Sensitivity (S n ) measures the fraction of actual positive examples that are correctly classified FN TP TP + = S n ; while specificity (S p ) measures the fraction of actual negative examples that are correctly classified FP TN TN + S . p = 3.4 Cross-validation To evaluate the robustness of the classifier, the normal methodology is to perform cross validation on the classifier. Ten fold cross validation has been proved to be statistically good enough in evaluating the performance of the classifier (Witten and Frank, 2000). In ten fold cross validation, the training set is equally divided into 10 different subsets. Nine out of ten of the training subsets are used to train the learner and the tenth subset is used as the test set. The procedure is repeated ten times, with a different subset being used as the test set. Results and Discussion We summarise our experimental results in Figure 1 and 2. The full analysis of this study is available in http://www.brc.dcs.gla.ac.uk/~actan/APBC2003. Figure 1. Accuracy vs Positive Predictive Value Figure 2. Specificity vs Sensitivity From the results, we observed that most of the individual learners tend to perform well either in accuracy or specificity. Probably this is due to the induced classifier being able to characterise the negative examples (most of the training sets have large ratio of negative examples compared to positive examples). Furthermore, the results suggest that combination approaches are in general better at minimising overfitting of the training data. We also observed from this experiment that boosting performs better than bagging. This is because attributes which are highly important in discriminating between classes are randomly removed by bagging; however they are preserved in boosting and thus contribute to the final voting scheme. The only individual learning system that perform better than the combined methods is Nave Bayes learning. This may suggest that Nave Bayes is capable of classifying instances based on simple prior probabilistic knowledge. In this study SVM does not perform well compared to other methods, probably due to the fact that training data are not separable in the vector space. 4.1 Rules-of-thumb In this section, we address the following questions by providing some suggested issues (rules-of-thumb) to be considered when answering them. (i) How does one choose which algorithm is best suitable for their data set? Ratio of the training data From these experiments, we observed that the division of the training data plays a crucial role in determining the performance of the algorithms. If the training TPs and TNs are almost equal in size, the algorithms tend to construct much better classifiers. This observation suggested that the classifier induced from equal size of TP and TN tend to be more robust in classifying the instances. Furthermore, the classifiers generated consider all the discriminative attributes that distinguish between two different classes. If the size of the TP set is small compared to that of TN, most probably the classifier will overfit the positive examples and thus perform poorly in the cross validation stages. Attributes Another factor that must be taken into consideration when choosing a learning method is the nature of the attributes. Generally, statistical methods (e.g. SVM, neural networks) tend to perform much better over multi-dimensions and continuous attributes. This is because the learning strategy embedded in these algorithms enables the learners to find a maximal margin that can distinguish different classes in the vector space. By contrast, rule-based systems (e.g. Decision trees, PART) tend to perform better in discrete / categorical attributes. The algorithms of these methods operate in a top-down manner where the first step is to find the most discriminative attribute that classifies different classes. The process is iterated until most of the instances are classified into their class. Credibility vs. Comprehensibility When choosing a machine learning technique, users need to ask themselves what they really want to "discover" from the data. If they are interested in generating understandable hypotheses, then a rule-base learning algorithm should be used instead of statistical ones. Most machine learning algorithms follow Occam's principle when constructing the final hypothesis. According to this principle, the algorithm tends to find the simplest hypotheses by avoiding overfitting the training data. But does this principle still hold in bioinformatics? In bioinformatics we often wish to explore data and explain results, and hence we are interested in applying intelligent systems to provide an insight to understand the relations between complex data. The question then arises as to whether we prefer a simple classifier or a highly comprehensible model. In general, there is a trade off between the credibility and comprehensibility of a model. Domingos (1999) suggested applying domain constraints as an alternative for avoiding overfitting the data. We agree with Muggleton et al. (1998) that when comparing the performance of learning systems in a bioinformatics context, the hypothesis with better explanatory power is preferable when there exist more than one hypotheses with statistical equivalent predictive accuracy. (ii) Are combined methods better than a single approach? From the experiments most of the combined methods perform better than the individual learner. This is because none of the individual methods can claim that they are superior to the others due to statistical, computational and representational reasons (Dietterich, 2000). Every learning algorithm uses a different search strategy. If the training data is too small, the individual learner can induce different hypotheses with similar performances from the search space. Thus, by averaging the different hypotheses, the combined classifier may produce a good approximation to the true hypotheses. The computational reason is to avoid local optima of individual search strategy. By performing different initial searches and combining the outputs, the final classifier may provide a better approximation to the true hypotheses. Lastly, due to the limited amount of training data, the individual classifier may not represent the true hypotheses. Thus, through considering different classifiers, it may be possible to expand the final classifier to an approximate representation of the true hypotheses. Ensemble learning has been an active research topic in machine learning but not in the bioinformatics community. Since most of the hypotheses induced are from incomplete biological data, it is essential to generate a good approximation by combining individual learners. (iii) How does one compare the effectiveness of a particular algorithm to the others? Predictive accuracy Most of the time, we can find in the literature reports that a learning scheme performs better than another in term of one model's accuracy when applied to a particular data set. From this study, we found that accuracy is not the ultimate measurement when comparing the learner's credibility. Accuracy is just the measurement of the total correctly classified instances. This measurement is the overall error rate, but there can be other measures of the accuracy of a classifier rule. If the training data set has 95 TNs and 5 TPs, by classifying all the instances into a negative class, the classifier still can achieve a 95% accuracy. But the sensitivity and the positive predicted value is 0% (both measurements evaluate the performance in classifying TPs). This means that although the accuracy of the classifier is 95% it still cannot discriminate between the positive examples and the negatives. Thus, when comparing the performance of different classifiers, accuracy as a measure is not enough. Different measures should be evaluated depending on what type of question that the user seeks to answer. See Salzberg (Salzberg, 1999) for a tutorial on comparing classifiers. Conclusions Machine learning has increasingly gained attention in bioinformatics research. With the availability of different types of learning methods, it has become common for the researchers to apply the off-shelf systems to classify and mine their databases. In the research reported in this paper, we have performed a comparison of different supervised machine learning techniques in classifying biological data. We have shown that none of the single methods could consistently perform well over all the data sets. The performance of the learning techniques is highly dependant on the nature of the training data. This study also shows that combined methods perform better than the individual ones in terms of their specificity, sensitivity, positive predicted value and accuracy. We have suggested some rules-of-thumb for the reader on choosing the best suitable learning method for their dataset. Acknowledgements We would like to thank colleagues in the Bioinformatics Research Centre for constructive discussions. We would also like to thank the anonymous reviewers for their useful comments. The University of Glasgow funded AC Tan's studentship. References BALDI, P. AND BRUNAK, S. (2001) Bioinformatics: The Machine Learning Approach, 2 nd Ed., MIT Press. Blake, C.L. AND Merz, C.J. (1998) UCI Repository of machine learning databases [http://www.ics.uci.edu/~mlearn/MLRepository.html] CAI, Y.-D. AND CHOU, K.-C. (1998) Artificial neural network model for predicting HIV protease cleavage sites in protein. Advances in Engineering Software, 29: 119-128. DIETTERICH, T.G. (2000) Ensemble methods in machine learning. In Proceedings of the First International Workshop on MCS, LNCS 1857: 1-15. DOMINGOS, P. (1999) The role of Occam's razor in knowledge discovery. Data Mining and Knowledge Discovery, 3: 409-425. HORTON, P. AND NAKAI, K. (1996) A probabilistic classification system for predicting the cellular localization sites of proteins. In Proceedings of Fourth International Conference on ISMB, p.109-115. AAAI / MIT Press. MITCHELL, T. (1997) Machine Learning. McGraw-Hill. MUGGLETON, S., SRINIVASAN, A., KING, R.D. AND STERNBERG, M.J.E. (1998) Biochemical knowledge discovery using inductive logic programming. In H. Motoda (Ed.) Proceedings of the First Conference on Discovery Science, Springer-Verlag. SALZBERG, S. (1999). On comparing classifiers: a critique of current research and methods. Data mining and knowledge discovery, 1: 1-12. SHAVLIK, J., HUNTER, L. & SEARLS, D. (1995). Introduction. Machine Learning, 21: 5-10. TOWELL, G.G., SHAVLIK, J.W. AND NOORDEWIER, M.O. (1990) Refinement of approximate domain theories by knowledge-based neural networks. In Proceedings of the Eighth National Conference on Artificial Intelligence, p. 861-866. AAAI Press. WITTEN, I.H. AND FRANK, E. (2000) Data Mining: Practical machine learning tools and techniques with java implementations. Morgan Kaufmann.

classification;Supervised machine learning;cross validation;performance evaluation;training data;biological data;supervised machine learning;machine learning;ensemble methods;bioinformatics

32

An expressive aspect language for system applications with Arachne

C applications, in particular those using operating system level services, frequently comprise multiple crosscutting concerns : network protocols and security are typical examples of such concerns. While these concerns can partially be addressed during design and implementation of an application, they frequently become an issue at runtime, e.g., to avoid server downtime. A deployed network protocol might not be efficient enough and may thus need to be replaced. Buffer overflows might be discovered that imply critical breaches in the security model of an application. A prefetching strategy may be required to enhance performance. While aspect-oriented programming seems attractive in this context, none of the current aspect systems is expressive and efficient enough to address such concerns. This paper presents a new aspect system to provide a solution to this problem. While efficiency considerations have played an important part in the design of the aspect language, the language allows aspects to be expressed more concisely than previous approaches. In particular, it allows aspect programmers to quantify over sequences of execution points as well as over accesses through variable aliases. We show how the former can be used to modularize the replacement of network protocols and the latter to prevent buffer overflows. We also present an implementation of the language as an extension of Arachne, a dynamic weaver for C applications. Finally, we present performance evaluations supporting that Arachne is fast enough to extend high performance applications , such as the Squid web cache.

INTRODUCTION Real-world applications typically comprise multiple crosscutting concerns. This applies, in particular, to C applications using operating system level services. We have exam-ined three concerns which are typical for this domain in the context of a large application, the open source web cache Squid [36]. More concretely, we have considered translation of network protocols (which may be necessary for efficiency reasons), insertion of checks for buffer overflows (which are at the heart of many of today's security issues), and introduction of prefetching strategies within the cache (which can be used to enhance efficiency of the web cache). We have found that all these concerns are scattered over large portions of the code of Squid. Hence, the three concerns are crosscutting in the sense of Aspect-Oriented Programming (AOP) [24] and aspects should therefore be a means of choice for their modularization . The concerns have three important characteristics. First, they must frequently be applied at runtime, e.g., in order to rapidly fix a buffer overflow and thus prevent security breaches without incurring server downtime. A dynamic aspect weaver is therefore needed. Second, they expose intricate relationships between execution points, e.g., network protocols are most concisely expressed in terms of sequences of execution points, not individual ones. The aspect system must therefore support expressive means for the definition of aspects, in particular pointcuts. Third, efficiency is crucial in the application domain we consider. To our knowledge, none of the current aspect systems for C meet these three requirements and is suitable for the modularization of such concerns. Moreover, requirements for dynamic weaving and efficiency often trade off with expressivity . Squid should be as efficient as possible and therefore exploit any suitable operating system and hardware particularity . Its code base is therefore difficult to understand and manipulate, thus hindering in particular modularization efforts . It is therefore highly questionable that the considered modularization problems can be solved without aspects. In this paper we propose a solution to the aspectization of such concerns of C applications. More concretely, we provide three main contributions. First, we provide a new expressive aspect language featuring a construct for quantification over sequences of execution points as well as over accesses to local aliases of global variables. We show how this aspect lan-27 guage permits concise expression of the considered concerns as aspects. Second, we present how the aspect language can be implemented efficiently through runtime weaving into binary code. Concretely, this is done by integrating the aspect language into our tool Arachne, a dynamic weaver for C applications . Furthermore, we present how Arachne improves on our previous work Dyner [32]. Finally, we give evidence that our approach meets strong efficiency requirements by showing performance evaluations in the context of Squid. The paper is structured as follows. Section 2 presents the motivating concerns we identified within Squid. Section 3 shows how to modularize these concerns as aspects and defines our aspect language. Section 4 describes its implementation within Arachne. Section 5 assesses the performance of our implementation. Section 6 describes related work. Section 7 concludes and suggests futures work. MOTIVATIONS Legacy C applications involve multiple crosscutting concerns . Many of them remain challenging, both in terms of expressiveness required to handle them properly in an aspect-oriented language and in terms of constraints posed on the weaver. This section describes three such concerns in C applications: switching the network protocol, buffer overflows and prefetching. The network protocol concern is typically scattered through the entire application. It is an issue when administrators discover at runtime that the retained protocol is not efficient enough. Likewise the security threats posed by buffer overflows is a real concrete problem for administrators. While guarding all buffers against overflows might decrease performance considerably, administrators are left with no other option than accepting the trade-off between security and performance chosen at application's design time. Prefetching is another well-known crosscutting concern [12]. Since prefetching aims at increasing performance , prefetching aspects make only sense with an efficient weaver. Yet, it is still difficult to modularize these three concerns in today's aspect-oriented language. In this section, we first describe the context in which the concerns arise before showing their crosscutting nature and finally explaining the lack in current aspect-oriented languages to handle them properly. 2.1 TCP to UDP protocol HTTP was essentially designed as a file transfer protocol running on top of TCP, a connection-oriented protocol ensuring communication reliability. While the average Web page size does not exceed 8 KB [4], the cost of retrieving a Web page is often dominated by data exchanged for control purposes of TCP rather than by the page content itself. This is not a new problem, many researches have already pointed out that TCP is not suitable for short-lived connections . While HTTP 1.1 has introduced persistent connections allowing a client to retrieve multiple pages from the same server through the same TCP connection, the number of simultaneous TCP connections is limited by operating systems. Servers have a strong incentive to close HTTP connections as soon as possible. Hence, despite the persistent connection mechanism, many studies conclude that TCP should be replaced by UDP to retrieve short pages [10, 29, 7]. In spite of its performance improvements, the number of legacy Web applications has prevented a wide adoption of this solution. Typical legacy Web applications have to be listen accept read write close write read close connect socket Server Client TCP Protocol socket bind close close socket Server Client UDP Protocol recvfrom sendto recvfrom socket bind Network Network sendto Time Figure 1: Typical usage of the TCP and UDP APIs. stopped to switch the protocol. The traditional approach to avoid depriving a subnetwork from Internet connectivity while stopping the cache is to swap the application between different machines. This approach is not only expensive in terms of hardware, it complicates the administrative task of the Web cache administrator and poses the problem of con-sistently transferring the runtime state of the application before restarting it. Stopping an e-commerce Web server means a loss of money and many small companies can not afford the cost of redundant servers. For a wide acceptance, a HTTP dialect using UDP as transport protocol should thus be deployable on demand at runtime. In addition, replacing TCP by UDP in an application is relatively difficult. The choice of a transport protocol is usually based on standards believed to be ever-lasting and made at an early design stage. Hence no particular effort is made to localize this design decision in a single piece of code. For example, despite a modularization effort, the TCP API provided by the operating system is used directly in 7 of the 104 ".c" source files of the Squid Web cache. As shown in Fig. 1, the TCP API is built around a set of C functions to be invoked sequentially by the application. In a properly written program, TCP functions are first used to establish the connection (typically with socket, connect, bind and listen), exchange data through the connection (typically with read and write) and then close it (typically close). UDP uses similar but less functions. UDP applications first direct the operating system to dedicate the appropriate resources to exchange data (typically with socket and bind), then exchange data through these resources (typically with sendto and recvfrom) before releasing them (typically with close). Hence, the problem is not only difficult because TCP-related function invocations are scattered but because the relative order of each invocation is important in order to map it onto the appropriate UDP function. This example is typical of protocol based APIs. When such an API is used in an undisciplined way, it becomes quickly impossible to replace it by another one. Today, aspect-oriented systems lack an appropriate sequencing construct in their language. Moreover, many do not provide the ability to weave aspects dynamically. 2.2 Buffer overflows In C, the size of an array is fixed at allocation time. According to ISO and ANSI standards [2], an invalid array access does not result in an immediate error but leads to an implementation-dependent behavior. Such behavior is increasingly exploited by hackers to circumvent security re-28 strictions [37]. It is therefore crucial for C programmers to ensure every access to an array to be valid. On the other hand, bound checking code is error prone: it is easy to forget to check an access and even when the access is checked, it is easy to compare the index locating the access with an inappropriate bound. Therefore, researchers have proposed to make compilers responsible for enforcing proper array access [22, 31]. The problem is that even the most efficient system (CRED [31]) slows down an application up to 130%. Moreover, most frequently used compilers like gcc do not support bound checking. Today, administrators discovering a buffer overflow in production software are left with no other option than stopping the application and restarting a bug free version. This was the solution chosen when a buffer overflow was discovered in Squid in [6]. While widely used, this solution suffers from three major drawbacks. First, it does not enforce continuous servicing since the service delivered by the application is not available during the update. Second, this solution entails an important information loss: an administrator has no means to learn whether the buffer overflow has been exploited by a hacker or not. Third, it misunderstands the performance trade-off, i.e. it is not necessary to check every array access, it is only necessary to perform enough checking to discourage hackers. Therefore, bound checking code should only run when an environment becomes hostile [23]. Bound checking code tends to crosscut the entire application . For example, properly written C functions accepting an array argument commonly take a second argument holding the array size: the first one allows the function to access the array while the second is used to ensure correctness of accesses. In Squid, bound checking code can be found in any of the 104 ".c" files of its source code. On the 57635 lines composing these ".c" files, at least 485 check bounds. This problem fails to be handled properly in current aspect languages as they lack the ability to trigger advices upon access made through the alias of a variable. Again, many aspect-oriented systems offer only static weaving capabilities preventing the administrator to choose the trade-off security/performance suiting his needs. 2.3 From fetching to prefetching Operations like retrieving a file on a local disk or over the Web can be sped up if the underlying software anticipates user requests and start to fetch documents beforehand. Such prefetching schemes distinguish themselves from each other in the way they predict future user requests. These "ora-cles" actually prevent a clean encapsulation of prefetching in a single module communicating with the rest of the application through well-defined interfaces since predictions are based on information meant to be private to other modules. In addition, it is very likely that there is no universal perfect oracle [19]. A statically linked prefetching module is therefore inappropriate, but prefetching modules along with the necessary oracles should be loaded and unloaded on the fly. Due to their crosscutting nature, prefetching modules including such oracles are better written with aspects [32]. Coady et al. have already pointed out the crosscutting nature of prefetching in the FreeBSD OS [12]. In our previous work considering the Squid Web cache, we reached a similar conclusion [32]. We have previously shown that this concern can be addressed with cflow-like constructs. Despite potential performance improvements, prefetching also increases resource consumption (e.g. network prefetching consumes local storage and bandwidth). When the pressure on resources is too high, prefetching computation competes for them against regular user requests, and slows down their treatment instead of speeding it up. In such cases, prefetching should therefore be, temporarily, disabled. Squid essentially manages file descriptors, a resource only available in a limited quantity. A file descriptor is used between the underlying operating system and applications to describe a network connection or a file on the disk. Squid's file descriptor management is based on a global variable that tracks the number of file descriptors currently in use. By comparing its value with the maximum number of file descriptors allowed by the operating system, it is possible to estimate that prefetching should be disabled or resumed. For this problem of file descriptor consumption, the current practice of checking if prefetching should be disabled or not within the advice, is a bad practice that impedes both readability and maintainability. A mechanism is needed within the aspect language to restraint the advice execution at times where the pressure on resources is too high. This problem were not addressed in our previous work. AN EXPRESSIVE ASPECT LANGUAGE FOR SYSTEM PROGRAMMING IN C While AOP seems to be the obvious choice to tackle the crosscutting concerns introduced above, none of the existing AO systems provides explicit support for some of their essential elements, in particular, join point sequences for protocols , and references to aliases which are local to a function. In this section we introduce a new aspect language for system programming in C that allows such crosscutting concerns to be expressed concisely. In order to make this point, we first revisit the examples by concisely aspectizing them using our language. (Note that our aspect language is expressive in the sense of enabling the concise definition of certain types of aspects, especially compared to other tools for system-level manipulations, but not necessarily more expressive than existing approaches in a language-theoretic sense.) We then define the join point model underlying our language precisely, followed by the definition of its syntax and informal semantics. Finally, we illustrate how its semantics can be formally defined in terms of a small-step operational semantics using the framework introduced in [14]. 3.1 Example crosscutting concerns revisited We now revisit the concerns discussed in section 2 in order to show our language in action and give evidence that it allows such concerns to be concisely modularized. The aspect shown in Fig. 2 translates transport protocols from TCP to UDP. A protocol defines a sequence of function calls, so the top-level operator of this aspect is seq. The sequence aspect syntactically consists of a list of pairs of pointcut and advice (separated by then). In the example , the TCP protocol starts with a call to socket() with three constant arguments: AF INET, SOCK STREAM and 0. When such a call is matched, the second parameter is replaced by SOCK DGRAM as required by the UDP protocol . The result of this transformed call, the file descriptor, is bound to fd by return(fd). Then the next call to connect () with the same file descriptor fd as its first parameter is matched. In this case the values of the other parameters 29 seq( call(int socket(int, int, int)) && args(AF INET, SOCK STREAM, 0) && return(fd) then socket(AF INET, SOCK DGRAM, 0); call(int connect(int, struct socketaddr, socklen t)) && args(fd, address, length) then returnZero(); // where int returnZero() { return 0; } ( call(size t read(int, void, size t)) && args(fd, readBuffer, readLength) then recvfrom(fd, readBuffer, readLength, 0, address, length); || call(size t write(int, void, size t)) && args(fd, writeBuffer, writeLength) then sendto(fd, writeBuffer, writeLength, 0, address, length); ) call(int close(int)) && args(fd) ; ) Figure 2: An Aspect for Switching Transport Protocols, from TCP to UDP seq( call(void malloc(size t)) && args(allocatedSize) && return(buffer) ; write(buf f er) && size(writtenSize) && if(writtenSize &gt; allocatedSize) then reportOverflow (); call(void free(void)) ) Figure 3: An Aspect for Detecting Buffer Overflow are bound to arguments address and length, and the original call is replaced by returnZero(). Indeed, there is no connect step in the UDP protocol. After that, calls to read() and write() (using the `or' on aspects: ||) on the same file descriptor fd are translated to UDP recvfrom() and sendto(), respectively. Note that sequences of such access are potentially translated (due to use of the repetition operator ). Finally, a call to close() on fd terminates the TCP protocol as well as the UDP protocol and thus is not modified (i.e., there is no then clause). This last step is required to free the variables used in the sequence (here, fd, address and length). Indeed, this aspect can use numerous (instances of these) variables when it deals with interleaved sequences, as each call to socket() creates a new instance of the sequence. The aspect shown in Fig. 3 detects buffer overflows. The corresponding sequence starts when the function malloc() returns the buffer address which is then bound to buffer. Then, each time this address is accessed (through a global variable or a local alias) the size of the data to be written is compared with the size of the initially allocated memory. If the former exceeds the latter, an overflow is indicated. The sequence ends when the memory is deallocated using free(). The aspect in Fig. 4 introduces prefetching in a web cache. The first controlflow phrase initializes prefetching when an HTTP response is built (clientBuildReply()) within the control flow of a client request (clientSendMoreData()). The until clause stops prefetching when the number of connection becomes too large, a situation where prefetching would effectively degrade performance. The second controlflow phrase analyzes hyperlinks in a page being transmitted (i.e., when comm write mbuf() is called within the control flow of clientSendMoreData()). Finally, the last call phrase pre-fetches hyperlinks analyzed by the second aspect. It does so by replacing the method call to clientWriteComplete() with retrieveHyperlinks(). Finally, note that the two require clauses at the top of the aspect declare the types of the global variables of the base program used in the aspects. 3.2 Join points A join point model defines the points in the execution of the base program to which pointcuts may refer. In our JP ::= callJP(val funId(val )) | readGlobalJP(varId, val) | readJP(@, val) | writeGlobalJP(varId, val, size) | writeJP(@, val, size) | controlflowJP(---f unId, cfEnd) | controlflowstarJP(---f unId, cfEnd) cf End ::= callJP(val funId(val )) | readGlobalJP(varId, val) | writeGlobalJP(varId, val, size) val ::= 0 | 1 | 2 | ... // int | @0 | @1 | @2 | ... // int* | ... // values of other C types Figure 5: Join point model case, join points are defined by JP in the grammar shown in Fig. 5. A join point is either: A call of a function callJP(v 1 funId( v 2 )) with function name funId, return value v 1 and a vector of arguments v 2 . A read access which comes in two variants: readGlobalJP(varId, v) denotes reading a global variable with name varId holding the value v ; readJP(@, v) denotes reading a global variable or a local alias with address @ holding the value v . Write access which also comes in two variants: writeGlobalJP(varId, v, size) denotes assignment to a global variable with name varId of the value v of size size. writeJP(@, v, size) denotes assignment to a global variable or a local alias with address @ of the value v of size size. A cflow expression controlflowJP(---f unId, c), where ---f unId = [funId 1 , .., funId n ] is a stack of function names, and c (either a function call or an access to a global variable) occurs within the body of function f unId n . Such a join point requires a call to f unId i+1 within the body of f unId i . A cflow expression controlflowstarJP(---f unId, c), where ---f unId = [funId 1 , .., funId n ] is a partial stack of function names, and c (either a function call or an access to a global variable) occurs within the control flow of function f unId n . Such a join point requires a call to f unId i+1 within the control flow of (i.e., not necessarily in the body of) f unId i . Two features of this join point model may be surprising at first sight: distinction of accesses to aliases from those to global variables and explicit representation of control flow 30 require N umber Of F d as int; require Squid M axF d as int; controlflow(call(void clientSendMoreData(void, char, size t)), call(HttpReply clientBuildReply(clientHttpRequest, char, size t)) && args( request, buf f er, buf f erSize )) then startPrefetching(request, buffer, bufferSize); && until(writeGlobal(int N umber Of F d) && if((N umber Of F d) 100/(Squid M axF d) 75) ; ) controlflow( call(void clientSendMoreData(void, char, size t)), call(void comm write mbuf(int, MemBuf, void, void)) && args(fd, mb, handler, handlerData) && if(! isP ref etch(handler)) ) then parseHyperlinks(fd, mb, handler, handlerData); call(void clientWriteComplete(int, char, size t, int, void)) && args(fd, buf, size, error, data) && if(! isP ref etch(handler)) then retrieveHyperlinks(fd, buf, size, error, data); Figure 4: An Aspect for Prefetching expressions. Both are motivated by our quest for efficiency and are grounded in strong implementation constraints in the context of dynamic weaving of binary C code: an access to a local alias is several magnitudes slower than that to a global variable and matching of control flow join points can be done using an atomic test on the implementation level. 3.3 Pointcuts We now present a pointcut language (see Fig. 6) that provides constructs to match individual join points. Primitive pointcuts are defined by PPrim and comprise three basic pointcuts matching calls, global variable accesses, and control flow join points. Primitive pointcuts can also be combined using a logical "or" noted ||. A call pointcut PCall selects all function call join points callJP(val funId(val )), i.e., all calls to a function matching the signature type funId(-type ), where the arguments of the function can be bound to pointcut variables using argument binder args( ----pattern ) and the return value can be bound to a pointcut variable using a return clause return( pattern ). The two constructs args( ----pattern ) and return( pattern ) can also provide pattern matching by using values (or already bound pointcut variables) in pattern. Pointcuts can also depend on a boolean condition using the if-constructor. A global access pointcut PAccGlobal selects either all read join points readGlobalJP(varId, val) or all write join points writeGlobalJP(varId, val, size) on the global base program variable varId. In these cases, the read or written value can be bound to a variable using value(pattern); in addition, the size of the written value can be bound with size(varName). Pattern matching can also be used for variable access. A control flow pointcut PCf of the form controlflow( PCallSig 1 , ..., PCallSig n , PCfEnd) matches all join points of the form controlflowJP(funId 1 , ..., funId n , cfEnd), where the function identifier in P CallSig i is f unId i . Similarly, a control flow pointcut may match a global variable access for a given stack configuration. The pointcuts of the form controlflowstar(. . . ) select calls or global variable accesses in a stack context allowing for calls that are not directly nested within one another. Finally, P Acc, an access pointcut for a global variable or all of its local aliases, matches all join points of the form readJP or writeJP. Asp ::= AspP rim [ && until( AspP rim ) ] | AspSeq [ && until( AspP rim ) ] AspP rim ::= P P rim Advice AspSeq ::= seq( AspP rim AspSeqElts AspSeqElt ) AspSeqElts ::= [AspSeqElts] AspSeqElt [ ] AspSeqElt ::= AspP rim | P Acc Advice | (AspSeqElt || AspSeqElt) Advice ::= [ then f unId(----pattern ) ] ; Figure 7: Aspect language 3.4 Aspect Language The aspect language we propose is defined in Fig. 7. Aspects Asp are either primitive AspP rim, or sequences of primitive aspects AspSeq. A primitive aspect AspPrim combines a primitive pointcut with an advice that will be applied to all join points selected by the pointcut. If the primitive pointcut has the form p 1 || p 2 , then all variables used in the advice have to be bound in both, p 1 and p 2 . An advice (Advice) is a C function call that replaces a join point in the base program execution (similarly to around in AspectJ). It must have the same return type as the join point it replaces: the type of the global variable in case of a read access, void for a write access and the return type of the function for a call. When the advice is empty (no then clause), the original join point is executed. The original join point can be skipped by calling an empty C function. A sequence aspect is composed of a sequence of primitive aspects. A sequence starts when the first primitive aspect matches. Then the second primitive aspect becomes active instead of the first one. When it matches, the third aspect becomes active instead of the second one. And so on, until the last primitive aspect in the sequence. All but the first and last primitive aspects can be repeated zero or multiple times by using : in this case, the primitive aspect is ac-31 P P rim ::= P Call | P AccGlobal | P Cf | P P rim || P P rim P Call ::= P CallSig [ && args( ----pattern ) ] [ && return( pattern ) ] [ && P If ] P CallSig ::= call( type f unId(-type ) ) P If ::= if( expr ) [ && P If ] P AccGlobal ::= readGlobal( type varId ) [ && value( pattern ) ] [ && P If ] | writeGlobal( type varId ) [ && value( pattern ) ] [ && size( pattern ) ] [ && P If ] P Cf ::= controlflow( P CallSigList, P Cf End ) | controlflowstar( P CallSigList, P Cf End ) P CallSigList ::= P CallSig [ , P CallSigList ] P Cf End ::= P Call | P AccGlobal P Acc ::= read( var ) [ && value( pattern ) ] [ && P If ] | write( var ) [ && value( pattern ) ] [ && size( pattern ) ] [ && P If ] pattern ::= var | val Figure 6: Pointcut language A ::= A | A || A ; parallelism A ::= a.A ; recursive definition (a Rec) | C I; A ; prefixing | C I; a ; end of sequence (a Rec) | C I; STOP ; halting aspect | A P A ; choice Figure 8: Tiny aspect language tive as long as the following one in the sequence does not match. Branching, i.e., a logical `or' between two primitive aspects, can be introduced in a sequence by the operator ||. An element of the sequence can also match a global variable of the base program and accesses to its local aliases, as soon as its address is known (i.e., a previous primitive pointcut has already bound its address to a pointcut variable). Hence, an aspect matching accesses cannot start a sequence. Every join point matching the first primitive pointcut of a sequence starts a new instance of the sequence. The different instances are matched in parallel. A primitive or a sequence aspect a can be used in combination with an expression until(a 1 ), to restrict its scope. In this case, once a join point has been matched by a, the execution of a proceeds as previously described until a 1 matches. To conclude the presentation of our language, note that it does not include some features, such as named pointcuts as arguments to controlflows and conjunctive terms, which are not necessary for the examples we considered but which could easily be added. (As an aside, note that such extensions of the pointcut language may affect the computability of advanced algorithmic problems, such as whether a pointcut matches some part of any base program [25].) 3.5 Towards a formal semantics for expressive aspects In the previous sections, we have given an informal semantics of our aspect language. We now illustrate how the aspect language could be formally defined by translating one of the example aspects into formal aspect language by extension of that used in the formal framework of [14]. The original formal language must be extended in order to deal with halting aspects, an unbounded number of sequential aspects and arbitrary join point predicates. The grammar of the extension, our tiny aspect language, is defined in Figure 8. In this language, aspect expressions A consists of parallel combinations of aspects, C is a join point predicate (similar to our pointcut language) expressed as a conjunction of a term pattern and possibly an expression from the constraint logic programming language CLP(R) [20]. An aspect A is either: A recursive definition. A sequence formed using the prefix operation C I ; X, where X is an aspect or a recursion variable and I a piece of code (i.e., an advice). A choice construction A 1 P A 2 which chooses the first aspect that matches a join point (the other is thrown away). If both match the same join point, A 1 is chosen. A parallel composition of two aspects A 1 || A 2 that cannot occur in choice construction. A halting aspect STOP. The semantics of the protocol translation aspect (from TCP to UDP) is given in Fig. 9. A sequence can have several instances. This is translated into the language A by the expression a 1 || ... which starts a new sequence a 1 once the first join point has been matched and continue to match the rest of the sequence in progress. The repetition operator is translated into recursion on variable the a 2 . The branching operator || is translated into the choice operator 32 a 1 . callJP(fd socket(AF INET, SOCK STREAM, 0)) socket(AF INET, SOCK DGRAM, 0); a 1 || ( callJP(a connect(fd, address, length)) returnZero(); a 2 . callJP(b close(fd)) skip; STOP P callJP(c read(fd, readBuffer, readLength)) recvfrom(fd, readBuffer, readLength, 0, address, length); a 2 P callJP(d write(fd, writeBuffer, writeLength)) recvfrom(fd, writeBuffer, writeLength, 0, address, length); a 2 Figure 9: Definition of the protocol translation using the tiny aspect language P. Finally, the last primitive aspect of the sequence occurs as the first aspect of a choice to get priority over the join points read and write because of the . Note that we use pattern matching in A and that an overbar marks the first occurrence of a variable (i.e., its definition not a use). Note that formal definitions such as that of the protocol translation aspect precisely define several important issues, in particular, when new instances of the sequence aspect are created, and disambiguate of potentially non-deterministic situations, e.g., when two pointcuts of consecutive primitive aspects in the sequence match at the same time. DYNAMIC WEAVING WITH ARACHNE Arachne is built around two tools, an aspect compiler and a runtime weaver. The aspect compiler translates the aspect source code into a compiled library that, at weaving time, directs the weaver to place the hooks in the base program. The hooking mechanisms used in Arachne are based on improved techniques originally developed for Dyner [32]. These techniques allow to rewrite the binary code of executable files on the fly i.e.without pausing the base program, as long as these files conform to the mapping defined by the Unix standard [35] between the C language and x86 assembly language . Arachne's implementation is structured as an open framework that allows to experiment with new kinds of join points and pointcut constructs. Another important difference between Arachne and Dyner is, that Dyner requires a compile time preparation of the base program, whereas Arachne does not. Hence Arachne is totally transparent for the base program while Dyner is not. 4.1 The Arachne Open Architecture The Arachne open architecture is structured around three main entities: the aspect compiler, the instrumentation kernel , and the different rewriting strategies. The aspect compiler translates the aspect source code into C before compiling it. Weaving is accomplished through a command line tool weave that acts as a front end for the instrumentation kernel. weave relays weaving requests to the instrumentation kernel loaded in the address space of the program through Unix sockets. Upon reception of a weaving request, the instrumentation kernel selects the appropriate rewriting strategies referred by the aspects to be woven and instruments the base program accordingly. The rewriting strategy consults the pointcut analysis performed by the aspect compiler to locate the places where the binary code of the base program needs to be rewritten. It finally modifies the binary code to actually tie the aspects to the base program. With this approach, the Arachne core is independent of a particular aspect, of the aspect language, of the particular processor architecture, and of a particular base program. In fact, all dependencies to aspect language implementation are limited to the aspect compiler. All dependencies to the operating system are localized in the instrumentation kernel and finally all dependencies to the underlying hardware architecture are modularized in the rewriting strategies. 4.1.1 The Arachne aspect compilation process The aspect compilation scheme is relatively straightforward : it transforms advices into regular C functions. Pointcuts are rewritten as C code driving hook insertions into the base program at weaving time. There are however cases where the sole introduction of hooks is insufficient to determine whether an advice should be executed. In this case, the aspect compiler generates functions that complement the hooks with dynamic tests on the state of the base program . These dynamic tests are called residues in AspectJ and the rewritten instructions within the base program the shadow [16]. Once the aspects have been translated into C, the Arachne compiler uses a legacy C compiler to generate a dynamically linked library (DLL) for the compiled aspects. 4.1.2 The Arachne weaving process From a user viewpoint, the Arachne weave and deweave command line programs the same syntax than Dyner's version . They both take two arguments. The first identifies the process to weave aspects in or deweave aspects from, and the second indicates the aspect DLL. However, Arachne can target potentially any C application running on the machine while Dyner was limited to applications compiled with it running on the machine. When Arachne's weave receives a request to weave an aspect in a process that does not contain the Arachne instrumentation kernel, it loads the kernel in the process address space using standard techniques [11]. The instrumentation kernel is transparent for the base program as the latter cannot access the resources (memory and sockets essentially) used by the former. Once injected , the kernel creates a thread with the Linux system call: clone. This thread handles the different weaving requests . Compared to the POSIX pthread create function, the usage of clone allows the instrumentation thread to prevent the base program to access its sockets. The instrumentation kernel allocates memory by using side effect free allocation routines (through the Linux mmap API). Because the allocation routines are side effect free, Arachne's memory is totally invisible to the base program. It is up to the aspect to use Arachne's memory allocation routines or base program specific allocation functions. This transparency turns out to be crucial in our experiments. Legacy applications such as Squid use dedicated resource management routines and expect any piece of code they run to use these routines. Failures will result in an application crash. After loading an aspect, the instrumentation kernel rewrites the binary code of the base program. These rewriting strategies are not included in the kernel and must be fetched on demand by each loaded aspect. 4.2 Rewriting strategies Rewriting strategies are responsible for transforming the binary code of the base program to effectively tie aspects to 33 shadow: rewriting site replaced by a x86 instruction x86 instruction x86 instruction x86 instruction execution flow generated at aspect compile time Aspect DLL Hooks generated at weaving time jump Binary code of the compiled base program and/or advices Residue (dynamic tests) Entry hook save registers Return hook Restore registers instruction(s) Relocated tailored updating registers Legacy base program Figure 10: Generic hook operations. the base program at weaving time. These strategies localize Arachne's main dependencies to the underlying hardware architecture. In general, rewriting strategies need to collect information about the base program. These information typically consist of the addresses of the different shadows, their size, the symbol (i.e.function or global variable name) they manipulate, their length etc. In order to keep compiled aspects independent from the base program, this information is gathered on demand at runtime. The mapping between a symbol name in the base program source code and its address in memory is inferred from linking information contained in the base program executable. However because these information can be costly to retrieve, Arachne collects and stores it into meta-information DLLs. these DLLs behave as a kind of cache and lessen the problem of collecting the information required to instrument the base program. To implement our aspect language, Arachne provides a set of eight rewriting strategies that might eventually use each other. 4.2.1 Strategies for call , readGlobal and writeGlobal In Arachne, call, readGlobal and writeGlobal allow an advice to be triggered upon a function call, a read on a global variable or a write respectively. While the implementation of readGlobal and writeGlobal in Arachne is close to the one in Dyner, Arachne implements the strategy for call by rewriting function invocations found in the base program. Dyner instead rewrites the function body of the callee. On the Intel architecture, function calls benefit from the direct mapping to the x86 call assembly instruction that is used by almost, if not all, compilers. Write and read accesses to global variables are translated into instructions using immediate, hard coded addresses within the binary code of the base program. By comparing these addresses with linking information contained in the base program executable , Arachne can determine where the global variable is being accessed. Therefore those primitive pointcuts do not involve any dynamic tests. The sole rewriting of the binary base program code is enough to trigger advice and residue 1 executions at all appropriate points. The size of the x86 call instruction and the size of an x86 jump (jmp) instruction are the same. Since the instruction performing an access to a global variable involves a hard coded address, x86 instructions that read or write a global 1 Residues (i.e. dynamic tests on the base program state) are required when these primitive pointcuts are combined with conditional pointcuts or when pattern matching is involved. variable have at least the size of a x86 jmp instruction. Hence at weaving time, Arachne rewrites them as a jmp instruction to a hook. Hooks are generated on the fly on freshly allocated memory. As shown in figure 10, hooks contain a few assembly instructions that save and restore the appropriate registers before and after an advice (or shadow) execution. A generic approach is to have hooks save the whole set of registers, then execute the appropriate residue and/or advice code before restoring the whole set of registers; finally the instructions found at the join point shadow are executed to perform the appropriate side effects on the processor registers . This is accomplished by relocating the instructions found at the join point shadow. Relocating the instructions makes the rewriting strategies handling read and write access to global variable independent from the instruction generated by the compiler to perform the access 2 . The limited number of x86 instructions used to invoke a function allows Arachne's rewriting strategy to exploit more efficient, relo-cation free, hooks. 4.2.2 Strategies for controlflow and controlflowstar Every time a C function is called, the Linux runtime creates an activation record on the call stack [35]. Like Dyner, Arachne's implementation of the rewriting strategy for controlflow uses the most deeply nested function call (or global read or write access) in the control flow pointcut as shadow. This shadow triggers a residue. This residue uses the activation record's chaining to check whether the remaining function calls of the control flow, are on the call stack maintained by the Linux runtime. An appropriate usage of hashtables that store the linking information contained in the base program executables can thereby decrease the cost of determining if a specific function is the caller of another to a pointer comparison. Therefore, the residue for a controlflow with n directly nested functions implies exactly n pointer comparisons. However, the residue worst case runtime for the indirect control flow operator controlflowstar that allows for not directly nested functions , is proportional to the base program stack depth. 4.2.3 Strategies for read and write read and write are new join points not included in Dyner that have been added to the latest version of Arachne. Their implementation relays on a page memory protection as allowed by the Linux operating system interface (i.e. mprotect) and the Intel processor specifications [18]. A read or write pointcut triggers a residue to relocate the bound variable into a memory page that the base program is not allowed to access and adds a dedicated signal handler. Any attempt made by the base program to access the bound variable identified will then trigger the execution of the previously added signal handler. This handler will then inspect the binary instruction trying to access the protected page to determine whether it was a read or a write access before eventually executing the appropriate advice. 4.2.4 Strategies for seq Like read and write, seq is a new language feature of Arachne. Dyner offers no equivalent construct. Arachne's rewriting strategy of this operator associates a linked list to 2 About 250 x86 instruction mnemonics can directly manipulate a global variable. This corresponds to more than one thousand opcodes. 34 every stage inside the sequence except the last one. Each stage in a sequence triggers a residue that updates these linked lists to reflect state transitions of currently matching execution flows. Upon matching of the first pointcut of the first primitive aspect in the seq, a node is allocated and added to the associated linked list. This node contains a structure holding variables shared among the different pointcuts within the sequence. Once a join point matches a pointcut of an primitive aspect denoting a stage in the sequence, Arachne consults every node in the linked list associated with the previous stage and executes the corresponding advice 3 . Arachne eventually updates the node and in the absence of a moves it to the list associated with the currently matched pointcut.If the matching pointcut corresponds to the end of the sequence, structures are not moved into another list but freed. Our aspect compiler includes an optimization where structures are allocated from a resizable pool and upon a sequence termination, structures are not freed but returned to the pool. 4.3 Arachne limitations Aggressive optimizations of the base program might prevent Arachne to seamlessly weave aspects. Two optimizations are not yet supported by Arachne. First if the compiler inlines a function in another one within the binary code of the base program, the Arachne weaver will fail to properly handle pointcuts referring to that function. Second, control flow pointcuts are based on the chaining of activation records. On the x86 architecture, in leaf functions, optimizing compilers sometimes do not maintain this chaining to free one register for the rest of the computation. This however has not been a problem during our experiments as we used the open source C compiler gcc. Arachne supports two of the three optimization levels proposed by gcc. Stripping that removes linking information and aggressive optimizations that break the interoperability between compilers and/or debuggers are incompatible with Arachne. In practice, Arachne can be used on applications compiled like squid with two of the three gcc optimization level. PERFORMANCE EVALUATION Aspect-oriented solutions will be used if the aspect sys-tem's language is expressive enough and if the aspect system overhead is low enough, for the task at hand. The purpose of this section is to study Arachne's performance. We first present the speed of each Arachne language construct and compare it to similar C language constructs. We then study the overhead of extending Squid with a prefetching policy. This case study shows that even if the cost of some Arachne aspect language constructs might be high compared to C language constructs, this overhead is largely amortized in real applications. 5.1 Evaluation of the language constructs This performance evaluation focuses on studying the cost of each construct of our aspect language. To estimate the cost for each construct of our aspect language, we wrote an aspect using this construct that behaves as an interpreter of 3 In case the previous stage pointcut was used with a star , Arachne examines nodes from linked list associated with the last two previous stages, and so on, until a not starred primitive aspect in the sequence is reached. Execution times (cycles) Arachne Native Ratio call 28 2.3% 21 1.9% 1.3 seq 201 0.5% 63 1.7% 3.2 cflow 228 1.6% 42 1.8% 5.4 readGlobal 2762 4.3% 1 0.2% 2762 read 9729 4.9% 1 0.6% 9729 Table 1: Speed of each language construct used to interpret the base program compared to a native execution. the base program. For example, to study the performance of readGlobal, we wrote an aspect whose action returns the value of the global variable referred in the pointcut, i.e., we wrote aspects behaving like the base program. For each of these aspects, we compare the time required to perform the operation matching the pointcut, in case the operation is interpreted by the woven aspect with the time required to carry out the operation natively (without the woven aspect). For example, to study the performance of readGlobal, we first evaluate the time needed to retrieve the global variable value through the code generated by the C compiler gcc without any aspect woven and compare this value to the time needed to retrieve the global variable value through the aspect once it has been woven in the base program. We express our measurements as a ratio between these two durations to abstract from the experimentation platform. This approach requires the ability to measure short periods of time. For instance, a global variable value is usually retrieved (readGlobal in our aspect language) in a single clock tick. Since standard time measurement APIs were not precise enough, our benchmarking infrastructure relies on the rdtsc assembly instruction [18]. This instruction returns the number of clock cycles elapsed since power up. The Pentium 4 processor has the ability to dynamically reorder the instructions it executes. To ensure the validity of our measurement, we thus insert mfence instructions in the generated code whose execution speed is being measured. An mfence forces the preceding instructions to be fully executed before going on. The pipeline mechanism in the Pentium 4 processor entails that the speed of a piece of assembly code depends from the preceding instructions. To avoid such hidden dependencies, we place the operation whose execution time is being measured in a loop. We use gcc to unroll the loop at compile time and we measure the time to execute the complete loop. This measure divided by the number of loop repetitions yields an estimation of the time required to execute the operation. The number of times the loop is executed is chosen after the relative variations of the measures ,i.e., we increased the number of repetitions until ten runs yields an average relative variation not exceeding 5%. To check the correctness of our experimental protocol, we measured the time needed to execute a nop assembly instruction , that requires one processor cycle according to the Intel specification. The measures of nop presented a relative variation of 1.6%. Table 1 summarizes our experimental results. Using the aspect language to replace a function that returns immediately is only 1.3 times slower than a direct, aspect-less, call to that empty function. Since the aspect compiler packages advices as regular C functions, and because a call pointcut involves no residue, this good result is not surprising. When 35 Controlflow 28 Cycles 228 Cycles 327 Cycles 424 Cycles 522 Cycles 1 2 5 3 4 10 20 30 Ratio with a normal function call Number of imbricated calls 1 2 5 3 4 Ratio with 3 calls Number of matching instances 5 10 200.6 Cycles 293.2 Cycles 380.8 Cycles 466.3 Cycles Sequence Ratio 1000 2000 3000 577 Cycles Figure 11: controlflow, seq, and read performances an access to a global variable is replaced by an advice execution , the hooks generated by the rewriting strategy need to prepare the processor to call the advice function. This increases the time spent in the hooks. In addition, while an access to a global variable is often performed by a single x86 instruction, an empty function is often composed of four instructions. Hence the relative cost of an aspect triggered upon a global variable access and a direct, aspect-less , access to a global variable is slightly higher than the corresponding ratio for functions. A seq of three invocations of empty functions is only 3.2 time slower than the direct, aspect-less, three successive functions calls. Compared to the pointcuts used to delimit the different stages, the seq overhead is limited to a few pointer exchanges between the linked lists holding the bound variable. On Intel x86, global variable accesses benefit from excellent hardware support. In the absence of aspects, a direct global variable read is usually carried out in a single unique cycle. To trigger the advice execution, the Arachne runtime has to save and restore the processor state to ensure the execution co-herency , as advices are packaged as regular C functions (see also 4.2.1). It is therefore not surprising that a global variable readGlobal appears as being 2762 times slower than a direct, aspect-less global variable read. read performance can be accounted in the same way: in the absence of aspect, local variables are accessed in a single unique cycle. The signal mechanism used in the read requires that the operating system detects the base program attempt to read into a protected memory page before locating and triggering the signal handler set up by Arachne, as shown in 4.2.3. Such switches to and from kernel space remain slow. Using read to read a local variable is 9729 times slower than retrieving the local variable value directly, without aspects. seq and controlflow can refer to several points in the execution of the base program (i.e. different stages for seq and different function invocations for the controlflow). The runtime of these pointcuts grows linearly with the number of execution points they refer to and with the number of matching instances. Figure 11 summarizes a few experimental results for controlflow and seq proving these points. 5.2 Case Study on a real application Since, depending on the aspect construct used, interpreting the base program with aspects can slow it down by a factor ranging between 1.3 and 9729, we studied Arachne's performance on a real world application, the Web cache Squid. Arachne Manual Top1 Top1 Diff Top2 Top2 (%) Throughput (request/s) 5.59 5.59 5.58 5.59 Response Time (ms) 1131.42 1146.07 1.2 -1 1085.31 1074.55 Miss response time (ms) 2533.50 2539.52 0.2 1.8 2528.35 2525.34 Hit response time (ms) 28.96 28.76 -0.6 3.8 30.62 31.84 Hit ratio 59.76 59.35 -0.6 0.7 61.77 62.22 Errors 0.51 0.50 -1.9 0 0.34 0.34 Table 2: Performances comparison between manual modification and Arachne, for prefechting policy integration in Squid We extended Squid with a prefetching policy [9]. As described in section 3.1, we implemented this policy as a set of aspects and made a second implementation of this policy by editing the Squid source code and recompiling it. This section compares the performance of these two implemen-tations using standard Web cache performance indicators: throughput, response time and hit ratio. Obtaining access traces adequate to study a Web cache performance is difficult. The trace must be long enough to fill the cache. Due to privacy issues, traces are usually not publicly available. Since traces do not include the content of the accessed pages, these pages must be downloaded again. In the meantime the page contents may have changed and even the URLs may have disappeared. Instead of traces, we based our evaluation on Web Polygraph [30]. Polygraph is a benchmarking tool developed by the Squid team and featuring a realistic HTTP and SSL traffic generator and a flexible content simulator. We filled up the cache and simulated a one day workload with its two request rate peaks observed in real life environments [30]. Table 2 shows results of our simulation. Measures have been made during the two request peaks. The hit time and the miss time, time needed to deliver a document present, respectively not present, in the cache are very similar. It shows that differences are imperceptible between the version of Squid extended by Arachne and the one extended manually (less than 1%). Hence, even if the cost of Arachne's aspect language constructs might seem high, they are largely amortized in real applications. To give a typical example observed on our experimental platform: in case of a cache hit, a 3.8 MB page was retrieved in a single second, the time spent in prefetching advices amounted to 1801 sec, and the time spent within Arachne to execute the hooks and dynamic tests to 0.45 sec. In a miss case, on the average, a client retrieved the same page in 1.3 seconds, 16679 sec were spent in the advices and 0.67 sec within Arachne itself. RELATED WORK Our work is directly related to other aspect weavers for C, approaches for expressive aspect languages, and dynamic weaving, in particular for C. In this section, we consider related work in each of these fields in turn. Apart from Dyner and Arachne, there are few aspect 36 weavers for C (or even C like languages); some noteworthy exceptions are AspectC [12] (no available implementation ), AspectC++ and [33]. All of these rely on source-code transformation and thus cannot apply aspects to running C applications as required by the applications we consider. Furthermore, none of these systems provides explicit support for aspects over join point sequences. There is quite a large body of work now on the notion of expressive aspect languages where "more expressive" typically compares to w.r.t. AspectJ's pointcut and advice models . Our work has been inspired by Event-based AOP [15], which aims at the definition of pointcuts in terms of arbitrary relations between events. Nevertheless, many other approaches to expressive aspect languages exist: e.g., data-flow relations [26], logic programming [13], process algebras [3], graphs [5], and temporal logics [1], have all been proposed as a basis for the definition of expressive aspect languages . However, few of these encompass dynamic weaving and only the latter has been applied to C code under efficiency considerations similar to our setting. Dynamic weaving is commonly realized in Java through preprocessing at load-time like [8] or through the JVM Debugging Interface [28]. These tools rely on bytecode rewriting techniques, have typically limited expressivity (some do not support field accesses) and incur a huge performance overhead. Dynamic weaving through modification at runtime is found infrequently for compiled languages. An exception for Java is JasCo [21] whose most recent version (0.7) supports dynamic weaving through the new instrumentation API of Java 5. Many instrumentation techniques have been proposed to rewrite binary code on the fly. In these approaches, difficulty issues range from the complexity to rewrite binary code to the lack of a well-defined relationship between source code and the compiler generated binary code. Hence many approaches work on an intermediate representation of the binary code and source language [34]. Producing this representation first and then regenerating the appropriate binary executable code has proven to be costly both in terms of memory consumption and in CPU time. A few other approaches have considered a direct rewriting of the binary code at runtime. Dyninst [17] and dynamic probes [27] allow programmers to modify any binary instruction belonging to an executable. Dyninst however relies on the Unix debugging API: ptrace. ptrace allows a third party process to read and write the base program memory. It is however highly inefficient: before using ptrace, the third party process has to suspend the execution of the base program and resume its execution afterwards. In comparison , Arachne uses ptrace at most once, to inject its kernel DLL into the base program process. In addition, Dyninst does not free the programmer from dealing with low level details. For example, it seems difficult to trigger an advice execution upon a variable access with Dyninst: the translation from the variable identifier to an effective address is left to the user. Worse, Dyninst does not grant that the manipulation of the binary instructions it performs will succeed. Dyninst uses an instrumentation strategy where several adjacent instructions are relocated. This is unsafe as one of the relocated instructions can be the target of branching instructions. In comparison, Arachne join point model has been carefully chosen to avoid these kind of issues; if an aspect can be compiled with Arachne, it can always be woven. CONCLUSION AND FUTURE WORK In this paper we have discussed three different crosscutting concerns which are typical for C applications using OS-level services and which frequently need to be applied at runtime. We have motivated that such concerns can be expressed as aspects and have defined a suitable aspect language . This language is more expressive than those used in other aspect weavers for C in that it provides support for aspects defined over sequences of execution points as well as for variable aliases. We have presented an integration of this language into Arachne, a weaver for runtime weaving of aspects in C applications. Finally, we have provided evidence that the integration is efficient enough to apply such aspects dynamically to high-performance applications, in particular the web cache "squid." As future work, we intend to investigate the suitability of the proposed aspect language for other C-applications. We also intend to investigate Arachne extension to the C++ language. Indeed, object-oriented programming heavily uses protocol-based interfaces collaboration (hence sequence aspects ). Along with its open architecture, extending Arachne to support C++, will pave the way to a relatively language independent aspect and weaving infrastructure. Finally, Arachne's toolbox should be extended with support for aspect interactions (e.g., analyses and composition operators). REFERENCES [1] R. A. Aberg, J. L. Lawall, M. S udholt, G. Muller, and A.-F. L. Meur. On the automatic evolution of an os kernel using temporal logic and AOP. In Proceedings of Automated Software Engineering (ASE'03), pages 196204. IEEE, 2003. [2] American National Standards Institute. ANSI/ISO/IEC 9899-1999: Programming Languages -- C. American National Standards Institute, 1430 Broadway, New York, NY 10018, USA, 1999. [3] J. H. Andrews. Process-algebraic foundations of aspect-oriented programming. In Proceedings of the 3rd International Conference on Metalevel Architectures and Separation of Crosscutting Concerns, volume 2192 of LNCS. Springer Verlag, Sept. 2001. [4] M. Arlitt and T. Jin. A workload characterization study of the 1998 world cup web site. IEEE Network, 14(3):3037, May 2000. [5] U. Amann and A. Ludwig. Aspect weaving by graph rewriting. In U. W. Eisenecker and K. Czarnecki, editors, Generative Component-based Software Engineering (GCSE), Erfurt, Oct. 1999. [6] CERT - Carnegie Mellon University. Vulnerability note vu#613459, Feb. 2002. published on line: http://www.kb.cert.org/vuls/id/613459. [7] H. Chen and P. Mohapatra. Catp: A context-aware transportation protocol for http. In International Workshop on New Advances in Web Servers and Proxy Technologies Held with ICDCS, 2003. [8] S. Chiba and K. Nakagawa. Josh: An open AspectJ-like language. In Proceedings of the third 37 international conference on Aspect-oriented software development, pages 102111. ACM Press, Mar. 2004. [9] K.-I. Chinen and S. Yamaguchi. An interactive prefetching proxy server for improvement of WWW latency. In Seventh Annual Conference of the Internet Society (INET'97), Kuala Lumpur, June 1997. [10] I. Cidon, A. Gupta, R. Rom, and C. Schuba. Hybrid tcp-udp transport for web traffic. In Proceedings of the 18th IEEE International Performance, Computing, and Communications Conference (IPCCC'99), pages 177184, Feb. 1990. [11] S. Clowes. Injectso: Modifying and spying on running processes under linux. In Black hat briefings, 2001. [12] Y. Coady, G. Kiczales, M. Feeley, and G. Smolyn. Using AspectC to improve the modularity of Path-Specific customization in operating system code. In V. Gruhn, editor, Proc. of the Joint 8th European Software Engeneering Conference and 9th ACM SIGSOFT Symposium on the Foundation of Software Engeneering (ESEC/FSE-01), volume 26, 5 of SOFTWARE ENGINEERING NOTES, pages 8898, New York, Sept. 1014 2001. ACM Press. [13] K. de Volder. Aspect-oriented logic meta programming. In P. Cointe, editor, Meta-Level Architectures and Reflection, 2nd International Conference on Reflection, volume 1616 of LNCS, pages 250272. Springer Verlag, 1999. [14] R. Douence, P. Fradet, and M. S udholt. A framework for the detection and resolution of aspect interactions. In Proceedings of the ACM SIGPLAN/SIGSOFT Conference on Generative Programming and Component Engineering (GPCE'02), volume 2487 of LLNCS, pages 173188. Springer-Verlag, Oct. 2002. [15] R. Douence, O. Motelet, and M. S udholt. A formal definition of crosscuts. In Proceedings of the 3rd International Conference on Metalevel Architectures and Separation of Crosscutting Concerns, volume 2192 of LNCS, pages 170186. Springer Verlag, Sept. 2001. [16] E. Hilsdale and J. Hugunin. Advice weaving in aspectj. In Proceedings of the 3rd international conference on Aspect-oriented software development, pages 2635. ACM Press, 2004. [17] J. K. Hollingsworth, B. P. Miller, M. J. R. Goncalves, O. Naim, Z. Xu, and L. Zheng. MDL: A language and compiler for dynamic program instrumentation. In IEEE Conference on Parallel Architectures and Compilation Techniques (PACT), pages 201213, Nov. 1997. [18] Intel Corportation. IA-32 Intel Architecture Software Developer's Manual. Intel Corportation, 2001. [19] V. Issarny, M. Ban^atre, B. Charpiot, and J.-M. Menaud. Quality of service and electronic newspaper: The Etel solution. Lecture Notes in Computer Science, 1752:472496, 2000. [20] J. Jaffar, S. Michaylov, P. J. Stuckey, and R. H. C. Yap. The clp( r ) language and system. ACM Trans. Program. Lang. Syst., 14(3):339395, 1992. [21] JasCo home page. http://ssel.vub.ac.be/jasco/. [22] R. Jones and P. Kelly. Backwards-compatible bounds checking for arrays and pointers in c programs. In M. Kamkar, editor, Proceedings of the Third International Workshop on Automatic Debugging, volume 2, pages 1326, May 1997. [23] A. D. Keromytis. "Patch on Demand" Saves Even More Time? IEEE Computer, 37(8):9496, 2004. [24] G. Kiczales, J. Lamping, A. Menhdhekar, C. Maeda, C. Lopes, J.-M. Loingtier, and J. Irwin. Aspect-oriented programming. In M. Aksit and S. Matsuoka, editors, Proceedings European Conference on Object-Oriented Programming, volume 1241, pages 220242. Jyvaskyla, Finland, June 1997. [25] K. J. Lieberherr, J. Palm, and R. Sundaram. Expressiveness and complexity of crosscut languages. Technical Report NU-CCIS-04-10, Northeastern University, Sept. 2004. [26] H. Masuhara and K. Kawauchi. Dataflow pointcut in aspect-oriented programming. In First Asian Symposium on Programming Languages and Systems (APLAS'03), 2003. [27] R. J. Moore. Dynamic probes and generalised kernel hooks interface for Linux. In USENIX, editor, Proceedings of the 4th Annual Linux Showcase and Conference, Atlanta, October 1014, 2000, Atlanta, Georgia, USA, Berkeley, CA, USA, 2000. USENIX. [28] A. Popovici, G. Alonso, and T. Gross. Just-in-time aspects: efficient dynamic weaving for Java. In Proceedings of the 2nd international conference on Aspect-oriented software development, pages 100109, Boston, Massachusetts, Mar. 2003. ACM Press. [29] M. Rabinovich and H. Wang. DHTTP: An efficient and cache-friendly transfer protocol for web traffic. In INFOCOM, pages 15971606, 2001. [30] A. Rousskov and D. Wessels. High-performance benchmarking with Web Polygraph. Software Practice and Experience, 34(2):187211, Feb. 2004. [31] O. Ruwase and M. S. Lam. A practical dynamic buffer overflow detector. In Proceedings of the 11th Annual Network and Distributed System Security Symposium. Internet Society, Feb. 2004. [32] M. Segura-Devillechaise, J.-M. Menaud, G. Muller, and J. Lawall. Web cache prefetching as an aspect: Towards a dynamic-weaving based solution. In Proceedings of the 2nd international conference on Aspect-oriented software development, pages 110119, Boston, MA, USA, Mar. 2003. ACM Press. [33] O. Spinczyk, A. Gal, and W. Schroeder-Preikschat. AspectC++: an aspect-oriented extension to the C++ programming language. In Proceedings of the Fortieth International Conference on Tools Pacific, pages 5360. Australian Computer Society, Inc., 2002. [34] A. Srivastava and A. Edwards. Vulcan: Binary transformation in a distributed environment. Microsoft Research Tech. Rpt. MSR-TR-2001-50, 2001. [35] U. S. L. System Unix. System V Application Binary Interface Intel 386 Architecture Processor Supplement. Prentice Hall Trade, 1994. [36] D. Wessels. Squid: The Definitive Guide. O'Reilly and Associates, Jan. 2004. [37] J. Wilander and M. Kamkar. A comparison of publicly available tools for dynamic buffer overflow prevention. In Proceedings of the 10th Network and Distributed System Security Symposium, pages 149162, San Diego, California, February 2003. 38

aspect language;buffer overflows;prefetching;sequence pointcut;system applications;binary code;dynamic weaving;Arachne;web cache;operating system;network protocol;C applications

33

An Index System for Quality Synthesis Evaluation of BtoC Business Website

It is important for successful electronic business to have a hi-quality business website. So we need an accurate and effective index system to evaluate and analyses the quality of the business website. In this paper, the evaluation index system following the `grey box' principle is proposed which considers both efficiency of business website and performance of electronic business system. Using R-Hierarchical clustering method to extract the typical indexes from sub-indexes is theoretically proved to have a rationality and effectiveness. Finally, the evaluation method is briefly discussed.

INTRODUCTION Business website is an online media between buyer and seller. A hi-quality website is crucial to a company for a successful e-business. What is a hi-quality business website? In terms of maintaining the website, what do we focus on so that the quality meets the users' needs? Apparently, using click-through rate to assess the popularity cannot objectively and accurately evaluate the quality of the business websites. Instead, we need to rely on scientific evaluation index system and methods. At present, there are many methods available for business website comparison or ranking, such as Usage Ranking, Purchase Comparison, Expert Opinion and Synthesis Evaluation etc. You can find both official authority and non-governmental organization that issue their power ranking. The former one is to monitor and regulate the market, such as CNNIC, which organized the competition for the Top Ten Websites in domestic. The latter one, such as Consumerreports ( www.consumerreports . org ), BizRate(www.bizrate.com), Forrester Research etc., is mainly to guide the web users' activity. These kinds of comparison or ranking have special value in getting reputation and increasing recognition of the business websites among the users, however, e-business enterprise can not improve the quality of their websites directly based on the results of these kinds of assessments. The main purpose of this paper is to develop an index system for quantitative evaluation of the BtoC websites, which dose not emphasize the income of the website but focus on evaluating of its synthesis quality. We hope that the applying of this index system will provide the technique developers and maintainers some references for designing, appraising and diagnosing their e-business system to improve its quality level, and to support managers to make decisions for operation of the websites. OVERVIEW OF PREVIOUS STUDIES Comparing to the fast growing of e-business websites in the world, currently we can rarely find the particular research on the evaluation index system of business website. QEM (The website quality evaluation method) proposed by Olsina and Godoy etc. in 1999 can be considered as one of the representative approaches. It based on the main factors to evaluate the quality of the websites, including functionality (global search, navigability, and content relativity), usability (website map, addresses directory), efficiency and reliability. In 2000, American researcher, Panla Solaman, presented e-SERVQUAL model based on the conventional service quality evaluation model SERVQUAL. It contains some factors like efficiency, deal completeness, reliability, privacy protection, responsiveness, recompense and contact etc. In the same year, another American researcher, Hauler, introduced an e-QUAL model which includes the factors of content, accessibility, navigability, design and presentation, responsiveness, background, personalization and customization, etc. In 2004, F.J. Miranda Gonzalez and T.M.Banegil Palacios developed an universal evaluation index system WIS (Web Assessment Index) that can be employed to assess websites by different organizations. It consists of four indexes of accessibility, navigability, speed and content. [1] However, the universal index system cannot measure a website exactly and absolutely due to the industry specialty, organizational characteristics and different usages. One of the representative researches is Mr. ZhongHai Li's paper about ergonomics standard of online store. It assesses the business websites by testing if the design of the website coincides with the shopping process of online consumers. This standard has five factors, such as search and browse, merchandise information, shopping cart, register and pay, service and support. [4] Another index system for small and medium business websites covers the factors of general features, design, promotion, information and the others. [5] Here we list our major findings from the previous researches: 2.1 Unreasonable Selection of the Index Some research consider not only the original design but also the factors such as promotion and income of business website. 75 Some evaluation systems have correlative or contradictive indexes. For example, it considers the download speed, at the same time, it requires the web designers not to excessively use flash and sounds to slow down the speed. 2.2 Unilateral Evaluation Most of the research takes the users' view to evaluate the function and design of website. It treats the business system as a `black box' and ignores the impact of system performance on the websites quality. But considering the factors of system performance alone is also not a complete evaluation for improving service quality of website. 2.3 Lack of a Complete Set of Quality Synthesis Evaluation System A complete set of tool to evaluate the websites must include the following important elements: categories, factors, weights, rankings standard and assessment model. So far, we have not seen any literature discussing complete set of evaluation index system aiming at the quality of BtoC websites. PRINCIPLE FOR THE QUALITY SYNTHESIS EVALUATION First, the three fundamental principles we need to follow are to be comprehensive, to be scientific and to be feasible. We should evaluate all the facets of the website from different dimensions and avoid missing value of important factors. Moreover, the definition of the evaluation index should be accurateobjective and logical so it can eliminate the impact on the evaluation result brought by the correlative indexes. Concurrently, we need reduce the quantity of indexes or adopt the simple ones which data is easier to be collected, and prevent from complicated calculation due to the excessive indexes. The main purpose of improving business websites is to serve the users better. They are concerned only about the websites' external attributes, such as content, function, presentation and browse speed, etc. So, evaluating only by taking their views cannot directly guide to develop, maintain and administrate the website. Just like treating the patient's symptom but not the disease itself, the technique developer or maintainer cannot radically improve the quality of their websites by correcting system structure and web design according to the evaluation result. Only after we adopt the `grey box' index system that considers both efficiency of business website and performance of e-business system, we can establish a quality synthesis evaluation index system to benefit the management of BtoC websites. QUALITY EVALUATION INDEXES FOR BUSINESS WEBSITE Selection of index items lays down the foundation for constructing evaluation index system. After we thoroughly analyze the evaluation objectives based on the characteristics of business website, we propose an initial index system includes 5 categories and totally 28 index items shown in the following Table 1. Table 1 Quality evaluation indexes for business websites Categories Indexes Function Effectiveness 1 Integrative Function 2 Interactive Function 3 Convenience 4 Service Personalization 5 Website Credibility 6 Business Authorization Business Information 7 Accuracy 8 Authoritativeness 9 Variety in Type 10 Inclusiveness 11 Uniqueness 12 Orderliness 13 Timeliness 14 Variety in Search Method 15 Search Effectiveness 16 Version Internationalization Website Design 17 User Interface Friendliness 18 Development Standardization 19 Website Uniqueness 20 Columns Originality 21 Website Structure Clarity 22 Page Style Consistency 23 Harmonization System Usability 24 System Stableness 25 Compatibility 26 System Security 27 Self-adaptability System Efficiency 28 System Speediness Website self-adaptability refers to capability of e-business system intelligently providing personalized service and dynamic optimizing system performance. System Efficiency refers to the ability that the system response quickly to the requests of numbers of web users. It can be measured through values of some quantitative indexes, such as response time, throughput or utilization rate, etc. OPTIMIZING THE EVALUATION INDEXES It is necessary for our initial evaluation system to optimize if it can be applied in practice. First, the indexes are more or less correlative which will affects the objectiveness of the evaluation. Second, there are too more indexes that will result in lower efficiency. Therefore, we try to extract and simplify the indexes by using R-Hierarchical clustering method. Generally, R indicates the coefficient of correlation between two items. R-Hierarchical clustering method is usually applied to cluster the indexes. The steps are described as following. 5.1 Calculate Coefficient of Correlation and Clustering It firstly treats every index as one cluster. So, we have 28 clusters. Then, coefficient of correlation is calculated between every two clusters by minimum-distance method. Next, the two clusters with the maximal coefficient of correlation are clustered into a new one. The same process is repeated until all the indexes are clustered into one. 76 5.2 Analyze the Clustering Process and Determine Clusters We analyze the variation of minimum coefficient of correlation during the clustering process to find the leap points. According to the number of leap points and the knowledge of special field, we can eventually determine how many clusters we need. The whole process is illustrated in the following Figure 1. Figure 1 The process of R-Hierarchical clustering Following the principle of simplification and feasibility and considering the characteristics of BtoC website, we cluster the 28 index items into 10. The precision rate is over 0.75. 5.3 Calculate Correlation Index and Extract the Representative Indexes First, we calculate the correlation index that is the average of R between one index and every other index in the same cluster. 1 2 2 = j j m r R mi in this formula is the number of the indexes in the cluster that index Xj belongs to. Then, we select the index with the maximal correlation index in the total 10 clusters individually and identify 10 of them as the most representative indexes. Finally, the weights of the indexes are derived by the expert grade method. The final indexes and their weights are shown in the following table 2. Table 2 The final indexes and their weights Category Weight Index Weight 1.1 Service Personalization 0.10 Function Effectiveness 0.22 1.2 Website Credibility 0.12 2.1 Information Inclusiveness 0.10 Business Information 0.18 2.2 Version Internationalization 0.08 3.1 Columns Originality 0.09 Website Design 0.28 3.2 Website Structure Clarity 0.10 3.3 Harmonization 0.09 4.1 System Stableness 0.10 System Usability 0.22 4.2 System Security 0.12 System Efficiency 0.10 5.1 System Speediness 0.10 CONCLUSION In this paper, we have proposed an index system for quality synthesis evaluation and diagnosis of the BtoC websites following the `Grey Box' evaluation principle, and scientifically determined and simplified the index items. Usually, factor analysis or principal component analysis is used to solve the problem of common-factor and multiple indexes. But these methods are only suitable for the quantitative indexes, and the evaluation process is not truly simplified. Because the new index is the linear function of some original ones, it still needs to calculate the value of new indexes by collecting all the values of the original ones. In our index system, most of index is descriptive one. So we have finalized the indexes by using the R-Hierarchical clustering method. It really has reduced the number of the evaluation indexes without losing the major information from the original indexes. Furthermore, it has effectively avoided the impact of common-factors on the evaluation result. Only the index of system efficiency can be measured through quantitative sub-indexes such as response time, etc. Most of depictive indexes are subjective and fuzzy. In view of this, we should use fuzzy comprehensive analysis method to evaluate to get more efficiency result. In our future work we are intended to propose an evaluation model and conduct evaluation to some famous domestic BtoC websites to prove if this index system is scientific and feasible. Moreover, we will improve this set of index system including evaluation model to make the whole set of index system more feasibility. REFERENCES [1] F.J. Miranda Gonzalez, T.M. Banegil Palacios, Quantitative evaluation of commercial web sites: an empirical study of Spanish firms, International Journal of Information Management, 24(2004)313-328 [2] Chang Liu, Kirk P. Amett, Exploring the factors associated with Web site success in the context of electronic commerce, Information & Management , 38 (2000) 23-33 [3] Evans, J. R., & King, V. E.. Business-to-business marketing and the World Wide Web: Planning, managing and assessing web sites. Industrial Marketing Management, 28(1999)343358 [4] Zhonghai Li, Jianqiao Liao, Hui Xiao; the analyze of work efficiency on webshop design in China; human work efficiency; 4(2002) 43-45 [5] Research a index system for evaluation of on enterprise website, http://www.365un.com/ xmb/viewthread.php?tid=1 998 77

Business website;B2C websites;System performance;representitive indexes;fuzzy analysis;R-Hierarchical clustering;Evaluation system;quality synthesis evaluation;quality evaluation index;R-Hierarchical clustering method;index optimization;e-commerce;clustering;correlation index

34

An Integrated Environment to Visually Construct 3D Animations

In this paper, we present an expressive 3D animation environment that enables users to rapidly and visually prototype animated worlds with a fully 3D user-interface. A 3D device allows the specification of complex 3D motion, while virtual tools are visible mediators that live in the same 3D space as application objects and supply the interaction metaphors to control them. In our environment, there is no intrinsic difference between user interface and application objects. Multi-way constraints provide the necessary tight coupling among components that makes it possible to seamlessly compose animated and interactive behaviors. By recording the effects of manipulations, all the expressive power of the 3D user interface is exploited to define animations. Effective editing of recorded manipulations is made possible by compacting all continuous parameter evolutions with an incremental data-reduction algorithm, designed to preserve both geometry and timing. The automatic generation of editable representations of interactive performances overcomes one of the major limitations of current performance animation systems. Novel interactive solutions to animation problems are made possible by the tight integration of all system components. In particular, animations can be synchronized by using constrained manipulation during playback. The accompanying video-tape illustrates our approach with interactive sequences showing the visual construction of 3D animated worlds. All the demonstrations in the video were recorded live and were not edited.

INTRODUCTION Modern 3D graphics systems allow a rapidly growing user community to create and animate increasingly sophisticated worlds. Despite their inherent three-dimensionality, these systems are still largely controlled by 2D WIMP user-interfaces. The lack of correlation between manipulation and effect and the high cognitive distance from users to edited models are the major drawbacks of this solution [13]. The inadequacy of user-interfaces based on 2D input devices and mindsets becomes particularly evident in the realm of interactive 3D animation. In this case, the low-bandwidth communication between user-interface and application and the restrictions in interactive 3D motion specification capabilities make it extremely difficult to define animations with straight-ahead actions. This inability to interactively specify the animation timing is a major obstacle in all cases where the spontaneity of the animated object's behavior is important [21; 35; 4]. In this paper, we present an expressive 3D animation environment that enables users to rapidly and visually prototype animated worlds with a fully 3D user-interface. A 3D device allows the specification of complex 3D motion, while virtual tools supply the interaction metaphors to control application objects. In our environment, there is no intrinsic difference between user interface and application objects. Multi-way constraints provide the necessary tight coupling among components that makes it possible to compose animated and interactive behaviors. By recording the effects of manipulations, all the expressive power of the 3D user interface is exploited to define animations. Effective editing of recorded manipulations is made possible by compacting all continuous parameter evolutions with our data-reduction algorithm, designed to preserve both geometry and timing. Novel interactive solutions to animation problems are made possible by the tight integration of all system components. In particular, animations can be synchronized using constrained manipulation during playback. In the following sections, we present an overview of the system, we make comparisons with related work, and we conclude with a view of future directions. The accompanying video-tape illustrates our approach with interactive sequences showing the visual construction of 3D animated worlds. All demonstrations in the video were recorded live and were not edited. SYSTEM OVERVIEW Our animation environment is built on top of VB2 [17; 18], a graphics architecture based on objects and constraints. During interaction, the user is the source of a flow of information propagating from input device sensors to manipulated models. Permission to make digital/hard copy of part or all 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, the copyright notice, the title of the publication and its date appear, and notice is given that copying is by permission of ACM, Inc. To copy otherwise, to republish, to post on servers, or to redistribute to lists, requires prior specific permission and/or a fee. 1995 ACM-0-89791-701-4/95/008...$3.50 395 VB2 applications are represented by a network of interrelated objects, and the maintenance of relationships is delegated to a constraint-based change propagation mechanism. Different primitive elements represent the various aspects of the system's state and behavior: active variables store the system's state, domain-independent hierarchical constraints [9] maintain multi way relations between active variables, daemons provide support for discrete simulation tasks, and indirect expressions allow constraints and daemons to dynamically locate their variables. Constraints are maintained using an efficient local propagation algorithm based on Skyblue [27; 17; 18]. The solver is domain independent and can maintain a hierarchy of multi-way, multi output dataflow constraints. The fact that constraint solving consists in performing method selection on the basis of constraint priorities and graph structure, without considering the variables' values, allows an effective application of a lazy evaluation strategy [17; 18]. The main drawback of such a local propagation algorithm is the limitation to acyclic constraint graphs. However, as noted by Sannella et al. [28], cyclic constraint networks are seldom encountered in the construction of user interfaces, and limiting the constraint solver to graphs without cycles gives enough efficiency and flexibility to create highly responsive complex interactive systems. In VB2 , the objects' internal constraint networks are designed so as to reduce the possibility of creating cyclic constraint graphs. Runtime introduction of a constraint that would create a cyclic graph causes an exception that can be handled to remove the offending constraint 1. The state manager behavior and the constraint solving techniques are detailed in [17; 18]. 2.2 Interaction The system's desktop configuration uses keyboard commands to trigger mode changes and animation playback, a Spaceball for continuous specification of spatial transformations, and a mouse for picking. Both hands are thus used simultaneously to input information. LCD shutter glasses provide binocular perception of the synthetic world. Since our main research goal is to explore the potentialities of 3D interaction, we do not provide a two dimensional graphical user interface. A 3D cursor, controlled by the Spaceball , is used to select and manipulate objects of the synthetic world. Direct manipulation and virtual tools are the two techniques used to input information. Both techniques involve using mediator objects that transform the cursor's movements into modifications of manipulated objects. Virtual tools are visible first class objects that live in the same 3D space as application objects and offer the interaction metaphor to control them. Their visual appearance is determined by a modeling hierarchy, while their behavior is controlled by an internal constraint network [18]. As in the real world, users configure their workspaces by selecting tools, positioning and orienting them in space, and binding them to application objects. At the moment of binding, the tool decides whether to accept the connection by checking if the application object contains all the needed information and by verifying that the constraint graph obtained by connecting the tool to the model can be handled by the underlying solver (i.e. it is acyclic). The binding mechanism is defined in a declarative way by using indirect constraints [18]. 1 VB2's current constraint solver [17; 28] is unable to find acyclic solutions of potentially cyclic constraint graphs. An algorithm that removes this limitation is presented in [36]. Information control Information display MODEL TOOL v1 v2 c1 c2 bound bound.v1 bound.v2 v1 v2 out_variable in_variable constraint in_out_variable direct reference indirect reference Instance (a) (b) Figure 1a. Design notation Figure 1b. Model and virtual tool When bound, the tool changes its visual appearance to a shape that provides information about its behavior and offers semantic feedback. During manipulation, the tool's and the application object's constraint networks remain continuously connected, so as to ensure information propagation. Multiple tools can be active simultaneously in the same 3D environment in order to control all its aspects. The environment's consistency is continuously ensured by the underlying constraint solver. The bi-directionality of the relationships between user-interface and application objects makes it possible to use virtual tools to interact with a dynamic environment, opening the door to the integration of animation and interaction techniques. 2.3 Animation By recording the effects of manipulations, animations can be sketched. In order to be able to edit the captured performance, a compact representation of continuous parameter evolution must be obtained. This representation must not only precisely approximate the shape of the initial parameter curves but also their timing. The data reduction algorithm must therefore treat the geometry and time components simultaneously in order to avoid the introduction of errors that would be difficult to control. We have developed an algorithm that incrementally builds, from the input sequence, a parametric B-spline preserving value and time of each input sample within a given tolerance. It is an incremental version of the Lyche and Mrken algorithm [22] that works in parallel with the interactive specification by considering only a small portion of the input curve at any time. Latency time and memory requirements for handling each portion of the curve are constant. Data reduction may therefore be performed concurrently with interactive parameter input, and the responsiveness of the application can be ensured when handling animations defined by any number of samples. The algorithm is presented in detail in [2; 4]. This performance-based approach complements key-framing by providing the ability to create animations with straight-ahead actions. It provides complete control over the animation shape and timing, while key-framing offers control only at a limited number of points. The mediation of virtual tools makes it possible to sketch the evolution of non-geometric attributes, while constrained or free motion can be specified with 3D devices. Since these devices offer continuous control of spatial transformations, subtle synchronizations between position and orientation components can be directly specified. In our environment, straight-ahead animations are defined by expressing the desire to record parameter evolution during interaction. This is done simply by pressing a different mouse button when starting an interaction task. A controller object is connected to each animatable model and is responsible for monitoring model state changes. While recording, all changes are handled by the controller to feed the animation tracks. Continuous tracks apply the data reduction 396 algorithm to the incoming information, while discrete tracks simply store a change value event. During playback, information propagates from the animation tracks through the controllers and down to the models. All connections are realized by bi-directional constraints. Since playback constraints are weaker than interaction constraints, the user can take control over animated models during playback. Animations involving synchronizations with the environment's evolution can thus be specified by interacting during playback [5]. Discrete Track Interaction Mediator Data reduction Data sampling Editing Animation recording Animation playback Continuous Track Application Object Animation Controller Figure 2. Interactive animation and playback RELATED WORK Constraint-based architectures have long been used for 2D graphics systems (see [28] for a survey). In the 3D graphics world, one-way constraints are commonly employed to maintain dependencies between components [20; 34; 37; 38]. This type of constraint cannot easily model mutual relations between objects, thus hindering the tight coupling between user-interface and application objects [28]. Our system uses instead multi-way local propagation constraints, which offer support for two-way communication between objects while remaining efficient enough to ensure the responsiveness of the system [17; 18; 27]. TBAG [14] also uses multi-way constraints maintained by Skyblue [27], but its functional approach concentrates more on modeling time varying behaviors than on creating interactive systems. Much effort has been spent in developing powerful numerical solvers for computer graphics (e.g. [7; 15; 16]). This work is complementary to ours, which focuses more on providing ways to interact with constrained environments. Such advanced solvers could replace local propagation in our system for the maintenance of numerical relationships. 3.2 Three-dimensional User Interfaces Much recent research has focused on obtaining rich interaction with 3D environments by means of advanced devices and 3D interaction metaphors [8; 10; 11; 13; 16; 19; 26; 30; 32]. 3D widgets or manipulators, similar to our virtual tools, are presented in [13; 32]. These works focused on providing support for 3D widget construction, while we concentrate more on the integration of multiple tools in a single dynamic environment. We are not aware of any attempts to apply the results of 3D interaction research to enhance animation capabilities. 3.3 Performance Animation A number of authors have proposed using live performances to drive computer animations (e.g. [1; 23; 33; 35]). We strive to bring the expressiveness of these approaches to general purpose animation systems running on graphics workstations. Instead of relying on advanced motion capture devices, we exploit our fully 3D user-interface to control the animated environment at a higher level of abstraction. The guiding approach proposed in [23] also seeks to provide better control of synthetic objects by raising the abstraction level of user interaction. That work concentrates on modeling complex behaviors in a discrete simulation framework, while we focus on providing intuitive user interfaces. A major limitation of current performance animation systems is the inability to build editable representations out of captured performances [35]. 3.4 Data Reduction Data reduction or curve fitting techniques have been successfully applied for the interactive specification of 2D or 3D curves or surfaces (e.g. [12; 24; 25; 29]). These techniques cannot be easily adapted to sketching animations of multi-dimensional parameters because they all exhibit one or more of the following problems: (i) restriction to 2D or 3D geometric constructions, (ii) lack of control on parameterization errors, and (iii) need to consider the entire input curve before reduction. An early attempt to use data reduction for animation is described in [29]. In that system, path geometry and path timing specifications were decoupled, loosing thus the advantages of performance approaches. Banks and Cohen [6] proposed for their drafting tool an incremental version of the Lyche and Mrken algorithm [22] that does not have the aforementioned drawbacks and could be used in a performance animation context. Their method shares with ours the idea of processing successive portions of the input curve which are then spliced together, but is unable to ensure constant latency times and memory needs [4]. CONCLUSIONS AND FUTURE WORK In this video-paper, we have presented an integrated environment for the rapid and visual prototyping of 3D animated worlds. Using our fully 3D user-interface, non-professional users can swiftly create complex animations with pose-to-pose and straight-ahead techniques. Thanks to automatic data-reduction, animations created by interactive performances can then be effectively edited. In our future work, we intend to develop new virtual tools and visualizations that will improve our 3D user interface for discrete and continuous track manipulation. To allow the system to adhere to timing requirements, we are developing time-critical techniques for controlling rendering complexity and constraint evaluation. ACKNOWLEDGMENTS The authors would like to thank Ronan Boulic for providing the walking engine used in the interactive sequences, Sally Kleinfeldt as well as Dean Allaman for helpful comments and suggestions, Angelo Mangili for technical help, and Michele Mller for doing the voice on the video. This research was conducted by the authors while at the Swiss Federal Institute of Technology in Lausanne. 397 REFERENCES [1] Baecker RM (1969) Picture-driven Animation. Proc. Spring Joint Computer Conference 34: 273-288. [2] Balaguer JF (1993) Virtual Studio: Un systme d'animation en environnement virtuel . PhD Thesis, Swiss Federal Institute of Technology in Lausanne. [3] Balaguer JF, Gobbetti E (1995) Animating Spaceland. To appear in IEEE Computer Special Isssue on Real-world Virtual Environments 28(7). [4] Balaguer JF, Gobbetti E (1995) Sketching 3D Animations. To appear in Proc. EUROGRAPHICS. [5] Balaguer JF, Gobbetti E (1995) Supporting Interactive Animation using Multi-way Constraints. Submitted for publication. [6] Banks M, Cohen E (1990) Real-time Spline Curves from Interactively Sketched Data. Proc. SIGGRAPH Symposium on Interactive 3D Graphics: 99-107 [7] Barzel R, Barr A (1988) A Modeling System Based on Dynamic Constraints. Proc. SIGGRAPH: 179-188. [8] Bier EA (1990) Snap-Dragging in Three Dimensions. Proc. SIGGRAPH Symposium on Interactive 3D Graphics : 193 204 . [9] Borning A, Freeman-Benson B, Wilson M (1992) Constraint Hierarchies. Lisp and Symbolic Computation 5(3): 221-268. [10] Butterworth J, Davidson A, Hench S, Olano TM (1992) 3DM: A Three Dimensional Modeler Using a Head-Mounted Display. Proc. SIGGRAPH Symposium on Interactive 3D Graphics: 135-138. [11] Card SK, Robertson GG, Mackinlay JD (1991) The Information Visualizer: An Information Workspace. Proc. SIGCHI : 181-188. [12] Chou JJ, Piegl LA (1992) Data Reduction Using Cubic Rational Splines. IEEE Computer Graphics and Applications 12(3): 60-68. [13] Conner DB, Snibbe SS, Herndon KP, Robbins DC, Zeleznik RC, Van Dam A (1992) Three-Dimensional Widgets. SIGGRAPH Symposium on Interactive 3D Graphics : 183 188 . [14] Elliott C, Schechter G, Yeung R, Abi-Ezzi S (1994) TBAG: A High Level Framework for Interactive, Animated 3D Graphics Applications. Proc. SIGGRAPH: 421-434. [15] Gleicher M (1993) A Graphics Toolkit Based on Differential Constraints. Proc. UIST : 109-120. [16] Gleicher M, Witkin A (1992) Through-the-Lens Camera Control. Proc. SIGGRAPH: 331-340. [17] Gobbetti E (1993) Virtuality Builder II: Vers une architecture pour l'interaction avec des modes sysnthtiques. PhD Thesis, Swiss Federal Institute of Technology in Lausanne. [18] Gobbetti E, Balaguer JF (1993) VB2: A Framework for Interaction in Synthetic Worlds. Proc. UIST: 167-178. [19] Herndon KP, van Dam A, Gleicher M (1994) Report: Workshop on the Challenges of 3D Interaction, CHI Bulletin, October. [20] Kass M (1992) CONDOR: Constraint-based Dataflow. Proc. SIGGRAPH: 321-330. [21] Lasseter J (1987) Principles of Traditional Animation Applied to 3D Computer Animation. Proc. SIGGRAPH: 35 44 . [22] Lyche T, Mrken K (1987) Knot Removal for Parametric B spline Curves and Surfaces. Computer Aided Geometric Design 4: 217-230. [23] McKenna M, Pieper S, Zeltzer D (1990) Control of a Virtual Actor: The Roach. Proc. SIGGRAPH Symposium on Interactive 3D Graphics: 165-174. [24] Plass M, Stone M (1983) Curve Fitting with Piecewise Parametric Cubics. Proc. SIGGRAPH: 229-239. [25] Pudet T (1994) Real Time Fitting of Hand Sketched Pressure Brushstrokes. Proc. EUROGRAPHICS: 205-220. [26] Sachs E, Roberts A, Stoops D (1990) 3-Draw: A Tool for Designing 3D Shapes. IEEE Computer Graphics and Applications 11(6): 18-26. [27] Sannella M (1994) Skyblue: A Multi-Way Local Propagation Constraint Solver for User Interface Construction. Proc. UIST : 137-146. [28] Sannella M, Maloney J, Freeman-Benson B, Borning A (1992) Multi-way versus One-way Constraints in User Interfaces . Software Practice and Experience 23(5): 529-566. [29] Schneider PJ (1988) Phoenix: An Interactive Curve Design System Based on the Automatic Fitting of Hand-Sketched Curves. Master's Thesis, University of Washington. [30] Shaw C, Green M (1994) Two-Handed Polygonal Surface Design. Proc. UIST : 212-215. [31] Shelley KL, Greenberg DP (1982) Path Specification and Path Coherence. Proc. SIGGRAPH: 157-166. [32] Strauss PS, Carey R (1992) An Object-Oriented 3D Graphics Toolkit. Proc. SIGGRAPH: 341-347. [33] Tice S (1993) VActor Animation Creation System. SIGGRAPH Tutorial 1. [34] Upson C, Fulhauber T, Kamins D, Laidlaw D, Schlegel D, Vroom J, Gurwitz R, van Dam A (1989) The Application Visualization System: A Computational Environment for Scientific Visualization. IEEE CG&A 9(4): 30-42. [35] Walters G (1993) Performance Animation at PDI. SIGGRAPH Tutorial 1. [36] Vander Zanden B (1995) An Incremental Algorithm for Satisfying Hierarchies of Multi-way, Dataflow Constraints . Technical Report, University of Tennessee, Knoxville. [37] Zeleznik RC, Conner DB, Wlocka MM, Aliaga DG, Wang NT, Hubbard PM, Knepp B, Kaufman H, Hughes JF, van Dam A (1991) An Object-Oriented Framework for the Integration of Interactive Animation Techniques. Proc. SIGGRAPH: 105-112. [38] Zeltzer D, Pieper S, Sturman DJ (1989) An Integrated Graphical Simulation Platform. Proc. Graphics Interface: 266-274. 398

animation synchronization;computer graphics;Object-Oriented Graphics;3d animation environment;data reduction;visualization;multi-way constrained architecture;human interaction;Data Reduction;3D Animation;Local Propagation Constraints;recording 3d manipulation;Virtual Tools;3D Widgets;3d user interface;3D Interaction;dynamic model

35

An Intensional Approach to the Specification of Test Cases for Database Applications

When testing database applications, in addition to creating in-memory fixtures it is also necessary to create an initial database state that is appropriate for each test case. Current approaches either require exact database states to be specified in advance, or else generate a single initial state (under guidance from the user) that is intended to be suitable for execution of all test cases. The first method allows large test suites to be executed in batch, but requires considerable programmer effort to create the test cases (and to maintain them). The second method requires less programmer effort, but increases the likelihood that test cases will fail in non-fault situations, due to unexpected changes to the content of the database. In this paper, we propose a new approach in which the database states required for testing are specified intensionally, as constrained queries, that can be used to prepare the database for testing automatically . This technique overcomes the limitations of the other approaches, and does not appear to impose significant performance overheads.

INTRODUCTION Modern information systems are typically organised as collections of independent application programs that communicate with one another by means of a central database. The database records the state of the organisation that the 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. ICSE'06, May 2028, 2006, Shanghai, China. Copyright 2006 ACM 1-59593-085-X/06/0005 ... $ 5.00. information system supports, while the application programs implement the business processes that manipulate the state. To take a simple but ubiquitous example, a database system might record details of customers, products and sales, while the application programs associated with it handle operations such as new product purchases and update of the product catalogue, as well as supporting decision making by generating reports regarding the most profitable product lines, names and addresses of loss-making customers, etc. In order to test such application programs, it is necessary to create test fixtures that simulate the presence of the rest of the information system. Fixtures for traditional test cases typically consist of in-memory objects and data structures that provide the inputs to the program being tested. This kind of fixture is also needed when testing database applications (especially when performing unit testing); however, since it is unrealistic (and often incorrect) to execute test cases against an empty database, we need to create additional fixture elements within the database itself. Current practice in the software industry is to maintain one or more test databases that can be used for testing individual programs. These databases can be artificially generated (e.g., using tools such as DBMonster 1 and DataFac-tory 2 ) or they may be subsets of the live database, taken as a snapshot at some recent point in time. Copies of the live data sets have the advantage that they are more likely to be representative of the patterns of data encountered in practice, while artificial data sets have the advantage that they can be made to embody specific characteristics (such as particular data skew patterns or volumes), which may be useful for load and stress testing. Both approaches, however, suffer from several disadvantages . The most significant problem occurs when none of the available test databases are suitable starting points for a particular test case. For example, suppose a particular test case executes a program which purges inactive customers, with the aim of verifying that the business rule forbidding deletion of customers with negative balances is correctly en-forced . If none of the test databases contains any inactive customers with negative balances, then the test case cannot be executed successfully. For a one-off test run, testing personnel can choose a database that is close to what is required , and manually update it so that it is suitable for use with the test case. But if a complete test suite is to be executed (possibly including test cases which themselves make modifications to the database state) then in the worst case 1 http://DBMonster.kernelpanic.pl 2 http://www.quest.com/datafactory 102 this manual intervention will be required in between every test case execution. This is clearly undesirable if test suites are large or time-consuming to execute, or if the test suite is to be run in batch (as in the case of overnight regression testing, for example). Current research in testing for database systems proposes two approaches to this problem. One of these is to include within the test case description a full (extensional) specification of the database state against which it is to be run (and of the database state that should be produced if the test has executed successfully) [13, 14]. This solution is exemplified by DBUnit 3 , an extension of the JUnit testing framework 4 that is designed for testing database applications written in Java. Each DBUnit test case is accompanied by an XML file describing the data set required for the test. Before each test run, DBUnit clears the database state and inserts the data described by the XML file. This approach has the advantage of simplicity, but it places a considerable burden on testing personnel, especially when complex database states are required. It is also inefficient, since the database must be continually destroyed and recre-ated between tests, even when significant parts of the database might have been reused by the succeeding tests. Moreover, maintenance of a large suite of such tests is extremely challenging , since any small change to the database schema may require corresponding changes to many test cases. The second approach that has been explored in the literature is more efficient in that it requires the creation of only one database state per test suite (rather than one per test case). It is exemplified by the AGENDA database testing toolkit [6, 7], which can automatically generate a database state given information about the schema, some data generation functions for individual attributes and some user-selected heuristics describing the kind of database state required . The AGENDA tool also generates test cases from a simple analysis of the program being verified. The user must then add preconditions to each test case that are checked just before it is executed and that will prevent a case from being executed against an inappropriate database state. This approach successfully relieves the user of the need to specify complete database states in full detail, but at a cost. The user must accept that some of the test cases may not be executed because the database state fails the precondition, even when it would require only a small change to bring the database into a suitable state for the test. Since only one database state is created per test suite, this problem of failed tests is likely to become more severe as the size of the test suite grows. There is also a potential inefficiency involved in generating test descriptions and inputs, and in creating the additional log tables and constraints/triggers needed by the AGENDA tool, for test cases that are not in fact going to be executed. Ideally, we would prefer to be able to combine the advantages of both these approaches, to give a form of database test case that is quick and natural to specify, and which maximises the number of cases within the suite that can be executed while minimising the number of full test databases that need to be maintained. Our thesis is that this can be achieved by allowing testing personnel to describe the database states involved in their test cases intensionally, in 3 http://www.dbunit.org 4 http://www.junit.org the form of declarative conditions that the input database must satisfy, and by providing a testing harness that can automatically adjust the input database so that the test conditions are satisfied [19]. In this paper, we present a language for specifying such intensional database tests, and describe its semantics and operational behaviour (Section 2). We present an algorithm for automatically modifying database states so that test preconditions are satisfied (Section 3), thus ensuring that all test cases can be executed without requiring any human intervention. We further describe how we have extended the JUnit testing framework to allow intensional database tests to be specified and executed in practice (Section 4). Finally, we present the results of an evaluation of the performance of the techniques (Section 5) and conclude (Section 6). SPECIFYING INTENSIONAL TESTS A conventional test case is typically modelled as a triple &lt; p, i, o &gt;, which denotes a test that executes program p with inputs (e.g., parameters) denoted by i. If no faults are encountered during the test execution, the output that will be produced is o. In the case of test cases for database applications , we must add two further elements--the specification of the database state against which p is to be executed, and some statement of the database state that should result from the execution of p if it is operating correctly according to its specification. For example, consider the example program mentioned in Section 1 that prunes inactive customer details from the database. For this test case, we require a database state that contains at least one inactive customer. This could easily be stated as a predicate logic condition over the database, assuming the obvious mapping between stored relations and predicates, e.g.: (custNo, lastOrderOn, a, b, c) customer (custNo, a, b, c, lastOrderOn) lastOrderOn &lt; today - 90 The program in question does not access any parts of the database other than the customer table. Therefore, we do not care what values the other tables contain and need not mention them in the intensional specification of the test. This approach works equally well for observing the results of the test. For example, when testing the customer pruning behaviour, we might require that no inactive customer with a non-negative balance should exist in the database after the test: ((custNum, lastOrderDate, a, b, c) customer (custNum, a, bal , c, lastOrderDate) lastOrderDate &lt; today - 90 bal &gt; 0) Effectively, the test case describes a set of valid (i.e., fault-free ) state transition for the database, as a classic pre/post-condition pair. This first-order-logic style of database specification does not work so well when we consider the testing problem in more depth, however. The problem is that we need to do more than test the input database for compliance with the requirements of the test case; we also need to extract information from it to be used to instantiate other elements 103 of the test case. For example, suppose we wish to test a program that deletes details of individual customers. Such programs typically require some input from the user, identifying the specific customer record that is to be deleted (e.g., by supplying the relevant customer code as a parameter). This could be achieved by requiring the tester to embed the customer code into the test case elements, as literal values. Alternatively, we could search for a suitable customer that already exists in the database, using a standard database query, and use the values from that in specifying the inputs for the test case. This would minimise the amount of work required to prepare the database for test execution (since we would be using data already present in the database), and it would also mean that test cases can be written very quickly, since the user does not need to specify every last detail of the data to be used. Under this approach, the specification of the input database state now has a dual role: it must state the condition that determines whether the database state is suitable for execution of the test case and it must also return bindings for the free variables that appear in the remaining components of the test case. For the latter purpose, we would prefer to use a straightforward query language, while for the former we require the ability to place conditions on the data. With a simple extension of a standard query language such as SQL, we can combine both these purposes in a single statement. For example, the following statement: ANY :cn GENERATED BY SELECT custNo FROM customer WHERE lastOrderDate &lt; today() - 90 AND balance &lt; 0 retrieves the customer code of some record that meets the given conditions (an inactive customer with negative balance ) from the database, and binds it to the variable :cn. It also places a cardinality constraint on the result of the query, that at least one such binding must exist (implied by the use of the keyword ANY). The variable :cn can then be used to specify other elements of the test case. The obvious usage in this example is in specifying the inputs to the program being tested, but it can also be used in describing the expected outputs of the program. In this example test case, the correct behaviour of the DeleteCustomer program is to reject the deletion of :cn, since customers with a negative balance cannot be purged from the database. We might therefore give the following specification of the desired output database state: AT LEAST 1 :cn2 GENERATED BY SELECT custNo FROM customer WHERE custNo = :cn Of course, not all test cases are best specified in terms of values retrieved from the database. For example, suppose that we wish to write test cases for a program that adds new customers to the database. The inputs to this program are the details of the new customer, and the precondition for one particular test case states that no customer should exist that has the same customer code as that of the customer being created. We cannot retrieve the customer details from the database in this case, as they have not yet been stored in it. Again, we could force the user to include the required values as literals in the test case, but ideally we would like to give &lt;CONDITION&gt; ::= &lt;TYPE&gt; &lt;BINDINGLIST&gt; GENERATED BY &lt;SELECT&gt; &lt;TYPE&gt; ::= ANY | NO | AT LEAST &lt;i&gt; | AT MOST &lt;i&gt; | EXACTLY &lt;i&gt; | ALL | FIRST &lt;i&gt; ::= {0-9} &lt;BINDINGLIST&gt; ::= &lt;BINDING&gt; { `,' &lt;BINDINGLIST&gt; } &lt;BINDING&gt; ::= {A-Z | a-z} &lt;SELECT&gt; ::= ... Figure 1: Simplified BNF Grammar for SQL Extensions more support to the process of test case generation. One way to achieve this is to allow user-defined data generator functions to be incorporated within queries as though they were relations. For example, the following expression states our requirements for this test case, while also binding the variables needed for input to the program: ANY :cn, :name, :addr, :bal GENERATED BY SELECT gc.custno, gc.name, gc.addr, 0 FROM genCustomerDetails() AS gc WHERE gc.custno NOT IN ( SELECT custno FROM customer WHERE balance &gt; 0) Here, the data generator function getCustomerDetails() is used as if it were a normal relation, whereas in fact the results it returns are computed on the fly. In fact, several of the main commercial database management systems already allow user-defined functions to be embedded in queries in this way, so this does not require a further extension of SQL. Figure 1 shows the minimal extensions that are needed to support all the kinds of constrained query shown above using the SQL99 standard [17]. 2.1 Test Case Semantics Clearly, the semantics of these intensional database test cases is more complex than for traditional extensional tests. However, we can define their semantics formally in terms of a mapping from intensional tests to sets of equivalent extensional database test cases. We first present a formal definition of the structure of our intensional test cases: Definition 1. An intensional database test case is a quintuple &lt; p, i, DB i , o, DB o &gt;, where: p is the program to be executed in the test, i is a tuple of n variables and literals that describes the inputs to be given to program p, where n is the number of parameters expected by p, DB i is a set of constrained queries that together specify the initial database state. o is a tuple of m variables and literal that describes the expected outputs from the program p. DB o is a set of constrained queries that together specify the conditions that must hold in the database state after execution of p if no fault has been encountered. 104 A constrained query has the form &lt; Q, min, max , vars &gt;, where Q is a standard relational algebra query, min and max describe the constraints on the cardinality of the query result set, and vars is the list of variables bound by the query result. A database test case is well-formed for use with a particular database schema iff: for every variable v that occurs free in i, DB i , o and DB o , there exists a query in DB i that provides a binding for v, for every query &lt; q, n, m, vs &gt; in DB i DB o , q is a well-formed query over that returns k-tuples, where |vs| = k, and there are no circular variable dependencies amongst the queries in DB i . We can now define a semantics for the intensional database test cases as follows. Every intensional test case is equivalent to a set of extensional test cases. An extensional test case defines a specific test run, in terms of actual inputs and outputs, rather than expressions denoting sets of inputs and outputs. The set of all possible extensional test cases is given by: P L n DB L DB where P is the set of all programs, L is the set of all literals , L n is the set of all n-tuples formed from L and DB is the set of all database states (relative to all schemas) 5 . The components of each extensional test are the program to be tested, the input values, the initial database state, the expected output and the expected final database state, respectively. An intensional test case is effectively a shorthand expression for a set of extensional test cases that are all derived from the same equivalence partition of the test case inputs. An intensional database test &lt; p, i, DB i , o, DB o &gt;, where DB i = {&lt; q i , n i , m i , v i &gt;} and DB o = {&lt; q o , n o , m o , v o &gt;}, is equivalent to the following set of extensional tests: {&lt; p, i[v i /v], db i , o[v i /v], db o &gt; | db i DB (n i |q i (db i )| m i ) v q i (db i ) db o DB (n o |(q o [v i /v])(db o )| m o )} We use the notation exp[ 1 / 2 ] to express the substitution of the values in 1 by the corresponding values in 2 whereever they occur in exp. Therefore, this expression denotes the set of extensional tests where the input database satisfies the constraints imposed by the initial constrained query, and where the bindings from execution of that query (here expressed as the tuple of variables v) are substituted into the 5 For simplicity of presentation, we assume that all programs require the same number of inputs (n). In practice, n can be the largest number of inputs required by any program, and the unused values can be filled with nulls. expressions defining the inputs, expected output and expected final database state before they too are evaluated 6 . The idea underlying this notion of an intensional test is that when any of its corresponding extensional sets are executed , the intensional test is itself deemed to have been executed. Thus, the use of intensional tests allows much greater freedom at test execution time, since we may choose any of the possible extensional tests, depending on which is closest to our starting environment. In the next section, we will consider the practical ramifications of this approach to testing, and describe how the semantics just described can be implemented in practice. DATABASE PREPARATION The execution of an intensional database test case consists of three distinct phases: 1) preparation of the environment for test execution; 2) execution of the test with the prepared inputs; and 3) capture and storage of the results, for later analysis. Since all the work of finding bindings for the variables in the test case specification is done in the preparation phase, the final two phases are straightforward and differ little from standard testing procedures. When program execution is complete, the constrained query that determines whether the test has been successful or not is evaluated against the database, and the output from the program is checked against what is expected. In the case of test failure, the details of the actual extensional test that was executed are recorded, for diagnosis purposes. The first phase, however, is more complex. If we were content to execute only those test cases which happen to be suitable for use with the initial database state, then the preparation phase would simply be a matter of executing the input constrained queries against the database and, if they are all successful, using the bindings thus produced to instantiate the remaining components of the test case. However, thanks to the declarative nature of our test case specifications, the testing framework can be pro-active in cases where the given database is not suitable for use by the test case, and can automatically generate a sequence of updates that will cause the constrained queries to produce the required number of bindings. In fact, this problem is similar (though not identical) to one that has been studied by the database and artificial intelligence communities for many years. It is known variously as the view update problem [9], the knowledge base update problem [12], and the transaction repair problem [10]. Many database systems have the capability to define views on top of the basic database. A view is a kind of virtual relation. To the user, it appears to be a normal relation, but it contains no stored data. Instead, the contents of the view are defined by a expression over other relations, and attempts to retrieve data from the view are converted into queries over these relations. To take a simple example for illustration , we might create a view called Debtors which appears to be a relation of the same name containing all customers with a negative balance. Attempts to retrieve Debtors is 6 For simplicity of presentation, we assume here that there is only one query in each of DB i and DB o . In practice, it may be necessary to include several queries, each producing different bindings and imposing different cardinality constraints. In this case, the constraints must be conjoined, and the full set of bindings can be retrieved by performing a natural join of all the queries, with join condition true. 105 converted into a query against the customer table with an added constraint on the balance. If views are truly to act as normal relations then it should be possible to update them as well query them. But what does it mean to update a virtual relation? In this case, the view update must be converted into a sequence of updates on the stored relations that will cause the desired change in the contents of the view itself. This is a non-trivial problem for realistic view languages, and becomes even more difficult when we move into the context of knowledge bases, where virtual relations can be defined using rules over other relations , and when we add integrity constraints that must be maintained by all updates [1, 2, 3, 4, 5, 8, 11]. Only in very narrow circumstances does a view update have a single translation into real updates [15, 18]. Various heuristics for selecting from amongst the possible translations have been proposed (of which the most common is to choose the update that results in the smallest change to the existing data set [2]), but in real applications user input is needed in order to identify the translation that corresponds most closely to the real world state that the database should reflect [10]. In the case of intensional database tests, we have a query (the constrained query that describes our requirements for the test) that does not produce the correct number of answers when executed against the test database. We need to find a sequence of updates to the base data that will cause our query to produce the number of answers we need. However , in this case, there is no requirement to find the set of updates that matches the state of reality -- any sensible update that satisfies the query conditions will be acceptable. This simplifies the problem considerably, removing the need for complex search procedures and for any user input. 3.1 The Preparation Algorithm One of the advantages of using a query-based language for test specification (as opposed to a predicate calculus-based language) is that we can make use of a very common and easy-to-analyse internal form for (relational) database queries, called relational algebra. This form provides a small number of operations on relations that can be combined to form complex queries. For example, the three most basic (and useful) relational algebra operators are: The projection operator, Atts R, which creates a relation from R by deleting all attributes not in Atts. For example, [Country] Customer produces a relation that contains just the countries that appear in the Customer relation. The selection operator, c R, which creates a relation that contains all the rows from relation R that satisfy the condition c. For example, bal &lt;0 Customer returns a relation containing details of all customers with negative balances. The join operator, R 1 c S, which creates a relation containing rows from the cross product of R and S that satisfy the join condition c. The query Debtor 1 dNo=iNo Inactive returns details of all debtors who are also inactive . Since the result of each relational algebra operator is itself a relation, together they form a closed algebra. This means that we can form arbitrarily complex queries by applying operators to the results of other operators. For example, a query which retrieves the customer number of all customers with a negative balance would be written as: [custNo] ( balance&lt;0 Customer ) A common way to visualise such expressions is as a tree of operators. The tree for the above query is shown in Figure 2. Figure 2: Relational Algebra Tree for Negative Balance Query. Our algorithm for preparing a database for testing is based around this notion of a relational algebra tree. We take the cardinality constraints from the test specification, and push them down through the nodes of the input database query tree, collecting up additional conditions as we go. When we reach a leaf node (i.e. a base relation), we make updates to the database so that the pushed-down constraints are satisfied for that relation. At each stage, we collect up the different kinds of constraint and push them further down into the tree. These constraint types are: Min and Max, the upper and lower bounds on the desired cardinality of the result set. SelC, the selection conditions on the relations that we are interested in. UAtts, the collection of attributes that are used in the constrained query, and that must be populated in any new data that we insert. We also build up a collection of queries that describe the data that has been prepared for testing so far, as we progress through the tree. We call these queries "bindings" (Bgs), since they give us values for the variables that occur within the selection and join conditions. At each stage, the bindings should contain one query for each leaf node that has so far been prepared. It is easiest to see how this works by considering a simple example, such as that shown in Figure 2. Let us assume we have a constrained query that requires at least one customer with negative balance to exist, and that our database does not currently contain any such customers. We begin at the root node of the tree, with only the cardinality constraints extracted from the test specification: Min = 1, Max = null, SelC = true, UAtts = , Bgs = The top node is a projection operator. Projection does not affect the cardinality of the result set, nor impose any conditions , but it does tell us something about the attributes used 106 Figure 3: Relational Algebra Tree Showing Multiple Joins by the query. We therefore add the projection attributes to UAtts and push the constraints down to the next node: Min = 1, Max = null, SelC = true, UAtts = {custNo}, Bgs = Next we must deal with the selection node. Selection nodes reduce the cardinality of their input, so we need to push down the selection conditions to ensure that any updates we may make affect the correct tuples. We also need to add any attributes appearing in the selection condition to UAtts: Min = 1, Max = null, SelC = balance &lt; 0, UAtts = {custNo, balance}, Bgs = The final node is the leaf node, representing the Customer relation. We construct a query from the conditions on that relation and execute it, to find out how many answers are currently in the database. In this case, there are none, so we need to insert a new Customer record with at least the custNo and balance attributes populated, and with a negative balance. If there are any integrity constraints on this relation, then we need to make sure they are also satisfied by the new data. We use the DBMonster data generator mentioned earlier to create the new data. It allows generation functions to be specified for attributes, and additional constraints to be placed on them. It will also maintain primary key, foreign key, non-null and domain constraints if configured appro-priately using the information present in the pushed-down constraints. Of course, this is a very simple example. In general, we can expect to have to deal with more complicated queries involving several joins, such as that shown in Figure 3. This relational algebra tree is equivalent to the following constrained query: ANY :orderNo, :productNo GENERATED BY SELECT o.orderno, p.productno FROM Order o, Orderdetail d, Product p WHERE o.orderno = d.orderno AND d.productno = p.productno AND p.price &gt; 50 which requires that at least one order must exist that involves the purchase of at least one product that costs more than 50. Joins complicate the process of preparing the database, because they introduce dependencies between the updates that take place at different leaf nodes. For example, imagine that we have processed the tree shown in Figure 3 as far as the leaf node representing the OrderDetail relation. Join operators further constrain the selection condition (by conjoining in their join condition), but add no other constraints . So, by the time we reach this leaf node, SelC will have been set to: o.orderno = d.orderno d.productno = p.productno We need to find out whether a suitable OrderDetail record exists within the database. However, in order to do this, we need to know something about what preparation actions were performed when the Product leaf node was processed. Maybe there were already plenty of 50-plus products in the catalogue, or maybe there were none and one had to be created. How is this information passed through to the OrderDetail node so that the correct tuple can be identified or created? In the current version of our algorithm, we have chosen to use the database itself to communicate these values. If there are many suitable Product records, then we can find one by querying the database directly once again. If a new product had to be created, then it will now be present in the database, so we can still retrieve it by querying. The information needed to construct these queries is present in the selection conditions that have been considered during the processing of the relational algebra tree up to this point. For example, in order to search for an OrderDetail tuple that is connected to a suitable Product, we need to issue the following query: SELECT d.* FROM OrderDetail d, Product p WHERE d.productno = p.productno AND p.price &gt; 50 This query cannot be constructed from only the constraints pushed-down from the parent nodes of the leaf node; instead, we need to collect up the constraints imposed by all nodes visited before the current node, so that they are available for query formation. This is done using the Bgs data structure mentioned earlier. Figure 4 presents the complete algorithm, showing the behaviour required for each different type of operator. The algorithm is presented as a side-effecting function which takes the constrained query that is to be satisfied by the database, and a set of initial conditions that state the required cardinality bounds and initialise SelC to true, UAtts to and Bgs to . The function returns a set of bindings, but these are discarded. The main task of the algorithm is carried out by the side-effecting updates that occur when leaf nodes are processed. DOT-UNIT TESTING FRAMEWORK The intensional database test language and accompanying preparation algorithm have been implemented within a testing tool, called DOT-Unit. This tool is part of a larger Data-Oriented Testing 7 framework that is under development at the University of Manchester [20]. DOT-Unit has been implemented as an extension to the JUnit testing framework 7 http://www.cs.man.ac.uk/ willmord/dot/ 107 Projection operator prepare( Atts Q, Min, Max, UAtts, SelC, Bgs) = prepare(Q, Min, Max, UAtts Atts, SelC, Bgs) Selection operator prepare( c Q, Min, Max, UAtts, SelC, Bgs) = prepare(Q, Min, Max, UAtts, SelC c, Bgs) Join operator prepare(Q 1 1 jc Q 2 , Min, Max, UAtts, SelC, Bgs) = prepare(Q 2 , Min, Max, UAtts, SelC jc, prepare(Q 1 , Min, Max, UAtts, SelC, Bgs)) Relation (leaf node) prepare(Rasv , Min, Max, UAtts, SelC, Bgs) Q = bindingQuery(v, SelC, Bgs) Execute Q to produce result set RS if |RS| &lt; Min then Invoke DBMonster to create (Min - |RS|) more instances of R that satisfy the conditions in Q else if |RS| &gt; Max then Delete the first (|RS| - Max) tuples in RS else No preparation updates needed return (Bgs binding(v, Q)) Figure 4: The Database Preparation Algorithm for the unit testing of Java applications [16]. We have subclassed the standard JUnit TestCase class, to create a dedicated DatabaseTestCase class for specifying and managing intensional database tests. DatabaseTestCase provides facilities for specifying pre-conditions on database state, generating and manipulating the bindings that are produced by such pre-conditions, and evaluating post-conditions on the database state after the test has been completed. The standard JUnit methods for determining the results of test execution on the in-memory fixture can also be used. Figure 5 shows an example DatabaseTestCase that includes two individual tests. The first verifies that when a customer with a non-negative balance is deleted, all customers with that customer number really do disappear from the database. The second uses a data generation function to propose attribute values for a new customer record (including a unique customer number), and checks that after the program has executed only one customer with the generated customer number exists. We use a prefixed colon to indicate variables that are shared amongst the test components -- a notation that will be familiar to many database programmers, since it is commonly used in various forms of embedded SQL. The shared variables acquire their values when the test harness evaluates the precondition (and performs any necessary database preparation steps). These values can then be accessed using the binding method, and can be used in arbitrarily complex assert conditions, as well as in instantiating the post-condition query. One of the main advantages of using the JUnit framework as the basis for the implementation of DOT-Unit is that it allows us to integrate our tool seamlessly into existing development environments, such as Eclipse 8 . Thus, DOT-Unit tests are executed in exactly the same way as a standard JUnit test case, and the results are displayed using the same interface components. This allows testing of database and non-database components to be interleaved in a convenient and natural manner. 8 http://www.eclipse.org EVALUATION The practicality of this intensional test case approach depends largely on the performance overhead imposed by the database preparation algorithm. If the time required to execute each individual test case is significantly higher using our approach than with DBUnit, say, then fewer tests will be able to be executed in the time available and the benefits of faster test development and fewer spurious test failures will be negated. To gain a handle on the degree of performance overhead to be expected from DOT-Unit, we made use of an existing extensional DB test suite that we created for earlier work [20]. This suite was designed for mp3cd browser 9 , an open-source Java/JDBC program that stories information about mp3 files in a MySQL 5.0 database 10 . The schema of the database consists of 6 relations with 22 attributes, 7 primary key constraints and 6 foreign key constraints. We created an equivalent intensional test suite, consisting of 20 test cases, from the extensional suite by converting each test case into DOT-Unit pre- and post-conditions. We also re-placed each hard-coded test parameter in the original tests into constrained query bindings. We wanted to investigate two specific aspects of the performance of DOT-Unit. First, we wanted to compare its performance with that of DBUnit over the equivalent test cases as the database size grows. Second, we wanted to gain some idea of what aspects of DB preparation and testing were dominating the performance of DOT-Unit. The results of the experiments we performed are presented below. All experiments were run on a Pentium-M 2.0GHz machine, with 1Gb RAM, running Ubuntu Linux. 5.1 Comparison with DBUnit At first sight, the extensional approach, as exemplified by DBUnit, would seem to be the more efficient method of the two, as the testing harness does not need to spend any time figuring out what updates need to be made prior to each test--it only needs to execute them. This does 9 http://mp3cdbrowser.sourceforge.net/mp3cd/ 10 http://www.mysql.com 108 public class ProgramTest extends DatabaseTestCase { public void testDeleteCustomer() { preCondition(&quot;ANY :cn GENERATED BY SELECT custNo FROM customer WHERE balance &gt; 0;&quot;); Program p = new Program(); p.deleteCustomer(binding(&quot;:cn&quot;)); postCondition(&quot;NO :cn2 GENERATED BY SELECT custno FROM customer WHERE custNo = :cn;&quot;); } public void testNewCustomer() { preCondition(&quot;ANY :cn, :name, :addr GENERATED BY SELECT gc.custNo, gc.name, gc.addr FROM genCustomerDetails() AS gc WHERE gc.custNo NOT IN (SELECT custNo FROM customer);&quot;); Program p = new Program(); boolean b = p.newCustomer(binding(&quot;:cn&quot;), binding(&quot;:name&quot;), binding(&quot;:addr&quot;)); assertTrue(b); postCondition(&quot;EXACTLY 1 :cn, :name, :addr GENERATED BY SELECT custno, name, addr FROM customer;&quot;); } } Figure 5: Example DOT-Unit Test Case not happen by accident, but because a human programmer has spent time earlier, deciding exactly what the database should look like for each test case. However, when writing DBUnit tests, it is common to try to reuse database descriptions for multiple test cases where possible, to reduce the amount of programming and maintenance time. In this case, some redundant updates will be made before each test case - updates that our extensional approach will not bother to make. It is also the case that DBUnit makes its updates blindly, whether they are needed or not, whereas the intensional approach will be able to reuse much of the existing database state for each new test case. Given this, it seems likely that the performance of DBUnit will be better when the database state required for each test case is relatively small, but that the situation will be reversed when the database state grows much larger. In order to gauge the point at which this change occurs, we ran our two test suites (extensional and intensional) with databases of varying sizes, and measured the execution time taken to execute the whole test suite. In each case, we generated initial database states of varying sizes at random - either populating the database directly (for the intensional test cases) or generating XML descriptions of the required state (for the extensional test cases). The results are shown in Figure 6. Figure 6: Comparison of Approaches as DB Size Increases To our surprise, although the performance of DOT-Unit was initially worse than that of DBUnit, it overtook its competitor at a comparatively small database size of around 20 tuples per relation. Obviously, this experiment is a little unfair to DBUnit, since programmers are unlikely to create database descriptions consisting of 1000s of tuples per relation . However, tests of this scale will be needed at some point in the development cycle, in order to verify the behaviour of the system on more realistic data sets. In order to assess the behaviour of DOT-Unit more pre-cisely , consider the graph in Figure 7, which shows the results at small databases sizes in more detail. It can be ob-served that the performance of DOT-Unit first improves and then begins to degrade again at a database size of around 50 tuples per relation. Figure 7: Detailed Comparison of Approaches One possible explanation for this initial improvement in performance is that, as the database size rises, so does the probability that the data needed for the test case is already present in the database. For the very small states, a lot of preparation work is required to create the needed data, whereas less work is needed for a more fully populated database. As the database size increases further, however, the costs of making the queries needed to test the preconditions and formulate the preparation updates rises, pushing up the time required for the entire preparation step. This 109 behaviour may be a peculiarity of the particular test suite used, of course, and further, more extensive studies will be required in order to completely characterise the performance of the DOT-Unit test harness. From these initial results, however, DOT-Unit appears to scale well relative to database size, and the execution times are of the same order of magnitude as those resulting from DBUnit. This suggests that the intensional approach may provide a good compromise between saving expensive programmer time in developing new test cases and expenditure of cheaper processing time in executing the test cases. 5.2 Effect of Constraint Complexity A further concern was the effect of increasing constraint complexity on the performance of DOT-Unit test cases. How much additional overhead is added for conditions involving a higher number of selection conditions and (most impor-tantly ) joins? In order to assess this, we grouped the test cases into three groups, according to their complexity: A: queries with one or more selections and no joins, B: queries with one or more selections and a join between two relations, C: queries with one or more selections and joins between three relations. This gave a test suite with 5 test cases in each of these categories, which we executed against a randomly generated database state with 500 tuples per relation that does not satisfy any of the test case pre-conditions. Figure 8 shows the results obtained for the three complexity categories. We measured the average time taken to execute the test cases in each category, including a breakdown of where the time is spent in each case: Test: the time required to execute the procedural aspects of the test case; Query: the time required to execute the query aspect of the test case condition; Prepare the time required to execute the preparation aspect of the test case condition. While the overall time required to execute the test cases rises as the complexity rises (unsurprisingly), the relative proportions of time spent in the various phases remains roughly the same. The preparation phase seems to account for slightly more than half of the time in each case, indicating that significant improvements could be achieved with a less-naive preparation algorithm. CONCLUSIONS We have presented a new approach to the specification of test cases for database systems that attempts to reduce the amount of manual intervention required in between test case runs while also minimising the number of spurious test failures due to inappropriate input database states. The approach has the further advantage that it sits naturally on top of test data sets taken from live databases, and this allows testing to be carried out using realistic data sets without requiring significant programmer effort to tailor the data set to the test cases. In effect, the intensional approach we have Figure 8: The Affect of Changing Constraint Complexity described allows software developers to trade programmer time for test execution time Our experience has indicated that intensional test cases are quick and natural to write for anyone who is familiar with SQL and database programming, although a study with an independent testing team would be necessary before we can make any strong claims in this regard. However , compared with what is involved in writing pure JDBC database test cases and DBUnit test cases, we found that the self-contained nature of the intensional test cases was a definite advantage. Writing DBUnit test cases requires the programmer to continually check that the test case is compatible with the database description. Moreover, since it is common to try to reuse database descriptions for multiple test cases by combining their requirements into one database state, it becomes very easy to break one test case by changing the database description in order to ready it for another. These problems do not arise with intensional testing, since all the information about the test case is present in a single file (the Java class file). We designed this first version of the preparation algorithm for simplicity and correctness rather than efficiency, and as such it performs rather stupidly in many cases. We are currently exploring options for improving the algorithm, including more intelligent selection of the order in which the relational algebra tree is traversed, alternating between passing query bindings and passing literal value bindings as is most efficient, and making use of modifications to existing tuples as well as simply adding and deleting tuples (both of which are comparatively expensive operations). The complexity of the conditions we can handle is at present limited by the capabilities of DBMonster, and can be expanded by development of a custom data generation facility. We also need to expand the range of queries that can be handled, beyond simple select-project-join queries. For example, standard SQL also allows aggregation and ordering within queries-both of which offer challenges in terms of automatic preparation . A further problem with our current algorithm is that it may sometimes fail to find a solution to the database preparation problem, even though one exists. This is due to the fact that updates are made at leaf nodes before the full set of constraints on those nodes has been encountered. It should 110 be possible to address the problem with more sophisticated querying techniques (this is an example of a fairly standard constrained search problem, after all), although this will add to the performance overhead. A thorough study of the trade-offs between spurious failures and more intelligent searching will need to be carried out before any concrete recommendations can be made. Finally, we note that where it is important to test large numbers of frame constraints (i.e. aspects of the original database state that are not affected by the execution of the program under test), it may be easier to express the test case using DBUnit, rather than cluttering up the intensional test with many such constraints. Our work presents a number of possible avenues for future work beyond the improvements mentioned above, of which the most urgent is the question of ordering of test cases within suites. This ordering can be in terms of reducing the cost of the modifications to database state or to maximise fault coverage. There is also the question of whether the modifications to database state should always persist between test cases or under certain conditions discarded. For example, a test case may specify that a relation be empty and to satisfy the condition the content is discarded. However , this relation may be required by later test cases and so by discarding its contents we increase the divide between the test state and the real world. This could be accomplished by either embedding the modifications inside of a transaction which can then be aborted or by using a hypothetical database engine. ACKNOWLEDGMENTS We thank Leonardo Mariani and the anonymous reviewers for comments on earlier drafts of this paper. David Willmor is supported by a research studentship from the UK Engineering and Physical Sciences Research Council. REFERENCES [1] M. Arenas, L. E. Bertossi, and J. Chomicki. Consistent query answers in inconsistent databases. In Proceedings of the 18th ACM SIGACT-SIGMOD-SIGART Symposium on Principles of Database Systems (PODS), pages 6879. ACM Press, 1999. [2] L. E. Bertossi and J. Chomicki. Query answering in inconsistent databases. In J. Chomicki, R. van der Meyden, and G. Saake, editors, Logics for Emerging Applications of Databases, pages 4383. Springer, 2003. [3] P. Bohannon, M. Flaster, W. Fan, and R. Rastogi. A cost-based model and effective heuristic for repairing constraints by value modification. In Proceedings of the SIGMOD Conference, pages 143154. ACM, 2005. [4] L. Bravo and L. E. Bertossi. Logic programs for consistently querying data integration systems. In G. Gottlob and T. Walsh, editors, Proceedings of the 18th International Joint Conference on Artificial Intelligence (IJCAI), pages 1015. Morgan Kaufmann, August 2003. [5] A. Cal`i, D. Lembo, and R. Rosati. On the decidability and complexity of query answering over inconsistent and incomplete databases. In Proceedings of the 22nd ACM SIGACT-SIGMOD-SIGART Symposium on Principles of Database Systems (PODS), pages 260271. ACM, June 2003. [6] D. Chays, S. Dan, P. G. Frankl, F. I. Vokolos, and E. J. Weber. A framework for testing database applications. In Proceedings of the International Symposium on Software Testing and Analysis (ISSTA), pages 147157, August 2000. [7] D. Chays, Y. Deng, P. G. Frankl, S. Dan, F. I. Vokolos, and E. J. Weyuker. An AGENDA for testing relational database applications. Software Testing, Verification and Reliability, 14(1):1744, 2004. [8] J. Chomicki and J. Marcinkowski. On the computational complexity of minimal-change integrity maintenance in relational databases. In L. E. Bertossi, A. Hunter, and T. Schaub, editors, Inconsistency Tolerance, volume 3300 of Lecture Notes in Computer Science, pages 119150. Springer, 2005. [9] S. S. Cosmadakis and C. H. Papadimitriou. Updates of relational views. Journal of the ACM, 31(4):742760, 1984. [10] S. M. Embury, S. M. Brandt, J. S. Robinson, I. Sutherland, F. A. Bisby, W. A. Gray, A. C. Jones, and R. J. White. Adapting integrity enforcement techniques for data reconciliation. Information Systems, 26(8):657689, 2001. [11] G. Greco, S. Greco, and E. Zumpano. A logical framework for querying and repairing inconsistent databases. IEEE Transactions on Knowledge and Data Engineering, 15(6):13891408, 2003. [12] A. Guessoum and J. W. Lloyd. Updating knowledge bases. New Generation Computing, 8(1):7189, 1990. [13] F. Haftmann, D. Kossmann, and A. Kreutz. Efficient regression tests for database applications. In Proceedings of the 2nd Biennial Conference on Innovative Data Systems Research (CIDR), pages 95106. Online Proceedings, January 2005. [14] G. M. Kapfhammer and M. L. Soffa. A family of test adequacy criteria for database-driven applications. In Proceedings of the 11th ACM SIGSOFT Symposium on Foundations of Software Engineering, pages 98107. ACM, September 2003. [15] R. Langerak. View updates in relational databases with an independent scheme. ACM Transactions on Database Systems (TODS), 15(1):4066, 1990. [16] P. Louridas. Junit: Unit testing and coding in tandem. IEEE Software, 22(4):12 15, July-Aug 2005. [17] J. Melton and A. R. Simon. SQL:1999 Understanding Relational Language Components. Morgan Kaufmann, 2002. [18] H. Shu. Using constraint satisfaction for view update. Journal of Intelligent Information Systems, 15(2):147173, 2000. [19] D. Willmor and S. M. Embury. Exploring test adequacy for database systems. In Proceedings of the 3rd UK Software Testing Research Workshop (UKTest), pages 123133. The University of Sheffield, September 2005. [20] D. Willmor and S. M. Embury. A safe regression test selection technique for databasedriven applications. In Proceedings of the 21st International Conference on Software Maintenance (ICSM), pages 421430. IEEE Computer Society, September 2005. 111

database testing;Efficient Testing;software testing;Seamless Integration;Query Based Language;Improvement for the Intensional Test Cases;DOT-UNIT;databases;Lesser Programmer Effort for Test Cases;Intensional Test Cases;Testing Framework;Testing for Database Systems;Performance Testing

36

Analysis of Soft Handover Measurements in 3G Network

A neural network based clustering method for the analysis of soft handovers in 3G network is introduced. The method is highly visual and it could be utilized in explorative analysis of mobile networks. In this paper, the method is used to find groups of similar mobile cell pairs in the sense of handover measurements. The groups or clusters found by the method are characterized by the rate of successful handovers as well as the causes of failing handover attempts. The most interesting clusters are those which represent certain type of problems in handover attempts. By comparing variable histograms of a selected cluster to histograms of the whole data set an application domain expert may find some explanations on problems. Two clusters are investigated further and causes of failing handover attempts are discussed.

INTRODUCTION Mobility management is a great challenge in current and future radio access networks. In third generation (3G) networks user experienced quality of service (QoS) under a move of mobile station (MS) from one mobile cell to another cell has been improved by implementing soft handover 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. MSWiM'06, October 26, 2006, Torremolinos, Malaga, Spain. Copyright 2006 ACM 1-59593-477-4/06/0010 ...$5.00. (SHO). Soft handover makes it possible to have connections on several base stations (BS) simultaneously. In this paper, a set of measurements which can be used for soft handover decision making are analyzed and compared with other measurements in which statistics of successful-ness of handover attempts have been collected. We do not know exactly the parameters of used SHO algorithm. SHOs are investigated only on basis of data set and some general knowledge of 3G systems. Mobile cell pairs with handovers (HO) are divided in groups using clustering algorithm. Cell pairs in which SHOs are similar with each other fall in same group. Different types of SHO failures are analyzed using clustering information and distributions of measurements in each cluster. In Section 2 the soft handover concept, the measurements and used neural network algorithm are shortly introduced. Analysis methods which have been used are described in Section 3. Preliminary results are shown and discussed in Section 4. Finally, some conclusions are drawn in the last section. BACKGROUND In this section, the basics of soft handover in 3G network is explained and the available data set is introduced. Neural network algorithm used in data clustering is also presented. 2.1 Soft handover Soft handover is a state of MS being connected to several BSs simultaneously. In GSM networks, a fixed threshold for handover from one cell to another is used. In 3G networks, each MS is connected to a network via a set of BSs called active set. Members of active set are updated on basis of measurements made by MS. The advantage of having connections on several BS simultaneously is realized when MS is moving towards another BS, the MS should have a connection at least on one BS all the time. In GSM system, the older connection has to be terminated before the new one can be setup. The connection setup phases are the most vulnerable steps in a call. The connection between MS and BS is setup in a beginning of a call or later when handover occurs. If the setup is not successful, it is useful to have an existing connection to another BS or otherwise the call will be abnormally terminated. Handover can occur due to signal quality reasons or when the traffic capacity in a cell has reached its maximum or is approaching it. In the latter case, traffic load in the network can be distributed more uniformly by handing over some users from the most crowded cells. The above method is 330 called cell breathing. Use of cell breathing without giving the information to the analyzer increases the complexity of the analysis and can mix up a lot in the analysis process. For a user soft handover means power saving (in uplink) and less abnormally terminated calls. For an operator lower MS transmitting powers mean less interference. When MS is in SHO, several BSs listen the same uplink channel, but all BSs have their own downlink channel. The offered diversity is resource consuming in downlink direction. There is a tradeoff between better QoS in mobility management and consumption of resources. Decision of soft handover is made in mobile station by comparing the signal-to-noise ratios of active and candidate BSs Common Pilot Channel (CPICH) [2]. Members of active set are selected on basis of powers of this pilot signal [5, 12, 16]. BSs which are not in the active set but next from it in the sense of measured quantity are in candidate set. Candidate set BSs are constantly monitored whether their offer better connection than cells in active set. Cells not in active or candidate set are monitored less frequently whether their can enter the candidate set. Cell is either added to the active set if the maximum amount of cells in the active set is not reached or cell replaces the cell which offers the lowest quality connection. Cells which are no more able to offer a connection which is good enough are removed from the active set. Thresholds are used in adding, replacing and removing BSs from active set by BSs in candidate set to avoid ping pong effect. This means that a value of measured quantity should be with a certain threshold better than the old one for changing cells in active set. If measurement which is only slightly better (i.e. with zero threshold) is enough for changing cells in sets, it is quite possible that the same change is performed in opposite direction very soon. Thus, the original update of the set was useless and resource consuming in the sense of all required signaling. 2.2 Data Three data sets of Key Performance Indicator (KPI) level measurements related on handover events are saved. Each set consists of measurements collected during one hour. KPI is considered as an important measure to be followed. It can be a measurement by itself or it has been computed from a group of raw counters [10]. One data vector consists of probabilities , means, sums and counters computed over one hour of one source target cell pair. Here, source refers on cell in active set and target on another cell which is measured and possibly added in active or candidate set. Measurements of target cell are compared with those of source cell. Handover decisions are made in MS on basis of measured and computed base stations received signal signal-to-noise ratios (E c /N 0 ). For each source and target cell pair mean of signal-to-noise ratio differences is computed using EcnoDiffMean = mean n[E c /N 0 ] target - [E c /N 0 ] source o . Mean value and number of made comparisons (EcnoDiffNum) are saved. Four bin pdfs of these measurements are also stored with bin centers in -6, -3, 0 and 3dB, correspond-ingly . In addition to E c /N 0 measurements, averages of received pilot signal power ratios between BS pairs (av rscp ratio) have been computed and stored in database. The time and probability of being in SHO with each other have also been measured. Time of target and source cell being in SHO with each other simultaneously is counted in variable t act. Then, at least one MS is in SHO having both source and target cell in its active set. The measurement is symmetric for a switch of source and target cells. Time of target cell being in SHO with source cell is stored in t act dir. Cell total time in SHO is saved in tot time sho. It has been counted over all the targets of fixed source cell. Probability of target and source being in same active set is stored in variable p act. Total number of SHO attempts to add target to active set is stored in SHO total att. Ratio of successful SHO attempts which lead to addition of target cell in active set is saved in add ratio. In addition to those above, the number of SHO failures is stored in pfail total and ratios of four different failure causes are saved. Failure occurs in setup or active time phase of SHO and it is either radio channel problem or not. Probability of cell being in monitored state is also measured (p4th 5th). All the measurements used in the analysis are shortly described in Table 1. A lot of data has been saved in data sets, but also some very important information is missing. Due to missing information on cell capacities, their locations and performed manual and automatic tuning operations on network configuration between successive data set saves, only preliminary analysis can be performed. The rest of the analysis process is described on theoretical level. 2.3 Self-Organizing Map Self-Organizing Map (SOM) [8] is an unsupervised neural network algorithm which adapts the codebook vectors of neurons so that they approximate the input data distribution . When the training has converged topological areas or domains corresponding to certain types of inputs can be found from the map. The topology and the size of the network is fixed before adaptation. In the SOM algorithm, the codebook vectors w j of the SOM are at first initialized. Then, the following steps are repeated for each input vector x: Find the index of best-matching or nearest codebook vector using i(x) = argmin||x - w j ||, in which j goes through all the neurons in the map. Next, the codebook vectors of winner neuron and its neighbors are updated using w j (t + 1) = w j (t) + h ij (x)(x(t) - w j (t)). Here, is the learning rate and h ij (x) is the neighborhood function centered around the winner neuron. Input sample x defines the winner neuron and the topological distance between indexes i and j defines how much the neuron is updated. Neighborhood function is typically Gaussian or bubble function i.e. function which decrease monotonically and even goes to zero when the distance increases. In this paper, a batch version of the SOM algorithm is used. In batch SOM, all codebook vectors of the SOM are computed after the best-matching units of all input data vectors have been found. The same data set is used several times. METHODS Handover related measurement from 3G network can be analyzed using standard data mining methods [1]. In this 331 Table 1: Measurements in the analysis. Data set has one sample vector for each source target cell pair. Variable Explanation Type EcnoDiffNum Computed E c /N 0 differences number EcnoDiffMean Computed E c /N 0 differences mean EcnoDiffPdf-6.0 -6 dB bin of E c /N 0 difference pdf ratio EcnoDiffPdf-3.0 -3 dB bin of E c /N 0 difference pdf ratio EcnoDiffPdf0.0 0 dB bin of E c /N 0 difference pdf ratio EcnoDiffPdf3.0 3 dB bin of E c /N 0 difference pdf ratio t act Target and source simultaneously in SHO mean t act dir Time of target being in SHO with source mean tot time sho Cell total time in SHO sum p act Target in active set of source ratio SHO total att SHO attempts to add Target to active set number add ratio Successful attempts leading to addition ratio pfail total Failures number pfail ini Setup phase failures due to non-radio ratio pfail ini radio Setup phase failures due to radio ratio pfail act Active time failures due to non-radio ratio pfail act radio Active time failures due to radio ratio p4th 5th Cell is in monitored state (=4th or 5th) ratio av rscp ratio Target / Source Received power ratio mean r fail Ratio pfail total / SHO total att ratio r EcnoDNum Ratio EcnoDiffNum / SHO total att ratio Variable defined in the analysis. study, methods presented in Figure 1 are used. At first, the miner have to decide what could be interesting in this data. The analysis task has to be defined. On basis of that the first choice of variables will be done. Next, the selected variables are preprocessed, in order to be able to use them in later analysis. In data mining tasks, variable selection and preprocessing are the most critical phases, because in this step the miner decides which variables are important and how should they be processed. The whole data mining process consists of several cycles performed repeatedly. The cycles include testing how different variable selections and preprocessing methods effect on final results. The process has inner loops in which some tasks or parameters are fixed on basis of selections made in outer loop. The inner loops are performed more frequently. Loops with more general task like the definition of mining task are repeated less frequently. When the mining task is defined the analyzer should be able to decide what is (s)he looking out for. Now, the analysis task is defined as finding groups of sim-ilarly behaving cell pairs in SHO situations. Importance of measurements can also be highlighted using proper weighting of variables. In addition to clustering, also other tasks for data analysis can be defined. One possibility is to try to find cells or cell pairs with anomalous behavior. Anomalies can also be found by clustering, but expert knowledge in variable selection and preprocessing steps are very important . Using different variables, preprocessing methods and weighting of variables different clustering results can be found. To find out which of them is useful, interpretation of clusters is needed. This can be done using histograms or rules defined by data samples falling in clusters. The results which have been found using clustering methods should be visualized together with spatial locations to be able to understand the usefulness of results. Methods should be performed repeatedly to analyze successive data sets under the knowledge of performed tuning operations. Thus, there is a possibility to find explanations to changing results. In this study, results of only one data set are shown, because more information on application domain is needed to be able to combine and compare successive clustering results. 3.1 Preprocessing Different preprocessing methods have been tested. The final method was selected on basis of histograms and the clusters which were found using the selected method. At the first step, the distributions are truncated. Outliers in the selected variables were replaced by their maximum permitted values. Two variables, pfail total and EcnoDiffNum, were scaled using the number of performed soft handover attempts (see Table 1). Logarithms of some of the variables were taken, but finally only scaled EcnoDiffNum was preprocessed with logarithmic function. Sample vectors with high amount of undefined measurements were canceled. Used clustering method (see section 3.2) allows using sample vectors in which some variables are undefined. However, they are not so useful when the rate of undefined values increases. Here, sample vectors with 15 or more missing values in 20 variables are canceled. In Figure 2 the histograms of the most interesting variables preprocessed using selected methods are visualized. Some of the variables have quite high peaks in distributions, but due to the origin of variables no other preprocessing have been performed. For example, handover failure reasons pfail ini, pfail ini radio, pfail act radio and pfail act sum up to unity. However, pfail act is not analyzed because it is zero all the time in the first data set. 332 Interpretation of clusters Clustering Preprocessing Parameter tuning Task definition Visualization with locations Variable selection Figure 1: Used data analysis method. Steps connected with solid arrows have been performed. 3.2 Clustering Cluster analysis is used to divide data vectors in groups. Data vectors falling in same cluster are similar with each other. Here, clustering is performed using a two-phase method [15]. In this method, data vectors are at first used to train a Self-Organizing Map. Neurons of the SOM adapt to incoming data so that the input data can in later analysis be represented by the codebook vectors of neurons. Number of these codebook vectors is much smaller than the number of original data vectors. Thus, computational complexity of the final crisp clustering algorithm is decreased. Another advantage of using a SOM based two-phase method instead of direct clustering of data vectors is the visualization capability of SOM. In addition to preprocessing, SOM algorithm provides another possibility to emphasize important properties of data. Larger weights in distance computation are given to the most important properties defined by the analyzer. Smaller or even zero weight can be given to those variables which are not used in organization of the SOM i.e. in building clusters . However, values of them can be compared to those with larger weights using various visualization methods. Weighting by variable importance can also be built into SOM training algorithm by utilizing learning distance metrics [7]. Figure 2: Logarithmic histograms after distribution cuts, logarithmic preprocessing of r EcnoDNum and scaling of all variables between [0,1] The codebook vectors are further clustered using k-means or some hierarchical clustering method. In this paper, Ward agglomerative clustering method has been used [4]. In the beginning of hierarchical clustering, each codebook vector is a cluster of its own. In the next step, the most similar clusters are combined and this is continued until all vectors are in same cluster. The clustering results form a tree structure called dendrogram. In visualization of a dendrogram, the clusters combined in each step and the distance between them are shown. Final clustering is selected by cutting this tree at certain level. The number of clusters can be selected manually or some cluster validation index can be utilized to find the optimum. In this paper, Davies-Bouldin validation index has been used [3]. Similar clustering methods have earlier been used in the analysis of both GSM and 3G network BTSs [9, 11, 13]. As a result of clustering, each data vector is represented by index of one neuron or by the codebook vector stored in that neuron. Furthermore, the neuron and the data vectors the neuron represents belong to same cluster. On basis of the clustering result, some clusters can be selected for more specific analysis. Cluster selection is usually done on basis of found higher values of some critical variables. It is possible to build a system in which rules are found for clusters [14, 9] and these are used to select interesting clusters au-tomatically . Here, interesting clusters are selected manually on basis of clusterwise variable mean values and histograms. RESULTS In this section, handover measurement data is used to train a Self-Organizing Map of size 17 12. Then, the codebook vectors of the SOM are clustered using hierarchical Ward method. Results of clustering are described and two clusters are then selected for more specific analysis. Characteristics of sample vectors falling in those clusters are studied using histograms. Only the most interesting variables are used to find the 333 nearest neuron of input data vector. These variables have nonzero mask which can also be considered as a weighting factor in a search for the best-matching neuron. Rest of the variables have zero mask, which means that they can be visualized and updated using SOM algorithm, but they do not have an effect on organization of the SOM and on selection of the cluster in which the sample belongs to. In Figure 3 all other component planes of SOM with positive mask are shown, except E c /N 0 difference distributions which are shown in Figure 6. In component plane visualization , distributions of components (or variables) of SOM codebook vectors are shown. Component values of one codebook vector are visualized using grayscaling and their locate in the same position at each plane. For example, values of one codebook vector are shown at upper right corner in each plane. Figure 3: Component planes of SOM with denor-malized scales. Shown variables have nonzero mask and they are not describing E c /N 0 difference distributions . Some component values which were not used in SOM training (i.e. they were masked out) are shown in Figure 4. Although, they have no effect on SOM organization, they are adapted to be able to compare their distributions even with those used in organizing the SOM. By visual comparison of variables in Figures 3 and 4, it can be seen that the total number of SHO attempts (SHO total att) and E c /N 0 difference measurements (EcnoDiffNum) is higher in upper part of the SOM. However , when the latter is scaled by total number of attempts, higher rate of measurements (r EcnoDNum) is in lower part of the map. Also, the total number of failuring SHO attempts (pfail total) is high in upper right corner, but scaling this by number of attempts tells us that the failure rate (r fail) in upper right corner is quite moderate. Instead, higher failure rates exists in both lower corners i.e. in clusters 5 and 8 (see Figure 5). Trained SOM codebook vectors are clustered using hierarchical Ward algorithm. The clustering result selected by Davies-Bouldin index is shown in Figure 5. Four bin E c /N 0 difference histograms are visualized on top of clustered SOM Figure 4: Denormalized component planes of variables which were not used in SOM training. in Figure 6. When component values of SOM (see Figures 3, 4 and 6) are compared with clustering result (see Figure 5) several types of source target pairs can be found. Most of them are behaving as expected, but some of them represent handover attempts with certain type of problems. Figure 5: SOM which is clustered using hierarchical Ward method and Davies-Bouldin validation index. To find out the most interesting clusters of the SOM for further investigations, distribution of data samples on SOM is visualized. In Figure 7a hits of all samples on SOM nodes are visualized and in Figure 7b hits of samples with SHO failure rate (r fail) larger than 22% are shown. Samples are distributed all over the map, only some edge nodes have slightly larger hit rate. Lower part of the map has more hits when samples with increased failure rate are considered. In Figure 8 hits of samples which represent two differ-334 Figure 6: EcnoDiff distributions on top of clustered SOM. In each SOM node a four bin E c /N 0 histogram is shown. ent types of SHO failures are shown. Samples are from cell pairs in which the rate of selected type of failures is larger than 75%. However, handover initialization failures due to some other reason than radio channel resources (i.e. pfail ini type failures) are obviously more frequent than failures due to radio channel initialization problems (pfail ini radio type failures). Cell pairs with SHO failures originating mainly from these two reasons are mapped on separate clusters. All SHO failures due to radio channel initialization are in cluster 9 (see Figures 5 and 8b) and most of all other initialization failures are in cluster 5 (see Figures 5 and 8a). In the following, these two clusters are studied in more detail. In Figures 9 and 10 histograms of samples which belong to clusters 5 and 9 are shown. These histograms should be compared with histograms of whole data set which were shown in Figure 2. In histograms of cluster 5 (see Figure 9), the average received signal power ratio (av rscp ratio) is slightly lower than in general. Distributions of three largest E c /N 0 difference measurement bins are completely different than corresponding distributions from the whole data set. In cluster 5 most of the samples have about 3dB E c /N 0 difference (EcnodiffPdf3.0) which means that at least this measurement makes successful SHOs possible and SHO should be performed. Exceptional E c /N 0 difference measurements (a) All (b) SHO failure rate &gt; 22% Figure 7: Sample vector hits on SOM nodes. Size of black hexagonal on SOM node denotes the number of hits. Maximum number of hits per node is shown above the plot. (a) pfail ini (b) pfail ini radio Figure 8: Hits of samples of two failure types. Samples of which more than 75% are failuring due to selected cause are counted. of this cluster can also be seen in Figure 6. All the failing cell pairs fail in initialization due to other than radio channel reasons (pfail ini). Total rate of failures is very high (r fail). One reason for high rate of failures can be that all the capacity is in use. In histograms of cluster 9 (see Figure 10), the average received power ratios are a bit higher than usual, but there are no samples with high rate of 3dB E c /N 0 differences (EcnoDiffPdf3 .0). However, in such a situation it should be possible to perform successful SHOs. The rate of initialization failures in radio channels (pfail ini radio) is higher than usually, but because only a small part of samples in this cluster have above mentioned problems the total SHO failure rate is not higher than usually. The total number of samples or cell pairs with high rate of initialization failures in SHO is so small, that it is impossible to make any further inferences from these clusters. It is possible to check histograms of only those samples which fulfill the failure rate criteria, but the number of samples is anyway quite low. 335 Figure 9: Histograms of data vectors of cluster 5. Cell pairs with high rate of radio channel initialization failures in SHO attempts vary from data set to another, but without any information on network topology and with uncomplete information on performed tuning operations, it is impossible to make any further inferences. Figure 10: Histograms of data vectors of cluster 9. CONCLUSIONS In this paper, a data analysis method based on a neural network has been presented. The method is utilized in data visualization and clustering. The presented method is only one possibility for finding data clusters. However, the benefits of the proposed method are the decrease in computational complexity due to used two-phase clustering algorithm and the visualization capability of the method. Thus, it is well suitable for this kind of explorative data analysis. It is desirable to find clusters with characteristics which differ from one cluster to another. In the presented method, selection of variables and variable weighting factors have been used to find interesting clusters. In the preprocessing phase, also the number of permitted undefined measurement values in sample vector has an effect on found clusters. Sample vectors with high rate of missing values are not so usable and describable as samples without them. Vectors with missing values can be used in the SOM training but the benefit of using them decreases when the rate of undefined values increases. In this study, histograms are used both when preprocessing methods are decided and when an interpretation for the found clusters are looked for. However, clusters can also be compared using other visual methods, finding limiting rules for variable values in clusters or comparing distributions of variable values in clusters using more sophisticated distribution comparison measures like Kullback-Leibler divergences [6]. The results which have been obtained using all available data sets differ slightly from each other, but due to uncomplete information on network configuration and parameter tuning, further inferences cannot be made. However, adding this information would offer interesting possibilities to continue this study. REFERENCES [1] P. Chapman, J. Clinton, T. Khabaza, T. Reinartz, and R. Wirth. CRISP-DM 1.0 step-by-step data mining guide. Technical report, CRISM-DM consortium, 2000. http://www.crisp-dm.org. [2] Y. Chen. Soft Handover Issues in Radio Resource Management for 3G WCDMA Networks. PhD thesis, Queen Mary, University of London, 2003. [3] D. Davies and D. Bouldin. A cluster separation measure. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1(2):224227, April 1979. [4] B. Everitt. Cluster Analysis. Arnold, 1993. [5] V. K. Garg. Wireless Network Evolution: 2G to 3G. Prentice-Hall, Inc., 2002. [6] S. Haykin. Neural Networks, a Comprehensive Foundation. Macmillan, 1999. [7] S. Kaski and J. Sinkkonen. Metrics that learn relevance. In Proceedings of the International Joint Conference on Neural Networks, volume 5, pages 547552, 2000. [8] T. Kohonen. Self-Organizing Maps. Springer-Verlag, Berlin, 1995. [9] J. Laiho, K. Raivio, P. Lehtim aki, K. H at onen, and O. Simula. Advanced analysis methods for 3G cellular networks. IEEE Transactions on Wireless Communications, 4(3):930942, May 2005. [10] J. Laiho, A. Wacker, and T. Novosad, editors. Radio Network Planning and Optimisation for UMTS. John Wiley & Sons Ltd., 2001. [11] P. Lehtim aki and K. Raivio. A SOM based approach for visualization of GSM network performance data. In IEA/AIE, pages 588598, 2005. [12] R. Prakash and V. Veeravalli. Locally optimal soft handoff algorithms. IEEE Transactions on Vehicular Technology, 52(2):347356, March 2003. [13] K. Raivio, O. Simula, and J. Laiho. Neural analysis of mobile radio access network. In IEEE International 336 Conference on Data Mining, pages 457464, San Jose, California, USA, November 29 - December 2 2001. [14] M. Siponen, J. Vesanto, O. Simula, and P. Vasara. An approach to automated interpretation of SOM. In N. Allinson, H. Yin, L. Allinson, and J. Slack, editors, Advances in Self-Organizing Maps, pages 8994. Springer, 2001. [15] J. Vesanto and E. Alhoniemi. Clustering of the self-organizing map. IEEE Transactions on Neural Networks, 11(3):586600, May 2000. [16] J. Zander. Radio Resource Management for Wireless Networks. Artech House, Inc., 2001. 337

Two-Phase Clustering Algorithm;data mining;mobility management;Key Performance Indicator of Handover;soft handover;Data Mining;Soft Handover;Visualization Capability;Neural Network Algorithm;neural networks;hierarchical clustering;Self-Organizing Map;Cluster Analysis;Histograms;3G network;Decrease in Computational Complexity

37

Aspect Oriented Programming for a component-based real life application: A case study

Aspect Oriented Programming, a relatively new programming paradigm, earned the scientific community's attention . The paradigm is already evaluated for traditional OOP and component-based software development with remarkable results. However, most of the published work, while of excellent quality, is mostly theoretical or involves evaluation of AOP for research oriented and experimental software . Unlike the previous work, this study considers the AOP paradigm for solving real-life problems, which can be faced in any commercial software. We evaluate AOP in the development of a high-performance component-based web-crawling system, and compare the process with the development of the same system without AOP. The results of the case study mostly favor the aspect oriented paradigm.

INTRODUCTION Aspect Oriented Programming, a relatively new programming paradigm introduced by Kiczales ([2]), recently earned the scientific community's attention. Having around six years of life, the paradigm was already presented in important conferences, and recently triggered the creation of several conferences and workshops to deal with it. The paradigm is already evaluated for traditional OOP and component-based software development and is found very promising. Several evaluations consider it to be the continuation of the OOP paradigm. However, most of the published work while of excellent quality is mostly theoretical or involves evaluation of AOP for research oriented and experimental software. Unlike previous works, this study considers the AOP paradigm for solving real-life problems, which need to be faced in any commercial software. We evaluated Aspect Oriented Programming in the development of 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. SAC'04 March 14-17, 2004, Nicosia, Cyprus Copyright 2004 ACM 1-58113-812-1/03/04 ... $ 5.00. a high-performance component-based web-crawling system, and compared the process with the development of the same system without AOP. The results of the case study, mostly favoring the aspect oriented paradigm, are reported in this work. This introduction is followed by an introduction to the AOP approach. We then describe the application that was used for our evaluation and proceed with a description of our evaluation scenario. We then present and comment our evaluation results. We continue with references to similar evaluation attempts, and, finally, we summarize the conclusions from our evaluation, and report on future work. ASPECT ORIENTED PROGRAMMING Aspect Oriented Programming, as proposed by Kiczales ([2]), is based on the aspectual decomposition. Aspectual decomposition is somewhat complementary to functional decomposition , and tries to overcome the limitation of functional decomposition to capture and represent crosscutting functionality. After separating the system to functional constructs , aspectual decomposition is applied to the design in order to catch the crosscutting concerns. Crosscutting functionality usually includes extra-functional requirements (e.g. timing constraints or logging facility to all the system components ). This functionality is usually replicated a number of times, spread over the whole system. There is no single point of reference where the developer can say that the aspectual functionality belongs and should be implemented. The main purpose of AOP is to capture the crosscutting concerns throughout the system, and promote them as first-class citizens, in order to enable the modeling and reusing of them. The high level goals of such an approach, as reported in various publications, follow: 1. AOP makes programming easier and faster, closer to the human perception ([2, 3, 7]). Developers understand the concept of crosscutting concerns and crosscutting functionality, and they use it in understanding the whole system. However, apart from AOP, there is no easy way to implement such a crosscutting concern . With AOP, aspects are closer to the human perception for crosscutting concerns and simplify the design and implementation of systems with such requirements . Aspects can even allow code reuse for the extra-functional requirements they implement, which usually crosscut the whole system. Thus, they make system implementation easier and faster. 1554 2004 ACM Symposium on Applied Computing 2. AOP makes programming less error-prone and easier to debug and maintain ([2, 3, 7, 6]). Not only the code becomes more modular, thus, easier to maintain and enhance, but also the goal for debugging is more easily gained (offered from the AOP inherent ability of automatic aspect invocation). Furthermore, AOP favors reusability and modular representation of crosscutting concerns, which make the code more readable and prevent tangling code. The AOP approach is already used in the implementation of several academic-oriented systems such as [4], but there is not much work reported on AOP relating with commercial environment. However, we strongly believe that AOP can enter the industrial environment, and that it has much to offer. We expect to witness that in the near future. THE HIGH PERFORMANCE COMPONENT-BASED WEB CRAWLER To evaluate the AOP paradigm, we chose a high performance component-based web crawler, which would serve the needs of our laboratory. However, it was important for us to make the crawler easily extensible and changeable in order to be able to reuse it in different projects. Furthermore, the crawler should not be characterized as experimental (e.g. unstable or with extremely complicated configuration) since it should be reusable in a number of different projects, and without needing to know the complete infrastructure. We also needed the crawler to be easily adjustable to different configurations, hardware, and network situations, because of the variety of our hardware, as this would be desired in a real-life application. This application was found suitable for our AOP evaluation , since it was of respectable size, which would give us the opportunity for better results. Furthermore, the non-experimental characterization of the current application , which is rarely the outcome in the academic environment , would ensure a more practical approach of our evaluation . For the same reason, the extra-functional requirements implemented for the evaluation, were carefully selected. It was important for us to keep the whole implementation and, consequently, the AOP evaluation not far from the commercial field, which we feel to be the important end-user of the programming paradigms. Having these points in mind, we decided to use the following design, comprising three basic multi-threaded components : (i) the database component, (ii) the crawling component , and (iii) the processing component. Figure 1: The architecture of a high-performance component-based web-crawling system. The database component was responsible for two tasks: (a) updating the database with the processed information, received from the processing component, and (b) feeding the crawling component with the necessary URLs to be crawled. Furthermore, as in all the components, a number of threads were running in parallel in each component, so that the fast devices like CPU and memory (as opposed to the usually slow devices like I/O and network) would be more efficiently utilized. The number of threads running in parallel in each component could be selected from the user, and also adjusted dynamically from each component for optimal performance . Selecting a very small number of threads, the user would let fast resources like processor and the memory rather unutilized, while selecting an overly large number of threads would result to large context switching overhead. The crawling component's responsibility was to download the URLs from the web and provide the processing component with the page information for further processing. Page information included the page's URL, IP address, and page text. Again, the crawling component ran a number of threads to maximize resource utilization. Finally, the processing component was responsible for receiving the page information from the crawling component and processing it, and passing the results to the database component for permanent storage. As in the other components , this component was also multi-threaded, thus utilizing the resources better. EVALUATION SCENARIO To evaluate AOP in the crawling project, we ran the following scenario: First, we set our metrics for the evaluation of AOP, trying to keep them as objective as possible; then, we designed the component-based web crawler and located the different functionalities that could be modeled as aspects . Following that, we implemented and tested the three components independently. The implementation up to that point did not include any of the functionalities identified as aspects in the earlier step. Finally, we tried to integrate the three components, and also include the extra functionality, implemented with and without AOP. Our selection for the metrics was mostly to favor (as much as possible) objective results. Our goal, as Murphy in [5] suggests, was to answer two important questions: (a) if AOP makes it easier to develop and change a certain category of software (usefulness), and (b) what is the effect of AOP in the software development (usability). For these reasons, we selected the following metrics: 1. We measure effectiveness of AOP for implementing the extra functionality, compared to traditional OOP. 2. We measure the learning curve of AOP methodology. 3. We measure time that took to complete the project with the two approaches, AOP and traditional OOP. 4. We measure complete lines of code for the added functionality with AOP and with traditional OOP. 5. We compare code tangling in the AOP and the traditional OOP model. 6. We report on the stability of the AOP model for creating component-based software. The types of functionality that we identified as being best modeled as aspects were the following ones: 1555 Logging : This functionality requires saving extended program execution trace to a file or printing it to the screen. The trace should include entrance and exit messages from the methods, exceptions thrown, and time of each event. Overloading checks : Since the crawling function is expensive in resources, we must constantly check for overloading in any of the resources, in order to avoid driving the machines to collapse. The two resources we had to monitor were the DNS server that was serving our crawler and the machine that was hosting our crawling database. Database optimizer : Even with the combination of the expensive high performance hardware and software that was used for the database server, we still needed to follow some optimization techniques to minimize the need for database connectivity. This was due to the heavy load that our database server experienced from the crawling function. The Logging aspect, the most common aspect in AOP, was mostly to help debugging during the developing stage of the application, but it would also be used for identifying bottlenecks (profiling) and performing optimizations to the components in a later stage. When the logging aspect was enabled, entering or exiting a method would print (to stderr) the method's name, the exact time, and some other useful information. Moreover, a method throwing an exception would result in invoking the logging aspect to print the exception with the method's name in stderr. The overloading checks were broken in two aspects, the DNS monitoring aspect and the database monitoring aspect . The DNS monitoring aspect was trying to adjust the number of active downloading threads according to the DNS server status. More to the point, the problem we faced was that the DNS server that was serving our crawler was shared with other machines, some of them running experimental software, doing extensive use of the DNS server for DNS resolution. This practically meant that the efficiency of the DNS server was dependent of the number of software clients using it in parallel. Running more than the appropriate (for each moment) downloading threads in our crawler (that were doing the DNS resolution) resulted in more DNS resolution requests that our DNS server could handle, and eventually, collapsing of our DNS server. On the other hand, underestimating our DNS server's abilities in low-usage hours would result in significantly lower crawling speed. For these reasons we constructed and used the DNS monitoring aspect, which would adjust the number of the downloading threads according to the running DNS load. Each DNS resolution was timed, and when discovering latency higher than expected , we were temporarily pausing some of the downloading threads (the pause time and the number of the threads that we were pausing were analogous to the latency), thus, causing less DNS lookups in a specific time. The database monitoring aspect's goal was to disable overloading in the database machine. A similar approach to the DNS monitoring aspect was used. We were monitoring the responses from our database server and when we were detecting overloading of the database we would pause some of the downloading threads. The reason that we could not predict the ideal number for the database component threads from the beginning was because of the variety of the web-pages. For example, a web-page with many new words (words that are for first time parsed from the crawler) would result in much database load, while words that are seen before from the crawler would result in much less (due to some optimizations, similar to those proposed in [1]). For these reasons, we constructed the database monitoring aspect to monitor database queries. The aspect would time every interaction with the database server and try to detect overloading. When the time demanded for the query was bigger than a threshold (all the queries we were executing were having the same average time for execution in normal circumstances), we would pause some of the downloading threads for some time, in order to allow the database server to complete its work without extra work added at the same time. Later on, the downloading threads would resume their work. These two last aspects would not contradict each other, since they were both doing the same action, pausing some of the downloading threads. However, the pause time and the number of the downloading threads to pause were not the same in the two cases. Each of the aspects was calculating the time and the number of threads to pause with a different algorithm. Finally, we also constructed the database optimizer aspect which acted as a database cache and released some of the database load. More specifically, for the parsing function we were making heavy usage of the crawling dictionary table from the database. That dictionary was matching every word we found up to the moment with its id number. The choice was to avoid needless and costly replication of data and enable saving the page text as numbers (smaller in storing size and faster in seeking). By keeping a memory cache of that table as in Brin's implementation ([1]), we would manage to get important workload off the database server and speed things up. More to the point, prior addressing the database for a word's serial number, we were querying an indexed structure in the local memory. If the query failed, we were then inserting the word in the database and in the RAM dictionary and continuing our work. This minimized the database interactions and boosted the complete process, since RAM access was enormously faster than access to the database. Processing English language pages with an average-size dictionary of 1 million words would result to around 99,9% success from the RAM table, thus, it would prevent querying the database very efficiently. EVALUATION RESULTS As already mentioned, these four aspects were implemented in two distinct ways: (a) injected in the program code, using standard OOP approach, and (b) modeled and implemented as aspects. The two versions were then compared and evaluated in the described metrics. The results from the evaluation were mostly in favor of the AOP methodology. While the developers were not long experienced in AOP, the new model boosted the implementation speed and helped in more modular software. Regarding effectiveness of the AOP approach compared to the traditional OOP approach, the two approaches were the same. We managed to add the extra functionality in both versions of the software (however, it was not always trivial to do so). In short, for the presented aspects there was always an AOP-oriented and an OOP-oriented solution 1556 available, and there was not a noticeable performance difference between the two. Regarding the time demanded to learn the AOP methodology , this was not significant. Both the developers that were working on the project were very experienced with OOP, but did not have previous practical experience with AOP. Fortunately for the project, they were able to learn AOP sufficiently without tutoring using only publicly available online sources in a single week. There was also another short overhead of one day for installing and getting familiar to an AOP-aware IDE (we used Eclipse with the AOP modules). The complete time that was required to finish the crawler was shorter in the AOP version (this time did not include the time spent for learning AOP however). Both the versions used the same core already developed (the three components demonstrated earlier) but they were continued com-pletely independently, without reusing knowledge or code from one version to the other (the nature of the two versions prohibited reusing knowledge or code anyway). The time demanded for completing the crawler with the aspects in the AOP version was 7 man-hours, while the OOP version demanded 10 man-hours in order to design and develop the code. Most of this time, in the case of the OOP version, was needed for locating the methods and putting the necessary code to them. For implementing the logging functionality for instance, in the OOP version there were 73 such methods counted, while AOP did not demand this task since the pointcuts were found automatically from the aspect definition. It was the developers' feeling that most of the man-hours spent in the OOP version of the crawler were wasted, because they were repeating trivial code in the application . Furthermore, as they said, the result in the OOP case was not satisfactory for them since, if they needed to change something in an aspect, they should relocate the aspect code from the beginning and this would be difficult to be done. We also measured the number of lines we needed to add in the two approaches to implement the extra functionality. For the logging aspect with the AOP approach, we needed less than 20 lines in one single file, while the same functionality for the traditional OOP version required 126 lines of code spread in eight different files (the number of lines for the AOP code also include the pointcuts definitions and the java include directives). This, apart from a time-demanding approach, also reveals important code tangling since we had to modify eight classes for a simple logging requirement. The other two aspects, the DNS monitoring and the database monitoring aspect, needed roughly the same number of lines for the two versions. To implement both the DNS monitoring and the database monitoring functionality, we needed around 30 lines for the pure OOP solution: (a) four lines for timing the DNS or the database query, (b) ten lines for checking for overloading, and proceeding in alternate behavior if overloading occurs, and finally (c) one line for invoking the check wherever needed. In the AOP solution, we were able to join the two concerns in a single aspect - something that we were unable to do in the OOP version - and reuse some of the code. The AOP version of the solution demanded roughly the same number of lines, around forty for both the concerns (the additional code was because of aspect and advice headers and the pointcuts definition). Finally, the database optimizer needed the same number of lines in the two versions, that is forty lines. These lines in the OOP version were split in three different places in the original database component file, while at the AOP approach the original file was kept intact and all the new code was in a single aspect-definition file. We also tried to capture the code tangling that occurs in the two versions, after the extra functionality is added. To do that, we found the distribution of the added code in the eight affected files. The OOP version of the logging aspect, as expected, was spread in all the eight files in seventy-three different places. The OOP version of the DNS monitoring aspect resulted in addition of code in one file only, the down-loader component file, in three different places. Similarly, the OOP version of the database monitoring aspect resulted in addition of code in the database component file, in three different places. Finally, the database optimizer aspect implementation , without AOP, also resulted in code addition in three different places (again, in the database component file). On the other hand, implementation of the four aspects with the AOP approach, as expected, created no code tangling . The complete code for the extra functionalities was included in the three (instead of four, since the DNS and the database monitoring concerns were implemented as a single aspect) aspect files. For the case of the database optimizer, this offered us another important advantage since we often needed to disable the database optimizer due to hardware (memory) limitations in weaker machines. While we did not take any provision for that, the AOP version, unlike the OOP version, enabled removal of the optimizer without changing any code. In the OOP version, the developer had to remove or modify some of the original code. Table 1 summarizes the results for the code size and code tangling for the four aspects: Finally, the implementation of AOP we used, combined with the IDE tool, were stable and did not cause us any unexpected problems (such as bugs in the compiler). While the crawling system was not extremely big, it did make extensive use of the machines' resources, and AspectJ compiled files did not face any trouble with that. AspectJ compiled files proved to work fine under pressure with the standard virtual machine, and aspects introduction was not causing a noticeable overhead to the machines. RELATED WORK Several publications try to evaluate AOP. Almost all of them report results similar to ours. However, while of excellent quality, most of the previous work we are aware of follow a theoretical approach or limit their hands-on evaluation for academic or experimental software. We will now briefly comment on some of them. Walker in [8] constructs several experiments and a case-study to evaluate AOP. The outcome of the evaluation is that AOP can help faster software development (programming , debugging, etc.) under certain conditions, while other cases make development of AOP less attractive. While of superb quality and significant importance, this work is limited to the evaluation of AOP based on a preliminary version of AspectJ, version 0.1. Since then, AOP and especially AspectJ, changed significantly confronting most of the limitations detected in the evaluation, and also powering the users with more functionalities. Furthermore, CASE tools and powerful IDE environments were developed to assist the 1557 Aspect # Lines of code # Places to add # Files to add OOP AOP OOP AOP OOP AOP Logging 126 19 73 1 8 1 DNS Monitoring 15 40 3 1 1 1 Database Monitoring 15 3 1 Database optimizer 45 45 3 1 1 1 Table 1: Size of added code and code tangling for implementation of the aspects. The two aspects, DNS and Database monitoring were easily joined to a single aspect in the AOP version developers in the process. Mendhekar in [4] also presents a case study, evaluating AOP in an image processing application. Although AOP was then still in infancy, this case study presents results very similar to ours. However, Mendhekar, being in Xerox labs where AOP was born, follows a more research-oriented approach during the evaluation. The evaluation uses an AOP implementation that cannot easily be used from people outside the Xerox environment. Also, being interested in performance , this work does not elaborate on various other important measures, such as the learning curve and the time that took the developer to complete. Several other important publications ([2, 3, 7, 6]) evaluate AOP from mostly a theoretical approach. Most of them also report results that favor AOP programming. Some of their results are reported in section 3 of this report. CONCLUSIONS During the construction of the component-based high-performance web crawler, we had the opportunity to evaluate the relatively new aspect oriented paradigm for building component-based systems. Having defined our extra functionality , we implemented and compared the two versions of the web crawler, the AOP and the OOP one. For the required extra functionality, both the paradigms proved able to implement a correct solution. The quantity of code (number of lines) that the developer needed to implement in the two versions was not of much difference, with the only exception of the logging aspect where the OOP implementation was much larger than the AOP one. Furthermore, in both the versions of the application there was no apparent performance difference. Both the versions were stable, even when working under high load and in varying system environments . The significant difference however between the two implementations was in the time required to develop and debug each of them, and the quality of the produced code. The AOP approach not only completed the system faster, but it also produced modular high quality code, while the traditional approach was creating the well-known spaghetti code. More specifically, the AOP version was having all the extra functionality apart of the code implementing the standard functional requirements. This not only kept the original components reusable in different implementations, but also prevented tangling the code, thus, making future maintenance easier. Furthermore, this enabled us to easily enable and disable the extra functionality, depending on the hardware resources available and on our requirements. Concluding, we have to report that the AOP model in general appears to favor the development of quality component-based software. The AOP model itself is able to boost the implementation speed without negatively affecting quality of the software. Moreover, the learning time of the model, judging from our experience, is not long. While not having much experience of AOP implementation languages, we were able to produce AOP-based code in no time. Finally, while AOP cannot offer any solution to problems unsolv-able from traditional approaches, and while AOP does not always target to less code, it can offer better and easier solutions to programs that are otherwise difficult to be implemented . Therefore, we can safely arrive to the conclusion that AOP has much to offer in component-based software development. We strongly believe that integration of AOP with component-based software is going to be the target of important research attempts in the near future and can produce some very interesting results, and we await for the introduction of AOP software in commercial component-based software products. REFERENCES [1] S. Brin and L. Page. The anatomy of a large-scale hypertextual Web search engine. Computer Networks and ISDN Systems, 30(17):107117, 1998. [2] G. Kiczales, J. Lamping, A. Menhdhekar, C. Maeda, C. Lopes, J.-M. Loingtier, and J. Irwin. Aspect-oriented programming. In Proceedings of the European Conference on Object-Oriented Programming (ECOOP), LNCS 1241, pages 220242, Springer-Verlag, 1997. [3] C. Lopes. D: A Language Framework for Distributed Programming. PhD thesis, College of Computer Science, Northeastern University, November 1997. [4] A. Mendhekar, G. Kiczales, and J. Lamping. RG: A case-study for aspect-oriented programming. Technical Report SPL97-009 P9710044, Xerox Palo Alto Research Center, Palo Alto, CA, USA, February 1997. [5] G. C. Murphy, R. J. Walker, and E. L. Baniassad. Evaluating emerging software development technologies: Lessons learned from assessing aspect-oriented programming. Technical Report TR-98-10, Department of Computer Science, University of British Columbia, 1998. [6] A. Navasa, M. A. Perez, J. Murillo, and J. Hernandez. Aspect oriented software architecture: a structural perspective. In Proceedings of the Aspect-Oriented Software Development, 2002, The Netherlands. [7] D. Shukla, S. Fell, and C. Sells. Aspect-oriented programming enables better code encapsulation and reuse. MSDN Magazine, http://msdn.microsoft.com/msdnmag/, March 2002. [8] R. J. Walker, E. L. A. Baniassad, and G. C. Murphy. An initial assessment of aspect-oriented programming. Technical Report TR-98-12, Department of Computer Science, University of British Columbia, Sept. 1998. 1558

case study;evaluation;Web Crawler Implementation;AOP;component-based application;Aspect Oriented Programming;development process experiment metrics;Software development process experiment;programming paradigm comparison;OOP;Object Oriented Programming;programming paradigms;Aspect Oriented Programming application

38

Attack-Resilient Hierarchical Data Aggregation in Sensor Networks

In a large sensor network, in-network data aggregation, i.e., combining partial results at intermediate nodes during message routing, significantly reduces the amount of communication and hence the energy consumed. Recently several researchers have proposed robust aggregation frameworks, which combine multi-path routing schemes with duplicate-insensitive algorithms, to accurately compute aggregates (e.g., Sum, Count, Average) in spite of message losses resulting from node and transmission failures. However, these aggregation frameworks have been designed without security in mind. Given the lack of hardware support for tamper-resistance and the unattended nature of sensor nodes, sensor networks are highly vulnerable to node compromises. We show that even if a few compromised nodes contribute false sub-aggregate values, this results in large errors in the aggregate computed at the root of the hierarchy. We present modifications to the aggregation algorithms that guard against such attacks, i.e., we present algorithms for resilient hierarchical data aggregation despite the presence of compromised nodes in the aggregation hierarchy. We evaluate the performance and costs of our approach via both analysis and simulation . Our results show that our approach is scalable and efficient.

INTRODUCTION In large sensor networks, computing aggregates in-network, i.e., combining partial results at intermediate nodes during message routing , significantly reduces the amount of communication and hence the energy consumed [11, 23]. An approach used by several data acquisition systems for sensor networks is to construct a spanning tree rooted at the querying node, and then perform in-network aggregation along the tree. Partial results propagate level-by-level up the tree, with each node awaiting messages from all its children before sending a new partial result to its parent. Tree-based aggregation approaches, however, are not resilient to communication losses resulting from node and transmission failures , which are relatively common in sensor networks [11, 22, 23]. Because each communication failure loses an entire subtree of readings, a large fraction of sensor readings are potentially un-accounted for at the querying node, leading to a significant error in the query answer. To address this problem, researchers have proposed the use of multi-path routing techniques for forwarding sub-aggregates [11]. For aggregates such as Min and Max which are duplicate-insensitive, this approach provides a fault-tolerant solution . For duplicate-sensitive aggregates such as Count and Sum, however, multi-path routing leads to double-counting of sensor readings , resulting in an incorrect aggregate being computed. Recently researchers [3, 12, 14] have presented clever algorithms to solve the double-counting problem associated with multi-path approaches. A robust and scalable aggregation framework called Synopsis Diffusion has been proposed for computing duplicate- sensitive aggregates such as Count and Sum. There are two primary elements of this approach - the use of a ring-based topology instead of a tree-based topology for organizing the nodes in the aggregation hierarchy, and the use of duplicate-insensitive algorithms for computing aggregates based on Flajolet and Martin's algorithm for counting distinct elements in a multi-set [5]. As presented, the Synopsis Diffusion aggregation framework does not include any provisions for security. Although we can easily prevent unauthorized nodes from launching attacks by augmenting the aggregation framework with authentication and encryption protocols [15, 24], compromised nodes present an entirely new set of security challenges. The lack of tamper-resistance and the unattended nature of many networks renders sensor nodes highly vulnerable to compromise. Standard authentication mechanisms cannot prevent a compromised node from launching attacks since all its keys are also compromised. In this paper, we present novel mechanisms for making the synopsis diffusion aggregation framework resilient to attacks launched by compromised nodes. We present counter-measures against attacks in which a compromised node attempts to change the aggregate value computed at the root of the hierarchy. In particular, we focus on an attack in which 71 a sensor node that is not a leaf node in the aggregation hierarchy relays a false sub-aggregate value to its parents. We refer to this attack as the falsified sub-aggregate attack. We show that if the synopsis diffusion approach is used to compute aggregates such as Count and Sum, an adversary can use the falsified sub-aggregate attack to cause the answer computed at the base station in response to a query to differ from the true value by an arbitrary amount. Moreover, we show that this attack can be launched with a high rate of success, even if only one or a small number of nodes are compromised. We present an approach in which the synopsis diffusion aggregation frameork is augmented with a set of countermeasures that mitigate the effect of the falsified sub-aggregate attack. In our approach , a subset of the total number of nodes in the network include an authentication code (MAC) along with their response to a query. These MACs are propagated to the base station along with the partial results that are computed at each level in the hierarchy. By verifying these MACs, the base station can estimate the accuracy of the final aggregate value it computes, and can filter out the effect of any false sub-aggregates contributed by compromised nodes. Thus, our approach can be used in conjunction with synopsis diffusion to compute basic aggregates such as Count and Sum despite the presence of compromised nodes in the aggregation hierarchy. The communication overhead of our approach depends upon the number of contributing nodes which send a MAC to the base station . We evaluate the performance and costs of our approach via both analysis and simulation. We show that our approach is scalable since the number of contributing nodes (and hence the average communication overhead) do not increase with network size. To further reduce the communication overhead, we describe a variation of our basic approach that trades communication costs for latency. BACKGROUND SYNOPSIS DIFFUSION FOR ROBUST AGGREGATION In this section, we provide a brief overview of the synopsis diffusion approach for robust aggregation [3, 14]. Figure 1 illustrates how the synopsis diffusion approach uses a rings topology for aggregation . R0 R1 R2 q C B A D Figure 1: Synopsis Diffusion over a rings topology In the query distribution phase, nodes form a set of rings around the querying node q based on their distance in hops from q. During the subsequent query aggregation period, starting in the outermost ring each node generates a local synopsis s = SG(v) where v is the sensor reading relevant to the query, and broadcasts it. (SG () is the synopsis generation function.) A node in ring R i will receive broadcasts from all the nodes in its range in ring R i +1 . It will then combine its own local synopsis with the synopses received from its children using a synopsis fusion function SF (), and then broadcast the updated synopsis. Thus, the fused synopses propagate level-by -level until they reach the querying node, who first combines the received synopses with its local synopsis using SF () and then uses the synopsis evaluation function SE () to translate the final synopsis to the answer to the query. The functions SG (), SF(), and SE() depend upon the target aggregation function, e.g. Count, Sum, etc. We now describe the duplicate-insensitive synopsis diffusion algorithms for the Count aggregate, i.e., the total number of nodes in the sensor network, and the Sum aggregate, i.e., the sum of the sensor readings of the nodes in the network. These algorithms are based on Flajolet and Martin's well-known probablistic algorithm for counting the number of distinct elements in a multi-set[5]. 2.1 COUNT In this algorithm, each node generates a local synopsis which is a bit vector ls of length k &gt; log n, where n is an upper bound on the nodes in the network. To generate its local synopsis, each node executes the function CT (X, k) given below, where X is the node's identifier and k is the length of ls in bits. CT () can be interpreted as a coin-tossing experiment (with a cryptographic hash function h (), modeled as a random oracle whose output is 0 or 1, simulating a fair coin-toss), which returns the number of coin tosses until the first heads occurs or k + 1 if k tosses have occurred with no heads occurring. In the local synopsis ls of node X , a single bit i is set to 1, where i is the output of CT (X, k). Thus ls is a bitmap of the form 0 i -1 1 with probability 2 -i . Algorithm 1 CT (X, k) i=1; while i &lt; k + 1 AND h(X, i) = 0 do i = i + 1; end while return i; The synopsis fusion function SF () is simply the bitwise Boolean OR of the synopses being combined. Each node fuses its local synopsis ls with the synopses it receives from its children by computing the bit-wise OR of all the synopses. Let S denote the final synopsis computed by the querying node by combining all the synopses received from its children and its local synopsis. We observe that S will be a bitmap of length k of the form 1 r -1 0 . The querying node can estimate Count from S via the synopsis evaluation function SE (): if r is the lowest-order bit in S that is 0, the count of nodes in the network is 2 r -1 /0.7735. The synopsis evaluation function SE () is based on Property 2 below. Intuitively, the number of sensor nodes is proportional to 2 r -1 since no node has set the rth bit while computing CT (X, k). We now present a few important properties of the final synopsis S computed at the querying node that have been derived in [5, 3], and that we will find useful in the rest of this paper. Let S [i], 1 i k denote the ith bit of S, where bits are numbered starting at the left. Property 1 For i &lt; log 2 n -2log 2 log 2 n , S [i] = 1 with probability 1. For i 3 2 log 2 n , S [i] = 0 with probability 1. This result implies that for a network of n nodes, we expect that S has an initial prefix of all ones and a suffix of all zeros, while only the bits around S [log 2 n ] exhibit much variation. This provides an estimate of the number of bits, k, required for a node's local synopsis. In practice, k = log 2 n + 4 bits are sufficient to represent S with high probability [5]. This result also indicates that the length of the prefix of all ones in S can be used to estimate n. Let r = 72 min {i|S[i] = 0}, i.e., r is the location of the leftmost zero in S. Then R = r -1 is a random variable representing the length of the prefix of all ones in the sketch. The following results hold for R. Property 2 The expected value of R, E (R) log 2 (n) where the constant is approximately 0.7735. This result implies that R can be used for an unbiased estimator of log 2 (n), and it is the basis for the synopsis evaluation function SE () which estimates n as 2 R /. Property 3 The variance of R, denoted as 2 R n , satisfies 2 R n = 2 R + Q(log 2 n ) + o(1), where constant R is approximately 1 .1213 and Q(x) is a periodic function with mean value 0 and period 1. This property implies that the standard deviation of R is approximately 1 .1213, i.e., the estimates of n derived from R will often be off by a factor of two or more in either direction. To reduce the standard deviation of R, Flajolet et al [5] proposed an algorithm named PCSA, where m synopses are computed in parallel and the new estimator ( R ) is the average of all individual R's of these synopses . For PCSA, the standard error in the estimate of n, i.e., n /n, is equal to 0 .78/m [5]. Property 4 In a network of n nodes, the expected number of nodes that will have the ith bit of their local synopsis ls [i] = 1 is n/2 i . This result implies that the expected number of nodes that contribute a 1 to the ith bit of S and the bits to the right of the ith bit in S (i.e., bits j , where i j k) is n/2 i -1 . 2.2 SUM Considine et al. [3] extended the Count algorithm described above for computing the Sum aggregate. The synopsis generation function SG () for Sum is a modification of that for Count while the fusion function SF () and the evaluation function SE() for Sum are identical to those for Count. To generate its local synopsis for a sensor reading v, a node X invokes the function CT () v times 1 and ORs the results. As a result, the local synopsis of a node is a bitmap of length k = log 2 u s + 4 where u s is an upper bound on the value of Sum aggregate. Unlike the local synopsis of a node for Count, more than one bit in the local synopsis of a node for Sum will be equal to 1. Count can be considered as a special case of Sum where each node's sensor reading is equal to one unit. Considine et al. [3] proposed an optimized version of SG () for Sum to make it suitable for a low-end sensor node, even if the sensed value v is high. Moreover, they showed that Properties 14 described above for Count also hold for Sum (with appropriate modifications). Similarly, as in the case of Count, the PCSA algorithm can be used to reduce the standard deviation of the estimate for Sum. ATTACKS ON SYNOPSIS DIFFUSION The Synopsis Diffusion aggregation framework does not include any provisions for security; as a result, it is vulnerable to many attacks that can be launched by unauthorized or compromised nodes. To prevent unauthorized nodes from eavesdropping on or participating in communications between legitimate nodes, we can augment the aggregation framework with any one of several recently proposed authentication and encryption protocols [15, 24]. However , compromised nodes pose an entirely new set of security challenges . Sensor nodes are often deployed in unattended environments, so they are vulnerable to physical tampering. Current sensor nodes 1 Each sensor reading is assumed to be an integer lack hardware support for tamper-resistance. Consequently, it is relatively easy for an adversary to compromise a node without being detected. The adversary can obtain confidential information (e.g., cryptographic keys) from the compromised sensor and repro-gram it with malicious code. A compromised node can be used to launch multiple attacks against the sensor application. These attacks include jamming at physical or link layer, other denial of service attacks like flooding, route disruption, message dropping, message modification, false data injection and many others. Standard authentication mechanisms cannot prevent these insider attacks since the adversary knows all the keying material possessed by the compromised nodes. In this paper, we focus on defending against an important subclass of these insider attacks which can potentially corrupt the final result of the aggregation query. Below we describe these attacks in the context of the Count and Sum aggregates. A compromised node M can corrupt the aggregate value computed at the root (i.e., the sink) of the hierarchical aggregation framework in three ways. First, M can simply drop aggregation messages that it is supposed to relay towards the sink. If M is located at a relatively high position in the aggregation hierarchy, this has the effect of omitting a large fraction of the set of sensor readings being aggregated. Second, M can falsify its own sensor reading with the goal of influencing the aggregate value. Third, M can falsify the sub-aggregate which M is supposed to compute based on the messages received from M's child nodes. The effect of the first attack in which a node intentionally drops aggregation messages is no different from the effect of transmission and node failures, which are common in sensor networks [7]. The synopsis diffusion approach employs multi-path routing for addressing these failures, and thus it also addresses message losses due to compromised nodes [3, 12, 14]. We refer to the second attack in which a sensor intentionally falsifies its own reading as the falsified local value attack . This attack is similar to the behavior of nodes with faulty sensors and can be addressed by well-studied approaches for fault tolerance such as majority voting and reputation-based frameworks [10, 6]. The third attack, however, in which a node falsifies the aggregate value it is relaying to its parents in the hierarchy is much more difficult to address, and is the main focus of this paper. We refer to this attack as the falsified sub-aggregate attack . The Falsified Sub-Aggregate Attack Since the sink estimates the aggregate based on the lowest-order bit r that is 0 in the final fused synopsis, a compromised node would need to falsify its own fused synopsis such that it would affect the value of r. It can accomplish this quite easily by simply inserting ones in one or more bits in positions j, where r j k, in its own fused synopsis which it broadcasts to its parents. Note that the compromised node does not need to know the true value of r; it can simply set some higher-order bits to 1 in the hope that this will affect the value of r computed by the sink. Since the synopsis fusion function is a bitwise Boolean OR, the resulting synopsis computed at the sink will reflect the contributions of the compromised node. Let r be the lowest-order bit that is 0 in the corrupted synopsis, whereas r is the lowest-order bit that is 0 in the correct synopsis. Then the sink's estimate of the aggregate will be larger than the correct estimate by a factor of 2 r -r . It is easy to see that, with the above technique, the compromised node can inject a large amount of error in the final estimate of the sink. We also observe that even a single node can launch this attack with a high rate of success because the use of multi-path routing in the synopsis diffusion approach makes it highly likely that the falsified synopsis will be propagated to the base station. If p is the 73 packet loss rate and if each node has parents in the aggregation hierarchy then the probability of success for this attack is (1 - p ) h , where the compromised node is h hops away from the sink. As an example, if p = 0.2, = 3, and h = 5 then the probability that the attack will succeed is 96%. On the other hand, it is very hard to launch an attack which results in the aggregate estimated at the sink being lower than the true estimate. This is because setting a bit in the falsified synopsis to 0 has no effect if there is another node X that contributes a 1 to the same position in the fused synopsis. To make this attack a success the attacker has to compromise all the possible paths from node X to the sink so that X 's 1 cannot reach the sink, which is hard to achieve. If there is more than one node which contributes to the same bit then it is even harder. As an example, in Count algorithm, half of the nodes are likely to contribute to the leftmost bit of the synopsis, one-fourth nodes of contribute to the second bit, and so on. There are bits in the synopsis to which only one or two nodes contribute but it is very hard to predict in advance which nodes will be contributing to these particular bits if the sink broadcasts along the query request a random seed to be used with the hash function in the synopsis generation phase. Hence, we can safely assume that this attack is extremely difficult to launch. In the rest of this paper, we restrict our discussion to the previous attack where the goal of the attacker is only to increase the estimate. PROBLEM DESCRIPTION & ASSUMPTIONS In a sensor network where some fraction of the nodes are potentially compromised, there are three sources that contribute to the error in the sink's estimate of the aggregate being computed: (i) error due to packet losses, (ii) error due to the approximation algorithm used, e.g., Flajolet and Martin's probabilistic algorithm [5], and (iii) error injected by compromised nodes. The first two types of error are already addressed by the synposis diffusion aggregation framework. Our paper is complementary to this previous work; our objective is to filter out the third type of error. In particular, we aim to make the synopsis diffusion approach resilient to the falsified local value attack and the falsified sub-aggregate attack , i.e., to enable the sink to get the "true" estimate of the aggregate being computed despite the presence of compromised nodes in the aggregation hierarchy. By "true" estimate we mean the estimate of the aggregate which the sink would compute if there were no compromised nodes. 4.2 Assumptions We now discuss our assumptions with respect to the sensor network and the adversary. System Assumptions We assume that the base station is located at the center of the sensor network, and nodes are deployed around the base station. However, our approach for attack-resilient aggregation does not depend upon this assumption. We assume that sensor nodes are similar to the current generation of sensor nodes, e.g., Mica2 motes [13], in their computational and communication capabilities and power resources, while the sink is a laptop class device supplied with long-lasting power. We assume that the sink has an estimate of the upper bound on the value of the Count aggregate. If the sink does not have any further knowledge, the upper bound of Count can be set to the total number of nodes deployed. We also assume that there exists an upper bound on the value of a sensor reading. The upper bound of Sum can be conservatively set to be equal to product of the upper bound of Count and the upper bound of a sensor reading. Previous works on the synopsis diffusion approach [3, 14] have made the same assumptions regarding the upper bounds for Count and Sum; these bounds provide an estimate of the length of the synopsis. Security Assumptions We assume that the sink cannot be compromised and it uses a protocol such as Tesla [15]) to authenticate its broadcast messages. We also assume that each node shares a pairwise key with the sink, which is used to authenticate the messages it sends to the sink. We assume that the adversary can compromise sensor nodes without being detected. If a node is compromised, all the information it holds will also be compromised. We use a Byzantine fault model, where the adversary can inject malicious messages into the network through the compromised nodes. We conservatively assume that all compromised nodes can collude, or are under the control of a single attacker. Notations The following notations are used in the description of our attack-resilient aggregation algorithms. BS refers to the base station, i.e., the sink. X is the identifier of a the sensor node whereas M represents a compromised node. K X is the pair-wise key X shares with the sink. m1|m2 denotes the concatenation of two message fields m1 and m2. MAC(K,m) is the message authentication code (MAC) of the message m generated using the key K. X Y : m denotes a one-hop delivery of message m from X to Y , while X : m denotes that X broadcasts message m to all of its one-hop neighbors, and X : m denotes that X broadcasts message m to all nodes in the network. ATTACK-RESILIENT AGGREGATION THE BASIC APPROACH In this section, we present an attack-resilient approach for computing the Count and Sum aggregates. In this approach we assume that the BS has an estimate of the lower bound and the upper bound of the aggregates. We will see that this approach is scalable only if the ratio of the upper bound to the lower bound is small. Despite this limitation, we discuss this approach in detail because it provides the background and motivation for our extended approach, which is discussed in Section 6. We first present the main idea underlying the basic approach and then present the detailed protocol for securing Count and Sum. 5.1 The Main Idea In our approach, nodes execute the synopsis diffusion aggregation algorithm as specified in [3, 14]. However, a subset of the nodes include along with their synopses a message authentication code (MAC) that can be used by the sink to verify the validity of their contribution to the aggregate function. The key observations behind the design of our approach are that In order to derive the correct estimate from the final synopsis (say S) computed at the sink, we need only to figure out the correct lowest order bit (say r) in S that is 0. The number of nodes contributing a 1 to bit j decreases exponentially as we move from the lowest order bit ( j = 1) to higher order bits of the synopsis. For example, in the case 74 of Count, on average, half the nodes in the network will contribute 2 to the leftmost bit of the synopsis, one-fourth of the nodes contribute to the second bit of the synposis, and so on. Thus, we expect that only a small number of nodes will contribute to the bits around the expected value of r. Each such node includes along with its response to an aggregation query a MAC computed using a pairwise key shared exclusively with sink. We demonstrate that these MACs enable the sink to filter out the contributions of the falsified sub-aggregates injected by the compromised nodes to the final aggregate. For our scheme to work, two issues need to be addressed. First, since the the value of r is not known before the execution of the query, we need to specify a criterion whereby a node can determine if it needs to include a MAC along with its synopsis. Second, this criterion should be designed so that the number of such nodes who include a MAC is minimized. In our basic approach, we assume that the BS has an estimate of the lower bound and the upper bound of Count which are denoted by l c and u c respectively. Based upon these bounds, the BS knows that bit r will lie between a and b, which are the bit positions in the synopsis S corresponding to l c and u c respectively, i.e., a = log 2 (l c ) and b = log 2 (u c ) (by Property 2 in Section 2). Thus, there is no need for the BS to verify the bits to the left of a; only nodes contributing to bits in the range a to b need to prove to the BS that their contribution to the synopsis S is valid. We refer to the collection of bits in the range a to b in synopsis S as the synopsis-edge as shown in Figure 2. It is easy to see that the length of the synopsis-edge is (log 2 ( u c l c )+1) bits. If we denote the number of nodes contributing to the synopsis-edge by , then, by Property 4 in Section 2, ( u c 2 a + . . . + u c 2 b ) 1 ( 2u c l c -1). The upper bound for Count (u c ) can be set to the total number of nodes deployed. The lower bound for Count (l c ) can be guessed depending on the the energy reserve of the sensor nodes and rate of energy expenditure. As an example, if 2000 nodes are deployed then u c = 2000 and l c = 1000 may be a safe estimate at the time of the Count query's execution. For this example, the length of the synopsis-edge is u c l c = 2 and the expected number of nodes contributing to synopsis-edge is less than 3.87. synopsis-edge corresponds to Lower Bound corresponds to Upper Bound Figure 2: Securing Count synopsis. To securely compute Count synopsis, the base station needs to verify only bits in the synopsis-edge. For the ease of presentation, we present the basic approach assuming that only one synopsis is computed. We can easily extend this approach to compute m synopses in parallel as in algorithm PCSA. 5.2 Securing Count To compute the Count aggregate securely, we extend the original Count algorithm discussed in Section 2 as follows. For the sake 2 For convenience, henceforth, we say that a node "contributes" to a position j in the synopsis S if bit j in its local synopsis is 1. of completeness, we first briefly describe the query dissemination phase, and then we present the aggregation procedure in detail. In the query dissemination phase, the BS broadcasts the name of the aggregation function, a random number (Seed) and the bit positions of the start and the end of the synopsis-edge, which are specified by a and b respectively. Each node will use the random number, Seed, as an input to the hash function in the synopsis generation procedure. In more concrete terms, a query packet that the BS broadcasts is as follows: BS : F agg , Seed, a, b, s,t, h where F agg is the name of the aggregation function (i.e. `Count'), s denotes the time when the aggregation phase will start, t represents the duration of one round i.e. t = T h , where h is the total number of hops and T is the duration of the aggregation phase (also called epoch ). Note that, as in the original Count algorithm discussed in Section 2, the epoch is sub-divided into a series of rounds, one for each hop, starting from the farthest hop. Tesla [15] can be used for authenticating the broadcast packet. In the aggregation phase, each node executes the synopsis generation function SG () and the synopsis fusion function SF() for Count as discussed in Section 2. In addition, each node checks whether it contributes to the synopsis-edge, and if so, it generates a MAC and forwards the MAC along with its fused synopsis. Specifically , if node X contributes to bit i in the synopsis-edge, it generates a MAC, M = MAC(K X , m) over the message m whose format is [X|i|Seed], where Seed is the random number which was dissem-inated in the query distribution phase. Each node X forwards to its parents its fused synopsis along with the set of MACs ( M ) it received from its child nodes and its own MAC if it generated one. The format of the message a node X forwards to its parents is as follows: X : S l | M , where S l is the fused synopsis computed by X . If the message does not fit into one packet, node X breaks it into several packets and forwards them. In Appendix A, we formally describe the algorithm (SecureCount) executed by each node in response to an aggregation query. After the BS receives the MACs, it checks their validity. In particular , for each message and MAC pair [m|MAC(K X , m)] where m is [X|i|Seed], the BS executes the synopsis generation function SG () of X and verifies whether node X really contributes to bit i in the synopsis-edge, and then checks whether the attached MAC is valid. If any of these tests fail, the corresponding MAC is discarded . After this verification step, the BS checks whether it has received at least one valid MAC for each bit in the synopsis-edge. The bits in the synopsis-edge for which the BS has not received a valid MAC are reset to 0. The bits at positions to the left of the synopsis-edge are set to 1. Finally, the BS computes the Count using the synopsis evaluation function SE (). Security Analysis The security of our approach follows from two facts: The sink can independently verify the output of SG() for a particular node X . This is because the output of SG () depends only upon the node id X , and the random seed included in the query message. Each bit that is set to 1 in the synopsis edge has an associated MAC that can be verified by the sink. This MAC is computed using a pairwise key that is known only to the contributing node and the sink, thus the MAC cannot be fabricated by an attacker (as long as it is reasonably long.) 75 Although a compromised node can falsely set some bits in its fused synopsis and forward false MACs corresponding to those bits, the sink will be able to discard any false MACs. This implies that the attacker cannot falsely increase the Count. On the other hand, the attacker may attempt to decrease the Count by dropping a genuine MAC (or by corrupting a genuine MAC) sent by a contributing node, but the genuine MAC is likely to reach BS via an alternate path. If BS receives at least one valid MAC for each 1 bit in the synopsis-edge, then BS obtains the true estimate of Count as discussed below. As discussed in Section 2, the synopsis diffusion approach uses a multi-path routing scheme to reduce the aggregation error due to packet losses resulting from node and link failures. The effect of packets being dropped by compromised nodes is simply to increase the overall packet loss rate, and this can be countered by an appropriate choice of , the number of parents of a node in the synopsis diffusion ring-based aggregation hierarchy. Specifically, if each node has more than parents, the total number of rings in the rings topology is h, and if the probability of a node being compromised is p then, on average, a contributing node's MAC will reach the BS with probability q, where q 1h h j =1 (1 - p ) j Here we have assumed that the contributing nodes are uniformly distributed over the rings in the hierarchy. As an example, if p = 0 .05, = 3, and h = 10 then q is greater than 0.999, i.e., the impact of the compromised nodes on the communication error is negligible . We also note that while deriving q we assumed that there is only one node which contributes to a particular bit in the synopsis. In reality, the expected number of nodes contributing to a bit increases exponentially as we move from the Rth bit, where R is the the length of the prefix of all ones in the synopsis S, to the lower-order bits, thereby increasing the probability that at least one MAC corresponding to a bit position reaches the sink. Computation and Communication Overhead Each contributing node computes one MAC. The expected number of contributing nodes is = 1 ( 2u c l c -1), which is independent of network size. Thus, only a subset of nodes will incur any computational overhead . With respect to communication overhead, the maximum number of MACs that any node will need to forward is . Thus this approach is scalable, and can be used in large-scale sensor networks as long as the ratio u c /l c is reasonably small. 5.3 Securing SUM We can extend the approach used for making the Count aggregate resilient to compromised nodes to the Sum aggregate. To derive the synopsis-edge for Sum we need to assume upper and lower bounds for the value of a sensor reading in addition to the upper and lower bounds for the number of sensor nodes. A node X sends to the BS a MAC, M = MAC(K X , m), only if it contributes to the synopsis-edge as in SecureCount. The format of the message m sent by a node is [X|A|Seed|v], where X is the node id, Seed is the random seed included in the broadcast query, A represents the collection of bits in the synopsis to which X contributes, and v is X 's sensed value. Security Analysis In the case of the Sum aggregate, the attacker could falsely set some bits in its synopsis not only by using a false node id but also using a false sensor reading. Although MACs from the contributing nodes enable the BS to verify the node Ids, the BS cannot verify the sensed value of a node. A compromised node can claim to have a large sensed value close to the upper bound u v to increase its chance of being able to contribute to the synopsis-edge. The following theorem (whose proof can be found in Appendix B) shows that this attack's impact is limited. Theorem 1. Let be the number of compromised nodes in a network of n nodes. Let u v and a v denote the upper bound and the average value of the sensor reading respectively. Let S be the final synopsis computed at the sink and let R be the length of the prefix of all ones in S. Let s denote the value of the Sum aggregate. If each compromised node claims that its sensed value is equal to the upper bound u v , and if ( u v ) &lt; s, then the probability Pr[S[R + 1] = 1] is proportional to the product of the fraction of compromised nodes in the network, /n. Note that if the compromised node contributes to the (R + 1)th bit BS's estimate of Sum doubles. Thus, the theorem shows that for a large network, as long the fraction of compromised nodes grows sub-linearly, the probability of this attack succeeding is small. For smaller networks, the probability of this attack succeeding depends upon the ratio /n and on the ratio u v /a v . As an example, if n = 1000, = 25, and u v /a v = 4, then Pr[S[R + 1] = 1] = 0.064. The impact of the attack is further reduced if we employ the PCSA algorithm in which m independent synopses are computed and the final estimator R is calculated by averaging these m esti-mators . As an example, to add an error of 40% to the final Sum, the attacker needs to set the R + 1-th bit in at least m 2 synopses. In the example above where Pr[S[R + 1] = 1] is 0.064, this probability is close to zero when m is 20. This example illustrates that this attack's impact is limited when ( u v n a v ) is small. On the other hand, when ( u v n a v ) is large, we cannot neglect the possibility that the attacker will succeed in injecting a significant error to the Sum computed at the sink. To address this scenario, we can use a scheme in which a node that contributes to the synopsis-edge needs an endorsement from at least neighbors attesting to the validity of its sensed value. We assume that the sensed values of one-hop neighbors are correlated so that one node can verify the reading of its neighbors. We assume that there are fewer than compromised nodes among the one hop neighbors of any node. Each contributing node X collects at least endorsements from its one-hop neighbors in the form of a MAC computed over the sensor reading using the pairwise key that the neighbor shares with the sink. Then X computes an XMAC [1] by XORing the collected MACs and X 's own MAC, and sends the XMAC to the BS. (Zhu et al. [25] use an identical scheme to reduce the total size of the MACs.) We also assume that BS has the knowledge to verify if a set of nodes are one-hop neighbors, which prevents the collusion attack. (We refer to this scheme as the XMAC-based scheme.) Computation and Communication Overhead The number of contributing nodes is less than 1 ( 2u s l s -1), where u s and l s are the upper bound and lower bound of Sum. As in the case of Count, is independent of the network size and thus this approach is scalable . With respect to worst case communication overhead, a node will need to forward at most MACs. THE EXTENDED APPROACH TRADING LATENCY FOR COMMUNICATION OVERHEAD When the ratio ( ) of the upper bound of the aggregate to the lower bound is high, the basic approach described in the previous section is not scalable because the worst case communication cost incurred by a node is proportional to . In this section, we describe an approach which has lower communication costs in comparison to the basic approach at the expense of greater latency. 76 6.1 Protocol Overview Our extended approach is based on the observation that the expected number of nodes that contribute to bits i, where R &lt; i k in the synopsis (k is the length of the synopsis) is very small. In fact, using Property 2 and Property 4 from Section 2, we can show that expected number of nodes contributing to the Rth and higher-order bits of S is less than 2 / 2.58. We use a sliding-window based approach in which the aggregation phase is divided into multiple epochs 3 . Starting at the right-most bit k, we proceed from right to left within the synopsis S using a bit window of w bits. In each epoch, only the nodes that contribute a 1 to the bits in S corresponding to the current position of the window, send MACs to the sink that can be used to verify their contribution to S. In other words, in the first epoch, only nodes that contribute a 1 to bits k to k -w+1 respond to the query. In epoch two, nodes that contribute to bits between k - w and k - 2w + 1 respond, and so on. The algorithm terminates when the querying node has determined that the remaining bits of S to the left of the current window are likely to be 1 with high probability. The design of this termination criterion is the main challenge in using this approach; we discuss the termination criterion and its analytical underpinnings in detail. Once the querying node determines that the algorithm can terminate it broadcasts a STOP message to the network to announce the end of the iterative aggregation procedure. 6.2 Protocol Operation The operation of the protocol is similar to that of the protocol used in the basic approach with some minor differences as follows. The query message broadcast to the network includes the window size w in addition to the other parameters. As in the original synopsis diffusion algorithm [3, 14], we assume that the time is syn-chronized among BS and the sensor nodes. Each node computes the start and end time of the current epoch, based on the window w. Further, although the MACs generated by nodes are sent to the BS over the course of multiple epochs, the fused synopsis computed by each node is forwarded to its parent in the first epoch. Thus, the BS can compute the aggregate at the end of the first epoch itself, although this aggregate may be erroneous in the presence of compromised nodes. 6.3 Termination Criterion The goal of our algorithm is to find r, the lowest-order bit in S that is 0. Further, recall that S is of the form 1 r -1 0 , where the bits at positions i &gt; r are highly likely to be 0. Thus, the intuition behind our termination criterion is simple: as we examine the bits of S moving from right to left, if we observe two consecutive 1's, i.e., if we observe the string "110", it is highly likely that the 0 is at the rth position. In fact, we can show analytically that the probability of this event is greater than 90% which follows from the following theorem. Theorem 2. Let F denote the event that the string "0s l 11" where s l represents any string of length l, l 0 appears in a synopsis S. The probability of the event F is less than 10%. (The proof is given in the appendix.) Further, we can take advantage of the fact that most applications will use the PCSA algorithm to reduce the approximation error in estimating R = r -1. Recall that in the PCSA algorithm m synopses are computed in parallel. Let R i denote the value of R estimated from the ith synopsis. Then, according to the PCSA algorithm, 3 The original synopsis diffusion algorithm [3, 14] takes one epoch to complete. the the expected value of the random variable R is estimated by averaging the individual values of R for each synopsis, i.e., E [R] = R = i =m i =1 R i . Although there is likely to be some variation among the R i , we know from Property 3 in Section 2 that the variation is expected to correspond to two bit position both to the left and the right of the true value of R. This suggests that there is a high degree of correlation between the R i for different synopses. Thus, in our window-based approach, we can increase our confidence that we have found the correct position of R, if we observe the bit pattern "11" in multiple synopses among the m that are being computed in parallel. Based on this intuition, our termination criterion consists of checking whether we have observed the string "11" in at least m out of the m synopses. 1 1 1 1 1 1 0 0 1st Synopsis 2nd Synopsis m-th Synopsis 3rd Synopsis 1st position of the window the aggregation after this window process stops the termination-test passes in this window 1 1 1 1 1 1 1 0 Figure 3: Each synopsis is divided into several windows of width w = 2 bits. After the termination criterion is satisfied, the base station broadcasts a STOP message and the aggregation phase stops after the next epoch. In each epoch, nodes which contribute to the corresponding window send a MAC to the base station. The MACs which correspond to the crossed bits are never sent. Our goal in selecting the threshold m is to reduce the likelihood of both a false positive, which means that the algorithm was terminated too early, and a false negative, which means that the algorithm terminated too late, i.e., after the sliding window had already crossed the true position of R. A false positive results in an over-estimate of R, whereas a false negative results in additional communication overhead. We now show that it is possible to find a suitable value for m such that the probability of false positive and the probability of false negative are both low. Theorem 3. Let G i denote the event that both bit i and (i + 1) in a synopsis S are 1. Let denote the expected value of the estimator R . Then, Pr [G ] = 0.3454, Pr[G +1 ] = 0.1315, and Pr[G +2 ] = 0 .0412. Because of space limitations, the proof of this theorem can be found in the appendix. If the sliding window in our algorithm is two bits wide, i.e, w = 2, from the definition of the false positive (FP), we get that the probability Pr [FP] is the probability that the event G +2 occurs in m or more synopses. Similarly, the probability of a false negative, Pr [FN] is the probability that the event G occurs in fewer than m synopses. For m = 20 (which is the typical value used in previous 77 work [3, 14], we find that the best value of m is 4 in which case Pr [FP] = 0.0082 and Pr[FN] = 0.0484. The same approach illus-trated here can be used to derive the appropriate threshold m for other window sizes. Figure 3 illustrates the operation of our algorithm for w = 2. Assume that the termination criterion is satisfied in epoch e. The BS broadcasts a STOP message which directs all nodes to terminate the aggregation phase. Note that by the time each node in the network receives the broadcast STOP message, many of the nodes will have already sent MACs corresponding to their contributions to the next epoch e + 1 of the algorithm. Thus, the effect of the termination criterion being satisfied in epoch e message is to terminate the aggregation after epoch e + 1. We can take advantage of this extra epoch to further increase our level of confidence in the estimated value of R. Let the bit position of the sliding window in epoch e correspond to bits and + 1. Instead of estimating R = + 1 because m out of m synopses had both bits and +1 equal to 1, we can now estimate R based on the observed value of R i for all m synopses. Our simulations show that estimate of the aggregate computed using our extended approach is close to the estimate computed using the original synopsis diffusion [3, 14] algorithm. Latency The number of epochs taken by our sliding window approach depends on the ratio ( ) of the upper bound of the aggregate to the actual value. If the upper bound is u and the actual value is , for a window of width w the number of epochs is equal to (log 2 u m -log 2 m w +2) = (log 2 w +2) Communication Overhead Theorem 3 implies that it is highly likely that the sliding window contains the position R when the termination criterion is satisfied. As discussed above, if the termination criterion is satisfied in epoch e, the aggregation completes after epoch e + 1. Thus, by property 2 and property 4 in Section 2, if m synopses are computed in parallel, the expected number of nodes which send a MAC varies in the range of (2.58 m) to (5.16m). Even if a sensor node contributes to more than one bits, it sends just one MAC validating all the bits. Note that the number of contributing nodes does not exceed this range even if the network size is increased. Our simulation results show that 85 MACs are sent on average when m = 20. We observe that the width of the window w determines a tradeoff between the communication overhead and the latency. If we divide the synopses into wider windows, the number of MACs sent and hence the communication overhead will increase while the latency of the aggregation process will decrease, and vice versa. 6.4 Discussion An alternative approach to the sliding window-based approach described above is one in which the base station computes the aggregate of interest in the first epoch using the original Synopsis Diffusion algorithm. It then broadcasts a message requesting only the nodes that contribute to the bit window that contains R to send the MACs authenticating their local synopses. If the BS success-fully verifies all the MACs it receives, then the protocol terminates at the end of the second epoch. However, if it does not receive the requested MACs or if one or more MACs are invalid, the BS executes the sliding-window protocol described above to compute the correct value of R. If the probability of compromised nodes being present in the network is low, then this alternative approach is preferable to the extended approach since it will have much lower latency on average. SIMULATION RESULTS In this section, we report on a detailed simulation study that examined the performance of our attack-resilient aggregation algorithms discussed in Sections 5 and 6. Our simulations were written using the TAG simulator developed by Madden et al. [11]. We added the attack-resilient functionality to the source code provided by Considine et al. [3] which simulates their multipath aggregation algorithms in the TAG simulator environment. 7.1 Simulation Environment For our basic experimental network topology, we used a regular 30 30 grid with 900 sensor nodes, where one sensor is placed at each grid point and the base station is at the center of the grid, as in [3]. The communication radius of each node is 2 unit allowing the nearest eight grid neighbors to be reached. The goal of our simulation experiments is to examine the communication overhead and accuracy of our scheme in the presence of packet losses, which are relatively frequent in sensor networks. We use a simple packet loss model in which packets are dropped with a fixed probability; this packet loss rate is assumed to include packets that are lost due to being dropped by compromised nodes. We do not model any additional attacks by compromised nodes, specifically the falsified subaggregate and the falsified local value attacks, in our simulation. This is because we have already shown that these attacks cannot affect the estimate of the aggregate computed at the sink. Consequently, these attacks simply have the effect of increasing the communication and computation overhead; in effect , they become a form of DOS or resource consumption attacks. We assign a unique id to each sensor, and we assume that the sensor reading is a random integer uniformly distributed in the range of 0 to 250 units. We compute 20 synopses in parallel using the PCSA algorithm as in the experiments reported in [3, 14]. We use the method of independent replications as our simulation methodology . Each simulation experiment was repeated 200 times with a different seed. The plots below show the 95% confidence intervals of the reported metric. 7.2 Results and Discussion Due to space constraints, we will only present the results of our extended approach for computing the Sum aggregate. Accuracy of our estimate In the first set of experiments, we validate our claim that our attack-resilient approach has the same accuracy in computing the true value of the aggregate as the original synopsis diffusion approach. Figure 4a plots the estimates of our approach and the synopsis diffusion approach as a function of the packet loss rate. We observe that the two estimates are indeed very close in all loss rate conditions. We observe that the average value of the sensor reading is approximately 125, i.e., the accurate Sum is 900 125 = 11250. Communication overhead We now compare the communication overhead of our approach to that of the original synopsis diffusion approach. Figure 4(b) plots the total number of bytes transmitted for computing the Sum aggregate. As discussed in Section 5.3, for preventing a node from using a false reading to generate its own local synopsis, we can adopt two approaches. In the first approach , we ignore the impact of the falsified local value attack; in the figure, this approach is labeled as ARSD (attack-resilient synopsis diffusion). The second approach requires the contributing node to include a XMAC, which corresponds to an endorsement from its neighbors, in the message; in the figure, this approach is labeled ARSD+XMAC. For ARSD+XMAC, each contributing node sends an authentication message which has two parts: the first part contains the ID (2 78 0 0.1 0.2 0.3 0.6 0.8 1 1.2 1.4 x 10 5 Link Loss Rate Estimated Sum ARSD SD Actual Sum 0 0.1 0.2 0.3 1 2 3 4 5 6 x 10 4 Link Loss Rate Total Byte Transmitted ARSD ARSD+XMAC SD a. Accuracy of Sum b. Byte overhead 0 2 4 6 8 10 12 0 2 4 6 8 Log (upper bound / actual Sum) Number of epochs ARSD (window width=2) 1000 2000 3000 4000 5000 40 60 80 100 120 Number of sensor nodes Number of Contributing nodes ARSD c. Latency d. Varying the network size Figure 4: Experimental Results bytes) of the contributing node and its sensed value (3 bytes), and the second part includes the IDs of the k neighbors and a XMAC (4 bytes). If the value of k is not more than 4 then a node needs 8 bytes to specify the identity of the neighbors whose MACs are used to generate the XMAC. Thus, the size of one authentication message is 17 bytes. For ARSD, the contributing node just needs to send its own MAC; no neighbor endorsement is needed, which reduces the authentication message size to 9 bytes. Figure 4(b) shows that the byte overhead of the ARSD+XMAC scheme is roughly 5 times larger than the original approach, whereas ARSD is 2.5 times larger than the original approach. One might expect that if loss rate is high our extended approach may take more time to stop because some MACs could be lost en-route, and, as a result, the communication overhead could increase. But Figure 4(b) demonstrates that the overhead of the extended approach does not increase with the loss rate. Latency As discussed in Section 6, the latency of the extended approach depends on the looseness of the base station's estimate of the upper bound of Sum. Figure 4(c) plots the number of epochs taken by our extended approach as a function of the ratio of the upper bound to the actual value of the aggregate. The figure shows that the number of epochs increases at logarithmic scale with the ratio of the upper bound to the actual Sum. We note, however, that the byte overhead of our scheme is independent of this ratio. Effect of network size In this experiment, we study the impact of the network size on the communication overhead of the extended approach. The communication overhead depends upon the number of contributing nodes that send a MAC to the base station, authenticating their synopsis. Recall from Section 6 that the expected number of contributing nodes is independent of the network size. Figure 4(d) confirms our analysis; we observe that the number of contributing nodes is more or less constant as the network size increases 4 . This figure thus illustrates the scalability of our approach for attack-resilient aggregation. RELATED WORK Several data aggregation protocols [11, 19, 23] have been proposed in the literature which efficiently fuse the sensed information en-route to the base station to reduce the communication overhead. Since packet losses and node failures are relatively common in sensor networks, several studies have investigated the design of robust aggregation algorithms. Considine et al. [3] and Nath et al. [14, 12] have presented robust aggregation approaches that combine the use of multi-path routing with clever algorithms that avoid double-counting of sensor readings. Jelasity et al. [9] proposed a robust gossip-based protocol for computing aggregates over network components in a fully decentralized fashion. They assume that nodes form an overlay network where any pair of nodes are considered to be neighbors, which makes this protocol impractical for sensor networks . We note that none of the above algorithms were designed with security in mind. Recently several researchers have examined security issues in aggregation. Wagner [17] examined the problem of resilient data aggregation in presence of malicious nodes, and provided guidelines for selecting aggregation functions in a sensor network. Buttyan et al. [2] proposed a model of resilient aggregation and analyzed the maximum deviation from the true value of the aggregate that an adversary could introduce while remaining undetected. The models used by by both Buttyan et al and Wagner assume that there is no in-network aggregation, that is, the aggregation is performed at the sink. Przydatek et al [16] present protocols that can be used by a trusted remote user to query a sensor network in which the base 4 The link loss rate is held at 20% in this set of experiments. 79 station may be compromised and the base station is the only aggregator . One of the protocols described by Przydatek et al is a robust approach for counting distinct elements in a data stream that can be used for estimating the size of the network, i.e., the Count aggregate . Their approach for counting distinct elements is similar to our scheme for Count in the sense that in both cases only a subset of elements need to be verified. The first secure in-network data aggregation protocol was designed by Hu and Evans [8]. Their protocol is effective only if no more than one node is compromised. Recently, Yang et al. [18] proposed SDAP, a secure hop-by-hop data aggregation protocol which can tolerate more than one compromised node. SDAP is a tree-based aggregation protocol with communication cost comparable with that of the ordinary aggregation protocols while it provides certain level of assurance on the trustworthiness of the aggregation result. As SDAP is a tree-based protocol, it is vulnerable to link loss and node failures which are relatively common in sensor networks , whereas our protocol is robust to this communication loss and, at the same time, secure against compromised nodes. We note that our work is related to the general problem of preventing false data injection. Du et al. [4] proposed a mechanism that allows the base station to check the aggregated values submit-ted by several designated aggregators, based on the endorsements provided by a certain number of witness nodes around the aggregators . Their scheme does not provide per-hop aggregation. Several other works [20, 21, 25] have also proposed solutions to prevent false data injection attacks in sensor networks, but they do not involve data aggregation. CONCLUSION In this paper, we investigated the security issues of synopsis diffusion framework in presence of compromised nodes. We showed that a compromised node can launch several simple attacks on the existing aggregation algorithms, which could significantly deviate the estimate of the aggregate. We also proposed modifications to the aggregation algorithms that guard against these attacks. Our analytical results and simulation results show that our approach is effective and it incurs minimal computation and communication overhead. In this paper, we assume that a sensor node has a security association only with the base station, and, as a result, the authentication messages cannot be processed in-network in our approach. To further reduce the communication overhead, we plan to exploit other security settings, e.g., local pairwise keys among nodes, as a part of our future work. REFERENCES [1] M. Bellare, R. Guerin, and P. Rogaway. XOR MACs: New methods for message authentication using finite pseudorandom functions. In Proc. of the 15th Annual International Cryptology Conference on Advances in Cryptology - CRYPTO'95 , pages 1528, 1995. [2] L. Buttyan, P. Schaffer, and I. Vajda. Resilient aggregation with attack detection in sensor networks. In Proc. of 2nd IEEE Workshop on Sensor Networks and Systems for Pervasive Computing , 2006. [3] J. Considine, F. Li, G. Kollios, and J. Byers. Approximate aggregation techniques for sensor databases. In Proc. of IEEE Int'l Conf. on Data Engineering (ICDE) , 2004. [4] W. Du, J. Deng, Y. S. Han, and P. Varshney. A pairwise key pre-distribution scheme for wireless sensor networks. In Proc. of the 10th ACM Conference on Computer and Communications Security (CCS '03). , 2003. [5] P. Flajolet and G. N. Martin. Probabilistic counting algorithms for data base applications. Journal of Computer and System Sciences , 31(2):182209, 1985. [6] S. Ganeriwal and M. B. Sribastava. Reputation-based framework for highly integrity sensor networks. In Proc. of ACM Workshop on Security of Sensor and Adhoc Networks (SASN) , Washington, DC, 2004. [7] D. Ganesan, R. Govindan, S. Shenker, and D. Estrin. Highly-resilient energy-efficient multipath routing in wireless sensor networks. Mobile Comuting and Communication Review , 4(5):1125, 2001. [8] L. Hu and D. Evans. Secure aggregation for wireless networks. In Proc. of Workshop on Security and Assurance in Ad hoc Networks. , 2003. [9] M. Jelasity, A. Montresor, and O. Babaoglu. Gossip-based aggregation in large dynamic networks. ACM Transactions on Computer Systems , 23(3):219252, 2005. [10] F. Koushanfar, M. Potkonjak, and A. Sangiovanni-Vincentelli. Fault tolerance techniques in wireless ad-hoc sensor networks. In Sensors 2002. Proceedings of IEEE , pages 1491 1496. [11] S. Madden, M. J. Franklin, J.M. Hellerstein, and W. Hong. TAG: A tiny aggregation service for ad hoc sensor networks. In Proc. of 5th USENIX Symposium on Operating Systems Design and Implementation , 2002. [12] A. Manjhi, S. Nath, and P. Gibbons. Tributeries and deltas : Efficient and robust aggregation in sensor network streams. In Proc. of ACM International Conference on Management of Data (SIGMOD) , 2005. [13] Mica Motes. http://www.xbow.com. [14] S. Nath, P. B. Gibbons, S. Seshan, and Z. Anderson. Synopsis diffusion for robust aggregation in sensor networks. In Proc. of the 2nd international conference on Embedded networked sensor systems (SenSys) , 2004. [15] A. Perrig, R. Szewczyk, V. Wen, D. Culler, and J. D. Tygar. SPINS: Security protocols for sensor networks. In Seventh Annual International Conference on Mobile Computing and Networks (MobiCOM) , 2001. [16] B. Przydatek, D. Song, and A. Perrig. SIA: Secure information aggregation in sensor networks. In Proc. of the 1st international conference on Embedded networked sensor systems (SenSys) , 2003. [17] D. Wagner. Resilient aggregation in sensor networks. In Proc. of ACM Workshop on Security of Sensor and Adhoc Networks (SASN) , 2004. [18] Y. Yang, X. Wang, S. Zhu, and G. Cao. SDAP: A secure hop-by-hop data aggregation protocol for sensor networks. In Proc. of ACM MOBIHOC, 2006. [19] Y. Yao and J. E. Gehrke. The cougar approach to in-network query processing in sensor networks. ACM SIGMOD Record, 31(2):918, September 2002. [20] Fan Ye, Haiyun Luo, Songwu Lu, and Lixia Zhang. Statistical en-route filtering of injected false data in sensor networks. In Proc. of IEEE Infocom, 2004. [21] W. Zhang and G. Cao. Group rekeying for filtering false data in sensor networks: A predistribution and local collaboration-based approach. Proc. of IEEE Infocom, 2005. [22] J. Zhao and R. Govindan. Understanding packet delivery performance in dense wireless sensor networks. In Proc. of 80 the 1st international conference on Embedded networked sensor systems (SenSys) , 2003. [23] J. Zhao, R. Govindan, and D. Estrin. Computing aggregates for monitoring sensor networks. In Proc. of the 2nd IEEE International Workshop on Sensor Network Protocols and Applications , 2003. [24] S. Zhu, S. Setia, and S. Jajodia. LEAP: Efficient security mechanisms for large-scale distributed sensor networks. In Proc. of the 10th ACM Conference on Computer and Communications Security (CCS '03). , 2003. [25] S. Zhu, S. Setia, S. Jajodia, and P. Ning. An interleaved hop-by-hop authentication scheme for filtering injected false data in sensor networks. In Proc. of IEEE Symposium on Security and Privacy , 2004. Appendix A. Below we describe the algorithm (SecureCount) executed by each node in response to a Count query. X represents the node Id. Algorithm 2 SecureCount(X , Seed, a, b) 1: M = {}; // M is initialized as an empty set 2: i = SG(X, Seed); // X contributes to bit i 3: if (a i b) then 4: m = [X|i|Seed]; 5: M = MAC(K X , m); 6: M = M M; 7: end if 8: S l = SF(); // S l is the fused synopsis at X 9: M = M C ; // C represents the set of MACs X received from // its child nodes 10: X : S l | M ; B. Proofs of Theorems We provide the proofs for the theorems present in the paper. Theorem 1. Let there be n nodes in the sensor network among which nodes are compromised. Let u v and a v denote the upper bound and the average value of the sensor reading respectively. Let S be the final synposis computed at the sink and let R be the length of the prefix of all ones in S. Let s denote the value of the Sum aggregate. If each compromised node claims that its sensed value is equal to the upper bound u v , and if ( u v ) &lt; s, then the probability Pr[S [R + 1] = 1] is proportional to the product of the fraction of compromised nodes in the network, /n, and the ratio u v /a v . P ROOF . By property 2 in Section 2, the expected value of the estimator R, for the Sum synopsis S, is log 2 (s), where s denotes the Sum. As a node X with sensed value v invokes the function CT() v times (in the synopsis generation phase), the probability that X does not contribute to bit i in S is (1 1 2 i ) u v . So, the probability (p) that a node with sensed value u v will contribute to the (R + 1)th bit is (1 -(11 2 R +1 ) u v ). After simplifying, we get p = 1 -(1- 1 2 s ) u v 1-(1- u v 2 s ) = u v 2 s The above approximation is valid as u v is smaller than s. If there are compromised nodes, then Pr (S[R + 1] = 1) is q = 1 -(1- p) p = u v 2 s = ( 1 2 ) ( n ) ( u v a v ) To prove Theorem 2, we first prove the following results. Lemma 1. Let E i , 1 i k -2 denote the event that the string "011" appears in a synopsis S from bit i to bit (i + 2) (i.e., S[i] = 0, S [i + 1] = 1, and S[i + 2] = 1), where k is the length of S. The maximum value of the probability (p i ) of the event E i is 0.037 for any value of i and for any value of Count (or Sum) shared by S. P ROOF . If function CT () is invoked once (ref. Section 2), then Pr [S[ j] = 1] = q j = 1 2 j , 1 j k. This probability increases if it is given that bit j , 1 j k will remain 0. Specifically, Pr [S[ j] = 1|S[j ] = 0] = q j 1 -q j = 1 (2 j ) (1 1 2 j ) . If is the total Count (or Sum) shared by synopsis S, then Pr[S[i] = 0 ] = (1 -q i ) , and Pr [S[i + 1] = 1, S[i + 2] = 1] = 1 -Pr[S[i+1] = 0]-Pr[S[i+2] = 0] + Pr[S[i + 1] = 0, S[i + 2] = 0] (1) = 1 -(1-q i +1 ) -(1-q i +2 ) + (1 -q i +1 -q i +2 ) = 1 -(11 2 i +1 ) -(11 2 i +2 ) + (1 1 2 i +1 1 2 i +2 ) So, the probability of the event E i is p i = Pr[S[i] = 0] Pr [S[i + 1] = 1, S[i + 2] = 1 | S[i] = 0] = (1 1 2 i ) [1 -(11 (2 i +1 ) (11 2i ) ) -(11 (2 i +2 ) (11 2i ) ) + (1 1 (2 i +1 ) (11 2i ) 1 (2 i +2 ) (11 2i ) ) ] (2) Note that if i &lt;&lt; log 2 (), the 1st factor is close to 0 and second factor is close to 1, making p i close to 0. On the other hand, if i &gt;&gt; log 2 (), the 1st factor is close to 1, but the 2nd factor are close to 0, again making p i close to 0. p i attains the highest value when i is close to log 2 (). We have numerically found that the maximum value of p i is 0.037, for any value of i or . Lemma 2. Let E denote the event that the string "011" appears in a synopsis S at any position. The probability of the event E is less than 0.099. P ROOF . E i denotes the event that "011" appears in a synopsis S where 0 is at the ith bit. We observe that the events E i , E i +1 , E i +2 are mutually exclusive, for any value of i. Following the same direction of Lemma 1, we can show that the probability that " 011s l 011" appears in synopsis S is close to zero, where s l represents any string of length l, l 0. So, the probability that two events E i and E j where j (i+3) can occur together is negligible, for any value of i. As a result, we can approximate that events E i s are mutually exclusive and hence the probability of event E is p = k i =1 p i , where p i is given by expression (2) and k is the length of S. We have numerically found that maximum value of p is 0.099. Lemma 3. Let F i denote the event that a string "0s i 11" appears in a synopsis S, and let F denote the general event that a string 81 "0s l 11", l 0 appears in S, where s l represents any string of length l . Pr [F] = Pr[F 0 ] P ROOF . As the string "011" is a special case of string "0s l 11" where l = 0, Pr[F] Pr[F 0 ]. On the other hand, if string s = "0s l 11", l 0 appears in S, string "011" must also appear as a substring of s . As an example, if s = "01011" where s l = "10", we can see "011" as a substring of s . Hence, Pr [F] Pr[F 0 ]. So, we get that Pr [F] = Pr[F 0 ]. Theorem 2. Let F denote the event that the string "0s l 11" where s l represents any string of length l, l 0 appears in a synopsis S. The probability of the event F is less than 10%. P ROOF . As the event F 0 in Lemma 3 is same as the event E in Lemma 2, we get that the probability of event F is less than 10%. Theorem 3. Let G i denote the event that both bit i and (i + 1) in a synopsis S are 1. Let denote the expected value of the estimator R . Then, Pr [G ] = 0.3454, Pr[G +1 ] = 0.1315, and Pr[G +2 ] = 0 .0412. P ROOF . The expected value of R is log 2 ( m ), which we denote by , where is the total Count (or Sum) shared by m synopses following the algorithm PCSA. If function CT () is invoked once (ref. Section 2), Pr [S[i] = 1] = q i = 1 m 1 2 i , 1 i k because synopsis S is selected with probability 1 m among m synopses . As is the total Count (or Sum) shared by all synopses, we get by using similar expression as (1) in Lemma 1 that Pr [G i ] = 1 -(11 m 2 i ) -(11 m 2 i +1 ) + (1 1 m 2 i 1 m 2 i +1 ) So, Pr [G ] = 1 -(11 m m ) -(11 m 2 m ) + (1 1 m m 1 m 2 m ) 1-e 1 -e 1 2 + e 3 2 = 0.3454 Similarly, we find Pr[G +1 ] = 0.1315, and Pr[G +2 ] = 0.0412. 82

sensor networks;node compromise prevention;falsified local value attack;in-network data aggregation;Attack resilient heirarchical data aggregation;Sum aggregate;falsified sub-aggregate attack;Attack-Resilient;Count aggregate;Sensor Network Security;robust aggregation;Synopsis Diffusion;Data Aggregation;synopsis diffusion aggregation framework;network aggregation algorithms;Hierarchical Aggregation

39

Automated Rich Presentation of a Semantic Topic

To have a rich presentation of a topic, it is not only expected that many relevant multimodal information, including images, text, audio and video, could be extracted; it is also important to organize and summarize the related information, and provide users a concise and informative storyboard about the target topic. It facilitates users to quickly grasp and better understand the content of a topic. In this paper, we present a novel approach to automatically generating a rich presentation of a given semantic topic. In our proposed approach, the related multimodal information of a given topic is first extracted from available multimedia databases or websites. Since each topic usually contains multiple events, a text-based event clustering algorithm is then performed with a generative model. Other media information, such as the representative images, possibly available video clips and flashes (interactive animates), are associated with each related event. A storyboard of the target topic is thus generated by integrating each event and its corresponding multimodal information. Finally, to make the storyboard more expressive and attractive, an incidental music is chosen as background and is aligned with the storyboard. A user study indicates that the presented system works quite well on our testing examples.

INTRODUCTION In the multimedia field, a major objective of content analysis is to discover the high-level semantics and structures from the low-level features, and thus to facilitate indexing, browsing, searching, and managing the multimedia database. In recent years, a lot of technologies have been developed for various media types, including images, video, audio and etc. For example, various approaches and systems have been proposed in image content analysis, such as semantic classification [1], content-based image retrieval [2] and photo album management [3]. There are also a lot of research focuses on video analysis, such as video segmentation [4], highlight detection [5], video summarization [6][7], and video structure analysis [8], applied in various data including news video, movie and sports video. Since audio information is very helpful for video analysis, many research works on audio are also developed to enhance multimedia analysis, such as audio classification [9], and audio effect detection in different audio streams [10]. Most recently, there are more and more approaches and systems integrating multimodal information in order to improve analysis performance [11][12]. The main efforts of the above mentioned research have focused on understanding the semantics (including a topic, an event or the similarity) from the multimodal information. That is, after the multimedia data is given, we want to detect the semantics implied in these data. In this paper, we propose a new task, Rich Presentation, which is an inverse problem of the traditional multimedia content analysis. That is, if we have a semantic topic, how can we integrate its relevant multimodal information, including image, text, audio and video, to richly present the target topic and to provide users a concise and informative storyboard? In this paper, the so-called "semantic topic" is a generic concept. It could be any keyword representing an event or events, a person's name, or anything else. For example, "World Cup 2002" and "US election" could be topics, as well as "Halloween" and "Harry Potter". In this paper, our task is to find sufficient information on these topics, extract the key points, fuse the information from different modalities, and then generate an expressive storyboard. Rich presentation can be very helpful to facilitate quickly grasping and better understanding the corresponding topic. People usually search information from (multimedia) database or the Internet. However, what they get is usually a bulk of unorganized information, with many duplicates and noise. It is tedious and costs a long time to get what they want by browsing the search results. If there is a tool to help summarize and integrate the multimodal information, and then produce a concise and informative storyboard, it will enable users to quickly figure out the overview contents of a topic that they want to understand. Rich presentation provides such a tool, and thus it could have many potential applications, such as education and learning, multimedia authoring, multimedia retrieval, documentary movie production, and information personalization. In this paper, we will present the approach to rich presentation. In order to produce a concise and informative storyboard to richly present a target topic, we need to answer the following questions. 1) How to extract the relevant information regarding the target 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. MM'05, November 611, 2005, Singapore. Copyright 2005 ACM 1-59593-044-2/05/0011...$5.00. 745 topic? 2) How to extract the key points from the relevant information and build a concise and informative storyboard? 3) How to fuse all the information from different modality? and 4) how to design the corresponding rendering interface? Storyboard Relevant Media Music Multiple Events Clustering Event summary (4w + time) Geographic information Relevant multimodal information Retrieval Text A Target Topic Rhythm Analysis Onset/Beat Sequence Strength confidence Media Association Representative images Relevant video clips Storyboard Generation Event presentation, multimodal information fusion, layout design Music and storyboard synchronization Rich Presentation User Interaction Fig. 1 The system framework of rich presentation of a target semantic topic. It is mainly composed of three steps, relevant multimodal information extraction, media analysis, and rich presentation generation. In this paper, we propose a number of novel approaches to deal with the above issues and also present an example system. Fig. 1 illustrates the proposed system framework of rich presentation. It is mainly composed of three steps, relevant multimodal information extraction, media analysis including multiple events clustering , representative media detection and music rhythm analysis; and the final storyboard generation and music synchronization. In the proposed system, given the semantic topic, the relevant information, including text, image, video and music, is first extracted from the available multimedia database or the web database . User interaction is also allowed to provide extra relevant material or give relevant feedback. Then, the information is summarized, with an event clustering algorithm, to give a concise representation of the topic and figure out the overview of the contents. Other multimedia materials, such as representative images (or image sequences) and geographic information, are subsequently associated with each event. In the next step, all the above information is integrated to generate a storyboard, in which each event is presented as one or multiple slides. An incidental music, which is also possibly relevant to the topic, is finally synchronized with the storyboard to improve its expressiveness and attractiveness. Thus, with these steps, a concise and informative rich presentation regarding the target topic is generated . The rest of the paper is organized as follows. Section 2 discusses the relevant information extraction corresponding to the target topic. Section 3 presents our approach to the topic representation, including multiple events clustering, event description, and representative media selection. Section 4 describes the approach to rich presentation generation, including storyboard generation, incidental music analysis and synchronization. Experiments and evaluations are presented in the Section 5. Conclusions are given in the Section 6. OBTAINING RELEVANT INFORMATION To obtain the multimodal information which is relevant to the input topic (keyword), generally, we could search them from various databases which have been indexed with the "state-of-the-art" multimedia analysis techniques. However, in current stage, there is lack of such publicly available multimedia databases. The public search engine like MSN or Google indexes all the Internet web-pages and can return a lot of relevant information, but the search results usually contain much noise. We could also build a private database for this system to provide more relevant and clean results, but it will be too much expensive to collect and annotate sufficient multimedia data for various topics. In order to obtain relatively accurate and sufficient data for an arbitrary topic, in our system, we chose to collect the relevant multimodal information of the given topic from the news websites such as MSNBC, BBC and CNN, instead of building an available database from the scratch. These news websites are usually well organized and managed; and contain various kinds of high quality information including text, image and news video clips. Although the news websites are used as the information sources in our system, other various multimedia databases can be also easily incorporated into the system if they are available. Instead of directly submitting the topic as a query and getting the returned results by using the search function provided by the websites, in our system, we crawled the news documents from these websites in advance and then build a full-text index. It enables us to quickly obtain the relevant documents, and also enable us to use some traditional information retrieval technologies, such as query expansion [13], to remove the query ambiguousness and get more relevant documents. In our approach, user interaction is also allowed to provide more materials relevant to the topic, or give relevant feedback on the returned results. For example, from the above websites, we can seldom find a music clip relevant to the target topic. In this case, users could provide the system a preferred music, which will be further used as incidental music to accompany with the storyboard presentation. Users could also give some feedbacks on the obtained documents. For example, if he gives a thumb-up to a document, the relevant information of the document needs to be presented in the final storyboard. On the other side, users could also thumb-down a document to remove the related information. TOPIC REPRESENTATION A semantic topic is usually a quite broad concept and it usually contains multiple events. For example, in the topic "Harry Potter", the publication of each book and the release of each movie could be considered as an event; while in the topic "World Cup 2002", each match could also be taken as an event. For each event, there are usually many documents reporting it. Therefore, in order to generate an informative and expressive storyboard to present the topic, it would be better to decompose the obtained information and cluster the documents into different events. However, event definition is usually subjective, different individuals may have different opinions. It is also confusing in which scale an event should be defined. Also take "World Cup" as an example, in a larger scale, "World Cup 2002" and "World Cup 2006" could also be considered as a big event. Therefore, due to the above vagueness, in this paper, we do not strictly define 746 each event of the target topic. Following our previous works on news event detection [14], an event is assumed as some similar information describing similar persons, similar keywords, similar places, and similar time duration. Therefore, in our system, an event is represented by four primary elements: who (persons), when (time), where (locations) and what (keywords); and event clustering is to group the documents reporting similar primary elements. As for the scale of event, in the paper, it could be adaptively determined by the time range of the obtained documents or the required event number. In this section, we present a novel clustering approach based on a generative model proposed in [14], instead of using traditional clustering methods such as K-means. After event clusters are obtained, the corresponding event summary is then extracted and other representative media is associated with each event. 3.1 Multiple Event Clustering To group the documents into different events, essentially, we need to calculate p(e j | x i ), which represents the probability that a document x i belongs to an event e j . Here, as mentioned above, an event e j (and thus the document x i describing the event) is represented by four primary elements: who (persons), when (time), where (locations) and what (keywords). That is, } , , , { / time keywords locations persons Docment Event = Assuming that a document is always caused by an event [14] and the four primary elements are independent, to calculate the probability p(e j | x i ), in our approach, we first determine the likelihood that the document x i is generated from event e j , p(x i | e j ) which could be further represented by the following generative model, ) | ( ) | ( ) | ( ) | ( ) | ( j i j i j i j i j i e time p e key p e loc p e name p e x p = (1) where name i , loc i , key i , and time i are the feature vectors representing persons, locations, keywords and time in the document x i , respectively. In our approach, the above entities are extracted by the BBN NLP tools [15]. The tool can extract seven types of entities, including persons, organizations, locations, date, time, money and percent. In our approach, the obtained organization entity is also considered as a person entity; and all the words except of persons, locations, and other stop-words are taken as keywords. In more detail, name i (similarly, loc i and key i ) is a vector &lt;c i1 , c i2 , ..., c iNp &gt;, where c in is the occurrence frequency of the person n appears in the document x i , and person n is the nth person in the person vocabulary, which is composed of all the persons appeared in all the obtained documents (similarly, we can define keyword vocabulary and location vocabulary). Assuming N p is the size of person vocabulary, p(name i |e j ) could be further expressed by = = p in N n c j n j i e person p e name p 1 ) | ( ) | ( (2) Since the person, location and keyword are discrete variables represented by words, and the probability of the location and keyword can be also defined similarly as that of the person in (2), in the flowing sections, we will not discriminate them and uniformly represent the probability p(person n | e j ) (correspond-ingly , the p(location n | e j ) and p(keyword n | e j )) as p(w n | e j ), which denotes the probability that the word w n appears in the event e j On the other hand, the time of an event usually lasts a continuous duration. It is also observed, especially in the news domain, that the documents about an event usually increases at the beginning stage of the event and then decreases at the end. Therefore, in our approach, a Gaussian model N(u j , j ) is utilized to roughly represent the probability p(time i | e j ), where u j and j is the mean and standard deviation, respectively. To this end, in order to estimate the probability p(e j | x i ), we need to estimate the parameters = {p(w n | e j ), u j , j , 1jK}, assuming K is the number of events (the selection of K is discussed in section 3.2). In our approach, the Maximum Likelihood is used to estimate the model parameters, as, = = = = = = M i K j j i j M i i e x p e p x p X p 1 1 1 * )) , | ( ) ( log( max arg )) | ( log( max arg )) | ( log( max arg (3) where X represents the corpus of the obtained documents; M and K are number of documents and events, respectively. Since it is difficult to derive a close formula to estimate the parameters, in our approach, an Expectation Maximization (EM) algorithm is applied to maximize the likelihood, by running E-step and M-step iteratively. A brief summary of these two steps is listed as follows, and more details can be found in [14]. In E-step, the posterior probability p(e j | x i ) is estimated as: ) ( ) ( ) | ( ) | ( ) ( ) ( ) 1 ( i t j t j i t i j x p e p e x p x e p = + (4) where the upper script (t) indicate the tth iteration. In M-step, the model parameters are updated, as, + + = = = + = + + M i N s t i j M i t i j t j n s i tf x e p N n i tf x e p e w p 1 1 ) 1 ( 1 ) 1 ( ) 1 ( ) ) , ( ) | ( ( ) , ( ) | ( 1 ) | ( (5) = = + = + + M i t i j M i i t i j t j x e p time x e p u 1 ) 1 ( 1 ) 1 ( ) 1 ( ) | ( ) | ( (6) = = + = + + + M i t i j M i tj i t i j t j x e p u time x e p 1 ) 1 ( 1 2 ) 1 ( ) 1 ( ) 1 ( 2 ) | ( ) ( ) | ( (7) where tf(i,n) is the term frequency of the word w n in the document x i and N is the corresponding vocabulary size. It is noted that, in (5), the Laplace smoothing [16] is applied to prevent zero probability for the infrequently occurring word. At last, the prior of each event is updated as: M x e p e p M i t i j t j = = + + 1 ) 1 ( ) 1 ( ) | ( ) ( (8) The algorithm can increase the log-likelihood consistently with the iterations; and then converge to a local maximum. Once the parameters are estimated, we can simply assign each document to an event, as following )) | ( ( max arg i j j i x e p y = (9) where y i is the event label of the document x i . 747 The advantage of this generative approach is that it not only considers the temporal continuity of an event, it also can deal with the issue that some events overlap in some time durations. In this case, the Gaussian model of the event time can also be overlapped through this data-driven parameter estimation. From this view, the event clustering is also like a Gaussian mixture model (GMM) estimation in the timeline. 3.2 Determining the Number of Events In the above approach to event clustering, the event number K is assumed known (as shown in (3)-(8)). However, the event number is usually very difficult to be determined a priori. In our approach, an intuitive way is adopted to roughly estimate the event number based on the document distribution along with the timeline. As mentioned above, it is assumed that each document is caused by an event, and the document number of an event changes with the development of the event. According to this property, each peak (or the corresponding contour) of the document distribution curve might indicate one event [14], as the Fig. 2 shows. Thus, we can roughly estimate the event number by simply counting the peak number. However, the curve is quite noisy and there inevitably exist some noisy peaks in the curve. In order to avoid the noisy peaks, in our approach, only the salient peaks are assumed to be relevant to the event number. To detect the salient peaks, we first smooth the document curve with a half-Hamming (raised-cosine) window, and then remove the very small peaks with a threshold. Fig.2 illustrates a smoothed document distribution with the corresponding threshold, collected on the topic "US Election" in four months. In experiments, the threshold is adaptively set as d d /2, where d and d are the mean and standard deviation of the curve, respectively. After the smoothing and tiny peaks removal, we further detect the valleys between every two contingent peaks. Thus, the range of an event (which is correlated to the corresponding peak) can be considered as the envelope in the two valleys. As shown in Fig2, the duration denoted by L i +R i is a rough range of the event correlated to the peak P i . Assuming an important event usually has more documents and has effects in a longer duration, the saliency of each peak is defined as, ) )( ( avr i i avr i i D R L P P S + = (10) where P i is the ith peak, L i and R i is the duration from the ith peak to the previous and next valley; P avr is the average peak value and D avr is average duration between two valleys in the curve. S i is the saliency value of the peak P i . It could also be considered as the normalized area under peak P i , and thus, it roughly represents the document number of the corresponding event. In our approach, the top K salient peaks are selected to determine the event number: } / { max arg ' 1 1 ' = = = i N i k i i k S S K (11) where ' i S is the sorted saliency value from large to small, N is total number of detected peaks and is a threshold. In our experiments, is set as 0.9, which roughly means that at least 90% documents will be kept in the further initialization of event clustering. This selection scheme is designed to guarantee there is no important information is missed in presentation. After the event number and initial clusters (the most salient peaks with their corresponding range) are selected, the event parameters could be initialized and then updated iteratively. 0 5 10 15 20 0 20 40 60 80 100 120 #Doc Threshold P i P i+1 P i-1 L i R i Peaks relevant to event Fig.2 Peak saliency definition. It also illustrates the smoothed document distribution (document number per day) with the corresponding threshold for tiny peak removal. Each peak P i is assumed to be correlated with each event. It is noted that some technology such as Bayesian Information Criteria (BIC) or minimum description length (MDL) [17] could be used to estimate the optimal event number, by searching through a reasonable range of the event number to find the one which maximizes the likelihood in (3). However, these algorithms take long time, and it is usually not necessary to estimate the exact event number in our scenario of rich presentation. Actually, in our system, the most important point of event clustering is that the clustered documents `really' represent the same event, rather than the event number, as observed in the experiments . Moreover, in the step of synchronization between the music and storyboard (in the section 4.2), the number of presented events may be further refined, based on the user's preference, in order to match the presentation duration with the music duration. 3.3 Event Description After obtaining the events and the corresponding documents, we not only need a concise event summary, but also need to extract some representative media to describe each event. 3.3.1 Event Summary A simple way to summarize an event is to choose some representative words on the persons, locations and keywords of the event. For example, for the event e j , the `leading actor' could be chosen as the person with the maximum p(person n | e j ), while the major location could be selected based on p(location n | e j ). However, such brief description might have a bad readability. Therefore, in order to increase the readability of the summary, in our system, we also provide an alterative way. That is, we choose a candidate document to represent an event. For example, the document with the highest p(x i |e j ) is a good candidate representative of the event e j . However, a document might be too long to be shown on the storyboard. Therefore, in our system, only the "title-brow" (the text between the news title and news body) of the document, which usually exists and is usually a good overview (summary) of the document based on our observation (especially true in our case of news document), is selected to describe the event. 748 I III II IV Fig. 3 The event template of the Storyboard, which illustrates (I) the representative media, (II)geographic information, (III) event summary, and (IV) a film strip giving an overview of the events in the temporal order. 3.3.2 Extracting Representative Media In the obtained documents describing an event, there are usually many illustrational images, with possible flashes and video clips. These media information is also a good representative of the corresponding event. However, since the obtained documents are directly crawled from the news websites, they usually contain many noisy multimedia resources, such as the advertisements. Moreover, there also possible exist some duplicate images in different documents describing the same event. Therefore, to extract the representative media from the documents, we need to remove noisy media and possible duplicate images. Before this, we also performed a pre-filtering to remove all the images smaller than 50 pixels in height or width. Noisy Media Detection. In our approach, a simple but efficient rule is used to remove the noisy media resources. We find almost all advertisements are provided by other agencies rather than these news websites themselves. That is, the hosts of advertisement resources are from different websites. Thus, in our approach, we extract the host names from the URLs of all multimedia resources, and remove those resources with different host name. Duplicate Detection. A number of image signature schemes can be adopted here to accomplish duplicate detection. In our implementation, each image is converted into grayscale, and down-sampled to 88. That is, a 64-byte signature for each image is obtained. Then the Euclidean distance of the 64-byte signature are taken as the dissimilarity measure. Images have sufficiently small distance are considered as duplicates. Once removing the noisy resources and duplicate images, we simply select the 1-4 large images from the top representative documents (with the top largest p(x i |e j )), and take them as representative media of the corresponding event. The exact number of the selected images is dependent on the document number (i.e., the importance) of the event and the total image number the event has. It is noted that, in our current system, we only associates images with each event. However, other media like video and flashes can be chosen in a similar way. RICH PRESENTATION GENERATION In the proposed system, the above obtained information, including event summary and representative media, are fused to generate a concise and informative storyboard, in order to richly present the target topic. In this section, we will first describe the storyboard generation for the target topic, by presenting each event with the multimodal information. Then, we present the approach to synchronizing the storyboard with an incidental music. 4.1 Storyboard Generation In our approach, a storyboard of a target topic is generated by presenting each event of the topic slide by slide. To describe an event, we have obtained the corresponding information including the person, time, location, event summary and other relevant images. Therefore, to informatively present each event, we need first to design an event template (i.e., an interface) to integrate all the information. Fig. 3 illustrates the event template used in our proposed system, with an example event in the topic `US Election". First, the template presents the representative images in the largest area (part I), since the pictures are more vivid than the words. As for each representative picture, the title and date of the document from which it is extracted is also illustrated. In the Fig.3, there are 4 pictures extracted from 3 documents. Then, the corresponding event summaries of these three documents are presented (part III), where each paragraph refers to the summary of one document. If a user is interested in one document, he can click on the corresponding title to read more details. Moreover, the geographic information of the event is shown with a map in the top-left corner (part II), to give users a view of the event location. The map is obtained from "MapPoint Location" service [18], which can return a 749 corresponding map based on user's location query. However, the mapping is usually difficult, especially when the event location is confusing so that the representative location is not accurately detected. For example, the event shown in the Fig 1 is mapped to Washington D.C. rather than New York where the republic convention is held, since Washington is the most frequently mentioned places in the documents. Finally, a film strip (part IV) is also presented, arranging each event in the temporal order, where each event is simply represented by a cluster of images, with the current event highlighted. It enables users to have a quick overview of the past and the future in the event sequence. By connecting various events slide by slide, we could get an informative storyboard regarding the target topic. In order to catch the development process of a topic, the events are ordered by their timestamps in the generated storyboard. 4.2 Synchronizing with Music To make the storyboard more expressive and attractive, and to provide a more relaxing way to read information, in the proposed system, we will accompany the storyboard with an incidental music and align the transitions between event slides with the music beats, following the idea in music video generation [19][20]. Sometimes, music could also provide extra information about the target topic. For example, when the target topic is a movie, the corresponding theme song could be chosen for the rich presentation . In this sub-section, we will present our approach to music analysis and synchronization with the storyboard. 4.2.1 Music Rhythm Analysis In the proposed system, we detect the onset sequences instead of the exact beat series to represent music rhythm. This is because the beat information is sometimes not obvious, especially in light music which is usually selected as incidental music. The strongest onset in a time window could be assumed as a "beat". This is reasonable since there are some beat positions in a time window (for example, 5 seconds); thus, the most possible position of a beat is the position of the strongest onset. The process of onset estimation is illustrated in Fig. 4. After FFT is performed on each frame of 16ms-length, an octave-scale filter-bank is used to divide the frequency domain into six sub-bands, including [0, 0 /2 6 ), [ 0 /2 6 , 0 /2 5 ), ..., [ 0 /2 2 , 0 /2], where 0 refers to the sampling rate. Acoustic Music Data FFT Sub-Band 1 Envelope Extractor Difference curve Onset Curve Sub-Band N Envelope Extractor Difference curve ... ... ... . . . . . . Fig. 4 The process of onset sequence estimation After the amplitude envelope of each sub-band is extracted by using a half-Hamming window, a Canny operator is used for onset sequence detection by estimating its difference function, ) ( ) ( ) ( n C n A n D i i = (12) where D i (n) is the difference function in the ith sub-band, A i (n) is the amplitude envelope of the ith sub-band, and C(n) is the Canny operator with a Gaussian kernel, ] , [ ) ( 2 2 / 2 2 c c i L L n e i n C = (13) where L c is the length of the Canny operator and is used to control the operator's shape, which are set as 12 and 4 in our implementation, respectively. Finally, the sum of the difference curves of these six sub-bands is used to extract onset sequence. Each peak is considered as an onset, and the peak value is considered as the onset strength. Based on the obtained onsets, an incidental music is further segmented into music sub-clips, where a strong onset is taken as the boundary of a music sub-clip. These music sub-clips are then used as the basic timeline for the synchronization in the next step. Thus, to satisfy the requirement that the event slide transitions of the storyboard should occur at the music beats, we just need to align the event slide boundaries and music sub-clip boundaries. To give a more pleasant perception, the music sub-clip should not be too short or too long, also it had better not always keep the same length. In our implementation, the length of music sub-clips is randomly selected in a range of [t min , t max ] seconds. Thus, the music sub-clips can be extracted in the following way: given the previous boundary, the next boundary is selected as the strongest onset in the window which is [t min , t max ] seconds away from the previous boundary. In the proposed system, users can manually specify the range of the length of the music sub-clip. The default range in the system is set as [12, 18] seconds, in order to let users have enough time to read all the information on each event slide. 4.2.2 Alignment Scheme To synchronize the transitions between different event slides and the beats of the incidental music, as mentioned above, we actually need to align the slide boundaries and music sub-clip boundaries. To satisfy this requirement, a straightforward way is to set the length of each event slide be equal to the corresponding length of the sub-music clip. However, as Fig. 5 illustrates, the number of event slides is usually not equal to the number of music sub-clip. In this case, in our proposed system, we provide two schemes to solve this problem. 1) Music Sub-clip Based. In this scheme, only the top N important events of the target topic are adaptively chosen and used in the rich presentation, where N is supposed as the number of music sub-clip in the corresponding incidental music, as the Fig.5 shows. Although a formal definition of event importance is usually hard and subjective, in our approach, the importance score of an event is simply measured by the number of documents reporting it, assuming that the more important the event, the more the corresponding documents. The assumption is quite similar as that in the definition of (10). 750 2) Specified Event Number Based. In this scheme, users can specify the number of the event he wants to learn. For example, a user could choose to show the top 30 important events or all the events. Thus, to accommodate all the events in the music duration, we will repeat the incidental music if it is needed and then fade out the music at the end. E1 E2 E3 E4 E5 E6 E8 E7 S2 S1 S3 ....... S4 S5 ....... Event Slide List Music Sub-Clip Fig. 5 Music and storyboard synchronization: a music sub-slip based scheme, that is, only the top important events are presented to match the number of music sub-clips. 4.2.3 Rendering After the alignment between storyboard and incidental music, in our system, fifteen common transition effects, such as cross-fade, wipe and dissolve, are also randomly selected to connect the event slides, producing a better rich presentation in final rendering. EVALUATIONS In this section, we evaluate the performance of the proposed approach to rich presentation and its key component, event clustering. In the experiments, we randomly select 8 topics of different types, including Earthquake, Halloween, Air Disaster, US Election, Nobel Prize, Britney Spears, David Beckham, and Harry Potter, from some hot news topics in the end of 2004 and beginning of 2005. Once the topic is selected, the topic name is used as a query and the relevant documents are collected from CNN, MSNBC and BBC. More details about the selected topics and the corresponding documents are shown in the Table 1, which lists the topic name, the time range of the collected documents, and the number of documents and its corresponding events. Table 1. A list of testing topics in the rich presentation evaluations No. Topic Time #doc #event 1 Earthquake 1995-2004 976 17 2 Halloween 1995-2004 762 9 3 Air Disaster 1995-2004 210 13 4 US Election 1995-2004 2486 -5 Britney Spears 2000-2004 1311 -6 Nobel Prize 1995-2004 186 -7 David Beckham 1995-2004 877 -8 Harry Potter 2000-2004 841 -Total --7649 It is noted that, in the table, only 3 topics have labeled events, while another 5 topics have not. This is because that, the labeling work of a topic is very subjective and usually hard for individuals to manually decide the event number of a given topic. Therefore, we only label the topics which are easily to be annotated based on the criterion in Topic Detection and Tracking (TDT) project [21]. For example, Halloween is a topic which is reported once a year, thus, each year's documents can be regarded as an event; as for Earthquake and Air Disaster, their events lists could be found from corresponding official websites. In the annotation, we remove the events which do not have or have few (less than 4) relevant documents, and also remove the documents not belonging to any events. After parsing the obtained documents, for each topic, we usually can obtain 3.8 images per document in average. With further duplicate detection, only 1.6 images per document are remained. Moreover, from each document, we could also obtain about 3.0 unique location entities and 2.8 unique name entities. Other words except of these entities are taken as keywords. Fig.6 shows a real representation of an example document with extracted entities in the XML format, from which the event clustering is performed. Fig. 6. XML representation of a document on "US Election" with extracted entities 5.1 Event Clustering As mentioned above, the evaluation of the approach to event clustering is evaluated on three topics, including Earthquake, Halloween , and Air Disaster, for which the corresponding event numbers are determined and the documents are labeled using a similar method in the TDT project. However, in the proposed appraoch, we actually do not estimate the optimal event number, but use a much larger one. Therefore, in order to better evaluate the performance of the event clustering algorithm and compare with its counterpart, we use the event number in the ground truth to initialize the cluster number in the proposed clustering algorithm. ..... &lt;URL&gt;http://news.bbc.co.uk/1/hi/world/americas/4071845.stm &lt;/URL&gt; &lt;Abstract&gt;The US battleground state of Ohio has certified the victory of President George W Bush's in last month's poll. &lt;/Abstract&gt; &lt;Date&gt; 2004/12/6 &lt;/Date&gt; &lt;NLPRESULT&gt; &lt;LOCATION&gt; &lt;entity&gt; Ohio &lt;/entity&gt; &lt;freq&gt;4&lt;/freq&gt; &lt;entity&gt; US &lt;/entity&gt; &lt;freq&gt; 2 &lt;/freq&gt; &lt;/LOCATION&gt; &lt;PERSON&gt; &lt;entity&gt; Bush &lt;/entity&gt; &lt;freq&gt; 3 &lt;/freq&gt; &lt;entity&gt;David Cobb&lt;/entity&gt; &lt;freq&gt;1&lt;/freq&gt; ... &lt;/PERSON&gt; ... &lt;DATE&gt; &lt;entity&gt; 6 December, 200&lt;/entity&gt; &lt;freq&gt; 1 &lt;/freq&gt; &lt;entity&gt; Friday &lt;/entity&gt; &lt;freq&gt; 2 &lt;/freq&gt; ... &lt;/DATE&gt; &lt;KEYWORDS&gt; ... &lt;entity&gt; recount &lt;/entity&gt; &lt;freq&gt;7&lt;/freq&gt; &lt;entity&gt; elect &lt;/entity&gt; &lt;freq&gt;3&lt;/freq&gt; &lt;entity&gt; America &lt;/entity&gt; &lt;freq&gt;3&lt;/freq&gt; &lt;entity&gt; poll &lt;/entity&gt; &lt;freq&gt;3&lt;/freq&gt; ... &lt;/KEYWORDS&gt; &lt;/NLPRESULT&gt; 751 In the experiments, K-means, which is another frequently used clustering algorithm (as well in TDT [22]), is adopted to compare with the proposed approach. The comparison results of two clustering approaches are illustrated in Table 2, with precision and recall for each topic. Table 2. The performance comparison between our approach and K-means on the event clustering Precision Recall K-means Ours K-means Ours Earthquake 0.74 0.87 0.63 0.74 Halloween 0.88 0.93 0.72 0.81 Air Disaster 0.57 0.68 0.55 0.61 Average 0.73 0.83 0.63 0.72 From Table 2, it can be seen that the results of our approach are significantly better than those of K-means, both on precision and recall. On the three testing topics, the average precision of our approach is up to 0.83 and the average recall achieves 0.72, which is 10% and 9% higher than those of K-means, respectively. By tracing the process of K-means, we find that K-means usually assigns documents far away from each other on the timeline into the same cluster, since the time information affects little in K-means . It also indicates the advantages of our approach with time modeling. The algorithms also show different performance on different kind topics. As for the "Air disaster", its performance is not as good as that of the other two, since the features (words and time) of its events are more complicated and intertwined in the feature space. As for the topics (4-8 in Table I) which could not have an objective evaluation, the clustering performance on these topics could be indirectly reflected by the subjective evaluation of the rich presentation presented in section 5.2. This is because users will be more satisfied when the grouped documents shown in each event slide really belong to the same event; while users are not satisfied if the documents from different events are mixed in one event slide. 5.2 Rich Presentation It is usually difficult to find a quantitative measure for rich presentation, since the assessment of the goodness of rich presentation is a strong subjective task. In this paper, we carry out a preliminary user study to evaluate the performance of the proposed rich presentation schemes. To indicate the performance of rich presentation, we design two measures in the experiments, including `informativeness' and `enjoyablity', following the criteria used in the work [7]. Here, the informativeness measures whether the subjects satisfy with the information obtained from the rich presentation; while enjoyablity indicates if users feel comfortable and enjoyable when they are reading the rich presentation. In evaluating the informativeness, we also provide the documents from which the rich presentation is generated. They are used as baseline, based on which the subjects can more easily evaluate if the important overview information contained in the documents is conveyed by the rich presentation. Moreover, in order to reveal the subjects' opinion on the design of the storyboard template, like the one shown in Fig 3, we also ask the subjects to evaluate the `interface design'. In the user study, 10 volunteered subjects including 8 males and 2 females are invited. The subjects are around 20-35 years old, have much experience on computer manipulation, and usually read news on web in their leisure time. We ask them to give a subjective score between 1 and 5 for each measure of the rich presentation of each testing topic (an exception is `interface design', which is the same for each rich presentation). Here, the score `1' to `5' stands for unsatisfied (1), somewhat unsatisfied (2), acceptable (3), satisfied (4) and very satisfied (5), respectively. In experiments, we first check with the `interface design' measure. We find 7 out of 10 subjects satisfy with the event template design and the left three also think it is acceptable. The average score is up to 3.9. An interesting observation is that, some subjects like the template design very much at the first glance, but they feel a little boring after they finish all the user study since every slide in the rich presentation of each topic has the same appearance. It hints us that we had better design different templates for different topics to make the rich presentation more attractive. As for the other two measures, we average the score across all the subjects to represent the performance for each topic, and list the detailed results in Table 3. It can be seen that the average score of both enjoyablity and informativeness achieves 3.7, which indicates that most subjects satisfy the provided overview information of the target topic, and they enjoy themselves when reading these rich presentations. Table 3. The evaluation results of rich presentation on each topic No. Topic Informative Enjoyable 1 Earthquake 4.3 3.2 2 Halloween 3.6 4.0 3 Air Disaster 4.0 3.4 4 US Election 4.1 4.0 5 Britney Spears 3.6 4.1 6 Nobel Prize 3.3 3.4 7 David Beckham 3.4 4.0 8 Harry Potter 3.3 3.4 Average 3.7 3.7 In the experiments, we find informativeness is highly depended on the correlation between the presented documents and the target topic. If the presented information is consistent with the topic, subjects usually give a high score for informativeness, such as those on Earthquake and US Election; otherwise, they will give a low score, like those on David Beckham and Nobel Prize. It indicates that it is quite important to provide users clean information of the target topic with less noise. However, in current system, the documents are crawled from web and inevitably contain many noises. It affects much on the performance of informativeness in the current system. We need to consider how to prone the information of the target topic in the future works. We also find that the enjoyablity score is usually related with informativeness. If the subjects do not get enough information from the rich presentation, they will be not enjoyable as well, such as the topics of Nobel Prize and Harry Potter. Enjoyablity is also topic-related, the subjects usually feel unconformable when they are facing with miserable topics, such as Earthquake and Air Disaster, although their informativeness is quite high. On the 752 contrary, users give a high score for enjoyablity on the interesting topics, such as Britney Spears and David Beckham, although their informative score is not high. This is because that there are usually many funny and interesting pictures in the presentation of these topics. Another finding is that users usually fell unenjoyable if the images and summaries in one event slide are not consistent with each other. From this view, the high enjoyablity score in our experiments also indicates that our event clustering algorithm works promisingly CONCLUSIONS To facilitate users to quickly grasp and go through the content of a semantic topic, in this paper, we have proposed a novel approach to rich presentation to generate a concise and informative storyboard for the target topic, with many relevant multimodal information including image, text, audio and video. In this approach, the related multimodal information of a given topic is first extracted from news databases. Then, the events are clustered, and the corresponding information, such as representative images, geographic information, and event summary, is obtained. The information is composed into an attractive storyboard which is finally synchronized with incidental music. A user study indicates that the presented system works well on our testing examples. There is still some room for improving the proposed approach. First, the proposed approach could be extended to other multimedia databases or more general websites. For example, some standard multimedia database like NIST TRECVID could provide a nice platform for the implementation and evaluation of event detection and rich presentation. Second, to integrate more relevant multimedia information (such as video clips and flashes) and more accurate information regarding the target topic is highly expected by users. Thus, more advanced information retrieval/ extraction techniques and other multimedia analysis techniques are needed to be exploited and integrated, such as relevance ranking, mapping schemes, important or representative video clips detection and video clip summarization. We also need to design a much natural way to incorporate video clips in the event template. Third, we also consider designing various storyboard templates for different kind of topics. For example, each topic may be belonging to different clusters such as politics, sports and entertainments, each of which can have a representative template. Forth, appropriate user interaction will be added to further make the storyboard more interactive and easy to control. Finally, a thorough evaluation will be implemented to evaluate the effect of each component in the framework and storyboard template. REFERENCES [1] A. Vailaya, M.A.T. Figueiredo, A. K. Jain, and H.-J. Zhang. "Image classification for content-based indexing". IEEE Transactions on Image Processing, Vol.10, Iss.1, 2001 [2] F. J., M.-J. Li, H.-J. Zhang, and B. Zhang. "An effective region-based image retrieval framework". Proc. ACM Multimedia'02, pp. 456-465, 2002 [3] J. Platt "AutoAlbum: Clustering Digital Photographs using Probabilistic Model Merging" Proc. IEEE Workshop on Content-Based Access of Image and Video Libraries, pp. 96 100, 2000. [4] A. Hanjalic, R. L. Lagendijk, J. Biemond, "Automated high-level movie segmentation for advanced video-retrieval systems", IEEE Trans on Circuits and Systems For Video Technology, Vol. 9, No. 4, pp. 580-588, 1999. [5] J. Assfalg and et al, "Semantic annotation of soccer videos: automatic highlights identification,&quot; CVIU'03, vol. 92, pp. 285-305, 2003. [6] A. Ekin, A. M. Tekalp, and R. Mehrotra, &quot;Automatic soccer video analysis and summarization,&quot; IEEE Trans. on Image Processing, 12(7), pp. 796-807, 2003. [7] Y. -F. Ma, L. Lu, H. -J. Zhang, and M.-J Li. "A User Attention Model for Video Summarization". ACM Multimeida'02, pp. 533-542, 2002. [8] L. Xie, P. Xu, S.F. Chang, A. Divakaran, and H. Sun, &quot;Structure analysis of soccer video with domain knowledge and hidden markov models,&quot; Pattern Recognition Letters, vol. 25(7), pp. 767-775, 2004. [9] L. Lu, H. Jiang, H. J. Zhang, "A Robust Audio Classification and Segmentation Method," Proc. ACM Multimedia'01, pp. 203-211, 2001 [10] R. Cai, L. Lu, H.-J. Zhang, and L.-H. Cai, "Highlight Sound Effects Detection in Audio Stream," Proc. ICME'03 Vol.3, pp.37-40, 2003. [11] Y. Rui, A. Gupta, and A. Acero, "Automatically Extracting Highlights for TV Baseball Programs", Proc. ACM Multi-media'00 , pp.105-115, 2000. [12] C. Snoek, and M. Worring. "Multimodal Video Indexing: A Review of the State-of-the-art". Multimedia Tools and Applications, Vol. 25, No. 1 pp. 5 35, 2005 [13] E.M. Voorhees, "Query expansion using lexical-semantic relations" Proc. ACM SIGIR Conference on Research and Development in Information Retrieval , pp 61 - 69, 1994 [14] Z.-W. Li, M.-J. Li, and W.-Y. Ma. &quot;A Probabilistic Model for Retrospective News Event Detection", Proc. SIGIR Conference on Research and Development in Information Retrieval, 2005 [15] D. M. Bikel, R. L. Schwartz, and R. M. Weischedel. "An Algorithm That Learns What's in a Name". Machine Learning, 34(1-3), 1999 [16] K. Nigam, A. McCallum, S. Thrun, and T. Mitchell. "Text Classification from Labeled and Unlabeled Documents using EM". Machine Learning, 39(2-3), 2000 [17] T. Hastie, R. Tibshirani, and J. Friedman. "The Elements of Statistical Learning: Data Mining, Inference and Prediction". Springer-Verlag, 2001 [18] MapPoint Web Service http://www.microsoft.com/mappoint/ products/ webservice/default.mspx [19] X.-S. Hua, L. Lu, H.-J. Zhang. &quot;Automated Home Video Editing&quot;, Proc. ACM Multimedia'03, pp. 490-497, 2003 [20] J. Foote, M. Cooper, and A. Girgensohn. "Creating Music Videos Using Automatic Media Analysis". ACM Multimedia'02, pp.553-560, 2002. [21] Topic Detection and Tracking (TDT) Project: http://www. nist.gov/speech/tests/tdt/ [22] J. Allan, R. Papka, and V. Lavrenko. "On-line New Event Detection and Tracking". Proc. SIGIR Conference on Research and Development in Information Retrieval 98, pp.37-45, 1998 753

documentary and movie;Rich presentation;events clustering;Communication and Multimedia;Representative Media;Images, videos and Audio Technologies;Rich video clips and flashes;Multi-modal information;Generate storyboard;storyboard;Subjective multiple events;multimedia fusion;High-level semantics;Event clustering;multimodality;multimedia authoring

4

A Database Security Course on a Shoestring

Database security has paramount importance in industrial, civilian and government domains. Despite its importance, our search reveals that only a small number of database security courses are being offered. In this paper, we share our experience in developing and offering an undergraduate elective course on database security with limited resources. We believe that database security should be considered in its entirety rather than being component specific. Therefore , we emphasize that students develop and implement a database security plan for a typical real world application . In addition to the key theoretical concepts, students obtain hands-on experience with two popular database systems . We encourage students to learn independently making use of the documentation and technical resources freely available on the Internet. This way, our hope is that they will be able to adapt to emerging systems and application scenarios.

INTRODUCTION Database systems are designed to provide efficient access to large volumes of data. However, many application domains require that the data access be restricted for security reasons. For example, an unauthorized access to 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. SIGCSE'06, March 1�5, 2006, Houston, Texas, USA. Copyright 2006 ACM 1-59593-259-3/06/0003... $ 5.00. a bank database can potentially cost millions of dollars. The federal Health Insurance Portability and Accountability Act (HIPAA) regulates the disclosure of information from a patient database, allowing access to health care providers, health plans, and health care clearinghouses, simultaneously protecting the privacy of patients. For obvious reasons, a Department of Defense (DoD) database needs to be protected from unauthorized access. Since many organizations increasingly entrust their information resources with database systems, in today's highly networked environment, the sensitive information can be at high risk unless there are security mechanisms in place to protect the data at the source itself. However, a large number of databases are incorrectly installed, configured, and maintained. This, in part, may be attributed to the lack of database security education in our computer science programs. We feel that a new undergraduate course on database security will help our students face the ever increasing challenges in this field. Our search shows that, despite the importance, only a handful database security courses are being offered. Most of the courses we found are graduate courses and are highly theoretical. We also found a few extension program courses, which are product specific. Although a large number of database courses exist at both undergraduate and graduate levels, we feel that, one reason for not offering database security courses may be the scarcity of textbooks, reference materials, and other resources. Realizing the importance of database security in computer science curriculum, [8] proposes adding a new module to the basic database course. Since the basic database course has already many topics to cover, we feel that the addition of new material will not completely serve the purpose. Further, we find it difficult to incorporate hands-on component to such a course. Similarly, a computer security course is too broad in scope, and rarely includes database security topics. Therefore, we decided to develop a new undergraduate level elective on database security. This paper is based on our experience of offering a database security course in Spring 2005. We have adjusted the contents and assignments in response to the feedback and course outcome. The modified version is presented here. Since many of our students seek industrial positions after graduation, we have designed our course to meet their needs with the right blend of theory and practice. The course objective is to develop an understanding of security aspects of databases, database administration, and database supported applications. We collected information from alumni as well as potential employers before finalizing the contents. 7 Although students were expected to gain hands-on experience with some popular databases in our course, we tried to focus on concepts rather than just syntax or product specific features. Often, students were asked to learn software packages on their own by reading the product documentation. We also offered an online feedback page for receiving anonymous student comments, which helped us know if students needed additional assistance. By encouraging students to learn and experiment on their own, we hope that they can easily apply the learned concepts to emerging application scenarios. This is particularly needed since today's work environment expects agility from employees to quickly master and develop software systems. To facilitate participation further, we asked students to research and make presentations choosing from a set of specified topics. Most impor-tantly , since many of our educational institutions are cash strapped, we designed our course to execute with a small budget. In the next section, we detail topics, which may be included in a database security course, with references that, we hope, will be useful for other instructors. We also discuss labs and assignments in detail. Finally, we conclude the paper with an account of lessons learned and future possibilities DATABASE SECURITY TOPICS Although a large number of topics can be included, we try to focus on a few important ones that, in our judgment, are likely to be immediately useful after graduation. We also include topics on securing the data within a database, as well as the security of database systems and operating systems as suggested in [8]. Our position is that database security should be considered in whole rather than adopting a piecemeal approach. However, we recognize that, in practice, it is often easy to overlook some aspects of database security. Therefore, we recommend that students develop a database security plan. We also include other relevant topics such as statistical database security, and security and privacy issues of data mining. Table 1 shows the schedule of topics for a typical sixteen week semester course on database security. Major labs and assignments are given in Table 2. The course begins with an "Introduction to database se-curity" , where the objective is to highlight the importance of database security and to motivate students to learn the rest of the topics. 2.1 Introducing Database Security One way to emphasize the importance of database security would be to reflect on the impact of not having security at all in application domains such as military, medical, financial , credit card, credit file, driving records, and insurance databases. Students may survey the incidents of database security breaches and evaluate the efforts to ensure database security by industry and government. Since database security is a combination of database technology and computer security, basics of both will be helpful . A discussion on security properties such as confiden-tiality , integrity, availability and non-repudiation should be included. Although an in-depth study of cryptography is not within the scope of this course, basics of secret key cryptography and public key cryptography will benefit students. A good reference book we found is "Data Security and Cryptogra-Week Topic 1 Course Overview and Introduction to Database Security, Basics of Data Security and Cryptography 2 Overview of Security Models 3 Access Control Models, Covert Channels and Inference Channels 4 MySQL Security 5 Oracle Security 6 Oracle Label Security 7 Developing a Database Security Plan 8 Spring Break 9 SQL Server Security 10 Security of Statistical Databases 11 Security and privacy issues of Data Mining 12 Database Applications Security, SQL Injection, Defensive Programming 13 Database Intrusion Prevention, Audit, Fault Tolerance and Recovery 14 Hippocratic Databases, XML Security 15 Network Security, Biometrics 16 Final Examination Week Table 1: Course Schedule phy" by Dorothy Denning [5]. Digital signatures, digital certificates and Public Key Infrastructure (PKI) [21] are other topics to consider. An overview of security and integrity models [4] will also be helpful at this point. This is the best time to introduce the computer security lingo such as subjects and objects. The difference between widely used access control techniques may also be highlighted. 2.2 Access Control Discretionary Access Control (DAC) mechanisms such as capabilities, profiles, access control lists, passwords, and permission bits may be discussed. Here we also introduce the operating system security aspects (using Windows R and Linux environments), and how they impact database security in general. Although details are not required until we introduce Oracle security, overview of Role-Based Access Control (RBAC) [6, 18] may be discussed. Unlike the above access control techniques, in Mandatory Access Control (MAC) the security is enforced by the system as dictated in the security policy, not by the owner of an object. Although there are many security models suggested for providing Mandatory security, Bell-LaPadula [2] model is probably the simplest to learn. Even when a system enforces Mandatory Access Control, information leakage through covert channels [11] and inference channels [13] may still be possible. A few examples will help students understand how the information leakage can take place through such means. Databases enforcing MAC often assign security classification levels for objects and security clearance levels for subjects. Access control is performed by the system based on these levels. A lab may be developed, where students 8 simulate a multilevel database on an ordinary database system . This means students will have to modify the schema to add additional fields for storing security classification levels. They also develop views for users having different clearance levels. Further, to support poly-instantiation, the primary key will have to be redefined to include security level to accommodate the possibility of the same key values existing at multiple security levels. Another topic of interest would be to explore how the Discretionary Access Control and the Mandatory Access Control can be combined and applied in some scenarios. 2.3 Securing Real Life Databases The candidate database systems we chose for hands-on experience were MySQL TM , Oracle R , and Microsoft R SQL Server TM . Because of time constraints, students were able to focus only on the first two databases, but an overview of SQL server security was also provided. 2.3.1 MySQL Security With more than six million installations worldwide [14], the simplicity and open source architecture make MySQL, probably, the first database to study. The primary source of information would be the MySQL manual itself (available from MySQL site [14]), particularly the section on "MySQL Access Privilege System". Another source, MySQL Security Handbook [22], explains MySQL security system and provides a few practical examples. Labs/Assignments 1 Multilevel Security � Poly-instantiation 2 MySQL Grant Privilege System 3 SQL Injection 4 Oracle Security � Basic Lab 5 Database Security Plan Development 6 Backend Development for B2C Application 7 Probability Distributions, Sampling 8 Statistical Databases - Breach of Security 9 Statistical Databases - Inference Protection Techniques 10 Data Mining Security - Reading and Presentation Table 2: Major Labs/Assignments MySQL Access Privilege System authenticates a user based on user name, host name, and password. Further, it ensures that users perform only permitted operations based on the privileges specified in the grant tables (namely, user, db, and host). The format and contents of these tables, therefore , are of particular importance. Since most of the critical information including the grant tables are stored on a default database named mysql, the security of mysql database is also crucial. Students should learn to apply the "principle of least privilege" when granting privileges in order to perform the task at hand. Each student was given a MySQL instance with root level access. Students were asked to create users and assign privileges while monitoring the privilege tables for changes. Students also experimented with the privilege system by man-ually modifying privilege tables. We created two person administrator-user teams for enabling the students to experience the system from both perspectives. Users were assigned certain tasks to perform. Some of the tasks given were specifically designed to understand the limitations of the MySQL privilege system. The role of the administrators was to grant privileges just sufficient for users to perform the task. Users could access the system in any manner they wish � in fact, users will be encouraged to expose the weaknesses in the privilege assignments. The administrators, on the other hand, controlled access based on need-to-know, at the same time trying not to be too restrictive for users to perform the required tasks. We found the users very excited to expose security weaknesses in the privilege assignment. Although administrators were a little embarrassed, they too were motivated by the exercise. For the the next lab session, students switched roles, i.e., those who were administrators became users and vice versa. MySQL supports data security by providing functions such as ENCRYPT, DES ENCRYPT, AES ENCRYPT, PASSWORD , OLD PASSWORD and ENCODE. Since these functions may not be safe under all circumstances, it would be useful to highlight the unsafe scenarios. Students may also learn how to use SSL for security, and simultaneously make sure that the system performance is not significantly impacted. Also useful would be to study how the authentication requirements may vary when using options such as REQUIRE SSL, REQUIRE ISSUER and REQUIRE X509. Even when using SSL, the data security can depend on the type of cipher and the key lengths used. Therefore, students may learn how to specify these parameters using the REQUIRE CIPHER option. Some privileges in MySQL, if not carefully used, can expose the system to high security risk. For example, FILE privilege may be misused to gain access to the system. Hence, a comprehensive study of unsafe privileges will be extremely useful. Even when the privilege system is correctly set up and maintained the entire privilege system can be circumvented using a MySQL startup option like --skip-grant-tables. On the other hand, some startup options make the server safer. Therefore, MySQL startup options and their security consequences must be discussed. Many web applications have MySQL database server deployed as the backend, and HTML based form acting as the front end. Since user input is used to generate SQL queries to interact with the database, if unchecked, malicious users or programs can inject unsafe SQL queries. Basic concepts of preventing SQL injection may also be discussed. Students may be asked to analyze a number of SQL queries for potential vulnerability. Other topics, which can be included are: using MySQL network scanner to detect MySQL servers on the network with default passwords, MySQL resource control, data backup and recovery, auditing, and firewalls. 2.3.2 Planning Database Security Since enforcing database security is extremely complex task with a large number of factors affecting the security of a database, the best way to approach the problem would be to systematically develop and implement a comprehensive database security plan. Therefore, in our course, we required that students develop a database security plan for a small Business-to-Consumer (B2C) E-Commerce application . See [19] Ch7 (available online) for detailed exposition on database security planning. Although the text is on Oracle security, the concepts can be applied to any database. 9 2.3.3 Oracle Security For security reasons, the computer science department was reluctant to grant administrative privileges to students on our Oracle server. Therefore, we ended up creating a separate Oracle instance for the course. For each student, we created one administrative account with DBA privileges, and then the students were allowed to create user accounts as needed, provided they follow a naming convention to avoid conflicting names. In addition to Oracle Security Handbook [12], we found the Oracle Database Administrators Guide [15] also useful. The guide is available online from the Oracle Database Documentation Library. First, we had an Oracle Security Basics Lab. Students were introduced to the Oracle security system through a series of tasks. The next lab was more advanced, and built up on the database security plan developed in a previous assignment. The task was to develop the backend for a small B2C E-Commerce application. Students were asked to create user accounts, roles, tables, views and triggers as required. The privileges were to be assigned by observing the "principle of least privilege", as per the security plan. Further, students may also be trained to perform some standard checks for security such as checking for default user accounts, default passwords, users having excessive privileges (e.g., DBA, ALTER SYSTEM, CREATE LIBRARY, CREATE ANY TRIGGER), security impact of WITH AD-MIN and WITH GRANT options on privileges, EXTER-NALLY authenticated users, and the existence of database links. Students may also learn how to display information on items such as triggers, views and externally authenticated users. A section on security issues of using default Oracle supplied roles will be useful. Other topics to include are: Transparent Network Substrate (TNS) security and listener management from remote machines and setting up listener passwords, buffer overflow attacks and prevention, auditing, and undocumented Oracle features. Students may also be introduced to reading security advisories and obtaining Oracle Critical Patch Updates (CPU). Recently, a large number of security breaches have been reported. Interestingly, however, many of these breaches were incidents of missing or stolen backup storage devices. Therefore, we feel appropriate to include a session on security and protection needs of exports, cold backups, hot backups, and disaster recovery sites. 2.3.4 Oracle Label Security Oracle Label Security provides built-in row level access control for high security applications. Essentially, Oracle adds a new field to each row for storing the row's sensitivity labels. Row access is granted or denied by comparing the user's identity and security clearance label with the row's sensitivity labels. Earlier, in Assignment 1, students have simulated a multilevel database. Therefore, the above concepts should be easy to learn at this point. As a source of information on Oracle Label Security Architecture , we used Oracle Label Security Administrator's Guide [16]. We covered levels, compartments, groups, session and row labels, label security algorithm, and management of label security using Oracle Internet Dictionary. 2.3.5 Microsoft SQL Server Security As we mentioned earlier at the beginning of this section, due to time constraints, we could not provide an extensive coverage of Microsoft SQL Server. We briefly discussed SQL Server security model, authentication mechanisms, authentication modes, and good security practices for SQL servers. Students presented information they gathered on SQL server vulnerabilities, security breaches, and prevention techniques. We found a few excellent articles on SQLServer-Central .com, an online community of DBAs, developers, and SQL server users. We also found SQL Server Developer Center (http://msdn.microsoft.com/sql) useful in providing a large number of resources in this area. 2.4 Statistical Security As for the rest of the course, this section is application oriented, giving the students the gist of the concepts they need to know and then putting them to work in the context of a real database. Thus, the first lab is a simulation based assignment designed as an introduction to probability distributions , expectation, spread, sampling methods, and sampling distributions of relevant statistics. We find that, even for students with prior coursework in probability and statistics , an assignment of this type is very beneficial. The second lab presents the task of setting up a sequence of queries, so that students can extract from a database what should have been secure information. At this point we introduce the main conceptual techniques for inference protection such as the lattice model and partitioning the database entities into populations. See [4], Ch 5 for details. The third major assignment aims at teaching inference protection techniques. Given a database, the students are asked to answer queries without disclosing sensitive information by applying restriction , perturbation, and combined techniques. 2.5 Security Issues of Data Mining Data mining may be misused to obtain confidential information from a database. So we believe, a course on database security should include an overview of security and privacy concerns of data mining. Organizations would like to share the data for operational convenience, at the same time prevent the mining of data for information they do not want to disclose. Likewise, private individuals would like to submit their personal information for data mining without compromising their privacy while keeping the key association rules intact. Secure data mining techniques appear similar to statistical security methods, however, their computational efficiency is a major concern. We found a number of interesting papers [3, 17, 20] that can be used for reading assignments and group discussions. 2.6 Other Topics Malicious users may bypass security mechanisms provided by an application by directly connecting to the database. Therefore, whenever possible stored procedures, and views must be used for providing data access. Database application security and defensive programming was briefly covered . Semi-structured nature of Extensible Markup Language (XML) documents make them ideal candidates for use in many applications including E-Business. Therefore, XML [7] security was also discussed. 10 Other topics of interest are: database intrusion detection and prevention [17], database fault tolerance and recovery, Hippocratic databases [1], network security [10], and biometrics [9]. RELATED COURSES Department of Computer Science at University of Alberta has offered an independent study on database security with topics such as security models, security mechanisms, intrusion detection systems, and statistical database protection. University of Maryland University College has a graduate level course on database security with theory and applications , including frameworks for discretionary and mandatory access control, data integrity, availability and performance, secure database design, data aggregation, data inference, secure concurrency control, and secure transaction processing. University of South Carolina and George Mason University offer graduate level elective courses on Database Security. Similar courses are offered at a few other institutions, but we do not discuss them here due to space constraints. Among the undergraduate courses we found, the closest one to what we offered is taught at University of Arkansas Little Rock. It provides database security theory and background on Oracle security environment. CONCLUSIONS Our new undergraduate elective course on database security covers basic concepts and provides practical experience on two popular databases. We emphasized that students develop a database security plan that, we hope, will encourage them to view the problem of ensuring database security as a task that needs to be carefully planned in whole rather than something that can be addressed in parts. The initial offering of the course had "Data Structures" as the only pre-requisite, because we wanted to keep the course open to a larger audience. Students, in general, were found be more motivated to follow through on course work than other courses we have taught. We received excellent numerical score as well as comments from students in the departmental student evaluations. We had a good mixture of students. All were computer science majors, and forty seven percent were honors students. Sixty seven percent of the class completed the course with an overall score of 80% or higher with all honors students falling into this category. However, it was felt that a few students lacked basics to fully grasp the material. Therefore, having a basic database technology course as the pre-requisite will help cover more of the suggested topics in depth. In closing, we hope that our experience shared herein will help other instructors develop and offer a similar course on database security with limited resources. REFERENCES [1] R. Agrawal, J. Kiernan, R. Srikant, and Y. Xu. Hippocratic Databases. In Proc. of the Very Large Data Bases (VLDB) Conference, Hong Kong, China, August 2002. [2] D. Bell and L. LaPadula. Secure Computer Systems: Mathematical Foundations. Technical Report ESD-TR-73-278, MITRE Corporation, 1973. [3] C. Clifton and D. Marks. Security and Privacy Implications of Data Mining. In Workshop on Data Mining and Knowledge Discovery, Montreal, Canada, February 1996. [4] S. Castano, M. G. Fugini, G. Martella, and P. Samarati. Database Security. Addison-Wesley & ACM Press, 1995. [5] D. E. Denning. Cryptography and Data Security. Addison-Wesley, 1982. [6] D. Ferraiolo and R. Kuhn. Role-Based Access Controls. In Proc. 15th NIST-NCSC National Computer Security Conference, Baltimore, MD, October 1992. [7] B. Dournaee. XML Security. RSA Press, Berkeley, CA, USA, 2002. [8] M. Guimaraes, H. Mattord, and R. Austin. Incorporating Security Components into Database Courses. In Proc. of the InfoSecCD Conference'04, Kennesaw, GA, September 2004. [9] A. Jain, L. Hong, and S. Pankanti. Biometric Identification. Commun. ACM, 43(2), 2000. [10] C. Kaufman, R. Perlman, and M. Speciner. Network Security: Private Communication in a Public World, Second Edition. Prentice-Hall, 2002. [11] B. W. Lampson. A Note on the Confinement Problem. Commun. ACM, 16(10), October 1973. [12] M. Theriault and A. Newman. Oracle Security Handbook : Implement a Sound Security Plan in Your Oracle Environment. Osborne McGraw-Hill, 2001. [13] M. Morgenstern. Security and Inference in Multi-Level Database and Knowledge-Base Systems. In ACM SIGMOD Conf. on the Management of Data, San Francisco, CA, May 1987. [14] http://www.mysql.com. [15] Oracle Database Administrator's Guide. Oracle Corporation, 2001. [16] Oracle Label Security Administrator's Guide. Oracle Corporation, 2003. [17] R. Agrawal and R. Srikant. Privacy-preserving Data Mining. In Proc. of the ACM SIGMOD Conference on Management of Data, Dallas, TX, May 2000. [18] R. Sandhu and Q. Munawer. How to do Discretionary Access Control Using Roles. In RBAC '98: Proceedings of the third ACM workshop on Role-based access control, Fairfax, VA, 1998. [19] M. Theriault and W. Heney. Oracle Security. O'Reilly & Associates, Inc., 1998. [20] V. Verykios, E. Bertino, I. Fovino, L. Provenza, Y. Saygin and Y. Theodoridis. State-of-the-art in Privacy Preserving Data Mining. SIGMOD Record, 33(1), 2004. [21] W. Ford and M. S. Baum. Secure Electronic Commerce: Building the Infrastructure for Digital Signatures and Encryption. Prentice Hall, 2000. [22] Wrox Author Team. MySQL Security Handbook. Wrox Press, 2003. 11

Database security course;Statistical Database Security;Statistical Security;Database security;High Risk of Sensitive Information;Database security plan;Security breaches;Labs;Database system;Undergraduate Database Security Course;Real Life Database Security;Database Security Education;XML Security;Undergraduate students;Cryptography;Secure information;Hands-on experience;Database Security Plan;Data access;Security Plan;Database Privacy;Administrators;Hands-on;Database Security Course;Undergraduate course;Access Privilege System;Database Security;Privacy Issues;Laboratory/Active Learning;Right Blend of Theory and Practice;Assignments;Real Life Databases Hands-on;MySQL Security;Topics;Importance of Database Security;Few Database Security Courses;Oracle Security

40

Automatic Extraction of Titles from General Documents using Machine Learning

In this paper, we propose a machine learning approach to title extraction from general documents. By general documents, we mean documents that can belong to any one of a number of specific genres, including presentations, book chapters, technical papers, brochures, reports, and letters. Previously, methods have been proposed mainly for title extraction from research papers. It has not been clear whether it could be possible to conduct automatic title extraction from general documents. As a case study, we consider extraction from Office including Word and PowerPoint. In our approach, we annotate titles in sample documents (for Word and PowerPoint respectively) and take them as training data, train machine learning models, and perform title extraction using the trained models. Our method is unique in that we mainly utilize formatting information such as font size as features in the models. It turns out that the use of formatting information can lead to quite accurate extraction from general documents. Precision and recall for title extraction from Word is 0.810 and 0.837 respectively, and precision and recall for title extraction from PowerPoint is 0.875 and 0.895 respectively in an experiment on intranet data. Other important new findings in this work include that we can train models in one domain and apply them to another domain, and more surprisingly we can even train models in one language and apply them to another language. Moreover, we can significantly improve search ranking results in document retrieval by using the extracted titles.

INTRODUCTION Metadata of documents is useful for many kinds of document processing such as search, browsing, and filtering. Ideally, metadata is defined by the authors of documents and is then used by various systems. However, people seldom define document metadata by themselves, even when they have convenient metadata definition tools [26]. Thus, how to automatically extract metadata from the bodies of documents turns out to be an important research issue. Methods for performing the task have been proposed. However, the focus was mainly on extraction from research papers. For instance, Han et al. [10] proposed a machine learning based method to conduct extraction from research papers. They formalized the problem as that of classification and employed Support Vector Machines as the classifier. They mainly used linguistic features in the model. 1 In this paper, we consider metadata extraction from general documents. By general documents, we mean documents that may belong to any one of a number of specific genres. General documents are more widely available in digital libraries, intranets and the internet, and thus investigation on extraction from them is 1 The work was conducted when the first author was visiting Microsoft Research Asia. 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. JCDL'05, June 711, 2005, Denver, Colorado, USA Copyright 2005 ACM 1-58113-876-8/05/0006...$5.00. 145 sorely needed. Research papers usually have well-formed styles and noticeable characteristics. In contrast, the styles of general documents can vary greatly. It has not been clarified whether a machine learning based approach can work well for this task. There are many types of metadata: title, author, date of creation, etc. As a case study, we consider title extraction in this paper. General documents can be in many different file formats: Microsoft Office, PDF (PS), etc. As a case study, we consider extraction from Office including Word and PowerPoint. We take a machine learning approach. We annotate titles in sample documents (for Word and PowerPoint respectively) and take them as training data to train several types of models, and perform title extraction using any one type of the trained models. In the models, we mainly utilize formatting information such as font size as features. We employ the following models: Maximum Entropy Model, Perceptron with Uneven Margins, Maximum Entropy Markov Model, and Voted Perceptron. In this paper, we also investigate the following three problems, which did not seem to have been examined previously. (1) Comparison between models: among the models above, which model performs best for title extraction; (2) Generality of model: whether it is possible to train a model on one domain and apply it to another domain, and whether it is possible to train a model in one language and apply it to another language; (3) Usefulness of extracted titles: whether extracted titles can improve document processing such as search. Experimental results indicate that our approach works well for title extraction from general documents. Our method can significantly outperform the baselines: one that always uses the first lines as titles and the other that always uses the lines in the largest font sizes as titles. Precision and recall for title extraction from Word are 0.810 and 0.837 respectively, and precision and recall for title extraction from PowerPoint are 0.875 and 0.895 respectively. It turns out that the use of format features is the key to successful title extraction. (1) We have observed that Perceptron based models perform better in terms of extraction accuracies. (2) We have empirically verified that the models trained with our approach are generic in the sense that they can be trained on one domain and applied to another, and they can be trained in one language and applied to another. (3) We have found that using the extracted titles we can significantly improve precision of document retrieval (by 10%). We conclude that we can indeed conduct reliable title extraction from general documents and use the extracted results to improve real applications. The rest of the paper is organized as follows. In section 2, we introduce related work, and in section 3, we explain the motivation and problem setting of our work. In section 4, we describe our method of title extraction, and in section 5, we describe our method of document retrieval using extracted titles. Section 6 gives our experimental results. We make concluding remarks in section 7. RELATED WORK Methods have been proposed for performing automatic metadata extraction from documents; however, the main focus was on extraction from research papers. The proposed methods fall into two categories: the rule based approach and the machine learning based approach. Giuffrida et al. [9], for instance, developed a rule-based system for automatically extracting metadata from research papers in Postscript. They used rules like "titles are usually located on the upper portions of the first pages and they are usually in the largest font sizes". Liddy et al. [14] and Yilmazel el al. [23] performed metadata extraction from educational materials using rule-based natural language processing technologies. Mao et al. [16] also conducted automatic metadata extraction from research papers using rules on formatting information. The rule-based approach can achieve high performance. However, it also has disadvantages. It is less adaptive and robust when compared with the machine learning approach. Han et al. [10], for instance, conducted metadata extraction with the machine learning approach. They viewed the problem as that of classifying the lines in a document into the categories of metadata and proposed using Support Vector Machines as the classifier. They mainly used linguistic information as features. They reported high extraction accuracy from research papers in terms of precision and recall. 2.2 Information Extraction Metadata extraction can be viewed as an application of information extraction, in which given a sequence of instances, we identify a subsequence that represents information in which we are interested. Hidden Markov Model [6], Maximum Entropy Model [1, 4], Maximum Entropy Markov Model [17], Support Vector Machines [3], Conditional Random Field [12], and Voted Perceptron [2] are widely used information extraction models. Information extraction has been applied, for instance, to part-of-speech tagging [20], named entity recognition [25] and table extraction [19]. 2.3 Search Using Title Information Title information is useful for document retrieval. In the system Citeseer, for instance, Giles et al. managed to extract titles from research papers and make use of the extracted titles in metadata search of papers [8]. In web search, the title fields (i.e., file properties) and anchor texts of web pages (HTML documents) can be viewed as `titles' of the pages [5]. Many search engines seem to utilize them for web page retrieval [7, 11, 18, 22]. Zhang et al., found that web pages with well-defined metadata are more easily retrieved than those without well-defined metadata [24]. To the best of our knowledge, no research has been conducted on using extracted titles from general documents (e.g., Office documents) for search of the documents. 146 MOTIVATION AND PROBLEM SETTING We consider the issue of automatically extracting titles from general documents. By general documents, we mean documents that belong to one of any number of specific genres. The documents can be presentations, books, book chapters, technical papers, brochures, reports, memos, specifications, letters, announcements, or resumes. General documents are more widely available in digital libraries, intranets, and internet, and thus investigation on title extraction from them is sorely needed. Figure 1 shows an estimate on distributions of file formats on intranet and internet [15]. Office and PDF are the main file formats on the intranet. Even on the internet, the documents in the formats are still not negligible, given its extremely large size. In this paper, without loss of generality, we take Office documents as an example. Figure 1. Distributions of file formats in internet and intranet. For Office documents, users can define titles as file properties using a feature provided by Office. We found in an experiment, however, that users seldom use the feature and thus titles in file properties are usually very inaccurate. That is to say, titles in file properties are usually inconsistent with the `true' titles in the file bodies that are created by the authors and are visible to readers. We collected 6,000 Word and 6,000 PowerPoint documents from an intranet and the internet and examined how many titles in the file properties are correct. We found that surprisingly the accuracy was only 0.265 (cf., Section 6.3 for details). A number of reasons can be considered. For example, if one creates a new file by copying an old file, then the file property of the new file will also be copied from the old file. In another experiment, we found that Google uses the titles in file properties of Office documents in search and browsing, but the titles are not very accurate. We created 50 queries to search Word and PowerPoint documents and examined the top 15 results of each query returned by Google. We found that nearly all the titles presented in the search results were from the file properties of the documents. However, only 0.272 of them were correct. Actually, `true' titles usually exist at the beginnings of the bodies of documents. If we can accurately extract the titles from the bodies of documents, then we can exploit reliable title information in document processing. This is exactly the problem we address in this paper. More specifically, given a Word document, we are to extract the title from the top region of the first page. Given a PowerPoint document, we are to extract the title from the first slide. A title sometimes consists of a main title and one or two subtitles. We only consider extraction of the main title. As baselines for title extraction, we use that of always using the first lines as titles and that of always using the lines with largest font sizes as titles. Figure 2. Title extraction from Word document. Figure 3. Title extraction from PowerPoint document. Next, we define a `specification' for human judgments in title data annotation. The annotated data will be used in training and testing of the title extraction methods. Summary of the specification: The title of a document should be identified on the basis of common sense, if there is no difficulty in the identification. However, there are many cases in which the identification is not easy. There are some rules defined in the specification that guide identification for such cases. The rules include "a title is usually in consecutive lines in the same format", "a document can have no title", "titles in images are not considered", "a title should not contain words like `draft', 147 `whitepaper', etc", "if it is difficult to determine which is the title, select the one in the largest font size", and "if it is still difficult to determine which is the title, select the first candidate". (The specification covers all the cases we have encountered in data annotation.) Figures 2 and 3 show examples of Office documents from which we conduct title extraction. In Figure 2, `Differences in Win32 API Implementations among Windows Operating Systems' is the title of the Word document. `Microsoft Windows' on the top of this page is a picture and thus is ignored. In Figure 3, `Building Competitive Advantages through an Agile Infrastructure' is the title of the PowerPoint document. We have developed a tool for annotation of titles by human annotators. Figure 4 shows a snapshot of the tool. Figure 4. Title annotation tool. TITLE EXTRACTION METHOD Title extraction based on machine learning consists of training and extraction. The same pre-processing step occurs before training and extraction. During pre-processing, from the top region of the first page of a Word document or the first slide of a PowerPoint document a number of units for processing are extracted. If a line (lines are separated by `return' symbols) only has a single format, then the line will become a unit. If a line has several parts and each of them has its own format, then each part will become a unit. Each unit will be treated as an instance in learning. A unit contains not only content information (linguistic information) but also formatting information. The input to pre-processing is a document and the output of pre-processing is a sequence of units (instances). Figure 5 shows the units obtained from the document in Figure 2. Figure 5. Example of units. In learning, the input is sequences of units where each sequence corresponds to a document. We take labeled units (labeled as title_begin, title_end, or other) in the sequences as training data and construct models for identifying whether a unit is title_begin title_end, or other. We employ four types of models: Perceptron, Maximum Entropy (ME), Perceptron Markov Model (PMM), and Maximum Entropy Markov Model (MEMM). In extraction, the input is a sequence of units from one document. We employ one type of model to identify whether a unit is title_begin, title_end, or other. We then extract units from the unit labeled with `title_begin' to the unit labeled with `title_end'. The result is the extracted title of the document. The unique characteristic of our approach is that we mainly utilize formatting information for title extraction. Our assumption is that although general documents vary in styles, their formats have certain patterns and we can learn and utilize the patterns for title extraction. This is in contrast to the work by Han et al., in which only linguistic features are used for extraction from research papers. 4.2 Models The four models actually can be considered in the same metadata extraction framework. That is why we apply them together to our current problem. Each input is a sequence of instances k x x x L 2 1 together with a sequence of labels k y y y L 2 1 . i x and i y represents an instance and its label, respectively ( k i , , 2 , 1 L = ). Recall that an instance here represents a unit. A label represents title_begin, title_end, or other. Here, k is the number of units in a document. In learning, we train a model which can be generally denoted as a conditional probability distribution ) | ( 1 1 k k X X Y Y P L L where i X and i Y denote random variables taking instance i x and label i y as values, respectively ( k i , , 2 , 1 L = ). Learning Tool Extraction Tool 2 1 1 2 1 2 22 21 2 22 21 1 12 11 1 12 11 nk n n k n n k k k k y y y x x x y y y x x x y y y x x x L L L L L L L L ) | ( max arg 1 1 mk m mk m x x y y P L L ) | ( 1 1 k k X X Y Y P L L Conditional Distribution mk m m x x x L 2 1 Figure 6. Metadata extraction model. We can make assumptions about the general model in order to make it simple enough for training. 148 For example, we can assume that k Y Y , , 1 L are independent of each other given k X X , , 1 L . Thus, we have ) | ( ) | ( ) | ( 1 1 1 1 k k k k X Y P X Y P X X Y Y P L L L = In this way, we decompose the model into a number of classifiers. We train the classifiers locally using the labeled data. As the classifier, we employ the Perceptron or Maximum Entropy model. We can also assume that the first order Markov property holds for k Y Y , , 1 L given k X X , , 1 L . Thus, we have ) | ( ) | ( ) | ( 1 1 1 1 1 k k k k k X Y Y P X Y P X X Y Y P = L L L Again, we obtain a number of classifiers. However, the classifiers are conditioned on the previous label. When we employ the Percepton or Maximum Entropy model as a classifier, the models become a Percepton Markov Model or Maximum Entropy Markov Model, respectively. That is to say, the two models are more precise. In extraction, given a new sequence of instances, we resort to one of the constructed models to assign a sequence of labels to the sequence of instances, i.e., perform extraction. For Perceptron and ME, we assign labels locally and combine the results globally later using heuristics. Specifically, we first identify the most likely title_begin. Then we find the most likely title_end within three units after the title_begin. Finally, we extract as a title the units between the title_begin and the title_end. For PMM and MEMM, we employ the Viterbi algorithm to find the globally optimal label sequence. In this paper, for Perceptron, we actually employ an improved variant of it, called Perceptron with Uneven Margin [13]. This version of Perceptron can work well especially when the number of positive instances and the number of negative instances differ greatly, which is exactly the case in our problem. We also employ an improved version of Perceptron Markov Model in which the Perceptron model is the so-called Voted Perceptron [2]. In addition, in training, the parameters of the model are updated globally rather than locally. 4.3 Features There are two types of features: format features and linguistic features. We mainly use the former. The features are used for both the title-begin and the title-end classifiers. 4.3.1 Format Features Font Size: There are four binary features that represent the normalized font size of the unit (recall that a unit has only one type of font). If the font size of the unit is the largest in the document, then the first feature will be 1, otherwise 0. If the font size is the smallest in the document, then the fourth feature will be 1, otherwise 0. If the font size is above the average font size and not the largest in the document, then the second feature will be 1, otherwise 0. If the font size is below the average font size and not the smallest, the third feature will be 1, otherwise 0. It is necessary to conduct normalization on font sizes. For example, in one document the largest font size might be `12pt', while in another the smallest one might be `18pt'. Boldface: This binary feature represents whether or not the current unit is in boldface. Alignment: There are four binary features that respectively represent the location of the current unit: `left', `center', `right', and `unknown alignment'. The following format features with respect to `context' play an important role in title extraction. Empty Neighboring Unit: There are two binary features that represent, respectively, whether or not the previous unit and the current unit are blank lines. Font Size Change: There are two binary features that represent, respectively, whether or not the font size of the previous unit and the font size of the next unit differ from that of the current unit. Alignment Change: There are two binary features that represent, respectively, whether or not the alignment of the previous unit and the alignment of the next unit differ from that of the current one. Same Paragraph: There are two binary features that represent, respectively, whether or not the previous unit and the next unit are in the same paragraph as the current unit. 4.3.2 Linguistic Features The linguistic features are based on key words. Positive Word: This binary feature represents whether or not the current unit begins with one of the positive words. The positive words include `title:', `subject:', `subject line:' For example, in some documents the lines of titles and authors have the same formats. However, if lines begin with one of the positive words, then it is likely that they are title lines. Negative Word: This binary feature represents whether or not the current unit begins with one of the negative words. The negative words include `To', `By', `created by', `updated by', etc. There are more negative words than positive words. The above linguistic features are language dependent. Word Count: A title should not be too long. We heuristically create four intervals: [1, 2], [3, 6], [7, 9] and [9, ) and define one feature for each interval. If the number of words in a title falls into an interval, then the corresponding feature will be 1; otherwise 0. Ending Character: This feature represents whether the unit ends with `:', `-', or other special characters. A title usually does not end with such a character. DOCUMENT RETRIEVAL METHOD We describe our method of document retrieval using extracted titles. Typically, in information retrieval a document is split into a number of fields including body, title, and anchor text. A ranking function in search can use different weights for different fields of 149 the document. Also, titles are typically assigned high weights, indicating that they are important for document retrieval. As explained previously, our experiment has shown that a significant number of documents actually have incorrect titles in the file properties, and thus in addition of using them we use the extracted titles as one more field of the document. By doing this, we attempt to improve the overall precision. In this paper, we employ a modification of BM25 that allows field weighting [21]. As fields, we make use of body, title, extracted title and anchor. First, for each term in the query we count the term frequency in each field of the document; each field frequency is then weighted according to the corresponding weight parameter: = f tf f t tf w wtf Similarly, we compute the document length as a weighted sum of lengths of each field. Average document length in the corpus becomes the average of all weighted document lengths. = f f f dl w wdl In our experiments we used 75 . 0 , 8 . 1 1 = = b k . Weight for content was 1.0, title was 10.0, anchor was 10.0, and extracted title was 5.0. EXPERIMENTAL RESULTS We used two data sets in our experiments. First, we downloaded and randomly selected 5,000 Word documents and 5,000 PowerPoint documents from an intranet of Microsoft. We call it MS hereafter. Second, we downloaded and randomly selected 500 Word and 500 PowerPoint documents from the DotGov and DotCom domains on the internet, respectively. Figure 7 shows the distributions of the genres of the documents. We see that the documents are indeed `general documents' as we define them. Figure 7. Distributions of document genres. Third, a data set in Chinese was also downloaded from the internet. It includes 500 Word documents and 500 PowerPoint documents in Chinese. We manually labeled the titles of all the documents, on the basis of our specification. Not all the documents in the two data sets have titles. Table 1 shows the percentages of the documents having titles. We see that DotCom and DotGov have more PowerPoint documents with titles than MS. This might be because PowerPoint documents published on the internet are more formal than those on the intranet. Table 1. The portion of documents with titles Domain Type MS DotCom DotGov Word 75.7% 77.8% 75.6% PowerPoint 82.1% 93.4% 96.4% In our experiments, we conducted evaluations on title extraction in terms of precision, recall, and F-measure. The evaluation measures are defined as follows: Precision: P = A / ( A + B ) Recall: R = A / ( A + C ) F-measure: F1 = 2PR / ( P + R ) Here, A, B, C, and D are numbers of documents as those defined in Table 2. Table 2. Contingence table with regard to title extraction Is title Is not title Extracted A B Not extracted C D 6.2 Baselines We test the accuracies of the two baselines described in section 4.2. They are denoted as `largest font size' and `first line' respectively. 6.3 Accuracy of Titles in File Properties We investigate how many titles in the file properties of the documents are reliable. We view the titles annotated by humans as true titles and test how many titles in the file properties can approximately match with the true titles. We use Edit Distance to conduct the approximate match. (Approximate match is only used in this evaluation). This is because sometimes human annotated titles can be slightly different from the titles in file properties on the surface, e.g., contain extra spaces). Given string A and string B: if ( (D == 0) or ( D / ( La + Lb ) &lt; ) ) then string A = string B D: Edit Distance between string A and string B La: length of string A Lb: length of string B : 0.1 + + + = t t n N wtf avwdl wdl b b k k wtf F BM ) log( ) ) 1 (( ) 1 ( 25 1 1 150 Table 3. Accuracies of titles in file properties File Type Domain Precision Recall F1 MS 0.299 0.311 0.305 DotCom 0.210 0.214 0.212 Word DotGov 0.182 0.177 0.180 MS 0.229 0.245 0.237 DotCom 0.185 0.186 0.186 PowerPoint DotGov 0.180 0.182 0.181 6.4 Comparison with Baselines We conducted title extraction from the first data set (Word and PowerPoint in MS). As the model, we used Perceptron. We conduct 4-fold cross validation. Thus, all the results reported here are those averaged over 4 trials. Tables 4 and 5 show the results. We see that Perceptron significantly outperforms the baselines. In the evaluation, we use exact matching between the true titles annotated by humans and the extracted titles. Table 4. Accuracies of title extraction with Word Precision Recall F1 Model Perceptron 0.810 0.837 0.823 Largest font size 0.700 0.758 0.727 Baselines First line 0.707 0.767 0.736 Table 5. Accuracies of title extraction with PowerPoint Precision Recall F1 Model Perceptron 0.875 0. 895 0.885 Largest font size 0.844 0.887 0.865 Baselines First line 0.639 0.671 0.655 We see that the machine learning approach can achieve good performance in title extraction. For Word documents both precision and recall of the approach are 8 percent higher than those of the baselines. For PowerPoint both precision and recall of the approach are 2 percent higher than those of the baselines. We conduct significance tests. The results are shown in Table 6. Here, `Largest' denotes the baseline of using the largest font size, `First' denotes the baseline of using the first line. The results indicate that the improvements of machine learning over baselines are statistically significant (in the sense p-value &lt; 0.05) Table 6. Sign test results Documents Type Sign test between p-value Perceptron vs. Largest 3.59e-26 Word Perceptron vs. First 7.12e-10 Perceptron vs. Largest 0.010 PowerPoint Perceptron vs. First 5.13e-40 We see, from the results, that the two baselines can work well for title extraction, suggesting that font size and position information are most useful features for title extraction. However, it is also obvious that using only these two features is not enough. There are cases in which all the lines have the same font size (i.e., the largest font size), or cases in which the lines with the largest font size only contain general descriptions like `Confidential', `White paper', etc. For those cases, the `largest font size' method cannot work well. For similar reasons, the `first line' method alone cannot work well, either. With the combination of different features (evidence in title judgment), Perceptron can outperform Largest and First. We investigate the performance of solely using linguistic features. We found that it does not work well. It seems that the format features play important roles and the linguistic features are supplements.. Figure 8. An example Word document. Figure 9. An example PowerPoint document. We conducted an error analysis on the results of Perceptron. We found that the errors fell into three categories. (1) About one third of the errors were related to `hard cases'. In these documents, the layouts of the first pages were difficult to understand, even for humans. Figure 8 and 9 shows examples. (2) Nearly one fourth of the errors were from the documents which do not have true titles but only contain bullets. Since we conduct extraction from the top regions, it is difficult to get rid of these errors with the current approach. (3). Confusions between main titles and subtitles were another type of error. Since we only labeled the main titles as titles, the extractions of both titles were considered incorrect. This type of error does little harm to document processing like search, however. 6.5 Comparison between Models To compare the performance of different machine learning models, we conducted another experiment. Again, we perform 4-fold cross 151 validation on the first data set (MS). Table 7, 8 shows the results of all the four models. It turns out that Perceptron and PMM perform the best, followed by MEMM, and ME performs the worst. In general, the Markovian models perform better than or as well as their classifier counterparts. This seems to be because the Markovian models are trained globally, while the classifiers are trained locally. The Perceptron based models perform better than the ME based counterparts. This seems to be because the Perceptron based models are created to make better classifications, while ME models are constructed for better prediction. Table 7. Comparison between different learning models for title extraction with Word Model Precision Recall F1 Perceptron 0.810 0.837 0.823 MEMM 0.797 0.824 0.810 PMM 0.827 0.823 0.825 ME 0.801 0.621 0.699 Table 8. Comparison between different learning models for title extraction with PowerPoint Model Precision Recall F1 Perceptron 0.875 0. 895 0. 885 MEMM 0.841 0.861 0.851 PMM 0.873 0.896 0.885 ME 0.753 0.766 0.759 6.6 Domain Adaptation We apply the model trained with the first data set (MS) to the second data set (DotCom and DotGov). Tables 9-12 show the results. Table 9. Accuracies of title extraction with Word in DotGov Precision Recall F1 Model Perceptron 0.716 0.759 0.737 Largest font size 0.549 0.619 0.582 Baselines First line 0.462 0.521 0.490 Table 10. Accuracies of title extraction with PowerPoint in DotGov Precision Recall F1 Model Perceptron 0.900 0.906 0.903 Largest font size 0.871 0.888 0.879 Baselines First line 0.554 0.564 0.559 Table 11. Accuracies of title extraction with Word in DotCom Precisio n Recall F1 Model Perceptron 0.832 0.880 0.855 Largest font size 0.676 0.753 0.712 Baselines First line 0.577 0.643 0.608 Table 12. Performance of PowerPoint document title extraction in DotCom Precisio n Recall F1 Model Perceptron 0.910 0.903 0.907 Largest font size 0.864 0.886 0.875 Baselines First line 0.570 0.585 0.577 From the results, we see that the models can be adapted to different domains well. There is almost no drop in accuracy. The results indicate that the patterns of title formats exist across different domains, and it is possible to construct a domain independent model by mainly using formatting information. 6.7 Language Adaptation We apply the model trained with the data in English (MS) to the data set in Chinese. Tables 13-14 show the results. Table 13. Accuracies of title extraction with Word in Chinese Precision Recall F1 Model Perceptron 0.817 0.805 0.811 Largest font size 0.722 0.755 0.738 Baselines First line 0.743 0.777 0.760 Table 14. Accuracies of title extraction with PowerPoint in Chinese Precision Recall F1 Model Perceptron 0.766 0.812 0.789 Largest font size 0.753 0.813 0.782 Baselines First line 0.627 0.676 0.650 We see that the models can be adapted to a different language. There are only small drops in accuracy. Obviously, the linguistic features do not work for Chinese, but the effect of not using them is negligible. The results indicate that the patterns of title formats exist across different languages. From the domain adaptation and language adaptation results, we conclude that the use of formatting information is the key to a successful extraction from general documents. 6.8 Search with Extracted Titles We performed experiments on using title extraction for document retrieval. As a baseline, we employed BM25 without using extracted titles. The ranking mechanism was as described in Section 5. The weights were heuristically set. We did not conduct optimization on the weights. The evaluation was conducted on a corpus of 1.3 M documents crawled from the intranet of Microsoft using 100 evaluation queries obtained from this intranet's search engine query logs. 50 queries were from the most popular set, while 50 queries other were chosen randomly. Users were asked to provide judgments of the degree of document relevance from a scale of 1to 5 (1 meaning detrimental, 2 bad, 3 fair, 4 good and 5 excellent). 152 Figure 10 shows the results. In the chart two sets of precision results were obtained by either considering good or excellent documents as relevant (left 3 bars with relevance threshold 0.5), or by considering only excellent documents as relevant (right 3 bars with relevance threshold 1.0) 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 P@10 P@5 Reciprocal P@10 P@5 Reciprocal 0.5 1 BM25 Anchor, Title, Body BM25 Anchor, Title, Body, ExtractedTitle Name All RelevanceThreshold Data Description Figure 10. Search ranking results. Figure 10 shows different document retrieval results with different ranking functions in terms of precision @10, precision @5 and reciprocal rank: Blue bar BM25 including the fields body, title (file property), and anchor text. Purple bar BM25 including the fields body, title (file property), anchor text, and extracted title. With the additional field of extracted title included in BM25 the precision @10 increased from 0.132 to 0.145, or by ~10%. Thus, it is safe to say that the use of extracted title can indeed improve the precision of document retrieval. CONCLUSION In this paper, we have investigated the problem of automatically extracting titles from general documents. We have tried using a machine learning approach to address the problem. Previous work showed that the machine learning approach can work well for metadata extraction from research papers. In this paper, we showed that the approach can work for extraction from general documents as well. Our experimental results indicated that the machine learning approach can work significantly better than the baselines in title extraction from Office documents. Previous work on metadata extraction mainly used linguistic features in documents, while we mainly used formatting information. It appeared that using formatting information is a key for successfully conducting title extraction from general documents. We tried different machine learning models including Perceptron, Maximum Entropy, Maximum Entropy Markov Model, and Voted Perceptron. We found that the performance of the Perceptorn models was the best. We applied models constructed in one domain to another domain and applied models trained in one language to another language. We found that the accuracies did not drop substantially across different domains and across different languages, indicating that the models were generic. We also attempted to use the extracted titles in document retrieval. We observed a significant improvement in document ranking performance for search when using extracted title information. All the above investigations were not conducted in previous work, and through our investigations we verified the generality and the significance of the title extraction approach. ACKNOWLEDGEMENTS We thank Chunyu Wei and Bojuan Zhao for their work on data annotation. We acknowledge Jinzhu Li for his assistance in conducting the experiments. We thank Ming Zhou, John Chen, Jun Xu, and the anonymous reviewers of JCDL'05 for their valuable comments on this paper. REFERENCES [1] Berger, A. L., Della Pietra, S. A., and Della Pietra, V. J. A maximum entropy approach to natural language processing. Computational Linguistics, 22:39-71, 1996. [2] Collins, M. Discriminative training methods for hidden markov models: theory and experiments with perceptron algorithms. In Proceedings of Conference on Empirical Methods in Natural Language Processing, 1-8, 2002. [3] Cortes, C. and Vapnik, V. Support-vector networks. Machine Learning, 20:273-297, 1995. [4] Chieu, H. L. and Ng, H. T. A maximum entropy approach to information extraction from semi-structured and free text. In Proceedings of the Eighteenth National Conference on Artificial Intelligence, 768-791, 2002. [5] Evans, D. K., Klavans, J. L., and McKeown, K. R. Columbia newsblaster: multilingual news summarization on the Web. In Proceedings of Human Language Technology conference / North American chapter of the Association for Computational Linguistics annual meeting, 1-4, 2004. [6] Ghahramani, Z. and Jordan, M. I. Factorial hidden markov models. Machine Learning, 29:245-273, 1997. [7] Gheel, J. and Anderson, T. Data and metadata for finding and reminding, In Proceedings of the 1999 International Conference on Information Visualization, 446-451,1999. [8] Giles, C. L., Petinot, Y., Teregowda P. B., Han, H., Lawrence, S., Rangaswamy, A., and Pal, N. eBizSearch: a niche search engine for e-Business. In Proceedings of the 26th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 413-414 , 2003. [9] Giuffrida, G., Shek, E. C., and Yang, J. Knowledge-based metadata extraction from PostScript files. In Proceedings of the Fifth ACM Conference on Digital Libraries, 77-84, 2000. [10] Han, H., Giles, C. L., Manavoglu, E., Zha, H., Zhang, Z., and Fox, E. A. Automatic document metadata extraction using support vector machines. In Proceedings of the Third ACM/IEEE-CS Joint Conference on Digital Libraries, 37-48, 2003. [11] Kobayashi, M., and Takeda, K. Information retrieval on the Web. ACM Computing Surveys, 32:144-173, 2000. [12] Lafferty, J., McCallum, A., and Pereira, F. Conditional random fields: probabilistic models for segmenting and 153 labeling sequence data. In Proceedings of the Eighteenth International Conference on Machine Learning, 282-289, 2001. [13] Li, Y., Zaragoza, H., Herbrich, R., Shawe-Taylor J., and Kandola, J. S. The perceptron algorithm with uneven margins. In Proceedings of the Nineteenth International Conference on Machine Learning, 379-386, 2002. [14] Liddy, E. D., Sutton, S., Allen, E., Harwell, S., Corieri, S., Yilmazel, O., Ozgencil, N. E., Diekema, A., McCracken, N., and Silverstein, J. Automatic Metadata generation & evaluation. In Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 401-402, 2002. [15] Littlefield, A. Effective enterprise information retrieval across new content formats. In Proceedings of the Seventh Search Engine Conference, http://www.infonortics.com/searchengines/sh02/02prog.html, 2002. [16] Mao, S., Kim, J. W., and Thoma, G. R. A dynamic feature generation system for automated metadata extraction in preservation of digital materials. In Proceedings of the First International Workshop on Document Image Analysis for Libraries, 225-232, 2004. [17] McCallum, A., Freitag, D., and Pereira, F. Maximum entropy markov models for information extraction and segmentation. In Proceedings of the Seventeenth International Conference on Machine Learning, 591-598, 2000. [18] Murphy, L. D. Digital document metadata in organizations: roles, analytical approaches, and future research directions. In Proceedings of the Thirty-First Annual Hawaii International Conference on System Sciences, 267-276, 1998. [19] Pinto, D., McCallum, A., Wei, X., and Croft, W. B. Table extraction using conditional random fields. In Proceedings of the 26th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 235-242 , 2003. [20] Ratnaparkhi, A. Unsupervised statistical models for prepositional phrase attachment. In Proceedings of the Seventeenth International Conference on Computational Linguistics. 1079-1085, 1998. [21] Robertson, S., Zaragoza, H., and Taylor, M. Simple BM25 extension to multiple weighted fields, In Proceedings of ACM Thirteenth Conference on Information and Knowledge Management, 42-49, 2004. [22] Yi, J. and Sundaresan, N. Metadata based Web mining for relevance, In Proceedings of the 2000 International Symposium on Database Engineering & Applications, 113-121 , 2000. [23] Yilmazel, O., Finneran, C. M., and Liddy, E. D. MetaExtract: An NLP system to automatically assign metadata. In Proceedings of the 2004 Joint ACM/IEEE Conference on Digital Libraries, 241-242, 2004. [24] Zhang, J. and Dimitroff, A. Internet search engines' response to metadata Dublin Core implementation. Journal of Information Science, 30:310-320, 2004. [25] Zhang, L., Pan, Y., and Zhang, T. Recognising and using named entities: focused named entity recognition using machine learning. In Proceedings of the 27th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 281-288, 2004. [26] http://dublincore.org/groups/corporate/Seattle/ 154

Digital Copies;metadata extraction;Metadata processing;Search ranking results;File Formats extraction;Information search and Retrieval;PowerPoint documents;information extraction;Precision extraction;File extraction;search;generic languages;machine learning;Microsoft Office Automation

41

Autonomous and Distributed Node Recovery in Wireless Sensor Networks

Intrusion or misbehaviour detection systems are an important and widely accepted security tool in computer and wireless sensor networks. Their aim is to detect misbehaving or faulty nodes in order to take appropriate countermeasures, thus limiting the damage caused by adversaries as well as by hard or software faults. So far, however, once detected, misbehaving nodes have just been isolated from the rest of the sensor network and hence are no longer usable by running applications. In the presence of an adversary or software faults, this proceeding will inevitably lead to an early and complete loss of the whole network. For this reason, we propose to no longer expel misbehaving nodes, but to recover them into normal operation. In this paper, we address this problem and present a formal specification of what is considered a secure and correct node recovery algorithm together with a distributed algorithm that meets these properties. We discuss its requirements on the soft- and hardware of a node and show how they can be fulfilled with current and upcoming technologies. The algorithm is evaluated analytically as well as by means of extensive simulations, and the findings are compared to the outcome of a real implementation for the BTnode sensor platform. The results show that recovering sensor nodes is an expensive, though feasible and worthwhile task. Moreover , the proposed program code update algorithm is not only secure but also fair and robust.

INTRODUCTION Wireless sensor networks (WSNs) consist of many wireless communicating sensor nodes. Essentially, these are mi-crocontrollers including a communication unit and a power supply, as well as several attached sensors to examine the environment. Sensor nodes typically have very limited computing and storage capacities and can only communicate with their direct neighbourhood. In addition, WSNs have to work unattended most of the time as their operation area cannot or must not be visited. Reasons can be that the area is inhospitable, unwieldy, or ecologically too sensitive for human visitation; or that manual maintenance would just be too expensive. More and more, WSN applications are supposed to operate in hostile environments, where their communication might be overheard and nodes can be removed or manipu-lated . Regarding attacks on sensor networks, one differentiates between so called outsider and insider attacks <A href="41.html#9">[21]. In the former, a potential attacker tries to disclose or influence a confidential outcome without participating in its computation ; for instance, by intercepting, modifying, or adding messages. In the latter, by contrast, an attacker appears as an adequate member of the WSN by either plausibly impersonating regular nodes or by capturing and compromising them. Cryptographic methods, such as encrypting or signing messages, are an effective protection against attacks from outside the network, but are of only limited help against insider attacks. Once an adversary possesses one or several valid node identities (including the associated keys), it can actively participate in the operations of the WSN and influence the computed results. Intrusion or misbehaviour detection systems (IDS), on the other hand, are an important and widely accepted security tool against insider attacks <A href="41.html#9">[18, 21]. They allow for the detection of malicious or failed nodes and the application of appropriate countermeasures. So far, however, once detected , misbehaving nodes have just been isolated from the rest of the network and hence are no longer usable by running applications. In the presence of an adversary or software faults, this proceding will inevitably result in an early and complete loss of the whole network. Therefore, not only the detection of misbehaving nodes is important, but also the selection and application of effective countermeasures. Their aim must not be to simply expel suspected nodes but to recover them into correct operation. In combination, the advantages of an IDS together with the appropriate recov-113 ery measures are manifold. Not only do they help in case of program faults (e.g., deadlocks or crashes) but even if an attacker manages to capture a node and to abuse it for his own purposes, there is a chance that the aberrant behaviour of this node will be detected and the node be recovered, thus nullifying the attack. However, due to the size of sensor networks, both the IDS functionality as well as the recovery measures should be autonomously executed by the involved nodes in a distributed and cooperative manner and without the need for central instances with extended functionality. Motivated by the above mentioned insights, this paper focuses on autonomous an distributed node recovery in wireless sensor networks and proposes three alternative countermeasures to node expelling; namely to switch a node off, to restart it, and to update its program code. We formally specify what we consider a secure and correct recovery algorithm , present a distributed algorithm which meets these properties, and reason why it can help to extend the overall lifetime of a sensor network. In addition, we discuss the limitations of the proposed countermeasures, show which hard-and software parts of a corrupted node must still work correctly to make them applicable, and explain how this can be achieved with current and upcoming technologies. More precisely, the contributions of this paper are as follows: We propose to no longer expel misbehaving nodes, but to either (i) switch them off, (ii) restart them, or (iii) update their program code. We give a formal specification of a secure and correct recovery algorithm. We present a provably secure and robust distributed node recovery algorithm. We discuss the requirements on the soft- and hardware of a node in order to make the countermeasures applicable and show how they can be fulfilled with current and upcoming technologies. The algorithm is evaluated analytically as well as by means of extensive simulations and the findings are compared to the outcome of a real implementation for the BTnode sensor platform. The results show that recovering sensor nodes is an expensive, though feasible and worthwhile task. Moreover , the proposed program code update algorithm is not only provably secure but also fair and robust. It distributes the update load equally over all participating nodes and terminates as long as at least one of them remains correct. The rest of this paper is organised as follows. Section <A href="41.html#2">1.1 presents the related work in the area of intrusion detection and node recovery in wireless sensor networks. Section <A href="41.html#3">2 states the required definitions and assumptions. Section <A href="41.html#4">3 specifies the proposed recovery algorithm whose correctness is proven in section <A href="41.html#6">4. The algorithm is evaluated in section <A href="41.html#6">5 and section <A href="41.html#8">6 concludes the paper. 1.1 Related Work In this section, we present related work in the area of intrusion detection in wireless sensor networks. Additionally, related work regarding program code updates in sensor networks is also discussed, as we propose program code updates as a mean to recover nodes. Intrusion Detection In recent years, intrusion detection systems for wireless sensor networks have become a major research issue and several approaches have been proposed. However, to our best knowledge, the only countermeasure applied so far was to (logically) exclude malicious nodes. Khalil, Bagchi, and Nina-Rotaru present a distributed IDS where nodes monitor the communication among their neighbours <A href="41.html#9">[14]. For each monitored node a malignity counter is maintained and incremented whenever the designated node misbehaves. Once a counter exceeds a predefined threshold, an according alert is sent to all neighbours and if enough alerts are received the accused node is revoked from the neighbourhood list. Hsin and Liu suggest a two-phase timeout system for neighbour monitoring which uses active probing to reduce the probability of false-positives <A href="41.html#9">[10]. A rule-based IDS, which proceeds in three phases, is proposed by da Silva et al. <A href="41.html#9">[5]. In the first phase, messages are overheard and the collected information is filtered and ordered . In the second phase, the detection rules are applied to the gathered data and each inconsistency counted as a failure. Finally, in the third phase, the number or failures is compared to the expected amount of occasional failures and if too high an intrusion alert is raised. Inverardi, Mostarda and Navarra introduce a framework which enables the automatic translation of IDS specifications into program code <A href="41.html#9">[12]. The so generated code is then installed on the sensor nodes in order to locally detect violations of the node interaction policies. In the approach by Herbert et al. <A href="41.html#9">[6], predefined correctness properties (invariants ) are associated with conditions of individual nodes or the whole network and program code to verify these invariants is automatically inserted during compilation. A reputation-based IDS framework for WSNs where sensor nodes maintain a reputation for other nodes is presented by Ganeriwal and Srivastava <A href="41.html#9">[8]. It uses a Bayesian formula-tion for reputation representation, update, and integration. Program Code Update The main difference between the already available reprogramming algorithms and the proposed recovery measures are that the former focus on the propagation of new program releases among all nodes of the network, whereas the aim of the latter is the local and autonomous update of a single node. Furthermore, most reprogramming mechanisms do not care about security at all, or rely on expensive public key cryptography. Kulkarni and Wang propose a multihop reprogramming service for wireless sensor networks which uses a sender selection algorithm to avoid collisions <A href="41.html#9">[16]. Impala, a middleware system for managing sensor systems is presented by Liu and Martonosi <A href="41.html#9">[17]. Its modular architecture supports updates to the running system. An application consists of several modules which are independently transferred; an update is complete if all its modules have been received. Jeong and Culler introduce an efficient incremental network programming mechanism <A href="41.html#9">[13]. Thanks to the usage of the Rsync algorithm, only incremental changes to the new program must be transferred. A secure dissemination algorithm to distribute new program releases among nodes is presented by Dutta et al. <A href="41.html#9">[7]. Program binaries are propagated as a sequence of data blocks of which the first is authenticated with the private key of 114 the base station and the subsequent ones by means of a hash chain. In order to improve the fault tolerance of the sensor network, nodes use a grenade timer to reboot period-ically . During the boot process neighboring nodes are asked whether a new program release is available and if so, its download is initiated. DEFINITIONS In this section, we define our assumptions regarding the observation of nodes and the network communication model. We specify the capabilities of a potential adversary, explain what we consider a correct recover algorithm, and discuss the requirements on the hard- and software of a sensor node. 2.1 Intrusion and Misbehaviour Detection Troughout this paper, we assume that the network is divided into N C so called observation clusters C i = (V i , E i ), 0 i &lt; N C of size n, n = |V i |. Within a cluster each node is connected to each other (v i , v j V k , v i = v j : {v i , v j } E k ) and observes the behaviour of its cluster neighbours. For the actual monitoring of the neighbours an arbitrary IDS can be used, as long as each node ends up with an (individual) decision about whether a certain node behaves correct or malicious. The set of malicious nodes in a cluster is denoted by M i and their number by t, t = |M i | n. 2.2 Network Model In the following, p s (p r ) denotes the probability that the sending (receiving) of a message fails. Thus, for 0 p s , p l &lt; 1 the resulting probability for an unsuccessful transmission (packet loss ratio, PLR) is p l := 1 - (1 - p s )(1 - p r ) = p s + p r + p s p r Additionally, we assume that there exists a constant upper bound p O(1) on the transmission time of a message. 2.3 Adversary Model We consider an omnipresent but computationally bounded adversary who can perform both outsider as well as insider attacks. This means that a potential adversary is able to intercept and create arbitrary messages but unable to decrypt or authenticate messages for which he does not possess the required keys. We further assume that nodes can be either logically (i.e., by exploiting a software bug) or physically captured. However, the time to compromise a node physically is considered non-negligible (i.e., it takes some time to move from node to node and to perform the physical manipulations ) and to not significantly decrease with the number of already captured nodes. 2.4 Hard- and Software Requirements To all presented recovery measures applies that they are only applicable if at least the therefore needed systems of the corrupt node in the following denoted as the recovery system still work correctly. In order to achieve this, one has to make sure that the recovery system is logically and, if feasible, physically protected. Logical Protection of the Recovery System Logical protection means that it should not be possible for a running application to prevent the execution of the recovery procedures. That is, if the program code running on a node has crashed or been corrupted by an adversary (e.g., by exploiting a security hole), this should not affect the integrity and availability of the recovery system. One mechanism to achieve this is to set up a hardware interrupt which cannot be suppressed or redirected by the application and by locating the dedicated interrupt routine in a write protected memory area. Consequently, on each interrupt request, control is handed over to the immutable interrupt routine an thus to the recovery system. A simple variant of this mechanism in which a grenade timer period-ically reboots the system and the bootloader is located in read only memory (ROM) is used by Dutta et al. <A href="41.html#9">[7]. Another approach would be to misuse some additionally available MCUs <A href="41.html#9">[23], for example the ARM CPU on the ARM-based Bluetooth module on the BTnode. Some of these MCUs are powerful enough to take on additional tasks like monitoring the main MCUs activities or rewriting the application memory. In case of the BTnode that extra MCU is directly responsible for communication and thus it would be guaranteed that it has access to all received packets as well. On more advanced systems, mechanisms as provided by Intel's protected mode (e.g., isolated memory areas, privilege levels, etc.) could be used to protect the recovery system more efficiently. Current technologies such as ARM's TrustZone <A href="41.html#9">[1] for embedded devices or Intel's LaGrande technology <A href="41.html#9">[11] go even further and enable a comprehensive protection of the CPU, memory, and peripherals from software attacks. Physical Protection of the Recovery System The physical protection of current sensor node platforms is very poor because of their focus on simple maintenance <A href="41.html#9">[9]. However, although it is generally agreed that entirely tamper-proof sensor nodes would be too expensive, current trends in the hardware development of embedded devices indicate that some level of physical protection will be available in the near future <A href="41.html#9">[20, 15]. Security mechanisms regarding the packaging of sensor nodes as, for instance, those proposed by FIPS 140-2 level 2 <A href="41.html#9">[19] could already significantly increase the cost for an adversary. For integrity and not confidentiality is the main concern with the recovery module , it has only to be protected against manipulations but not against unintended disclosure or side-channel attacks. In fact, it would be sufficient to have mechanisms which render a node useless if the case of the recovery system was opened; complete tamper resistance is not required. 2.5 Correct Node Recover Algorithms A node recovery algorithm for a cluster C i = (V i , E i ) is considered correct if the following liveness and safety properties hold: L1 If all correct nodes (V i \ M i ) accuse a node m V i to be faulty or malicious, its recovery process will finally be initiated with high probability. L2 Once the recovery process for a node m V i has been initiated, it will eventually terminate as long as there remain at least k 1 correct nodes V i \ M i . S1 If no more than n-1 3 correct nodes (i.e., a minority) accuse another correct node v V i \ M i the recovery process will not be initiated. 115 S2 After the recovery process, a node m V i must either (i) be halted, (ii) contain the same program code as before, or (iii) contain the correct program code. The two liveness properties L1 and L2 ensure that each malicious node is recovered if its aberrant behaviour is detected by enough neighbours. Safety property S1 is required to make sure that a node is only recovered if a majority of correct nodes accuses it and property S2 ensures that things are not worsened by applying the recovery process. DISTRIBUTED NODE RECOVERY In this section, we present a distributed node recovery algorithm which is autonomously executed within an observation cluster. The supported recovery measures are: node shutdown, node restart, and program code update. As long as the recovery module of an otherwise faulty or malicious node is still intact, it is tried to recover it by restarting it or updating its program code; or to at least eliminate its interfering influence by turning it off. If a node does not respond to any of these measures, it is still possible to logically expell it; preferably by means of a reliable majority decision <A href="41.html#9">[22] to avoid inconsistencies among the cluster members. 3.1 Description of the Recovery Procedure The proposed recovery algorithm consists of two phases. In the first, so called accusation phase, nodes accuse all neighbours which are regarded as being malicious. If a node is accused by at least two third of its neighbours it initiates the second, so called recovery phase, during which the actual countermeasures are executed. To simplify the cooperative program code update, the program memory of a node is divided into F frames f i , 0 i &lt; F of size f s. Additionally, for each frame f i its corresponding hash value h i := h(f i ) is computed. Accusation Phase Recovery Phase Round 1 Round 2 Round 3 Figure 1: Schematic depiction of a recovery procedure which performs a program update as the countermeasure. Accusation Phase Nodes which conclude that one of their neighbours behaves maliciously, send it an authenticated accusation mes<A href="41.html#4">sage <A href="41.html#4">1 <A href="41.html#4">. The proposed countermeasure depends on the observed aberration and can be either of type shutdown, reset, or update, if the node should be halted, restarted, or its program code updated, respectively. Accusation messages have to be ac-knowledged and are resent up to r times otherwise. 1 For simplicity, it is assumed that nodes can accuse their neighbours at any time. However, if the recovery module is only active from time to time, nodes could of course also actively ask for (pending) accusations. In case that a program update is requested, the accusation messages also include a list of the sender's F frame hash values h i . They represent the current state of its program memory and are required to deduce the correct program code. Therefore, for each frame f i not only its hash value h i but also a counter c i , which is initialised with zero, is stored. Upon reception of a accusation message, each included hash value is compared to the already stored one and if they are equal, c i is incremented by one. If they differ and c i &gt; 0 the counter is decremented by one; otherwise (i.e., they are not equal and c i = 0) the stored hash value is replaced with the received value. This procedure ensures that, for 3t &lt; n - 1, every h i will contain the hash value of the correct program code frame after 2(n-1) 3 accusations have been received (see Proof <A href="41.html#6">4). Recovery Phase When a node m has received 2(n-1) 3 accusations of a certain recovery type, the corresponding measure is initiated. In the non trivial case of a distributed program code update , the correct program code has therefore to be down-loaded from the neighboring nodes. Otherwise, the node is just rebooted or shutdown and no further communication or coordination is required. The autonomous program code transfer is performed in rounds of which each starts with the broadcasting of an authenticated update request message by the accused node m. Essentially, the message contains a list of so called frame descriptors (u i , Q i ), consisting of a node id u i and a set of requested frame numbers Q i := {r 0 , r 1 , . . . , r |Q|-1 }. Upon reception of a valid request, a node v seeks for descriptors which contain its own id (i.e., u i = v). If present, for each requested frame number r j Q i the corresponding program code frame is sent back to m with an update message. All received program code fragments f i , in turn, are verified by m using the stored hash values h i . Valid code fragments are copied into the program <A href="41.html#4">memory <A href="41.html#4">2 and the frame marked as updated. If for a duration of round no update messages arrive although there are still some outstanding frames, a new update request message is broadcasted and the next round initiated. As soon as all frames have been received, the node is rebooted and thus the new program code activated. In order to distribute the transfer load equally among all participating nodes and to ensure that the update procedure terminates if at least one correct node is available, the frame descriptors are determined as follows: First, the n - 1 participating nodes are ordered such that id(v 0 ) &lt; id(v 1 ) &lt; . . . &lt; id(v n-2 ). Next, the F memory frames are divided into n - 1 sectors of length l := F n-1 . Finally, to each node one such fragment is assigned per update round in a round robin fashion. Thus, in round i node v j , is responsible for the segment s := j + i mod (n - 1), that is, for the frames sl to min((s + 1)l - 1, F - 1). In the first round, for example, the first node is responsible for the first l frames, the second node for the second l frames and so on. In the second round, however, the assignment is rotated by one and thus the outstanding frames of the first sector are now requested from the second node. This process has to be continued until all required frames have been received. 2 On most sensor node platforms, new code is not directly written into program memory but into a therefore available Flash memory and installed during a subsequent reboot. 116 Extensions and Optimisations Even though not all but only the subset of modified program code frames has to be requested, updating a node is still a time consuming and expensive task. Consequently, the amount of update load that a specific node can cause should be restricted, for instance by limiting the number of update messages that are sent to it. To further reduce the load for the participating nodes, the F hash values h i in an accusation message can be replaced by the hash value h := h(h 0 ||h 1 || . . . ||h F -1 ). Once the correct value h has been determinded using the corresponding counter c in analogy to the above mentioned algorithm, the actual hash values can be requested from the neighbours in a second step and verified with h. In order to decrease the total number of required accusation messages, more than one recovery measure per message should be allowed. Alternatively, the measures could be hierarchically organised, having the type update also counting as a reboot or shutdown request. 3.2 Algorithms Listing 1: Algorithm for an accusing node v. v a r a c c r e t r i e s [ n - 1 ] : = {0, . . . ,0} a c c f a i l e d [ n - 1 ] : = {false, . . . ,false} n u m u p d a t e s [ n - 1 ] : = {0, . . . ,0} upon m i s b e h a v i o r d e t e c t i o n o f n o d e m c h o o s e an a p p r o p r i a t e a c c u s a t i o n -t y p e a m i f a m = acc update s e n d a c c u s a t i o n , v, m, , a m , {h(f 0 ), . . . , h(f F -1 )} t o m e l s e s e n d a c c u s a t i o n , v, m, , a m t o m s t a r t t i m e r A m upon r e c e p t i o n o f a c c u s a t i o n a c k , m, , a f r o m m s t o p t i m e r A m a c c r e t r i e s [ m ] := 0 upon t i m e o u t o f t i m e r A m i f a c c r e t r i e s [ m ] &lt; max acc retries a c c r e t r i e s [ m ] := a c c r e t r i e s [ m ] + 1 s e n d a c c u s a t i o n , v, m, , a m , {h(f 0 ), . . . , h(f F -1 )} t o m s t a r t t i m e r A m e l s e a c c f a i l e d [ m ] := true upon r e c e p t i o n o f u p d a t e r e q u e s t , m, , R f r o m m i f n u m u p d a t e s [ m ] &lt; m a x u p d a t e s and (u, {r 0 , . . . , r k }) R n u m u p d a t e s [ m ] := n u m u p d a t e s [ m ] + 1 r i , 0 i k s e n d u p d a t e , v, m, r i , f r i t o m Listing 2: Algorithm for the accused node m. v a r u p d a t i n g := false n u m a c c r e s e t := 0 n u m a c c u p d a t e := 0 n u m a c c s h u t d o w n := 0 s t a r t n o d e := 0 a c c r e s e t r e c v d [ n - 1 ] := {0, . . . ,0} a c c u p d a t e r e c v d [ n - 1 ] := {0, . . . ,0} a c c s h u t d o w r e c v d [ n - 1 ] := {0, . . . ,0} f r a m e u p d a t e d [ F - 1 ] := {false, . . . ,false} f r a m e d i g e s t [ F - 1 ] := {h(f 0 ), . . . , h(f F -1 )} f r a m e c o u n t [ F - 1 ] := {0, . . . ,0} upon r e c e p t i o n o f a c c u s a t i o n , v, m, ,acc reset f r o m v s e n d a c c u s a t i o n a c k , m, ,acc reset t o v i f n o t u p d a t i n g and n o t a c c r e s e t r e c v d [ v ] a c c r e s e t r e c v d [ v ] := true n u m a c c r e s e t : = n u m a c c r e s e t + 1 i f n o t u p d a t i n g and n u m a c c r e s e t 2(n-1) 3 r e s e t n o d e upon r e c e p t i o n o f a c c u s a t i o n , v, m, ,acc shutdown f r o m v s e n d a c c u s a t i o n a c k , m, ,acc shutdown t o v i f n o t u p d a t i n g and n o t a c c s h u t d o w n r e c v d [ v ] a c c s h u t d o w n r e c v d [ v ] := true n u m a c c s h u t d o w n : = n u m a c c s h u t d o w n + 1 i f n o t u p d a t i n g and n u m a c c s h u t d o w n 2(n-1) 3 shutdown n o d e f u n c t i o n s e t u p u p d a t e r e q u e s t ( ) k := 0 R := {} f o r 0 i &lt; n , i = m w := ( s t a r t n o d e + i) mod n Q := {} f o r 0 j &lt; F n i f n o t f r a m e u p d a t e d [ k ] Q := Q {k} k := k + 1 i f Q = {} R := R {(w, Q)} s t a r t n o d e := s t a r t n o d e + 1 r e t u r n R upon r e c e p t i o n o f a c c u s a t i o n , v, m, ,acc update , {h 0 , . . . , h F -1 } f r o m v s e n d a c c u s a t i o n a c k , m, ,acc update t o v i f n o t u p d a t i n g and n o t a c c u p d a t e r e c v d [ v ] a c c u p d a t e r e c v d [ v ] := true n u m a c c u p d a t e : = n u m a c c u p d a t e + 1 f o r 0 i &lt; F i f f r a m e d i g e s t [ i ] = h i f r a m e c o u n t [ i ] := f r a m e c o u n t [ i ] + 1 e l s e i f f r a m e c o u n t [ i ] &gt; 0 f r a m e c o u n t [ i ] := f r a m e c o u n t [ i ] - 1 e l s e f r a m e d i g e s t [ i ] := h i i f n o t u p d a t i n g and n u m a c c u p d a t e 2(n-1) 3 R : = s e t u p u p d a t e r e q u e s t ( ) b r o a d c a s t u p d a t e r e q u e s t , m, , R s t a r t t i m e r U u p d a t i n g = true upon t i m e o u t o f t i m e r U R : = s e t u p u p d a t e r e q u e s t ( ) b r o a d c a s t u p d a t e r e q u e s t , m, , R s t a r t t i m e r U upon r e c e p t i o n o f u p d a t e , v, m, i, f f r o m v r e s e t t i m e r U 117 i f h(f ) = f r a m e d i g e s t [ i ] and n o t f r a m e u p d a t e d [ i ] u p d a t e memory f r a m e i f r a m e u p d a t e d [ i ] := true i f i, 0 i &lt; F f r a m e u p d a t e d [ i ] r e s e t n o d e PROOF OF CORRECTNESS In this section, we proof the correctness of the proposed algorithm with respect to the specifications of section <A href="41.html#3">2. Theorem 1. Given the network and adversary model specified in section <A href="41.html#3">2, the proposed recovery algorithm is correct and fulfils the properties L1, L2, S1, and S2 if the recovery module of the accused node m is intact, if h() is a secure hashfunction, and if less than one third of the participating nodes are malicious (i.e., 3t &lt; n - 1). In order to prove Theorem <A href="41.html#6">1 we have to show that the properties L1, L2, S1, and S2 hold. We therefore first prove some helper Lemmas. Lemma 1. If all correct nodes accuse a node m, its recovery process will be initiated with high probability. Proof. The probability that less than 2(n-1) 3 accusations are received is equal to the probability that more than n-1 3 messages are either not sent or lost. Assuming that the t malicious nodes do not participate in the distributed update, at least n-1 3 -t+1 accusations must get lost. Given 0 p l &lt; 1, the probability for this is (p r l ) n-1 3 -t+1 + (p r l ) n-1 3 -t+2 + . . .+(p r l ) n-1 2(n-1)+3t 3 (p r l ) n-1 3 -t+1 . It holds that c &gt; 1 : r 1 such that 2(n-1)+3t 3 ,, p n-1 3 -t+1 l r &lt; n -c . Thus, the node m gets 2(n-1) 3 accusations w.h.p. and the recovery process is initiated. Lemma 2. Once the recovery process for a node m has been initiated it will eventually terminate as long as there remain at least k 1 correct nodes. Proof. In order that a frame is updated in a specific round, the dedicated request as well as its actual transmission must succeed. The probability that this is the case is (1 - p l ) 2 . With only one correct node (k = 1) the expected number of update rounds per frame a can be described as a Markov chain described by the expression a = (a + 1)(1 - (1 - p l ) 2 ) + (1 - p l ) 2 with the solution a = 1 (1-p l ) 2 . The overall expected number of rounds is thus aF = F (1-p l ) 2 O(1). In each round at most one request and F updates are transmitted, leading to an upper bound for its duration of (F + 1) p O(1). Altogether, the expected worst case duration is F (F +1) (1-p l ) 2 p O(1). Lemma 3. If no more than n-1 3 correct nodes accuse another correct node v M the recovery process will not be initiated. Proof. From each node only one accusation is accepted and thus the number of valid accusations is at most n-1 3 +t &lt; 2(n-1) 3 . Lemma 4. At the start of a program code update the target node m has stored the correct hash value h i for all frames, given that all correct nodes have loaded the same program code. Proof. Let's assume that there is a hash value h i which is not correct when the program code update starts. As a stored hash value is only substituted if the dedicated counter c i is zero, the node must have received at least as many wrong values as correct ones. From each node only one accusation is accepted, thus of the a 2(n-1) 3 received values at most t &lt; n-1 3 are false. It follows that at least a - t &gt; 2(n-1) 3 n -1 3 = n-1 3 &gt; t values must be correct, which contradicts the assumption that at least as many false as correct hash values were received. The properties L1, L2, and S1 are proven by Lemma <A href="41.html#6">1, <A href="41.html#6">2, and <A href="41.html#6">3, respectively. If the accused node is turned off or restarted, property S2 holds by definition. Otherwise, if the program code is updated, Lemma <A href="41.html#6">4 and property L2 guarantee that only correct code frames are installed and that the procedure finally terminates. EVALUATION In this section we provide an analytical evaluation of the proposed algorithm and present the findings of extensive simulations as well as of a real implementation for the BTnodes . The evaluated metrics are: (a) number of update rounds, (b) update load for the accused nodes, (c) update load for the other participating nodes, and (d) update duration . 5.1 Analytical Evaluation Number of Update Rounds In order to update a frame it is required that the dedicated request as well as the actual frame itself are successfully transmitted. The expected fraction of erroneous updates is therefore 1 - (1 - p l ) 2 . If one further assumes that the t malicious nodes do not participate in the program update, the fraction increases to 1 n -1-t n-1 (1 - p l ) 2 . Thus, the expected number of outstanding frames after the first and second update round are E f (1) = F (1 n -1-t n-1 (1 - p l ) 2 ) and E f (2) = E f (1)(1-n -1-t n-1 (1-p l ) 2 ) = F (1-n -1-t n-1 (1-p l ) 2 ) 2 , respectively. In general, the expected number of outstanding frames after i &gt; 0 rounds is: E f (i) = E f (i - 1) ,, 1 - n - 1 - t n - 1 (1 - p l ) 2 = F ,, 1 - n - 1 - t n - 1 (1 - p l ) 2 i Consequently, the expected number of update rounds is E r log(0.5) - log(F ) log(1 n -1-t n-1 (1 - p l ) 2 ) For reliable connections (p l = 0) we get E r log(0.5)-log(F ) log(1/3) O(1). In a worst case scenario a continuous sequence of frames is assigned to the t malicious nodes and thus at least t + 1, that is O(t) = O(n) rounds are required. Moreover, for a fixed p l , the expected number of rounds is (almost) independent of the cluster size n as 2 3 &lt; n-1-t n-1 1 for a fixed t, 0 t &lt; n-1 3 . 118 Update Load The expected amount of data (in bytes) to transfer for the accused node is E tm = E r -1 X i=0 ,, C req + C mac (n - 1) + C sel (n - 1) E f (i) F (C req + C mac (n - 1) + C sel (n - 1))E r and E tv = E r -1 X i=0 E f (i) n - 1 (1 - p l ) (C update + f s) F n - 1 (1 - p l ) (C update + f s) E r for the other participating nodes. In the above expressions (n - 1) E f (i) F is the expected number of addressed nodes and E f (i) n-1 (1 - p l ) expresses the expected number of successfully requested frames per node. Update Duration The total number of sent messages E m is bound by (F + 1)E r O(E r ) as there are only one request and no more that F update messages per round. Thus, the expected value of E m is in O(1) for reliable connections and in O(n) in the worst case. More precisely, the expected number of messages is given by E m = E r -1 X i=0 ,, 1 + E f (i) n - 1 (1 - p l ) = E r + (n - 1 - t)E tv C update + f s Neglecting the delays caused by the involved software routines , a good approximation for the update duration can be achieved by considering the overall transfer time and the delays caused by the round timeouts. The expected time to transfer all messages is E tm +(n-1-t)E tv B + E m mac whereas the overhead of the round timer is given by (E r - 1) round , resulting in a total update duration of E d E tm + (n - 1 - t)E tv B + E m mac + (E r - 1) round Parametrisation For the comparison of the analytical results with the simulation and implementation of the algorithm, the following parameters were used: Baudrate B 19.2 kBit/s B-MAC preamble mac 100 ms Round timeout round 3 s Request header size C req 12 Bytes Update header size C update 11 Bytes Frame selector site C sel 10 Bytes MAC size C mac 20 Bytes Frame size f s 1024 Bytes 5.2 Simulation The simulation of the algorithm was carried out with the Java based JiST/SWANS simulator <A href="41.html#9">[2]. In order to make the results comparable to the real BTnode implementation, the radio module was set up according to the characteristics of 0 10 20 30 40 50 60 70 80 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 packet loss ratio Expected number of update rounds F=100, n=10, t=0 (analytic) F=100, n=10, t=2 (analytic) F=100, n=10, t=0 (simulated) F=100, n=10, t=2 (simulated) Figure 2: Expected number of rounds to update a node. 0 50 100 150 200 250 300 350 400 450 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 packet loss ratio Expected duration (in sec.) to update a node F=100, n=10, t=0 (analytic) F=100, n=10, t=2 (analytic) F=100, n=10, t=0 (simulated) F=100, n=10, t=2 (simulated) Figure 3: Expected duration to update a node. the Chipcon CC1000 transceiver <A href="41.html#9">[4] and B-MAC was chosen as the data link layer protocol. The complete parametrisa-tion of the simulation is given in the table below: Transmission frequency 868 MHz Transmission power 5 dBm Receiver sensitivity -100 dBm Memory size 100 kByte Number of nodes 10 Deployment area 20 x 20 m (u.r.d.) 5.3 Implementation In addition to the above mentioned simulations, the algorithm was also implemented for the BTnodes, a wireless sensor platform running NutOS <A href="41.html#9">[3]. A detailed description of the created software is omitted due to space reasons but can be found in <A href="41.html#9">[22]. The implementation was evaluated by randomly distributing a cluster of 10 nodes in a field of 20 x 20 m, whereupon each node in turn initiated a complete program update. Altogether, over 100 recovery procedures where measured. 5.4 Results The packet loss ratio has, as expected, a significant effect on all evaluated metrics and each of them increases exponen-tially if the ratio worsens. The number of nodes, in contrast, 119 0 2 4 6 8 10 12 14 16 18 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 packet loss ratio Expected amount of data to transfer (in kBytes) for the updating node F=100, n=10, t=0 (analytic) F=100, n=10, t=2 (analytic) F=100, n=10, t=0 (simulated) F=100, n=10, t=2 (simulated) Figure 4: Expected update load for the accused node. 0 5 10 15 20 25 30 35 40 45 50 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 packet loss ratio Expected amount of data to transfer (in kBytes) for the participating nodes F=100, n=10, t=0 (analytic) F=100, n=10, t=2 (analytic) F=100, n=10, t=0 (simulated) F=100, n=10, t=2 (simulated) Figure 5: Expected update load for the participating nodes. has for a fixed packet loss ratio almost no negative impact on the evaluated metrics, showing that the algorithm itself scales well. Furthermore, the results show that the update algorithm is fair and equally distributes the update load over all participating nodes. Update Rounds and Update Duration Whilst the expected number of update rounds (see Figure <A href="41.html#7">2) is only of secondary importance, the update duration (see Figure <A href="41.html#7">3) is of major interest for the feasibility of the algorithm . The faster a node recovery is completed, the sooner the network is operable again. Even though the update duration almost triples from 50 to 150 s if the packet loss ratio increases from 0 to 40 percent, it is still in a range which most WSN application should be able to cope with. Update Load In a cluster of 10 nodes the update load for the accused node (see Figure <A href="41.html#8">4) is 0.5 to 3.5 kByte for 0 p l 0.4 and thus considerably smaller than for the other participating nodes (see Figure <A href="41.html#8">5) with a load of 12 to 24 kByte. However, the latter is, as expected, inverse proportional to the cluster size (see Figure <A href="41.html#8">7): the larger a cluster and the lower the number of malicious nodes, the smaller the expected update load per participating node. 0 50 100 150 200 250 5 10 15 20 25 30 number of nodes Expected duration (in sec.) to update a node F=100, plr=0.1, t=0 F=100, plr=0.5, t=0 F=100, plr=0.1, t=n/3 F=100, plr=0.5, t=n/3 Figure 6: Influence of the cluster size on the update duration . 0 10 20 30 40 50 60 70 80 90 5 10 15 20 25 30 number of nodes Expected amount of data to transfer (in kBytes) for the participating nodes F=100, plr=0.1, t=0 F=100, plr=0.5, t=0 F=100, plr=0.1, t=n/3 F=100, plr=0.5, t=n/3 Figure 7: Influence of the cluster size on the expected update load for the participating nodes. Implementation In the experiments conducted with the BTnode implementation , the average number of update rounds was five, the update duration about 100 s ( 10 s), and the load for the participating nodes about 15 kByte ( 1 kByte). Applied to the analytical model this would mean that the gross packet loss ratio was roughly 20%. SUMMARY AND CONCLUSIONS In this paper, we presented an autonomous and distributed recovery algorithm for sensor networks. The algorithm allows for bringing malicious or failed nodes back into normal operation or, at least, for securely shutting them down. Particularly in remote or unwieldy areas, such as deserts, the bottom of the sea, mountains, or even on planets in outer space, where redeployment is expensive and sensor nodes cannot easily be exchanged or maintained, the application of a node recovery system is most likely to extend the lifetime of the whole network. The results of the simulation and analytical analysis were confirmed by the real BTnode implementation. They show that recovering sensor nodes is as any form of reprogramming an expensive, though feasible task. Moreover, the proposed program code update algorithm is not only prov-120 ably secure but also fair and robust. It distributes the update load equally over all participating nodes and terminates as long as at least one of the nodes remains correct. To all presented recovery measures applies that they are only applicable if at least the therefore needed systems of the corrupt node still work correctly. However, although it is generally agreed that entirely tamper-proof sensor nodes are too expensive, current trends in the hardware development of embedded devices indicate that at least some logical and physical protection (e.g., CPUs which support isolated memory areas or automatic memory erason if a node is tempered with) will be available in the near future. We discussed how these upcoming technologies can be exploited to protect the recovery mechanisms of a sensor node and what is already feasible with existing systems. REFERENCES [1] T. Alves and D. Felton. TrustZone: Integrated Hardware and Software Security. ARM Ltd, July 2004. [2] R. Barr, Z. J. Haas, and R. van Renesse. Jist: an efficient approach to simulation using virtual machines: Research articles. Softw. Pract. Exper., 35(6):539576, 2005. [3] J. Beutel, O. Kasten, and M. Ringwald. Poster abstract: Btnodes a distributed platform for sensor nodes. In Proceedings of the 1st International Conference on Embedded Networked Sensor Systems, pages 292 293, Los Angeles, California, USA, Jan. 2003. ACM Press. http://www.btnode.ethz.ch/. [4] Chipcon AS, Oslo, Norway. Single Chip Very Low Power RF Transceiver, Rev. 2.1, Apr. 2002. http://www.chipcon.com/. [5] A. P. R. da Silva, M. H. T. Martins, B. P. S. Rocha, A. A. F. Loureiro, L. B. Ruiz, and H. C. Wong. Decentralized intrusion detection in wireless sensor networks. In Q2SWinet '05: Proceedings of the 1st ACM international workshop on Quality of service & security in wireless and mobile networks, pages 1623, New York, NY, USA, 2005. ACM Press. [6] S. B. Douglas Herbert, Yung-Hsiang Lu and Z. Li. Detection and repair of software errors in hierarchical sensor networks. To appear in IEEE conference on Sensor Networks and Ubiquitous Trustworthy Computing (SUTC), June 2006. [7] P. K. Dutta, J. W. Hui, D. C. Chu, and D. E. Culler. Towards secure network programming and recovery in wireless sensor networks. Technical Report UCB/EECS-2005-7, Electrical Engineering and Computer Sciences University of California at Berkeley, Oct. 2005. [8] S. Ganeriwal and M. B. Srivastava. Reputation-based framework for high integrity sensor networks. In SASN '04: Proceedings of the 2nd ACM workshop on Security of ad hoc and sensor networks, pages 6677, New York, NY, USA, 2004. ACM Press. [9] 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, Jan. 2005. [10] C. Hsin and M. Liu. A distributed monitoring mechanism for wireless sensor networks. In WiSE '02: Proceedings of the 3rd ACM workshop on Wireless security, pages 5766, New York, NY, USA, 2002. ACM Press. [11] Intel Corporation. LaGrande Technology Architectural Overview, Sept. 2003. [12] P. Inverardi, L. Mostarda, and A. Navarra. Distributed IDSs for enhancing security in mobile wireless sensor networks. AINA, 2:116120, 2006. [13] 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 (SECON), 2004. [14] I. Khalil, S. Bagchi, and C. Nina-Rotaru. Dicas: Detection, diagnosis and isolation of control attacks in sensor networks. securecomm, 00:89100, 2005. [15] P. Kocher, R. Lee, G. McGraw, and A. Raghunathan. Security as a new dimension in embedded system design. In DAC '04: Proceedings of the 41st annual conference on Design automation, pages 753760, New York, NY, USA, 2004. ACM Press. Moderator-Srivaths Ravi. [16] S. S. Kulkarni and L. Wang. Mnp: Multihop network reprogramming service for sensor networks. icdcs, 00:716, 2005. [17] T. Liu and M. Martonosi. Impala: a middleware system for managing autonomic, parallel sensor systems. SIGPLAN Not., 38(10):107118, 2003. [18] S. Northcutt and J. Novak. IDS: Intrusion Detection-Systeme. mitp Verlag Bonn, 2001. [19] N. B. of Standards. Security Requirements for Cryptographic Modules. National Bureau of Standards, Dec. 2002. [20] S. Ravi, A. Raghunathan, P. Kocher, and S. Hattangady. Security in embedded systems: Design challenges. Trans. on Embedded Computing Sys., 3(3):461491, 2004. [21] E. Shi and A. Perrig. Designing secure sensor networks. IEEE Wireless Communication Magazine, 11(6):3843, Dec. 2004. [22] M. Strasser. Intrusion detection and failure recovery in sensor networks. Master's thesis, Department of Computer Science, ETH Zurich, 2005. [23] H. Vogt, M. Ringwald, and M. Strasser. Intrusion detection and failure recovery in sensor nodes. In Tagungsband INFORMATIK 2005, Workshop Proceedings, LNCS, Heidelberg, Germany, Sept. 2005. Springer-Verlag. 121

sensor networks;countermeasures;intrusion detection;Wireless Sensor Networks;Node Recovery;Intrusion Detection;node recovery;IDS;sensor nodes;security;distributed algorithm

42

Bayesian Online Classifiers for Text Classification and Filtering

This paper explores the use of Bayesian online classifiers to classify text documents. Empirical results indicate that these classifiers are comparable with the best text classification systems. Furthermore, the online approach offers the advantage of continuous learning in the batch-adaptive text filtering task.

INTRODUCTION Faced with massive information everyday, we need automated means for classifying text documents. Since handcrafting text classifiers is a tedious process, machine learning methods can assist in solving this problem[15, 7, 27]. Yang & Liu[27] provides a comprehensive comparison of supervised machine learning methods for text classification. In this paper we will show that certain Bayesian classifiers are comparable with Support Vector Machines[23], one of the best methods reported in [27]. In particular, we will evaluate the Bayesian online perceptron[17, 20] and the Bayesian online Gaussian process[3]. For text classification and filtering, where the initial training set is large, online approaches are useful because they allow continuous learning without storing all the previously 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'02, August 11-15, 2002, Tampere, Finland. Copyright 2002 ACM 1-58113-561-0/02/0008 ... $ 5.00. seen data. This continuous learning allows the utilization of information obtained from subsequent data after the initial training. Bayes' rule allows online learning to be performed in a principled way[16, 20, 17]. We will evaluate the Bayesian online perceptron, together with information gain considerations, on the batch-adaptive filtering task[18]. CLASSIFICATION AND FILTERING For the text classification taskdefined by Lewis[9], we have a set of predefined categories and a set of documents. For each category, the document set is partitioned into two mutually exclusive sets of relevant and irrelevant documents. The goal of a text classification system is to determine whether a given document belongs to any of the predefined categories . Since the document can belong to zero, one, or more categories, the system can be a collection of binary classifiers , in which one classifier classifies for one category. In Text REtrieval Conference (TREC), the above taskis known as batch filtering. We will consider a variant of batch filtering called the batch-adaptive filtering[18]. In this task, during testing, if a document is retrieved by the classifier, the relevance judgement is fed backto the classifier. This feedbackcan be used to improve the classifier. 2.1 Corpora and Data For text classification, we use the ModApte version of the Reuters-21578 corpus 1 , where unlabelled documents are removed. This version has 9,603 training documents and 3,299 test documents. Following [7, 27], only categories that have at least one document in the training and test set are retained. This reduces the number of categories to 90. For batch-adaptive filtering, we attempt the taskof TREC-9 [18], where the OHSUMED collection[6] is used. We will evaluate on the OHSU topic-set, which consists of 63 topics. The training and test material consist of 54,710 and 293,856 documents respectively. In addition, there is a topic statement for each topic. For our purpose, this is treated as an additional training document for that topic. We will only use the title, abstract, author, and source sections of the documents for training and testing. 2.2 Representation There are various ways to transform a document into a representation convenient for classification. We will use the 1 Available via http://www.daviddlewis.com/resources/ testcollections/reuters21578. 97 bag-of-words approach, where we only retain frequencies of words after tokenisation, stemming, and stop-words removal . These frequencies can be normalized using various schemes[19, 6]; we use the ltc normalization: l i,d = 1 + log 2 T F i,d t i = log 2 N n i ltc i,d = l i,d t i j {terms in d} (l j,d t j ) 2 , where the subscripts i and d denote the ith term and the dth document respectively, T F i,d is the frequency of the ith term in the dth document, n i is the document-frequency of the ith term, and N is the total number of documents. 2.3 Feature Selection Metric Given a set of candidate terms, we select features from the set using the likelihood ratio for binomial distribution advocated by Dunning[5]: = R t +R t N R t +R t N t +N t N N t +N t R t R t +N t R t N t R t +N t N t R t R t +N t R t N t R t +N t N t , where R t (N t ) is the number of relevant (non-relevant) training documents which contain the term, R t (N t ) is the number of relevant (non-relevant) training documents which do not, and N is the total number of training documents. Asymptotically, -2 ln is 2 distributed with 1 degree of freedom. We choose terms with -2 ln more than 12.13, i.e. at 0.05% significance level. More details on the feature selection procedures will be given in section 4. 2.4 Performance Measures To evaluate a text classification system, we use the F 1 measure introduced by van Rijsbergen[22]. This measure combines recall and precision in the following way: Recall = number of correct positive predictions number of positive examples Precision = number of correct positive predictions number of positive predictions F 1 = 2 Recall Precision Recall + Precision . For ease of comparison, we summarize the F 1 scores over the different categories using the micro- and macro-averages of F 1 scores[11, 27]: Micro-avg F 1 = F 1 over categories and documents Macro-avg F 1 = average of within-category F 1 values. The micro- and macro-average F 1 emphasize the performance of the system on common and rare categories respectively . Using these averages, we can observe the effect of different kinds of data on a text classification system. In addition, for comparing two text classification systems, we use the micro sign-test (s-test) and the macro sign-test (S-test), which are two significance tests first used for comparing text classification systems in [27]. The s-test compares all the binary decisions made by the systems, while the S-test compares the within-category F 1 values. Similar to the F 1 averages, the s-test and S-test compare the performance of two systems on common and rare categories respectively. To evaluate a batch-adaptive filtering system, we use the T9P measure of TREC-9[18]: T9P = number of correct positive predictions Max(50, number of positive predictions) , which is precision, with a penalty for not retrieving 50 documents BAYESIAN ONLINE LEARNING Most of this section is based on workby Opper[17], Solla & Winther[20], and Csat o & Opper[3]. Suppose that each document is described by a vector x, and that the relevance indicator of x for a category is given by label y {-1, 1}, where -1 and 1 indicates irrelevant and relevant respectively. Given m instances of past data D m = {(y t , x t ), t = 1...m}, the predictive probability of the relevance of a document described by x is p(y|x, D m ) = da p(y|x, a)p(a|D m ), where we have introduced the classifier a to assist us in the prediction. In the Bayesian approach, a is a random variable with probability density p(a|D m ), and we integrate over all the possible values of a to obtain the prediction. Our aim is to obtain a reasonable description of a. In the Bayesian online learning framework[16, 20, 17], we begin with a prior p(a|D 0 ), and perform incremental Bayes' updates to obtain the posterior as data arrives: p(a|D t +1 ) = p(y t +1 |x t +1 , a)p(a|D t ) da p(y t +1 |x t +1 , a)p(a|D t ) . To make the learning online, the explicit dependence of the posterior p(a|D t +1 ) on the past data is removed by approximating it with a distribution p(a|A t +1 ), where A t +1 characterizes the distribution of a at time t + 1. For example , if p(a|A t +1 ) is a Gaussian, then A t +1 refers to its mean and covariance. Hence, starting from the prior p 0 (a) = p(a|A 0 ), learning from a new example (y t +1 , x t +1 ) comprises two steps: Update the posterior using Bayes rule p(a|A t , (y t +1 , x t +1 )) p(y t +1 |x t +1 , a) p(a|A t ) Approximate the updated posterior by parameterisation p(a|A t , (y t +1 , x t +1 )) p(a|A t +1 ), where the approximation step is done by minimizing the Kullback-Leibler distance between the the approximating and approximated distributions. The amount of information gained about a after learning from a new example can be expressed as the Kullback-Leibler distance between the posterior and prior distributions [25]: IG(y t +1 , x t +1 |D t ) = da p(a|D t +1 ) log 2 p(a|D t +1 ) p(a|D t ) da p(a|A t +1 ) log 2 p(a|A t +1 ) p(a|A t ) , where instances of the data D are replaced by the summaries A in the approximation. 98 To simplify notation henceforth, we use p t (a) and . . . t to denote p(a|A t ) and averages taken over p(a|A t ) respectively. For example, the predictive probability can be rewritten as p(y|x, D t ) p(y|x, A t ) = da p(y|x, a)p t (a) = p(y|x, a) t . In the following sections, the scalar field h = a x will also be used to simplify notation and calculation. 3.1 Bayesian Online Perceptron Consider the case where a describes a perceptron. We then define the likelihood as a probit model p(y|x, a) = ya x 0 , where 2 0 is a fixed noise variance, and is the cumulative Gaussian distribution (u) = 1 2 u d e 2 / 2 . If p 0 (a) is the spherical unit Gaussian, and p t (a) is the Gaussian approximation, Opper[16, 17] and Solla & Winther[20] obtain the following updates by equating the means and covariances of p(a|A t +1 ) and p(a|A t , (y t +1 , x t +1 )): a t +1 = a t + s t +1 h t ln p(y t +1 |h) t C t +1 = C t + s t +1 s T t +1 2 h 2 t ln p(y t +1 |h) t , where s t +1 = C t x t +1 , p(y t +1 |h) t = y t +1 h t t +1 , 2 t +1 = 2 0 + x T t +1 C t x t +1 and h t = a T t x t +1 . 3.1.1 Algorithm Training the Bayesian online perceptron on m data involves successive calculation of the means a t and covariances C t of the posteriors, for t {1, ..., m}: 1. Initialize a 0 to be 0 and C 0 to be 1 (identity matrix), i.e. a spherical unit Gaussian centred at origin. 2. For t = 0, 1, ..., m - 1 3. y t +1 is the relevance indicator for document x t +1 4. Calculate s t +1 , t +1 , h t and p(y t +1 |h) t 5. Calculate u = y t+1 h t t+1 and = 1 2 exp( 1 2 u 2 ) 6. Calculate h t ln p(y t +1 |h) t = y t+1 t+1 1 p (y t+1 |h) t 7. Calculate 2 h 2 t ln p(y t +1 |h) t = - 1 2 t +1 u p(y t +1 |h) t + p(y t +1 |h) t 2 8. Calculate a t +1 and C t +1 The prediction for datum (y, x) simply involves the calculation of p(y|x, a) m = p(y|h) m . 3.2 Bayesian Online Gaussian Process Gaussian process (GP) has been constrained to problems with small data sets until recently when Csat o & Opper[3] and Williams & Seeger[24] introduced efficient and effective approximations to the full GP formulation. This section will outline the approach in [3]. In the GP framework, a describes a function consisting of function values {a(x)}. Using the probit model, the likelihood can be expressed as p(y|x, a) = ya(x) 0 , where 0 and are described in section 3.1. In addition, p 0 (a) is a GP prior which specifies a Gaussian distribution with zero mean function and covariance/kernel function K 0 (x, x ) over a function space. If p t (a) is also a Gaussian process, then Csat o & Opper obtain the following updates by equating the means and covariances of p(a|A t +1 ) and p(a|A t , (y t +1 , x t +1 )): a t +1 = a t + s t +1 h t ln p(y t +1 |h) t C t +1 = C t + s t +1 s T t +1 2 h 2 t ln p(y t +1 |h) t , where s t +1 = C t k t +1 + e t +1 , p(y t +1 |h) t = y t +1 h t t +1 , 2 t +1 = 2 0 + k t +1 + k T t +1 C t k t +1 and h t = a(x t +1 ) t = a T t k t +1 Notice the similarities to the updates in section 3.1. The main difference is the `kernel trick' introduced into the equations through k t +1 = K 0 (x t +1 , x t +1 ) and k t +1 = (K 0 (x 1 , x t +1 ), . . . , K 0 (x t , x t +1 )) T New inputs x t +1 are added sequentially to the system via the (t + 1)th unit vector e t +1 . This results in a quadratic increase in matrix size, and is a drawbackfor large data sets, such as those for text classification. Csat o & Opper overcome this by introducing sparseness into the GP. The idea is to replace e t +1 by the projection ^ e t +1 = K -1 t k t +1 , where K t = {K 0 (x i , x j ), i, j = 1 . . . t}. This approximation introduces an error t +1 = (k t +1 - k T t +1 K -1 t k t +1 ) h t ln p(y t +1 |h) t , which is used to decide when to employ the approximation. Hence, at any time the algorithm holds a set of basis vectors . It is usually desirable to limit the size of this set. To accommodate this, Csat o & Opper describe a procedure for removing a basis vector from the set by reversing the process of adding new inputs. For lackof space, the algorithm for the Bayesian Online Gaussian Process will not be given here. The reader is re-ferred to [3] for more information. 99 EVALUATION In this evaluation, we will compare Bayesian online perceptron , Bayesian online Gaussian process, and Support Vector Machines (SVM)[23]. SVM is one of the best performing learning algorithms on the Reuters-21578 corpus[7, 27]. The Bayesian methods are as described in section 3, while for SVM we will use the SV M light package by Joachims[8]. Since SVM is a batch method, to have a fair comparison, the online methods are iterated through the training data 3 times before testing. 2 4.1.1 Feature Selection For the Reuters-21578 corpus, we select as features for each category the set of all words for which -2 ln &gt; 12.13. We further prune these by using only the top 300 features. This reduces the computation time required for the calculation of the covariances of the Bayesian classifiers. Since SVM is known to perform well for many features, for the SVM classifiers we also use the set of words which occur in at least 3 training documents[7]. This gives us 8,362 words. Note that these words are non-category specific. 4.1.2 Thresholding The probabilistic outputs from the Bayesian classifiers can be used in various ways. The most direct way is to use the Bayes decision rule, p(y = 1|x, D m ) &gt; 0.5, to determine the relevance of the document described by x. 3 However, as discussed in [10, 26], this is not optimal for the chosen evaluation measure. Therefore, in addition to 0.5 thresholding, we also empir-ically optimise the threshold for each category for the F 1 measure on the training documents. This scheme, which we shall call MaxF1, has also been employed in [27] for thresholding kNN and LLSF classifiers. The difference from our approach is that the threshold in [27] is calculated over a validation set. We do not use a validation set because we feel that, for very rare categories, it is hard to obtain a reasonable validation set from the training documents. For the Bayesian classifiers, we also perform an analyti-cal threshold optimisation suggested by Lewis[10]. In this scheme, which we shall call ExpectedF1, the threshold for each category is selected to optimise the expected F 1 : E [F 1 ] i D (1 - p i ) if |D + | = 0 2 iD+ p i | D + | + iD p i otherwise, where is the threshold, p i is the probability assigned to document i by the classifier, D is the set of all test documents , and D + is the set of test documents with probabilities higher than the threshold . Note that ExpectedF1 can only be applied after the probabilities for all the test documents are assigned. Hence the classification can only be done in batch. This is unlike the first two schemes, where classification can be done online. 4.1.3 Results and Discussion 2 See section A.2 for discussion on the number of passes. 3 For SVM, to minimise structural risks, we would classify the document as relevant if w x + b &gt; 0, where w is the hyperplane, and b is the bias. 4 See section A.3 for discussion on the jitter terms ij . Table 1: Description of Methods Description 4 SVM-1 K 0 = x i x j + 1 SVM-2 K 0 = (x i x j + 1) 2 SVM-R1 K 0 = exp( 1 2 |x i - x j | 2 ) Perceptron 0 = 0.5, one fixed feature (for bias) GP-1 0 = 0.5, K 0 = x i x j + 1 + 10 -4 ij GP-2 0 = 0.5, K 0 = (x i x j + 1) 2 + 10 -4 ij GP-R1 0 = 0.5, K 0 = exp( 1 2 |x i - x j | 2 ) + 10 -4 ij Table 2: Micro-/Macro-average F 1 0.5 MaxF1 ExpectedF1 SVM a -1 86.15 / 42.63 86.35 / 56.92 SVM a -2 85.44 / 40.13 86.19 / 56.42 SVM a -R1 84.99 / 37.61 86.63 / 53.14 SVM b -1 85.60 / 52.03 85.05 / 52.43 SVM b -2 85.60 / 50.53 84.50 / 50.49 SVM b -R1 85.75 / 50.52 84.65 / 51.27 Perceptron 85.12 / 45.23 86.69 / 52.16 86.44 / 53.08 GP-1 85.08 / 45.20 86.73 / 52.12 86.54 / 53.12 GP-2 85.58 / 47.90 86.60 / 52.19 86.77 / 55.04 GP-R1 85.18 / 44.88 86.76 / 52.61 86.93 / 53.35 Table 1 lists the parameters for the algorithms used in our evaluation, while Table 2 and 3 tabulate the results. There are two sets of results for SVM, and they are labeled SVM a and SVM b . The latter uses the same set of features as the Bayesian classifiers (i.e. using the -2 ln measure), while the former uses the set of 8,362 words as features. Table 2 summarizes the results using F 1 averages. Table 3 compares the classifiers using s-test and S-test. Here, the MaxF1 thresholds are used for the classification decisions. Each row in these tables compares the method listed in the first column with the other methods. The significance levels from [27] are used. Several observations can be made: Generally, MaxF1 thresholding increases the performance of all the systems, especially for rare categories. For the Bayesian classifiers, ExpectedF1 thresholding improves the performance of the systems on rare categories . Perceptron implicitly implements the kernel used by GP-1, hence their similar results. With MaxF1 thresholding, feature selection impedes the performance of SVM. In Table 2, SVM with 8,362 features have slightly lower micro-average F 1 to the Bayesian classifiers. However, the s-tests in Table 3 show that Bayesian classifiers outperform SVM for significantly many common categories . Hence, in addition to computing average F 1 measures, it is useful to perform sign tests. As shown in Table 3, for limited features, Bayesian classifiers outperform SVM for both common and rare categories. Based on the sign tests, the Bayesian classifiers outperform SVM (using 8,362 words) for common categories, and vice versa for rare categories. 100 Table 3: s-test/S-test using MaxF1 thresholding SVM a -1 SVM a -2 SVM a -R1 SVM b -1 SVM b -2 SVM b -R1 Pptron GP-1 GP-2 GP-R1 SVM a -1 / &lt; / / / / / / / / SVM a -2 / / &gt; / / / / / / / &gt; SVM a -R1 &gt; / / / / / &lt; / &gt; &lt; / &gt; / &lt; / SVM b -1 / &lt; / / / &gt; / &gt; / &lt; / / &lt; / &lt; SVM b -2 / / / / / / &lt; / &lt; / / SVM b -R1 / / / &lt; / &lt; / / &lt; / / / Perceptron / / &gt; / &lt; / &gt; / &gt; / &gt; / / / GP-1 / / &gt; / &lt; / / &gt; / / / / GP-2 / / / / &gt; / / / / / GP-R1 / / &lt; &gt; / / &gt; / / / / / " " or " " means P-value 0.01; "&gt;" or "&lt;" means 0.01 &lt; P-value 0.05; " " means P-value &gt; 0.05. The last observation suggests that one can use Bayesian classifiers for common categories, and SVM for rare ones. 4.2 Filtering on OHSUMED In this section, only the Bayesian online perceptron will be considered. In order to avoid numerical integration of the information gain measure, instead of the probit model of section 3.1, here we use a simpler likelihood model in which the outputs are flipped with fixed probability : p(y|x, a) = + (1 - 2) (ya x) , where (x) = 1 x &gt; 0 0 otherwise. The update equations will also change accordingly, e.g. p(y t +1 |h) t = + (1 - 2) y t +1 h t t +1 , 2 t +1 = x T t +1 C t x t +1 and h t = a T t x t +1 . Using this likelihood measure, we can express the information gained from datum (y t +1 , x t +1 ) as IG(y t +1 , x t +1 |D t ) log 2 + y t +1 h t +1 t +1 log 2 1 -log 2 p(y t +1 |h) t , where 2 t +1 = x T t +1 C t +1 x t +1 and h t +1 = a T t +1 x t +1 . We use = 0.1 in this evaluation. The following sections will describe the algorithm in detail. To simplify presentation , we will divide the batch-adaptive filtering taskinto batch and adaptive phases. 4.2.1 Feature Selection and Adaptation During the batch phase, words for which -2 ln &gt; 12.13 are selected as features. During the adaptive phase, when we obtain a feedback, we update the features by adding any new words with -2 ln &gt; 12.13. When a feature is added, the distribution of the perceptron a is extended by one dimension: a a 0 C C 0 0 1 . 4.2.2 Training the classifier During the batch phase, the classifier is iterated through the training documents 3 times. In addition, the relevant documents are collected for use during the adaptive phase. During the adaptive phase, retrieved relevant documents are added to this collection. When a document is retrieved, the classifier is trained on that document and its given relevance judgement. The classifier will be trained on irrelevant documents most of the time. To prevent it from "forgetting" relevant documents due to its limited capacity, whenever we train on an irrelevant document, we would also train on a past relevant document. This past relevant document is chosen succes-sively from the collection of relevant documents. This is needed also because new features might have been added since a relevant document was last trained on. Hence the classifier would be able to gather new information from the same document again due to the additional features. Note that the past relevant document does not need to be chosen in successive order. Instead, it can be chosen using a probability distribution over the collection. This will be desirable when handling topic-drifts. We will evaluate the effectiveness of this strategy of retraining on past retrieved relevant documents, and denote its use by +rel. Though its use means that the algorithm is no longer online, asymptotic efficiency is unaffected, since only one past document is used for training at any instance. 4.2.3 Information Gain During testing, there are two reasons why we retrieve a document. The first is that it is relevant, i.e. p(y = 1 |x, D t ) &gt; 0.5, where x represents the document. The second is that, although the document is deemed irrelevant by the classifier, the classifier would gain useful information from the document. Using the measure IG(y, x|D t ), we calculate the expected information gain IG(x|D t ) = {-1,1} p(y = |x, D t ) IG(y = , x|D t ). 101 0 10 20 30 40 50 60 70 80 90 100 0 0.2 0.4 0.6 0.8 1 N ret Target number of documents = 50 Figure 1: versus N ret tuned for T9P A document is then deemed useful if its expected information gain is at least . Optimizing for the T9P measure (i.e. targeting 50 documents), we choose to be = 0.999 1 + exp - N ret - 50.0 10 -1 + 0.001, where N ret is the total number of documents that the system has retrieved. Figure 1 plots against N ret . Note that this is a kind of active learning, where the willingness to tradeoff precision for learning decreases with N ret . The use of this information gain criteria will be denoted by +ig. We will test the effectiveness of the information gain strategy , against an alternative one. The alternative, denoted by +rnd, will randomly select documents to retrieve based on the probability U = 0 if N ret &gt;= 50 50-N ret 293856 otherwise, where 293,856 is the number of test documents. 4.2.4 Results and Discussion Table 4 lists the results of seven systems. The first two are of Microsoft Research Cambridge and Fudan University respectively . These are the only runs in TREC-9 for the task. The third is of the system as described in full, i.e. Bayesian online perceptron, with retraining on past retrieved relevant documents, and with the use of information gain. The rest are of the Bayesian online perceptron with different combinations of strategies. Besides the T9P measure, for the sake of completeness, Table 4 also lists the other measures used in TREC-9. Taken together, the measures show that Bayesian online perceptron , together with the consideration for information gain, is a very competitive method. For the systems with +rel, the collection of past known relevant documents is kept. Although Microsoft uses this same collection for its query reformulation, another collection of all previously seen documents is used for threshold adaptation. Fudan maintains a collection of past retrieved documents and uses this collection for query adaptation. 5 [18] reports results from run ok9bfr2po, while we report results from the slightly better run ok9bf2po. 0 2 4 6 8 10 12 14 16 18 20 40 50 60 70 80 90 100 110 120 130 Average number of relevant documents retrieved Average number of features Pptron+rel+ig Pptron+ig Pptron+rnd Pptron Figure 2: Variation of the number of features as relevant documents are retrieved. The plots for Pptron+rel+ig and Pptron+ig are very close. So are the plots for Pptron+rnd and Pptron. In a typical operational system, retrieved relevant documents are usually retained, while irrelevant documents are usually discarded. Therefore +rel is a practical strategy to adopt. Figure 2 plots the average number of features during the adaptive phase. We can see that features are constantly added as relevant documents are seen. When the classifier is retrained on past documents, the new features enable the classifier to gain new information from these documents. If we compare the results for Pptron+rel and Pptron in Table 4, we find that not training on past documents causes the number of relevant documents retrieved to drop by 5%. Similarly, for Pptron+rel+ig and Pptron+ig, the drop is 8%. Table 5 breaks down the retrieved documents into those that the classifier deems relevant and those that the classifier is actually querying for information, for Pptron+ig and Pptron+rnd. The table shows that none of the documents randomly queried are relevant documents. This is not surprising, since only an average of 0.017% of the test documents are relevant. In contrast, the information gain strategy is able to retrieve 313 relevant documents, which is 26.1% of the documents queried. This is a significant result. Consider Pptron+ig. Table 4 shows that for Pptron, when the information gain strategy is removed, only 731 relevant documents will be retrieved. Hence, although most of the documents queried are irrelevant, information gained from these queries helps recall by the classifier (i.e. 815 documents versus 731 documents), which is important for reaching the target of 50 documents. MacKay[13] has noted the phenomenon of querying for irrelevant documents which are at the edges of the input space, and suggested maximizing information in a defined region of interest instead. Finding this region for batch-adaptive filtering remains a subject for further research. Comparing the four plots in Figure 2, we find that, on average, the information gain strategy causes about 3% more features to be discovered for the same number of relevant documents retrieved. A consequence of this is better recall. 102 Table 4: Results for Batch-adaptive filtering optimized for T9P measure. Microsoft 5 Fudan Pptron+rel+ig Pptron+ig Pptron+rnd Pptron+rel Pptron Total retrieved 3562 3251 2716 2391 2533 1157 1057 Relevant retrieved 1095 1061 1227 1128 732 772 731 Macro-average recall 39.5 37.9 36.2 33.3 20.0 20.8 20.0 Macro-average precision 30.5 32.2 35.8 35.8 21.6 61.9 62.3 Mean T9P 30.5 31.7 31.3 29.8 19.2 21.5 20.8 Mean Utility -4.397 -1.079 15.318 15.762 -5.349 18.397 17.730 Mean T9U -4.397 -1.079 15.318 15.762 -5.349 18.397 17.730 Mean scaled utility -0.596 -0.461 -0.025 0.016 -0.397 0.141 0.138 Zero returns 0 0 0 0 0 8 0 Table 5: Breakdown of documents retrieved for Pptron+ig and Pptron+rnd. The numbers for the latter are in brackets. Relevant Not Relevant Total # docs retrieved by perceptron classifier proper 815 (732) 378 (345) 1193 (1077) # docs retrieved by information gain (or random strategy) 313 (0) 885 (1456) 1198 (1456) Total 1128 (732) 1263 (1801) 2391 (2533) CONCLUSIONS AND FURTHER WORK We have implemented and tested Bayesian online perceptron and Gaussian processes on the text classification problem , and have shown that their performance is comparable to that of SVM, one of the best learning algorithms on text classification in the published literature. We have also demonstrated the effectiveness of online learning with information gain on the TREC-9 batch-adaptive filtering task. Our results on text classification suggest that one can use Bayesian classifiers for common categories, and maximum margin classifiers for rare categories. The partitioning of the categories into common and rare ones in an optimal way is an interesting problem. SVM has been employed to use relevance feedbackby Drucker et al [4], where the retrieval is in groups of 10 documents . In essence, this is a form of adaptive routing. It would be instructive to see how Bayesian classifiers perform here, without storing too many previously seen documents. It would also be interesting to compare the merits of incremental SVM[21, 1] with the Bayesian online classifiers. Acknowledgments We would like to thank Lehel Csat o for providing details on the implementation of the Gaussian process, Wee Meng Soon for assisting in the data preparation, Yiming Yang for clarifying the representation used in [27], and Loo Nin Teow for proof-reading the manuscript. We would also like to thankthe reviewers for their many helpful comments in improving the paper. REFERENCES [1] G. Cauwenberghs and T. Poggio. Incremental and decremental support vector machine learning. In T. K. Leen, T. G. Dietterich, and V. Tresp, editors, NIPS 2000, volume 13. The MIT Press, 2001. [2] D. Cox and E. Snell. Analysis of Binary Data. Chapman & Hall, London, 2nd edition, 1989. [3] L. Csat o and M. Opper. Sparse representation for Gaussian process models. In T. K. Leen, T. G. Dietterich, and V. Tresp, editors, NIPS 2000, volume 13. The MIT Press, 2001. [4] H. Drucker, B. Shahrary, and D. C. Gibbon. Relevance feedbackusing support vector machines. In Proceedings of the 2001 International Conference on Machine Learning, 2001. [5] T. E. Dunning. Accurate methods for the statistics of surprise and coincidence. Computational Linguistics, 19(1):6174, 1993. [6] W. Hersh, C. Buckley, T. Leone, and D. Hickam. OHSUMED: An interactive retrieval evaluation and new large test collection for research. In Proceedings of the 17th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 192201, 1994. [7] T. Joachims. Text categorization with support vector machines: Learning with many relevant features. In Proceedings of the European Conference on Machine Learning (ECML), pages 137142, 1998. [8] T. Joachims. Making large-scale SVM learning practical. In B. Sch olkopf, C. Burges, and A. Smola, editors, Advances in Kernel Methods -- Support Vector Learning, chapter 11. The MIT Press, 1999. [9] D. D. Lewis. Representation and Learning in Information Retrieval. PhD thesis, Department of Computer and Information Science, University of Massachusetts at Amherst, 1992. [10] D. D. Lewis. Evaluating and optimizing automomous text classification systems. In Proceedings of the 18th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 246254, 1995. [11] D. D. Lewis, R. E. Schapire, J. P. Callan, and R. Papka. Training algorithms for linear text classifiers. In Proceedings of the 19th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 298306, 1996. [12] D. J. Mackay. Bayesian interpolation. Neural Computation, 4(3):415447, 1991. [13] D. J. Mackay. Information-based objective functions for active data selection. Neural Computation, 4(4):590604, 1992. 103 [14] R. M. Neal. Monte Carlo implementation of Gaussian process models for Bayesian regression and classification. Technical Report CRG-TR-97-2, Department of Computer Science, University of Toronto, January 1997. [15] H. T. Ng, W. B. Goh, and K. L. Low. Feature selection, perceptron learning, and a usability case study for text categorization. In Proceedings of the 20th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 6773, 1997. [16] M. Opper. Online versus offline learning from random examples: General results. Physical Review Letters, 77:46714674, 1996. [17] M. Opper. A Bayesian approach to online learning. In D. Saad, editor, On-Line Learning in Neural Networks. Combridge University Press, 1998. [18] S. Robertson and D. A. Hull. The TREC-9 filtering trackfinal report. In Proceedings of the 9th Text REtrieval Conference (TREC-9), pages 2540, 2001. [19] G. Salton and C. Buckley. Term-weighting approaches in automatic text retrieval. Information Processing and Management, 24(5):513523, 1988. [20] S. A. Solla and O. Winther. Optimal perceptron learning: an online Bayesian approach. In D. Saad, editor, On-Line Learning in Neural Networks. Combridge University Press, 1998. [21] N. A. Syed, H. Liu, and K. K. Sung. Incremental learning with support vector machines. In Proceedings of the Workshop on Support Vector Machines at the International Joint Conference on Artificial Intelligence (IJCAI-99), 1999. [22] C. van Rijsbergen. Information Retrieval. Butterworths, London, 1979. [23] V. N. Vapnik. The Nature of Statistical Learning Theory. Springer, New York, 1995. [24] C. K. Williams and M. Seeger. Using the Nystr om method to speed up kernel machines. In T. K. Leen, T. G. Dietterich, and V. Tresp, editors, NIPS 2000, volume 13. The MIT Press, 2001. [25] O. Winther. Bayesian Mean Field Algorithms for Neural Networks and Gaussian Processes. PhD thesis, University of Copenhagen, CONNECT, The Niels Bohr Institute, 1998. [26] Y. Yang. A study on thresholding strategies for text categorization. In Proceedings of the 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 137145, 2001. [27] Y. Yang and X. Liu. A re-examination of text categorization methods. In Proceedings of the 22nd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 4249, 1999. APPENDIX A. ON THE CHOICE OF PARAMETERS A.1 Likelihood model MacKay[12] has suggested the evidence frameworkfor model selection. Here, we calculate the evidence on the training Table 6: Micro-/Macro-avg F 1 (MaxF1 thresholds) and Avg log-evidence on Reuters-21578 for different likelihood models, using Bayesian online perceptron. Micro-/Macro-avg F 1 Avg log-evidence Logit 86.48 / 52.75 -45.02 Probit 86.69 / 52.16 -34.32 Flip 85.94 / 53.00 -368.8 Table 7: Micro-/Macro-avg F 1 (MaxF1 thresholds) and Avg log-evidence on Reuters-21578 for different passes over the training data, using Bayesian online perceptron. Passes Micro-/Macro-avg F 1 Avg log-evidence 1 87.08 / 52.87 -35.56 2 86.92 / 52.63 -34.36 3 86.69 / 52.16 -34.32 4 86.62 / 52.75 -34.54 5 85.22 / 46.93 -34.69 data using the final posterior for a: p(D m ) = m t =1 p(y t |x t , a) m . Table 6 illustrates this for selecting the likelihood measure for the text classification task, using the Bayesian online perceptron. In the table, the probit model follows the formulation in section 3.1 with 0 = 0.5, logit model is esti-mated by the probit model with 0 = 1.6474[2], and the flip noise model is as described in section 4.2. Although their F 1 averages are similar, the evidences show that the probit model with 0 = 0.5 is a more likely model. The small evidence for the flip noise model is because much information is lost through the threshold function . A.2 Effects of multiple passes over data Using the evidence measure defined in section A.1, Table 7 illustrates the effects of different number of passes over training data for Bayesian online perceptron. Treating the number of passes as a parameter for the algorithm, we see that having 3 passes over the data gives the highest average evidence, although there is no significant difference between 2, 3, or 4 passes. Similar results hold for the Gaussian process for the 3 different kernels. Hence, in section 4.1, we choose to use 3 passes for all the Bayesian algorithms. A.3 Jitter term The addition of the jitter term 10 -4 ij (where ij = 1 if i = j, and 0 otherwise) for Gaussian process for classification is recommended by Neal[14]. This term improves the conditioning of the matrix computations while having a small effect on the model. From our preliminary experiments , without the jitter term, the matrix operations in Bayesian online Gaussian process become ill-conditioned. A.4 Sizes of the basis vectors sets The sizes of the sets of basis vectors for GP in section 4.1 are limited to less than or equal to the number of features selected. This is because, as noted by Csat o & Opper[3], for a feature space of finite dimension M, no more than M basis vectors are needed, due to linear dependence. 104

Text Classification;perceptron;text classification;filtering;information gain;Bayesian online classifiers;Online;Machine Learning;continous learning;machine learning;Text Filtering;Gaussian process;Bayesian

43

Beyond PageRank: Machine Learning for Static Ranking

Since the publication of Brin and Page's paper on PageRank, many in the Web community have depended on PageRank for the static (query-independent) ordering of Web pages. We show that we can significantly outperform PageRank using features that are independent of the link structure of the Web. We gain a further boost in accuracy by using data on the frequency at which users visit Web pages. We use RankNet, a ranking machine learning algorithm, to combine these and other static features based on anchor text and domain characteristics. The resulting model achieves a static ranking pairwise accuracy of 67.3% (vs. 56.7% for PageRank or 50% for random).

INTRODUCTION Over the past decade, the Web has grown exponentially in size. Unfortunately, this growth has not been isolated to good-quality pages. The number of incorrect, spamming, and malicious (e.g., phishing) sites has also grown rapidly. The sheer number of both good and bad pages on the Web has led to an increasing reliance on search engines for the discovery of useful information. Users rely on search engines not only to return pages related to their search query, but also to separate the good from the bad, and order results so that the best pages are suggested first. To date, most work on Web page ranking has focused on improving the ordering of the results returned to the user (query-dependent ranking, or dynamic ranking). However, having a good query-independent ranking (static ranking) is also crucially important for a search engine. A good static ranking algorithm provides numerous benefits: Relevance : The static rank of a page provides a general indicator to the overall quality of the page. This is a useful input to the dynamic ranking algorithm. Efficiency : Typically, the search engine's index is ordered by static rank. By traversing the index from high-quality to low-quality pages, the dynamic ranker may abort the search when it determines that no later page will have as high of a dynamic rank as those already found. The more accurate the static rank, the better this early-stopping ability, and hence the quicker the search engine may respond to queries. Crawl Priority : The Web grows and changes as quickly as search engines can crawl it. Search engines need a way to prioritize their crawl--to determine which pages to re-crawl , how frequently, and how often to seek out new pages. Among other factors, the static rank of a page is used to determine this prioritization. A better static rank thus provides the engine with a higher quality, more up-to -date index. Google is often regarded as the first commercially successful search engine. Their ranking was originally based on the PageRank algorithm [5][27]. Due to this (and possibly due to Google's promotion of PageRank to the public), PageRank is widely regarded as the best method for the static ranking of Web pages. Though PageRank has historically been thought to perform quite well, there has yet been little academic evidence to support this claim. Even worse, there has recently been work showing that PageRank may not perform any better than other simple measures on certain tasks. Upstill et al. have found that for the task of finding home pages, the number of pages linking to a page and the type of URL were as, or more, effective than PageRank [32]. They found similar results for the task of finding high quality companies [31]. PageRank has also been used in systems for TREC's "very large collection" and "Web track" competitions, but with much less success than had been expected [17]. Finally, Amento et al. [1] found that simple features, such as the number of pages on a site, performed as well as PageRank. Despite these, the general belief remains among many, both academic and in the public, that PageRank is an essential factor for a good static rank. Failing this, it is still assumed that using the link structure is crucial, in the form of the number of inlinks or the amount of anchor text. In this paper, we show there are a number of simple url- or page- based features that significantly outperform PageRank (for the purposes of statically ranking Web pages) despite ignoring the Copyright is held by the International World Wide Web Conference Committee (IW3C2). Distribution of these papers is limited to classroom use, and personal use by others. WWW 2006, May 2326, 2006, Edinburgh, Scotland. ACM 1-59593-323-9/06/0005. 707 structure of the Web. We combine these and other static features using machine learning to achieve a ranking system that is significantly better than PageRank (in pairwise agreement with human labels). A machine learning approach for static ranking has other advantages besides the quality of the ranking. Because the measure consists of many features, it is harder for malicious users to manipulate it (i.e., to raise their page's static rank to an undeserved level through questionable techniques, also known as Web spamming). This is particularly true if the feature set is not known. In contrast, a single measure like PageRank can be easier to manipulate because spammers need only concentrate on one goal: how to cause more pages to point to their page. With an algorithm that learns, a feature that becomes unusable due to spammer manipulation will simply be reduced or removed from the final computation of rank. This flexibility allows a ranking system to rapidly react to new spamming techniques. A machine learning approach to static ranking is also able to take advantage of any advances in the machine learning field. For example, recent work on adversarial classification [12] suggests that it may be possible to explicitly model the Web page spammer's (the adversary) actions, adjusting the ranking model in advance of the spammer's attempts to circumvent it. Another example is the elimination of outliers in constructing the model, which helps reduce the effect that unique sites may have on the overall quality of the static rank. By moving static ranking to a machine learning framework, we not only gain in accuracy, but also gain in the ability to react to spammer's actions, to rapidly add new features to the ranking algorithm, and to leverage advances in the rapidly growing field of machine learning. Finally, we believe there will be significant advantages to using this technique for other domains, such as searching a local hard drive or a corporation's intranet. These are domains where the link structure is particularly weak (or non-existent), but there are other domain-specific features that could be just as powerful. For example, the author of an intranet page and his/her position in the organization (e.g., CEO, manager, or developer) could provide significant clues as to the importance of that page. A machine learning approach thus allows rapid development of a good static algorithm in new domains. This paper's contribution is a systematic study of static features, including PageRank, for the purposes of (statically) ranking Web pages. Previous studies on PageRank typically used subsets of the Web that are significantly smaller (e.g., the TREC VLC2 corpus, used by many, contains only 19 million pages). Also, the performance of PageRank and other static features has typically been evaluated in the context of a complete system for dynamic ranking, or for other tasks such as question answering. In contrast, we explore the use of PageRank and other features for the direct task of statically ranking Web pages. We first briefly describe the PageRank algorithm. In Section 3 we introduce RankNet, the machine learning technique used to combine static features into a final ranking. Section 4 describes the static features. The heart of the paper is in Section 5, which presents our experiments and results. We conclude with a discussion of related and future work. PAGERANK The basic idea behind PageRank is simple: a link from a Web page to another can be seen as an endorsement of that page. In general, links are made by people. As such, they are indicative of the quality of the pages to which they point when creating a page, an author presumably chooses to link to pages deemed to be of good quality. We can take advantage of this linkage information to order Web pages according to their perceived quality. Imagine a Web surfer who jumps from Web page to Web page, choosing with uniform probability which link to follow at each step. In order to reduce the effect of dead-ends or endless cycles the surfer will occasionally jump to a random page with some small probability , or when on a page with no out-links. If averaged over a sufficient number of steps, the probability the surfer is on page j at some point in time is given by the formula: + = j i i i P N j P B F ) ( ) 1 ( ) ( (1) Where F i is the set of pages that page i links to, and B j is the set of pages that link to page j. The PageRank score for node j is defined as this probability: PR(j)=P(j). Because equation (1) is recursive, it must be iteratively evaluated until P(j) converges (typically, the initial distribution for P(j) is uniform). The intuition is, because a random surfer would end up at the page more frequently, it is likely a better page. An alternative view for equation (1) is that each page is assigned a quality, P(j). A page "gives" an equal share of its quality to each page it points to. PageRank is computationally expensive. Our collection of 5 billion pages contains approximately 370 billion links. Computing PageRank requires iterating over these billions of links multiple times (until convergence). It requires large amounts of memory (or very smart caching schemes that slow the computation down even further), and if spread across multiple machines, requires significant communication between them. Though much work has been done on optimizing the PageRank computation (see e.g., [25] and [6]), it remains a relatively slow, computationally expensive property to compute. RANKNET Much work in machine learning has been done on the problems of classification and regression. Let X={x i } be a collection of feature vectors (typically, a feature is any real valued number), and Y ={y i } be a collection of associated classes, where y i is the class of the object described by feature vector x i . The classification problem is to learn a function f that maps y i =f(x i ), for all i. When y i is real-valued as well, this is called regression. Static ranking can be seen as a regression problem. If we let x i represent features of page i, and y i be a value (say, the rank) for each page, we could learn a regression function that mapped each page's features to their rank. However, this over-constrains the problem we wish to solve. All we really care about is the order of the pages, not the actual value assigned to them. Recent work on this ranking problem [7][13][18] directly attempts to optimize the ordering of the objects, rather than the value assigned to them. For these, let Z={&lt;i,j&gt;} be a collection of pairs of items, where item i should be assigned a higher value than item j. The goal of the ranking problem, then, is to learn a function f such that, ) ( ) ( , , j i f f j i x x Z &gt; 708 Note that, as with learning a regression function, the result of this process is a function (f) that maps feature vectors to real values. This function can still be applied anywhere that a regression-learned function could be applied. The only difference is the technique used to learn the function. By directly optimizing the ordering of objects, these methods are able to learn a function that does a better job of ranking than do regression techniques. We used RankNet [7], one of the aforementioned techniques for learning ranking functions, to learn our static rank function. RankNet is a straightforward modification to the standard neural network back-prop algorithm. As with back-prop, RankNet attempts to minimize the value of a cost function by adjusting each weight in the network according to the gradient of the cost function with respect to that weight. The difference is that, while a typical neural network cost function is based on the difference between the network output and the desired output, the RankNet cost function is based on the difference between a pair of network outputs. That is, for each pair of feature vectors &lt;i,j&gt; in the training set, RankNet computes the network outputs o i and o j . Since vector i is supposed to be ranked higher than vector j, the larger is o j -o i , the larger the cost. RankNet also allows the pairs in Z to be weighted with a confidence (posed as the probability that the pair satisfies the ordering induced by the ranking function). In this paper, we used a probability of one for all pairs. In the next section, we will discuss the features used in our feature vectors, x i . FEATURES To apply RankNet (or other machine learning techniques) to the ranking problem, we needed to extract a set of features from each page. We divided our feature set into four, mutually exclusive, categories: page-level (Page), domain-level (Domain), anchor text and inlinks (Anchor), and popularity (Popularity). We also optionally used the PageRank of a page as a feature. Below, we describe each of these feature categories in more detail. PageRank We computed PageRank on a Web graph of 5 billion crawled pages (and 20 billion known URLs linked to by these pages). This represents a significant portion of the Web, and is approximately the same number of pages as are used by Google, Yahoo, and MSN for their search engines. Because PageRank is a graph-based algorithm, it is important that it be run on as large a subset of the Web as possible. Most previous studies on PageRank used subsets of the Web that are significantly smaller (e.g. the TREC VLC2 corpus, used by many, contains only 19 million pages) We computed PageRank using the standard value of 0.85 for . Popularity Another feature we used is the actual popularity of a Web page, measured as the number of times that it has been visited by users over some period of time. We have access to such data from users who have installed the MSN toolbar and have opted to provide it to MSN. The data is aggregated into a count, for each Web page, of the number of users who viewed that page. Though popularity data is generally unavailable, there are two other sources for it. The first is from proxy logs. For example, a university that requires its students to use a proxy has a record of all the pages they have visited while on campus. Unfortunately, proxy data is quite biased and relatively small. Another source, internal to search engines, are records of which results their users clicked on. Such data was used by the search engine "Direct Hit", and has recently been explored for dynamic ranking purposes [20]. An advantage of the toolbar data over this is that it contains information about URL visits that are not just the result of a search. The raw popularity is processed into a number of features such as the number of times a page was viewed and the number of times any page in the domain was viewed. More details are provided in section 5.5. Anchor text and inlinks These features are based on the information associated with links to the page in question. It includes features such as the total amount of text in links pointing to the page ("anchor text"), the number of unique words in that text, etc. Page This category consists of features which may be determined by looking at the page (and its URL) alone. We used only eight, simple features such as the number of words in the body, the frequency of the most common term, etc. Domain This category contains features that are computed as averages across all pages in the domain. For example, the average number of outlinks on any page and the average PageRank. Many of these features have been used by others for ranking Web pages, particularly the anchor and page features. As mentioned, the evaluation is typically for dynamic ranking, and we wish to evaluate the use of them for static ranking. Also, to our knowledge, this is the first study on the use of actual page visitation popularity for static ranking. The closest similar work is on using click-through behavior (that is, which search engine results the users click on) to affect dynamic ranking (see e.g., [20]). Because we use a wide variety of features to come up with a static ranking, we refer to this as fRank (for feature-based ranking). fRank uses RankNet and the set of features described in this section to learn a ranking function for Web pages. Unless otherwise specified, fRank was trained with all of the features. EXPERIMENTS In this section, we will demonstrate that we can out perform PageRank by applying machine learning to a straightforward set of features. Before the results, we first discuss the data, the performance metric, and the training method. 5.1 Data In order to evaluate the quality of a static ranking, we needed a "gold standard" defining the correct ordering for a set of pages. For this, we employed a dataset which contains human judgments for 28000 queries. For each query, a number of results are manually assigned a rating, from 0 to 4, by human judges. The rating is meant to be a measure of how relevant the result is for the query, where 0 means "poor" and 4 means "excellent". There are approximately 500k judgments in all, or an average of 18 ratings per query. The queries are selected by randomly choosing queries from among those issued to the MSN search engine. The probability that a query is selected is proportional to its frequency among all 709 of the queries. As a result, common queries are more likely to be judged than uncommon queries. As an example of how diverse the queries are, the first four queries in the training set are "chef schools", "chicagoland speedway", "eagles fan club", and "Turkish culture". The documents selected for judging are those that we expected would, on average, be reasonably relevant (for example, the top ten documents returned by MSN's search engine). This provides significantly more information than randomly selecting documents on the Web, the vast majority of which would be irrelevant to a given query. Because of this process, the judged pages tend to be of higher quality than the average page on the Web, and tend to be pages that will be returned for common search queries. This bias is good when evaluating the quality of static ranking for the purposes of index ordering and returning relevant documents. This is because the most important portion of the index to be well-ordered and relevant is the portion that is frequently returned for search queries. Because of this bias, however, the results in this paper are not applicable to crawl prioritization. In order to obtain experimental results on crawl prioritization, we would need ratings on a random sample of Web pages. To convert the data from query-dependent to query-independent, we simply removed the query, taking the maximum over judgments for a URL that appears in more than one query. The reasoning behind this is that a page that is relevant for some query and irrelevant for another is probably a decent page and should have a high static rank. Because we evaluated the pages on queries that occur frequently, our data indicates the correct index ordering, and assigns high value to pages that are likely to be relevant to a common query. We randomly assigned queries to a training, validation, or test set, such that they contained 84%, 8%, and 8% of the queries, respectively. Each set contains all of the ratings for a given query, and no query appears in more than one set. The training set was used to train fRank. The validation set was used to select the model that had the highest performance. The test set was used for the final results. This data gives us a query-independent ordering of pages. The goal for a static ranking algorithm will be to reproduce this ordering as closely as possible. In the next section, we describe the measure we used to evaluate this. 5.2 Measure We chose to use pairwise accuracy to evaluate the quality of a static ranking. The pairwise accuracy is the fraction of time that the ranking algorithm and human judges agree on the ordering of a pair of Web pages. If S(x) is the static ranking assigned to page x, and H(x) is the human judgment of relevance for x, then consider the following sets: )} ( ) ( : , { y H x H y x &gt; = p H and )} ( ) ( : , { y S x S y x &gt; = p S The pairwise accuracy is the portion of H p that is also contained in S p : p p p H S H = accuracy pairwise This measure was chosen for two reasons. First, the discrete human judgments provide only a partial ordering over Web pages, making it difficult to apply a measure such as the Spearman rank order correlation coefficient (in the pairwise accuracy measure, a pair of documents with the same human judgment does not affect the score). Second, the pairwise accuracy has an intuitive meaning: it is the fraction of pairs of documents that, when the humans claim one is better than the other, the static rank algorithm orders them correctly. 5.3 Method We trained fRank (a RankNet based neural network) using the following parameters. We used a fully connected 2 layer network. The hidden layer had 10 hidden nodes. The input weights to this layer were all initialized to be zero. The output "layer" (just a single node) weights were initialized using a uniform random distribution in the range [-0.1, 0.1]. We used tanh as the transfer function from the inputs to the hidden layer, and a linear function from the hidden layer to the output. The cost function is the pairwise cross entropy cost function as discussed in section 3. The features in the training set were normalized to have zero mean and unit standard deviation. The same linear transformation was then applied to the features in the validation and test sets. For training, we presented the network with 5 million pairings of pages, where one page had a higher rating than the other. The pairings were chosen uniformly at random (with replacement) from all possible pairings. When forming the pairs, we ignored the magnitude of the difference between the ratings (the rating spread) for the two URLs. Hence, the weight for each pair was constant (one), and the probability of a pair being selected was independent of its rating spread. We trained the network for 30 epochs. On each epoch, the training pairs were randomly shuffled. The initial training rate was 0.001. At each epoch, we checked the error on the training set. If the error had increased, then we decreased the training rate, under the hypothesis that the network had probably overshot. The training rate at each epoch was thus set to: Training rate = 1 + Where is the initial rate (0.001), and is the number of times the training set error has increased. After each epoch, we measured the performance of the neural network on the validation set, using 1 million pairs (chosen randomly with replacement). The network with the highest pairwise accuracy on the validation set was selected, and then tested on the test set. We report the pairwise accuracy on the test set, calculated using all possible pairs. These parameters were determined and fixed before the static rank experiments in this paper. In particular, the choice of initial training rate, number of epochs, and training rate decay function were taken directly from Burges et al [7]. Though we had the option of preprocessing any of the features before they were input to the neural network, we refrained from doing so on most of them. The only exception was the popularity features. As with most Web phenomenon, we found that the distribution of site popularity is Zipfian. To reduce the dynamic range, and hopefully make the feature more useful, we presented the network with both the unpreprocessed, as well as the logarithm, of the popularity features (As with the others, the logarithmic feature values were also normalized to have zero mean and unit standard deviation). 710 Applying fRank to a document is computationally efficient, taking time that is only linear in the number of input features; it is thus within a constant factor of other simple machine learning methods such as nave Bayes. In our experiments, computing the fRank for all five billion Web pages was approximately 100 times faster than computing the PageRank for the same set. 5.4 Results As Table 1 shows, fRank significantly outperforms PageRank for the purposes of static ranking. With a pairwise accuracy of 67.4%, fRank more than doubles the accuracy of PageRank (relative to the baseline of 50%, which is the accuracy that would be achieved by a random ordering of Web pages). Note that one of fRank's input features is the PageRank of the page, so we would expect it to perform no worse than PageRank. The significant increase in accuracy implies that the other features (anchor, popularity, etc.) do in fact contain useful information regarding the overall quality of a page. Table 1: Basic Results Technique Accuracy (%) None (Baseline) 50.00 PageRank 56.70 fRank 67.43 There are a number of decisions that go into the computation of PageRank, such as how to deal with pages that have no outlinks, the choice of , numeric precision, convergence threshold, etc. We were able to obtain a computation of PageRank from a completely independent implementation (provided by Marc Najork) that varied somewhat in these parameters. It achieved a pairwise accuracy of 56.52%, nearly identical to that obtained by our implementation. We thus concluded that the quality of the PageRank is not sensitive to these minor variations in algorithm, nor was PageRank's low accuracy due to problems with our implementation of it. We also wanted to find how well each feature set performed. To answer this, for each feature set, we trained and tested fRank using only that set of features. The results are shown in Table 2. As can be seen, every single feature set individually outperformed PageRank on this test. Perhaps the most interesting result is that the Page-level features had the highest performance out of all the feature sets. This is surprising because these are features that do not depend on the overall graph structure of the Web, nor even on what pages point to a given page. This is contrary to the common belief that the Web graph structure is the key to finding a good static ranking of Web pages. Table 2: Results for individual feature sets. Feature Set Accuracy (%) PageRank 56.70 Popularity 60.82 Anchor 59.09 Page 63.93 Domain 59.03 All Features 67.43 Because we are using a two-layer neural network, the features in the learned network can interact with each other in interesting, nonlinear ways. This means that a particular feature that appears to have little value in isolation could actually be very important when used in combination with other features. To measure the final contribution of a feature set, in the context of all the other features, we performed an ablation study. That is, for each set of features, we trained a network to contain all of the features except that set. We then compared the performance of the resulting network to the performance of the network with all of the features. Table 3 shows the results of this experiment, where the "decrease in accuracy" is the difference in pairwise accuracy between the network trained with all of the features, and the network missing the given feature set. Table 3: Ablation study. Shown is the decrease in accuracy when we train a network that has all but the given set of features. The last line is shows the effect of removing the anchor, PageRank, and domain features, hence a model containing no network or link-based information whatsoever. Feature Set Decrease in Accuracy PageRank 0.18 Popularity 0.78 Anchor 0.47 Page 5.42 Domain Anchor, PageRank & Domain 0.10 0.60 The results of the ablation study are consistent with the individual feature set study. Both show that the most important feature set is the Page-level feature set, and the second most important is the popularity feature set. Finally, we wished to see how the performance of fRank improved as we added features; we wanted to find at what point adding more feature sets became relatively useless. Beginning with no features, we greedily added the feature set that improved performance the most. The results are shown in Table 4. For example, the fourth line of the table shows that fRank using the page, popularity, and anchor features outperformed any network that used the page, popularity, and some other feature set, and that the performance of this network was 67.25%. Table 4: fRank performance as feature sets are added. At each row, the feature set that gave the greatest increase in accuracy was added to the list of features (i.e., we conducted a greedy search over feature sets). Feature Set Accuracy (%) None 50.00 +Page 63.93 +Popularity 66.83 +Anchor 67.25 +PageRank 67.31 +Domain 67.43 711 Finally, we present a qualitative comparison of PageRank vs. fRank. In Table 5 are the top ten URLs returned for PageRank and for fRank. PageRank's results are heavily weighted towards technology sites. It contains two QuickTime URLs (Apple's video playback software), as well as Internet Explorer and FireFox URLs (both of which are Web browsers). fRank, on the other hand, contains more consumer-oriented sites such as American Express, Target, Dell, etc. PageRank's bias toward technology can be explained through two processes. First, there are many pages with "buttons" at the bottom suggesting that the site is optimized for Internet Explorer, or that the visitor needs QuickTime. These generally link back to, in these examples, the Internet Explorer and QuickTime download sites. Consequently, PageRank ranks those pages highly. Though these pages are important, they are not as important as it may seem by looking at the link structure alone. One fix for this is to add information about the link to the PageRank computation, such as the size of the text, whether it was at the bottom of the page, etc. The other bias comes from the fact that the population of Web site authors is different than the population of Web users. Web authors tend to be technologically-oriented, and thus their linking behavior reflects those interests. fRank, by knowing the actual visitation popularity of a site (the popularity feature set), is able to eliminate some of that bias. It has the ability to depend more on where actual Web users visit rather than where the Web site authors have linked. The results confirm that fRank outperforms PageRank in pairwise accuracy. The two most important feature sets are the page and popularity features. This is surprising, as the page features consisted only of a few (8) simple features. Further experiments found that, of the page features, those based on the text of the page (as opposed to the URL) performed the best. In the next section, we explore the popularity feature in more detail. 5.5 Popularity Data As mentioned in section 4, our popularity data came from MSN toolbar users. For privacy reasons, we had access only to an aggregate count of, for each URL, how many times it was visited by any toolbar user. This limited the possible features we could derive from this data. For possible extensions, see section 6.3, future work. For each URL in our train and test sets, we provided a feature to fRank which was how many times it had been visited by a toolbar user. However, this feature was quite noisy and sparse, particularly for URLs with query parameters (e.g., http://search .msn.com/results.aspx?q=machine+learning&form=QBHP). One solution was to provide an additional feature which was the number of times any URL at the given domain was visited by a toolbar user. Adding this feature dramatically improved the performance of fRank. We took this one step further and used the built-in hierarchical structure of URLs to construct many levels of backoff between the full URL and the domain. We did this by using the set of features shown in Table 6. Table 6: URL functions used to compute the Popularity feature set. Function Example Exact URL cnn.com/2005/tech/wikipedia.html?v=mobile No Params cnn.com/2005/tech/wikipedia.html Page wikipedia.html URL-1 cnn.com/2005/tech URL-2 cnn.com/2005 ... Domain cnn.com Domain+1 cnn.com/2005 ... Each URL was assigned one feature for each function shown in the table. The value of the feature was the count of the number of times a toolbar user visited a URL, where the function applied to that URL matches the function applied to the URL in question. For example, a user's visit to cnn.com/2005/sports.html would increment the Domain and Domain+1 features for the URL cnn.com/2005/tech/wikipedia.html. As seen in Table 7, adding the domain counts significantly improved the quality of the popularity feature, and adding the numerous backoff functions listed in Table 6 improved the accuracy even further. Table 7: Effect of adding backoff to the popularity feature set Features Accuracy (%) URL count 58.15 URL and Domain counts 59.31 All backoff functions (Table 6) 60.82 Table 5: Top ten URLs for PageRank vs. fRank PageRank fRank google.com google.com apple.com/quicktime/download yahoo.com amazon.com americanexpress.com yahoo.com hp.com microsoft.com/windows/ie target.com apple.com/quicktime bestbuy.com mapquest.com dell.com ebay.com autotrader.com mozilla.org/products/firefox dogpile.com ftc.gov bankofamerica.com 712 Backing off to subsets of the URL is one technique for dealing with the sparsity of data. It is also informative to see how the performance of fRank depends on the amount of popularity data that we have collected. In Figure 1 we show the performance of fRank trained with only the popularity feature set vs. the amount of data we have for the popularity feature set. Each day, we receive additional popularity data, and as can be seen in the plot, this increases the performance of fRank. The relation is logarithmic: doubling the amount of popularity data provides a constant improvement in pairwise accuracy. In summary, we have found that the popularity features provide a useful boost to the overall fRank accuracy. Gathering more popularity data, as well as employing simple backoff strategies, improve this boost even further. 5.6 Summary of Results The experiments provide a number of conclusions. First, fRank performs significantly better than PageRank, even without any information about the Web graph. Second, the page level and popularity features were the most significant contributors to pairwise accuracy. Third, by collecting more popularity data, we can continue to improve fRank's performance. The popularity data provides two benefits to fRank. First, we see that qualitatively, fRank's ordering of Web pages has a more favorable bias than PageRank's. fRank's ordering seems to correspond to what Web users, rather than Web page authors, prefer. Second, the popularity data is more timely than PageRank's link information. The toolbar provides information about which Web pages people find interesting right now, whereas links are added to pages more slowly, as authors find the time and interest. RELATED AND FUTURE WORK Since the original PageRank paper, there has been work on improving it. Much of that work centers on speeding up and parallelizing the computation [15][25]. One recognized problem with PageRank is that of topic drift: A page about "dogs" will have high PageRank if it is linked to by many pages that themselves have high rank, regardless of their topic. In contrast, a search engine user looking for good pages about dogs would likely prefer to find pages that are pointed to by many pages that are themselves about dogs. Hence, a link that is "on topic" should have higher weight than a link that is not. Richardson and Domingos's Query Dependent PageRank [29] and Haveliwala's Topic-Sensitive PageRank [16] are two approaches that tackle this problem. Other variations to PageRank include differently weighting links for inter- vs. intra-domain links, adding a backwards step to the random surfer to simulate the "back" button on most browsers [24] and modifying the jump probability ( ) [3]. See Langville and Meyer [23] for a good survey of these, and other modifications to PageRank. 6.2 Other related work PageRank is not the only link analysis algorithm used for ranking Web pages. The most well-known other is HITS [22], which is used by the Teoma search engine [30]. HITS produces a list of hubs and authorities, where hubs are pages that point to many authority pages, and authorities are pages that are pointed to by many hubs. Previous work has shown HITS to perform comparably to PageRank [1]. One field of interest is that of static index pruning (see e.g., Carmel et al. [8]). Static index pruning methods reduce the size of the search engine's index by removing documents that are unlikely to be returned by a search query. The pruning is typically done based on the frequency of query terms. Similarly, Pandey and Olston [28] suggest crawling pages frequently if they are likely to incorrectly appear (or not appear) as a result of a search. Similar methods could be incorporated into the static rank (e.g., how many frequent queries contain words found on this page). Others have investigated the effect that PageRank has on the Web at large [9]. They argue that pages with high PageRank are more likely to be found by Web users, thus more likely to be linked to, and thus more likely to maintain a higher PageRank than other pages. The same may occur for the popularity data. If we increase the ranking for popular pages, they are more likely to be clicked on, thus further increasing their popularity. Cho et al. [10] argue that a more appropriate measure of Web page quality would depend on not only the current link structure of the Web, but also on the change in that link structure. The same technique may be applicable to popularity data: the change in popularity of a page may be more informative than the absolute popularity. One interesting related work is that of Ivory and Hearst [19]. Their goal was to build a model of Web sites that are considered high quality from the perspective of "content, structure and navigation, visual design, functionality, interactivity, and overall experience". They used over 100 page level features, as well as features encompassing the performance and structure of the site. This let them qualitatively describe the qualities of a page that make it appear attractive (e.g., rare use of italics, at least 9 point font, ...), and (in later work) to build a system that assists novel Web page authors in creating quality pages by evaluating it according to these features. The primary differences between this work and ours are the goal (discovering what constitutes a good Web page vs. ordering Web pages for the purposes of Web search), the size of the study (they used a dataset of less than 6000 pages vs. our set of 468,000), and our comparison with PageRank. y = 0.577Ln(x) + 58.283 R 2 = 0.9822 58 58.5 59 59.5 60 60.5 61 1 10 100 Days of Toolbar Data P a i r w i s e A c c u r a c y Figure 1: Relation between the amount of popularity data and the performance of the popularity feature set. Note the x-axis is a logarithmic scale. 713 Nevertheless, their work provides insights to additional useful static features that we could incorporate into fRank in the future. Recent work on incorporating novel features into dynamic ranking includes that by Joachims et al. [21], who investigate the use of implicit feedback from users, in the form of which search engine results are clicked on. Craswell et al. [11] present a method for determining the best transformation to apply to query independent features (such as those used in this paper) for the purposes of improving dynamic ranking. Other work, such as Boyan et al. [4] and Bartell et al. [2] apply machine learning for the purposes of improving the overall relevance of a search engine (i.e., the dynamic ranking). They do not apply their techniques to the problem of static ranking. 6.3 Future work There are many ways in which we would like to extend this work. First, fRank uses only a small number of features. We believe we could achieve even more significant results with more features. In particular the existence, or lack thereof, of certain words could prove very significant (for instance, "under construction" probably signifies a low quality page). Other features could include the number of images on a page, size of those images, number of layout elements (tables, divs, and spans), use of style sheets, conforming to W3C standards (like XHTML 1.0 Strict), background color of a page, etc. Many pages are generated dynamically, the contents of which may depend on parameters in the URL, the time of day, the user visiting the site, or other variables. For such pages, it may be useful to apply the techniques found in [26] to form a static approximation for the purposes of extracting features. The resulting grammar describing the page could itself be a source of additional features describing the complexity of the page, such as how many non-terminal nodes it has, the depth of the grammar tree, etc. fRank allows one to specify a confidence in each pairing of documents. In the future, we will experiment with probabilities that depend on the difference in human judgments between the two items in the pair. For example, a pair of documents where one was rated 4 and the other 0 should have a higher confidence than a pair of documents rated 3 and 2. The experiments in this paper are biased toward pages that have higher than average quality. Also, fRank with all of the features can only be applied to pages that have already been crawled. Thus, fRank is primarily useful for index ordering and improving relevance, not for directing the crawl. We would like to investigate a machine learning approach for crawl prioritization as well. It may be that a combination of methods is best: for example, using PageRank to select the best 5 billion of the 20 billion pages on the Web, then using fRank to order the index and affect search relevancy. Another interesting direction for exploration is to incorporate fRank and page-level features directly into the PageRank computation itself. Work on biasing the PageRank jump vector [16], and transition matrix [29], have demonstrated the feasibility and advantages of such an approach. There is reason to believe that a direct application of [29], using the fRank of a page for its "relevance", could lead to an improved overall static rank. Finally, the popularity data can be used in other interesting ways. The general surfing and searching habits of Web users varies by time of day. Activity in the morning, daytime, and evening are often quite different (e.g., reading the news, solving problems, and accessing entertainment, respectively). We can gain insight into these differences by using the popularity data, divided into segments of the day. When a query is issued, we would then use the popularity data matching the time of query in order to do the ranking of Web pages. We also plan to explore popularity features that use more than just the counts of how often a page was visited. For example, how long users tended to dwell on a page, did they leave the page by clicking a link or by hitting the back button, etc. Fox et al. did a study that showed that features such as this can be valuable for the purposes of dynamic ranking [14]. Finally, the popularity data could be used as the label rather than as a feature. Using fRank in this way to predict the popularity of a page may useful for the tasks of relevance, efficiency, and crawl priority. There is also significantly more popularity data than human labeled data, potentially enabling more complex machine learning methods, and significantly more features. CONCLUSIONS A good static ranking is an important component for today's search engines and information retrieval systems. We have demonstrated that PageRank does not provide a very good static ranking; there are many simple features that individually out perform PageRank. By combining many static features, fRank achieves a ranking that has a significantly higher pairwise accuracy than PageRank alone. A qualitative evaluation of the top documents shows that fRank is less technology-biased than PageRank; by using popularity data, it is biased toward pages that Web users, rather than Web authors, visit. The machine learning component of fRank gives it the additional benefit of being more robust against spammers, and allows it to leverage further developments in the machine learning community in areas such as adversarial classification. We have only begun to explore the options, and believe that significant strides can be made in the area of static ranking by further experimentation with additional features, other machine learning techniques, and additional sources of data. ACKNOWLEDGMENTS Thank you to Marc Najork for providing us with additional PageRank computations and to Timo Burkard for assistance with the popularity data. Many thanks to Chris Burges for providing code and significant support in using training RankNets. Also, we thank Susan Dumais and Nick Craswell for their edits and suggestions. REFERENCES [1] B. Amento, L. Terveen, and W. Hill. Does "authority" mean quality? Predicting expert quality ratings of Web documents. In Proceedings of the 23 rd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 2000. [2] B. Bartell, G. Cottrell, and R. Belew. Automatic combination of multiple ranked retrieval systems. In Proceedings of the 17th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 1994. [3] P. Boldi, M. Santini, and S. Vigna. PageRank as a function of the damping factor. In Proceedings of the International World Wide Web Conference, May 2005. 714 [4] J. Boyan, D. Freitag, and T. Joachims. A machine learning architecture for optimizing web search engines. In AAAI Workshop on Internet Based Information Systems, August 1996. [5] S. Brin and L. Page. The anatomy of a large-scale hypertextual web search engine. In Proceedings of the Seventh International Wide Web Conference, Brisbane, Australia, 1998. Elsevier. [6] A. Broder, R. Lempel, F. Maghoul, and J. Pederson. Efficient PageRank approximation via graph aggregation. In Proceedings of the International World Wide Web Conference, May 2004. [7] C. Burges, T. Shaked, E. Renshaw, A. Lazier, M. Deeds, N. Hamilton, G. Hullender. Learning to rank using gradient descent. In Proceedings of the 22 nd International Conference on Machine Learning, Bonn, Germany, 2005. [8] D. Carmel, D. Cohen, R. Fagin, E. Farchi, M. Herscovici, Y. S. Maarek, and A. Soffer. Static index pruning for information retrieval systems. In Proceedings of the 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 43-50, New Orleans, Louisiana, USA, September 2001. [9] J. Cho and S. Roy. Impact of search engines on page popularity. In Proceedings of the International World Wide Web Conference, May 2004. [10] J. Cho, S. Roy, R. Adams. Page Quality: In search of an unbiased web ranking. In Proceedings of the ACM SIGMOD 2005 Conference. Baltimore, Maryland. June 2005. [11] N. Craswell, S. Robertson, H. Zaragoza, and M. Taylor. Relevance weighting for query independent evidence. In Proceedings of the 28 th Annual Conference on Research and Development in Information Retrieval (SIGIR), August, 2005. [12] N. Dalvi, P. Domingos, Mausam, S. Sanghai, D. Verma. Adversarial Classification. In Proceedings of the Tenth International Conference on Knowledge Discovery and Data Mining (pp. 99-108), Seattle, WA, 2004. [13] O. Dekel, C. Manning, and Y. Singer. Log-linear models for label-ranking. In Advances in Neural Information Processing Systems 16. Cambridge, MA: MIT Press, 2003. [14] S. Fox, K S. Fox, K. Karnawat, M. Mydland, S. T. Dumais and T. White (2005). Evaluating implicit measures to improve the search experiences. In the ACM Transactions on Information Systems, 23(2), pp. 147-168. April 2005. [15] T. Haveliwala. Efficient computation of PageRank. Stanford University Technical Report, 1999. [16] T. Haveliwala. Topic-sensitive PageRank. In Proceedings of the International World Wide Web Conference, May 2002. [17] D. Hawking and N. Craswell. Very large scale retrieval and Web search. In D. Harman and E. Voorhees (eds), The TREC Book. MIT Press. [18] R. Herbrich, T. Graepel, and K. Obermayer. Support vector learning for ordinal regression. In Proceedings of the Ninth International Conference on Artificial Neural Networks, pp. 97-102. 1999. [19] M. Ivory and M. Hearst. Statistical profiles of highly-rated Web sites. In Proceedings of the ACM SIGCHI Conference on Human Factors in Computing Systems, 2002. [20] T. Joachims. Optimizing search engines using clickthrough data. In Proceedings of the ACM Conference on Knowledge Discovery and Data Mining (KDD), 2002. [21] T. Joachims, L. Granka, B. Pang, H. Hembrooke, and G. Gay. Accurately Interpreting Clickthrough Data as Implicit Feedback. In Proceedings of the Conference on Research and Development in Information Retrieval (SIGIR), 2005. [22] J. Kleinberg. Authoritative sources in a hyperlinked environment. Journal of the ACM 46:5, pp. 604-32. 1999. [23] A. Langville and C. Meyer. Deeper inside PageRank. Internet Mathematics 1(3):335-380, 2004. [24] F. Matthieu and M. Bouklit. The effect of the back button in a random walk: application for PageRank. In Alternate track papers and posters of the Thirteenth International World Wide Web Conference, 2004. [25] F. McSherry. A uniform approach to accelerated PageRank computation. In Proceedings of the International World Wide Web Conference, May 2005. [26] Y. Minamide. Static approximation of dynamically generated Web pages. In Proceedings of the International World Wide Web Conference, May 2005. [27] L. Page, S. Brin, R. Motwani, and T. Winograd. The PageRank citation ranking: Bringing order to the web. Technical report, Stanford University, Stanford, CA, 1998. [28] S. Pandey and C. Olston. User-centric Web crawling. In Proceedings of the International World Wide Web Conference, May 2005. [29] M. Richardson and P. Domingos. The intelligent surfer: probabilistic combination of link and content information in PageRank. In Advances in Neural Information Processing Systems 14, pp. 1441-1448. Cambridge, MA: MIT Press, 2002. [30] C. Sherman. Teoma vs. Google, Round 2. Available from World Wide Web (http://dc.internet.com/news/article.php/ 1002061), 2002. [31] T. Upstill, N. Craswell, and D. Hawking. Predicting fame and fortune: PageRank or indegree?. In the Eighth Australasian Document Computing Symposium. 2003. [32] T. Upstill, N. Craswell, and D. Hawking. Query-independent evidence in home page finding. In ACM Transactions on Information Systems. 2003. 715

anchor text;relevance;Web pages;pairwise accuracy;fRank;popularity data;dynamic ranking;search engines;PageRank;static ranking;Static ranking;static features;RankNet

44

Black-Box Constructions for Secure Computation

It is well known that the secure computation of non-trivial functionalities in the setting of no honest majority requires computational assumptions. We study the way such computational assumptions are used. Specifically, we ask whether the secure protocol can use the underlying primitive (e.g., one-way trapdoor permutation) in a black-box way, or must it be nonblack-box (by referring to the code that computes this primitive)? Despite the fact that many general constructions of cryptographic schemes (e.g., CPA-secure encryption ) refer to the underlying primitive in a black-box way only, there are some constructions that are inherently nonblack-box. Indeed, all known constructions of protocols for general secure computation that are secure in the presence of a malicious adversary and without an honest majority use the underlying primitive in a nonblack-box way (requiring to prove in zero-knowledge statements that relate to the primitive). In this paper, we study whether such nonblack-box use is essential. We present protocols that use only black-box access to a family of (enhanced) trapdoor permutations or to a homomorphic public-key encryption scheme. The result is a protocol whose communication complexity is independent of the computational complexity of the underlying primitive (e.g., a trapdoor permutation) and whose computational complexity grows only linearly with that of the underlying primitive. This is the first protocol to exhibit these properties.

INTRODUCTION It is a known fact that most cryptographic tasks require the use of computational hardness assumptions. These assumptions typically come in two types: specific assumptions like the hardness of factoring, RSA, discrete log and others, and general assumptions like the existence of one-way functions , trapdoor permutations and others. In this paper, we refer to general assumptions and how they are used. Specifically , we consider an intriguing question regarding how secure protocols utilize a primitive that is assumed to carry some hardness property. Here again, there is a clear distinction between two types of uses: 1. Black-box usage: a protocol (or construction) uses a primitive in a black-box way if it refers only to the input/output behavior of the primitive. 1 For example, if the primitive is a trapdoor permutation, then the protocol may sample a permutation and its domain, and may compute the permutation and its inverse (if the trapdoor is given). Beyond this, no reference is made to the primitive. In particular, the code used to compute the permutation (or carry out any other task) is not referred to by the protocol. The vast majority of constructions in cryptography are black-box. 2. Nonblack-box usage: a protocol (or construction) uses a primitive in a nonblack-box way if it refers to the code for computing its functionality. A typical example of a nonblack-box construction is where a Karp reduction is applied to the circuit computing the function , say, in order to prove an N P zero-knowledge proof, as in [14]. A rich and fruitful body of work, initiated by [16], attempts to draw the borders between possibility and impossibility for black-box constructions in cryptography. While many of the relations between primitives are well understood, there are still some important tasks for which the only constructions that we have rely on nonblack-box access to the assumed primitive, yet the existence of a black-box construction is 1 It is typically also required that the security proof of the construction is black-box in the sense that an adversary breaking the protocol can be used as an oracle in order to break the underlying primitive. See, e.g., [11, 12, 29] for a comprehensive treatment of black-box reductions in cryptography. 99 not ruled out. In particular, all known general constructions of multiparty protocols that are secure in the presence of malicious adversaries and without an honest majority , originating from [15], use nonblack-box access to the assumed primitive. 2 (We note that by "general construc-tions" , we mean constructions that can be used to securely compute any functionality.) Another notable example of this phenomenon is the case of public-key encryption that is secure against chosen-ciphertext attacks [7, 30, 23]; here too, all known constructions are nonblack-box. The above phenomenon begs the following question: Is it possible to construct general protocols for secure computation without an honest majority and with malicious adversaries, given only black-box access to a "low-level" primitive? Answering the above question is of interest for the following reasons. First, it is of theoretical interest to understand whether or not nonblack-box access to a primitive is necessary for these tasks. An answer to this question would enhance our understanding of how hardness assumptions can (or must) be used. Second, as we have mentioned, the nonblack-box use of the underlying primitive is typically utilized in order to apply a Karp reduction for the purpose of using a (general) zero-knowledge proof. Such reductions are highly inefficient and are unlikely to be very useful in practice. Furthermore, in these protocols the communication complexity depends on the complexity of computing the primitive and the computational complexity grows more than linearly with that of the primitive. (An exception to this rule is the communication-efficient compiler presented in [26], which relies on the communication-efficient arguments of [20, 25]. However, the computational complexity of the protocol of [26] is even worse than the GMW protocol [15].) To illustrate the type of inefficiency resulting from current nonblack-box constructions, consider the following hypothetical scenario. Suppose that, due to major advances in cryptanalytic techniques, the security parameter must be large enough so that all basic cryptographic primitives require a full second of computation on a fast CPU. In such a case, would it still be possible to carry out a distributed task like oblivious transfer? Current nonblack-box techniques (e.g., the GMW protocol [15]) require parties to prove in zero-knowledge statements that involve the computation of the underlying primitive, say a trapdoor permutation . These zero-knowledge protocols, in turn, invoke cryptographic primitives for any gate of a circuit computing a trapdoor permutation. Since (by our assumption) a trapdoor permutation takes one second to compute, its circuit implementation contains trillions of gates, thereby requiring the protocol trillions of second to run. In contrast, a black-box construction of oblivious transfer from the trapdoor permutation primitive would make the number of invocations of the primitive independent of the complexity of 2 We stress that the above discussion is only true when considering general assumptions. Furthermore, it is only true when considering "low-level primitives" like trapdoor permutations. Specifically, there do exist constructions of secure multiparty protocols that use only black-box access to an oblivious transfer primitive [18]. However, since it is not known how to construct oblivious transfer using only black-box access to, say trapdoor permutations, the overall construction obtained does not use its "low-level" primitive in a black-box way. implementing the primitive, thus making oblivious transfer feasible even in the hypothetical scenario described above. We conclude that the current nonblack-box use of the underlying primitives constitutes an obstacle to efficiency. It is therefore of great interest to know whether or not it is possible to obtain solutions to these tasks that do not suffer from this obstacle. (We note that the inefficiency of nonblack-box constructions here is quite ironic because in many areas of cryptography, black-box constructions have been shown to have inherent computational limitations [21, 10].) Despite the above, we stress that the focus of this paper is not on efficiency, but rather on the theoretical question of whether or not it is possible to obtain the aforementioned black-box constructions. We believe this question to be interesting in its own right. Our results. We show how to construct general secure multiparty computation (for the case of no honest majority and malicious adversaries), given black-box access to either homomorphic encryption schemes or enhanced trapdoor permutations (see [13, Appendix C.1] for the definition of enhanced trapdoor permutations). We note that all known general constructions for this task from "low-level" primitives rely on either enhanced trapdoor permutations or homomorphic encryption schemes. However, they all use them in an inherently nonblack-box way. This is the case even for protocols that implement very simple functionalities, such as oblivious transfer. We prove the following: Theorem 1.1. There exist protocols for securely computing any multiparty functionality without an honest majority and in the presence of static malicious adversaries, that rely only on black-box access to a family of enhanced trapdoor permutations or to a homomorphic encryption scheme. We remark that nonblack-box access is not typically used when considering semi-honest adversaries [32, 15]. Rather, the nonblack-box access is utilized in known protocols in order to have the parties prove (in zero-knowledge) that they are correctly following the protocol specification. This is necessary for preventing a malicious adversary from effec-tively deviating from the protocol instructions. We note also that in the case of an honest majority, it is possible to securely compute any functionality information-theoretically, and without any hardness assumption [2, 5]. Thus, no primitive at all is needed. For this reason, we focus on the case of no honest majority (including the important two-party case) and malicious adversaries. Techniques. In order to prove Theorem 1.1, we begin by constructing oblivious transfer protocols that use only black-box access to enhanced trapdoor permutations or homomorphic encryption schemes, but provide rather weak security guarantees. We then "boost" the security of these protocols in order to obtain protocols that are secure in the presence of malicious adversaries. Constructions until today that have followed this paradigm work by first obtaining protocols that are secure in the presence of semi-honest adversaries, and then boosting them so that they are secure in the presence of malicious adversaries. However, it is not known how to carry out this "boosting" in a black-box way (and, indeed, it has been conjectured that malicious oblivious transfer cannot be constructed from semi-honest oblivious transfer in a black-box way [24]). Since we wish to make our construction black-box, we take a different route. 100 Protocol number Security for corrupted sender Security for corrupted receiver 3.1, 3.3 Private for defensible sender Private for defensible receiver 4.1 Private for defensible sender Secure for malicious receiver 5.1 Secure for malicious sender Private for defensible receiver In Theorem 6.1 Secure for malicious sender Secure for malicious receiver Table 1: The progression of our constructions: each protocol uses the previous one as a subprotocol. Specifically, we begin by introducing the notion of a defensible adversary. In order to describe this notion, we describe what a defense is: a defense is an input and random-tape that is provided by the adversary after the protocol execution concludes. A defense is good if the honest party upon that input and random-tape would have sent the same messages as the adversary sent. Such a defense is a supposed "proof" of honest behavior. However, the adversary need not actually behave honestly and can construct its defense retroactively (after the execution concludes). A protocol is said to be private in the presence of defensible adversaries if privacy is preserved in the event that an adversary provides a good defense. However, in the case that the adversary doesn't provide a good defense, nothing is guaranteed, and the entire honest party's input may be learned. This notion is therefore rather weak. We note that the oblivious transfer protocol of [8] is not secure under this notion. However, it can be efficiently modified into one that is secure under this notion. It is also possible to efficiently construct such an oblivious transfer protocol from homomorphic encryption. Importantly, we show that it is possible to construct oblivious transfer that is secure in the presence of malicious adversaries from oblivious transfer that is private in the presence of defensible adversaries. Furthermore, this construction is black-box. As we have mentioned, we start by constructing oblivious transfer protocols that are private in the presence of defensible adversaries. We present two such protocols: one that uses black-box access to a family of enhanced trapdoor permutations, and one that uses black-box access to a homomorphic public-key encryption scheme. Next, we construct from the above oblivious transfer protocol a new oblivious transfer protocol that is still private in the presence of defensible senders, but is secure in the presence of malicious receivers (where security is "full security" according to the ideal/real simulation paradigm). This is achieved using the so-called cut-and-choose technique. That is, many oblivious transfer executions (using random inputs) are run, and the receiver is asked to present a defense for its behavior in half of them. If it indeed presents a good defense, then we are guaranteed that it behaved somewhat honestly in most of the executions. We stress that this step is novel, because the requirements on a protocol that is secure according to the ideal/real simulation paradigm are much stricter than when only privacy is guaranteed. Indeed, some efficient protocols for oblivious transfer from the literature [27, 1, 17] are private for both (malicious) parties, but are not fully secure for either party. Nevertheless, we are able to boost both the resilience of the protocol (from a defensible to a malicious adversary) and its security guarantee (from privacy to full simulation-based security). Next, we "reverse" the oblivious transfer protocol (i.e., by switching the sender and receiver roles) in order to obtain a protocol with reversed security properties. Specifically , this next protocol is secure in the presence of malicious senders and private in the presence of defensible receivers. At this point, we reapply our security boosting technique in order to obtain a protocol that is "fully secure"; that is, a protocol that is secure in the presence of malicious senders and receivers. See Table 1 for the series of oblivious transfer protocols that we construct. Needless to say, each protocol uses its subprotocol in a black-box way. Finally, having constructed secure oblivious transfer protocols using only black-box access to primitives, it suffices to apply the well-known result of Kilian [18, 19] that shows that any functionality can be securely computed using black-box access to a secure oblivious transfer protocol. This therefore yields Theorem 1.1, as desired. Related work. Recently, in [6], it was shown that it is possible to construct constant-round protocols for the setting of an honest majority, that use only black-box access to the assumed primitive. As we have mentioned, in the setting of an honest majority, it is possible to construct information-theoretically secure protocols (which are, by triviality, black-box ). Nevertheless, there are no known (general) constant-round protocols for the information-theoretic setting, and so [6] relates to this issue. We remark that the techniques used in [6] and here are vastly different, due to the inherent differences between the setting of an honest majority and that of no honest majority. Organization. Due to lack of space in this abstract, we present only brief sketches of the definitions and proofs. Complete details appear in the full version of the paper. We often write OT as shorthand for oblivious transfer. DEFINITIONS We denote by P 1 (1 n , x 1 , 1 ) , P 2 (1 n , x 2 , 2 ) the transcript of an execution between parties P 1 and P 2 with a security parameter n, where P i has input x i and random-tape i . For brevity, we will sometimes omit the security parameter 1 n . The message sent by party P i (on the above inputs) after having received the series of incoming messages is denoted by P i ( x i , i ; ). Stated otherwise, P i ( x i , i ; ) denotes the next message function of P i . Let t = P 1 ( x 1 , 1 ) , P 2 ( x 2 , 2 ) . Then, denote the th message sent by P i in t by sent P i ( t) and the first messages received by P i in t by received P i 1,..., ( t). We also denote the output of P i in an execution by output P i P 1 ( x 1 , 1 ) , P 2 ( x 2 , 2 ) . In our presentation, we assume familiarity with the standard definitions of secure computation; see [13, Chapter 7] for a full treatment. In this work, we consider malicious adversaries (i.e., adversaries that may arbitrarily deviate from the protocol specification), and static corruptions (meaning that the set of corrupted parties is fixed before the protocol execution begins). We use a non-uniform formulation of adversaries here and therefore, without loss of generality, assume that they are 101 deterministic. However, this is not essential and all of our proofs hold for the uniform model of computation. Black-box access to primitives. In this paper, we consider constructions of protocols that use only black-box access to an underlying primitive. This can be easily formalized by defining oracles that provide the functionality of the primitive. For example, a trapdoor permutation can be defined by an oracle that samples a function description along with a trapdoor, an oracle that is given the function description and samples a random value from the domain, an oracle that is given the function description and a point in the domain and computes the permutation, and an oracle that is given the trapdoor and a point in the domain and computes the permutation inverse. It is easy to see that our protocols rely on the underlying primitive in a black-box way. We will therefore not burden the presentation by formally defining these oracles. We remark that we also construct protocols that use subprotocols in a black-box way. This can be formalized by just looking at the input/output behavior of the protocol. We will not formalize this. It suffices for our result to note that if the subprotocol uses the underlying primitive in a black-box way, then the protocol (that uses the subprotocol) also uses the underlying primitive in a black-box way. Again, this is easy to verify for all of our protocols. In addition to using the underlying primitive in a black-box way, our proofs of security are also black-box. Therefore, our reductions are what are typically called "fully black-box" [29]. 2.2 Defensible Adversarial Behavior We introduce the notion of defensible adversarial behavior . Loosely speaking, an adversary that exhibits defensible behavior may arbitrarily deviate from the protocol specification . However, at the conclusion of the protocol execution, the adversary must be able to justify or defend its behavior by presenting an input and a random-tape such that the honest party (with this input and random-tape) would behave in the same way as the adversary did. A protocol is "private" under defensible adversarial behavior if it is "private" in the presence of such adversaries. We stress that if an adversary behaves maliciously and cannot provide a good defense, then no security guarantees are given. We now define the notion of a good defense. Intuitively, a defense is an "explanation" of an adversary's behavior during the protocol execution. Such an explanation consists of an input and random-tape, and the defense is "good" if an honest party, given that input and random-tape, would have sent the same messages as the adversary did during the protocol execution. The formal definition follows. Definition 2.1. (good defense for t): Let t be the transcript of an execution of a protocol = (P 1 , P 2 ) between an adversary A (say, controlling P 1 ) and the honest party (say P 2 ). Then, we say that the pair ( x 1 , 1 ) constitutes a good defense by A for t in , denoted (x 1 , 1 ) = defense A ( t), if for every it holds that sent A ( t) = P 1 ( x 1 , 1 ; received A 1,..., -1 ( t)). In other words, every message sent by A in the execution is such that the honest party P 1 with input ( x 1 , 1 ) would have sent the same message. 2.3 Security of OT Protocols The starting point of our constructions is an oblivious transfer protocol [28, 8] that is private in the presence of a defensible receiver or sender. Recall that an oblivious transfer protocol involves a sender S with two input strings s 0 and s 1 , and a receiver R with an input bit r {0, 1}. Very informally, an oblivious transfer protocol has the property that the sender learns nothing about the receiver's bit r and the receiver obtains s r , but learns nothing about s 1-r . (The variant of oblivious-transfer that we use here is usually referred to as "1-out-of-2 OT".) We begin by presenting the formal definition of oblivious transfer that is private in the presence of a defensible receiver and then proceed to define privacy in the presence of a defensible sender. Non-trivial protocols. One technicality that must be dealt with is that a protocol that does nothing is trivially "private" in that it does not reveal anything about the par-ties' inputs. Of course, such a protocol is also useless. In order to make sure that the oblivious transfer protocols that we construct are "useful", we define the notion of a non-trivial oblivious transfer protocol. Such a protocol has the property that if both the sender and receiver are honest, then the receiver will receive its output as designated by the oblivious transfer functionality f((s 0 , s 1 ) , r) = (, s r ) (where denotes the empty output). Privacy for random inputs in the presence of a defensible receiver. We now define privacy for defensible receivers. Recall that the receiver in an oblivious transfer protocol is supposed to obtain one of the pair ( s 0 , s 1 ) in the execution. However, the other value must remain secret. When considering defensible adversaries, the requirement is that, as long as the adversary can provide a good defense, it can only learn one of the values. Recall that, by Definition 2.1, a party's defense includes its input (in this case, the bit r of the receiver, meaning that it wishes to obtain the value s r ). We therefore require that a defensible receiver can learn nothing about s 1-r when its defense contains the input value r. Due to technical reasons in our proofs later on, we define privacy only for the case that the sender's inputs are uniformly distributed bits. Fortunately, this will suffice for our constructions. We define an experiment for a protocol and an adversary A modelled by a polynomial-size family of circuits {A n } nN . Informally, the experiment begins by choosing a random pair of bits ( s 0 , s 1 ) to be used for the sender's input. The adversary's aim is to guess the value of the input that it doesn't receive as output. Experiment Expt rec ( A n ): 1. Choose s 0 , s 1 R {0, 1} uniformly at random. 2. Let S be a uniformly distributed random tape for S and let t = S(1 n , s 0 , s 1 , S ) , A n . 3. Let (( r, r ) , ()) be the output of A n ( t). (The pair ( r, r ) constitute A n 's defense and is its guess for s 1-r .) 4. Output 1 if and only if ( r, r ) is a good defense by A n for t in , and = s 1-r . Notice that by A's defense, it should have received s r . The challenge of the adversary is therefore to guess the value of s 1-r ; if it cannot do this, then the sender's privacy is preserved. 102 Definition 2.2. (privacy for random inputs in the presence of a defensible receiver): Let = (S, R) be a non-trivial oblivious transfer protocol. We say that is private for random inputs in the presence of a defensible receiver if for every polynomial-size family of circuits A = {A n } nN controlling R, for every polynomial p() and for all sufficiently large n's Pr [Expt rec ( A n ) = 1] &lt; 12 + 1 p(n) . Remark. The definition of Expt rec only considers the case that the inputs of the sender are uniformly distributed. We stress that this is a very weak definition. However, the reasons that we make this restriction are because (a) it suffices for our construction of "fully secure" oblivious transfer (see Protocol 4 .1), and more importantly, (b) without this restriction we were unable to prove the privacy of Protocol 3 .3 for defensible receivers (see Section 3.2). We stress that this restriction is not made when considering security in the presence of malicious parties. Privacy in the presence of a defensible sender. In an oblivious transfer protocol, the sender is not supposed to learn anything about the receiver's input. When considering a defensible sender, this means that the sender should not be able to simultaneously present a good defense of its behavior and make a correct guess as to the value of the receiver's input. We stress that this privacy requirement only needs to hold when the sender outputs a good defense; in all other cases, there may be no privacy whatsoever. The exact definition is formulated in a similar way as above. Security. The definitions above refer only to "privacy", meaning that the adversary can learn nothing more about the honest party's input than what is revealed by the output. However, these definitions say nothing about the simulata-bility of the protocols in question. In particular, a protocol that is private by one of the above definitions may not be secure according to the real/ideal simulation paradigm (see [13, Chapter 7] for these definitions). When we mention security in this paper, we refer to security according to the ideal/real model paradigm. PRIVACY FOR DEFENSIBLE SENDERS AND DEFENSIBLE RECEIVERS In this section we show how to construct oblivious transfer protocols that are private for defensible senders and receivers . We present two protocols: one based on homomorphic encryption and one based on enhanced trapdoor permutations . Importantly, both protocols access the underlying primitive in a black-box way only. 3.1 Bit OT from Homomorphic Encryption We assume the existence of a public-key encryption scheme ( G, E, D) that is indistinguishable under chosen-plaintext attacks and has the following homomorphic property: 1. The plaintext is taken from a finite Abelian group determined by the public key. For notational convenience , we assume here that the group is an "additive" group Z q ; however, the same construction works for "multiplicative" groups as well. 2. Given any public-key pk generated by the key generation algorithm G and any two ciphertexts c 1 = E pk ( m 1 ) and c 2 = E pk ( m 2 ), it is possible to efficiently compute a random encryption of the sum E pk ( m 1 + m 2 ). Consequently, it is also possible to efficiently compute E pk ( m 1 ) for any known integer . We also assume that ( G, E, D) has no decryption errors. Such encryption schemes can be constructed under the quadratic -residuosity, decisional Diffie-Hellman and other assumptions ; see [1, 17] for some references. The following protocol is implicit in [22]. Protocol 3.1. Inputs: The sender S has a pair of bits (s 0 , s 1 ); the receiver R has a bit r. The protocol: 1. The receiver R chooses a pair of keys (pk, sk) G(1 n ), computes c = E pk ( r) and sends c and p k to S. 2. The sender S uses the homomorphic property and its knowledge of s 0 and s 1 to compute a random encryption c = E pk ((1 - r)s 0 + rs 1 ). 3. R computes and outputs s r = D sk ( c ). Before proving security, note that if S and R are both honest, then R receives the correct output. For example, if r = 0, then c = E pk (1 s 0 + 0 s 1 ) = E pk ( s 0 ) and so R receives the correct value after decryption. Claim 3.2. Assume that the encryption scheme (G, E, D) is indistinguishable under chosen-plaintext attacks and has no decryption errors. Then, Protocol 3 .1 is a non-trivial oblivious transfer protocol that is private in the presence of defensible senders and private for random inputs in the presence of defensible receivers. Privacy in the presence of a defensible (or even malicious) sender follows from the fact that the sender's view consists only of a single encryption under E, and this encryption is secure. Privacy with respect to a defensible receiver follows since the existence of a proper defense implies that c is indeed an encryption of 0 or 1. This, in turn, guarantees that c is a random encryption of s r . Hence, again, privacy follows from the security of E. 3.2 Bit OT from Enhanced Trapdoor Permutations The following protocol is a modified version of [8] that is private in the presence of defensible adversaries. We stress that the original protocol of [8] is completely insecure in the presence of defensible adversaries. The construction uses any family of enhanced trapdoor permutations. Informally speaking, a family of trapdoor permutations is comprised of a function-sampling algorithm I, a domain-sampling algorithm D f , an algorithm F for computing the permutation and an algorithm F -1 for inverting the permutation (given the trapdoor). Such a family is called enhanced if it is hard to invert a random value y even when given the coins used by the domain-sampling algorithm to sample y. See [13, Appendix C.1 and Section 7.3] for a full definition. In the sequel, we will abuse notation and refer to the random coins used by D f as its input. We note that the enhanced property 103 is used in all constructions of oblivious transfer from trapdoor permutations. Indeed it has been shown that black-box constructions of oblivious transfer from plain trapdoor permutations is impossible [9]. We will require that I is errorless, meaning that for every series of random coins provided to I, the description of the function output is indeed a permutation. We call this errorless function sampling, or just errorless sampling. The protocol uses a perfectly binding commitment scheme C. We denote a commitment to a using randomness by C(a; ). For simplicity, we assume that in order to commit to a string a of length n, it suffices to use a random string that is also of length n. Such a commitment scheme can be obtained using black-box access to any trapdoor permutation or homomorphic encryption scheme. Protocol 3.3. Inputs: The sender S has a pair of random bits (s 0 , s 1 ); the receiver R has a bit r. Auxiliary information: The description of a family of (enhanced) trapdoor permutations ( I, D f , F, F -1 ) and a hard-core bit B for the family. The protocol: 1. The receiver R chooses 1 , R {0, 1} n and sends c = C( 1 ; ) to the sender S. 2. S chooses a trapdoor permutation pair (i, t) I(1 n ) and a random 2 R {0, 1} n , and sends i and 2 to R. 3. R computes y 1-r = D f ( 1 2 ); i.e., y 1-r is obtained by running the domain sampling algorithm with coins 1 2 . In addition, R chooses R {0, 1} n , obtains x r = D f ( ) and computes y r = f i ( x r ). Finally, R sends (y 0 , y 1 ) to S. 4. S uses t to compute 0 = B(f -1 i ( y 0 )) s 0 and 1 = B(f -1 i ( y 1 )) s 1 . S sends ( 0 , 1 ) to R. 5. R computes and outputs s r = B(x r ) r . Note that the only difference between Protocol 3.3 and the protocol of [8] is that in [8], the value y 1-r is chosen singlehandedly by the receiver, whereas here the value is chosen mutually using a (weak non-simulatable) coin-tossing protocol. (Indeed, in the protocol of [8] a cheating receiver can just choose a value y 1-r for which it knows the preimage. The receiver will then learn both s 0 and s 1 . Note also that a defensible receiver can also easily cheat in the protocol of [8] because it can send any value y 1-r and not the value that equals D f ( 1 2 ). In particular, it can send a value y 1-r for which it knows its preimage x 1-r under f i , and can still claim in its defense that its coins are such that y 1-r was sampled directly.) Claim 3.4. Assume that (I, D f , F, F -1 ) is a family of enhanced one-way trapdoor permutations and that the scheme C is perfectly binding and computationally hiding. Then, Protocol 3 .3 is a non-trivial oblivious transfer protocol that is private in the presence of defensible receivers and private for random inputs in the presence of defensible senders. Intuitively, a corrupted sender cannot guess the value of r from ( y 0 , y 1 ) because these values are identically distributed. This actually only holds as long as the function f i chosen by the sender is really a permutation from the family. (Otherwise , it may be possible to distinguish y r which is generated by computing f i ( x r ) from y 1-r which is randomly chosen from the domain.) The fact that the function is really a permutation is "proven" in the defense, and so if a good defense is provided, y r and y 1-r are identically distributed. We therefore have that the only way a defensible sender can learn the value of r is from the commitments. However, this involves distinguishing between c = C(D -1 f ( y 0 ) 2 ) and c = C(D -1 f ( y 1 ) 2 ) which is hard due to the hiding property of commitments. (Notice that y 1-r = D f ( 1 2 ) and so c = C( 1 ) = C(D -1 f ( y 1-r ) 2 ). Therefore, the problem of guessing r reduces to the problem of distinguishing such commitments.) As for privacy in the presence of a defensible receiver R : intuitively, if R behaves so that it can present a good defense, then it is unable to compute B(f -1 ( y 1-r )) because it has no freedom in choosing y 1-r . That is, R must choose y 1-r = 1 2 and so it cannot know the preimage f -1 ( y 1-r ). This implies that it can only learn the sender's bit s r . ACHIEVING SECURITY AGAINST A MALICIOUS RECEIVER In this section we construct a bit oblivious transfer protocol that is secure in the presence of a malicious receiver and private in the presence of a defensible sender. We stress that the security achieved for malicious receivers is according to the ideal/real model definition of security for secure computation. Our construction uses black-box access to an oblivious transfer protocol that is private for defensible receivers and senders (like those constructed in the previous section). Thus, in this section we show how to boost the security guarantee from privacy in the presence of a defensible receiver to security in the presence of a malicious receiver . The guarantee regarding a corrupted sender remains unchanged. Protocol 4.1. Inputs: The sender S has a pair of bits (s 0 , s 1 ); the receiver R has a bit r. The protocol: 1. The receiver R chooses 2n uniformly distributed bits r 1 , . . . , r 2n R {0, 1}. 2. The sender S chooses 2n pairs of random bits s 0 i , s 1 i R {0, 1} for i = 1, . . . , 2n. 3. S and R run 2n parallel executions of a bit oblivious transfer protocol that is private in the presence of defensible receivers and defensible senders. In the i th execution, S inputs (s 0 i , s 1 i ) and R inputs r i . Let t 1 , . . . , t 2n be the transcripts that result from these executions. 4. S and R run a secure two-party coin-tossing protocol (that accesses a one-way function in a black-box way) for generating a random string of length n: q = q 1 , . . . , q n . 3 The string q is used to define a set of indices Q {1, . . . , 2n} of size n in the following way: Q = {2i - q i } n i=1 . (Thus, for n = 3 and q = 010 we have that Q = {2, 3, 6}.) 3 Sequential executions of the coin-tossing protocol of [3] can be used. The security of this has been proven formally in [13]. 104 5. For every i Q, the receiver R provides a defense ( r i , i r ). 6. S checks that for every i Q, the pair (r i , i r ) constitutes a good defense by R for t i . If not, then S aborts and halts. Otherwise, it continues to the next step. 7. For every j / Q, the receiver R computes j = r r j (where r is R's initial input) and sends { j } j / Q to S. 8. S computes 0 = s 0 j / Q s j j and 1 = s 1 j / Q s 1-j j , and sends ( 0 , 1 ) to R. 9. R computes and outputs s r = r j / Q s r j j . We note that the sender's inputs to the executions of the oblivious transfer subprotocol in Protocol 4.1 are uniformly distributed. Therefore, it suffices to use Protocol 3.3, even though it has only been proven "private" for the case of uniformly distributed sender inputs. We stress that our proof below of Protocol 4.1 relies on the fact that the sender's inputs are single bits. 4 Claim 4.2. Assume that is a non-trivial oblivious transfer protocol that is private for random inputs in the presence of defensible senders and receivers. Then, Protocol 4 .1 is a non-trivial oblivious transfer protocol that is secure in the presence of malicious receivers and private in the presence of defensible senders. Proof Sketch: We first demonstrate the non-triviality property; that is, we show that if S and R are honest, then R receives s r , as required. To see this, first note that by the non-triviality of , the receiver R obtains all of the bits s r j j , and in particular all s r j j for j / Q. Now, if r = 0, then R sets j = r j for every j / Q. Therefore, R will compute s 0 = 0 j / Q s r j j = 0 j / Q s j j . This computation is correct because S computed 0 = s 0 j / Q s j j . In contrast, if r = 1, then j = 1 r j for every j, which is equivalent to r j = 1 j . Thus, once again, R's computation of j / Q s r j j when computing s 1 equals S's computation of j / Q s 1-j j when computing 1 , and R will obtain 1 . Privacy in the presence of defensible senders. We present only the idea behind the proof that Protocol 4.1 is private in the presence of a defensible sender A. Intuitively, if protocol is private in the presence of a defensible sender, then a defensible adversary here cannot learn any of the r i values in the execution (apart from those explicitly revealed by R when it provides its defenses). Therefore, the j = r j r values that it receives reveal nothing of the receiver's input r, because for all j / Q, the value r j is not learned. Security in the presence of malicious receivers. We present an almost full proof that Protocol 4.1 is secure in the presence of malicious receivers. The intuition behind 4 This is due to our definition of "oblivious transfer that is private for defensible adversaries". It is possible to define a stronger notion of defensible adversaries that is sufficient for proving that Protocol 4.1 is secure even when the sender's inputs are strings of an arbitrary length. However, we were not able to prove that Protocol 3.3 is private for defensible adversaries under this stronger notion (in contrast to Protocol 3.1 that can be proven secure under the stronger notion). this proof is that the cut-and-choose technique forces an adversarial receiver A to be able to provide a good defense for most of the oblivious transfer executions (or be caught with high probability). In particular, there must be at least one j / Q for which A could have provided a good defense. This implies that there exists some j for which A cannot predict the value of s 1-r j j with any non-negligible advantage. Since s 1-r is masked by s 1-r j j , it follows that A also learns nothing about s 1-r . We stress that the above intuition shows that a malicious A cannot learn anything about s 1-r . However , we actually need to prove a much stronger claim in that the protocol is secure for a malicious R , as defined via the ideal/real model simulation paradigm. We present our analysis in the so-called "hybrid model", where the honest parties use a trusted party to compute the coin-tossing functionality for them. We now describe the simulator Sim for A = {A n }: 1. For each i = 1, . . . , 2n, simulator Sim chooses random pairs s 0 i , s 1 i R {0, 1} and plays the honest sender in with these inputs, where A n plays the receiver. 2. Sim chooses a random string q R {0, 1} n and hands it to A n as if it is the output of the coin-tossing functionality , as sent by the trusted party. Let Q be the index set derived from q. Upon receiving back pairs ( r i , i r ) for i Q, simulator Sim checks that they all constitute good defenses, respectively. If not, then it aborts (just like the honest sender). 3. Sim rewinds A n to the beginning of the previous step and chooses a new random string q with associated index set Q . (We stress that q is independent of q.) Sim hands q to A n and sees if it replies with pairs ( r i , i r ) that are good defenses, for all i Q . Sim repeats this process with a new q until A n indeed replies with pairs ( r i , i r ) that are good defenses, for all i Q . If Q = Q, then Sim outputs fail. Otherwise it proceeds to the next step. 4. Given that Q = Q (and |Q | = |Q|), there exists at least one index j such that j / Q but j Q. For such a j, Sim computes r = r j j and sends r to the trusted party. (Note that r j is obtained from the defense ( r j , j r ) that was received from A n after it was sent the query set Q. In contrast, j is the value received from A n after rewinding; i.e., when the query set was Q .) 5. Upon receiving back a bit s r from the trusted party, Sim computes 0 and 1 as follows: (a) If r = 0, then 0 = s 0 j / Q s j j and 1 R {0, 1}. (b) If r = 1, then 0 R {0, 1} and 1 = s 1 j / Q s 1-j j . Sim sends ( 0 , 1 ) to A n and output whatever A n does. We proceed to prove that the joint output of Sim and the honest sender S in the ideal model is computationally indistinguishable from the joint output of A n and S in the real model. Actually, since the honest S has no output from the protocol, it suffices here to show that the output of Sim in the ideal model is computationally indistinguishable from the output of A n in the real model. We first claim that apart from the pair ( 0 , 1 ), the view of A n in the simulation with 105 Sim is statistically close to its view in a real execution with S; the only difference being in the case that Sim outputs fail. This can be seen as follows: if A n does not send good defenses after receiving q, then Sim aborts, just as the honest S would (and in this case the simulation is perfect). If A n does send good defenses, then Sim continues until it finds another (independent) q for which A n also replies with good defenses. It is not hard to see that this yields a distribution that is the same as in a real execution, except when q = q, in which case Sim outputs fail. However, this event (that it provides good defenses on q and then the next time that it provides good defenses is again on q) can happen with probability only 2 -n . We therefore have that in the simulation by Sim, the adversary A n 's partial view up until the point that it receives ( 0 , 1 ) is statistically close to its view in a real execution with S. We now show that A n 's full view is computationally indistinguishable. To do this, we consider a modified ideal-model simulator Sim who receives the sender S's input pair ( s 0 , s 1 ). Simulator Sim works in exactly the same way as Sim, except that it computes 1-r as an honest sender would instead of choosing it uniformly. By the above argument, it follows that the distribution generated by Sim in the ideal model is statistically close to the distribution generated by a real execution between S and A n . (Recall that Sim already generates r in the same way as an honest S, and therefore so does Sim .) It remains to show that the distribution generated by Sim is computationally indistinguishable to that generated by Sim. The only difference between Sim and Sim is in the generation of 1-r : simulator Sim generates it "honestly", whereas Sim chooses it uniformly. As mentioned above, intuitively, indistinguishability follows from the fact that at least one s 1-r j j masks the value of s 1-r . Formally, we show that if this "fake" 1-r can be distinguished from a real one, then we can construct a defensible receiver ~ A n that can break the oblivious transfer protocol . That is, we show that if the output generated by Sim and Sim can be distinguished with non-negligible probability, then it is possible for a defensible adversary ~ A n to succeed in the experiment of Definition 2.2 with non-negligible advantage , with respect to the subprotocol . Assume by contradiction that there exists a distinguisher D, a polynomial p() and infinitely many n's such that |Pr[D(output Sim ) = 1] - Pr[D(output Sim ) = 1] | 1 p(n) . Without loss of generality, assume that Pr[ D(output Sim ) = 1] - Pr[D(output Sim ) = 1] 1 p(n) . (1) We now use the above to construct a defensible adversary ~ A = { ~ A n }. Adversary ~ A n begins its attack by starting the simulation of Protocol 4.1, according to Sim's strategy. Specifically, ~ A n chooses s 0 , s 1 R {0, 1} and runs the simulation strategy of Sim with A n up until the point where 0 and 1 are sent. The simulation is the same as Sim, except for the following difference: ~ A n begins by choosing j R {1, . . . , 2n} and internally invokes A n , simulating an execution of Protocol 4.1. Then, all of the oblivious transfers subexecutions of , except for the j th one, are run internally with ~ A n playing the honest sender ( ~ A n also chooses the s 0 i and s 1 i values as S would); in contrast, the messages of the j th execution of the oblivious transfer protocol are forwarded between ~ A n 's external sender and the internal A n playing the receiver. Following the oblivious transfer executions , ~ A n runs the honest sender in the coin-tossing protocol to generate q and thus Q as required. If j / Q, then ~ A n outputs fail and halts. Otherwise, ~ A n receives back the defenses ; since j Q, the j th defense is included. If ( r j , j r ) is not a good defense, then ~ A n outputs fail and halts. Otherwise , it stores ( r j , j r ) and continues like Sim by rewinding A n and generating a new q and Q . If j Q , then once again ~ A n outputs fail and halts. Otherwise, it continues like Sim (using the j chosen above for which it is given that j Q and j / Q ). ~ A n continues in the same way that Sim does up until (but not including) the point at which ( 0 , 1 ) must be sent. Now, ~ A n computes ( 0 , 1 ) as follows. First, note that ~ A n knows the values ( s 0 , s 1 ) and s 0 i , s 1 i for all i = j (because it chose them). However, the values s 0 j and s 1 j are not known to ~ A n because these are the values used by the external sender with whom it interacts. Nevertheless, the (good) defense provided by A n is enough to obtain the value s r j j . This holds because given the transcript of the j th oblivious transfer execution and the input and random-tape of the receiver, it is possible to derive s r j j . The only value unknown to ~ A n is therefore s 1-r j j . Therefore, ~ A n is able to compute r like the honest sender. In contrast, it cannot honestly compute 1-r . Rather, ~ A n guesses the value of s 1-r j j R {0, 1} randomly, and then computes 1-r using s 1-r , all of the s i values that it knows (i.e., all apart from s 1-r j j ), and the uniformly chosen s 1-r j j . In order to determine its output, ~ A n obtains the output of A n and runs the distinguisher D (from Eq. (1)) on this output; let b be the bit output by D. Then, ~ A n sets = s 1-r j j b. (Recall that is ~ A n 's guess for the "not-received" bit used by the honest sender. The motivation for this guess is that by Eq. (1), D outputs 1 with higher probability on Sim (when the bit is random) than on Sim (when the bit is correct). Thus, when D outputs 1, we flip ~ A n 's guess for s 1-r j j .) Finally, ~ A n outputs the defense ( r j , j r ) from above and the bit . We proceed to analyze the probability that ~ A n succeeds in Expt rec . First, note that unless ~ A n outputs fail, the view of A n when interacting with ~ A n above is identical to its view in the simulation by Sim. This is due to the fact that ~ A n follows Sim's strategy, except for two differences. The first difference is that in the j th execution of the oblivious transfer protocol is run externally. However, since Sim plays the role of an honest receiver in all of the executions, this makes no difference to A n 's view. The second difference is in how 1-r is computed: Sim chooses it uniformly, whereas ~ A n computes it as described above. Clearly, the distribution generated is the same because ~ A n uses a uniformly distributed s 1-r j j , and thus 1-r is also uniformly distributed. Now, denote the inputs of the honest sender that ~ A n interacts with by (~ s 0 , ~s 1 ). Using the facts that (a) ~ A n generates the exact same distribution as Sim, (b) ~ A n sets = s 1-r j j b (where b is D's output bit), and (c) ~ A n presents a good defense every time that it does not output fail, we have that Pr Expt rec ( ~ A n ) = 1 | output ~ A n = fail (2) = Pr D(output Sim ) s 1-r j j = ~ s 1-r j . 106 (Recall that Expt rec ( ~ A n ) = 1 if ~ A n presents a good defense and = ~s 1-r j .) In contrast to the above, conditioned on the event that s 1-r j j = ~ s 1-r j (i.e., the event that ~ A n guessed correctly), the result is an execution that is distributed exactly according to Sim . (Recall that the only difference between Sim and Sim is with respect to the computation of 1-r .) That is, Pr D(output Sim ) s 1-r j j = ~ s 1-r j | s 1-r j j = ~ s 1-r j = Pr D(output Sim ) s 1-r j j = ~ s 1-r j | s 1-r j j = ~ s 1-r j = Pr [ D(output Sim ) = 0] where the last equality is just due to the fact that s 1-r j j = ~ s 1-r j . Now, recalling that s 1-r j j is chosen uniformly by ~ A n (and so equals ~ s 1-r j with probability exactly 1 /2), we have: Pr D(output Sim ) s 1-r j j = ~ s 1-r j = 1 2 Pr D(output Sim ) s 1-r j j = ~ s 1-r j | s 1-r j j = ~ s 1-r j + 1 2 Pr D(output Sim ) s 1-r j j = ~ s 1-r j | s 1-r j j = ~ s 1-r j = 1 2 Pr [D(output Sim ) = 0] + 1 2 Pr D(output Sim ) = 1 | s 1-r j j = ~ s 1-r j = 1 2 (1 - Pr [D(output Sim ) = 1]) + 1 2 Pr D(output Sim ) = 1 | s 1-r j j = ~ s 1-r j = 1 2 + 1 2 Pr D(output Sim ) = 1 | s 1-r j j = ~ s 1-r j - 12 Pr[D(output Sim ) = 1] . Recalling again that when s 1-r j j = ~ s 1-r j the output of Sim is the same as Sim , we have that 1 2 + 1 2 Pr D(output Sim ) = 1 | s 1-r j j = ~ s 1-r j - 12 Pr[D(output Sim ) = 1] = 1 2 + 1 2 Pr D(output Sim ) = 1 | s 1-r j j = ~ s 1-r j + 1 2 Pr D(output Sim ) = 1 | s 1-r j j = ~ s 1-r j - Pr [D(output Sim ) = 1] = 1 2 + Pr [D(output Sim ) = 1] - Pr [D(output Sim ) = 1] . Combining the above with Equations (1) and (2), we have that for infinitely many n's Pr Expt rec ( ~ A n ) = 1 | output ~ A n = fail = Pr D(output Sim ) s 1-r j j = ~ s 1-r j 12 + 1 p(n) . Recall now that ~ A n outputs fail if A n does not output a good defense, if j / Q, or if j Q . We first claim that A n must output a good defense with non-negligible probability. This follows simply from the fact that when A n does not output a good defense, the execution is truncated and the distributions generated by Sim and Sim are identical. Therefore, Eq. (1) implies that for infinitely many n's, A n outputs a good defense with probability at least 1 /p(n). Next, recall that ~ A n chooses the sets Q and Q randomly (under the constraints prescribed in the protocol). Thus, with probability exactly 1 /4, j Q and j / Q (because the probability that a given j is in a specified set is exactly 1/2). We conclude that with non-negligible probability, ~ A n does not output fail, and thus Pr[Expt rec ( ~ A n ) = 1] is non-negligible. It remains to show that Sim runs in expected polynomial-time . Aside from the rewinding stage, all work takes a fixed polynomial amount of time. Regarding the rewinding stage, we have the following. Let p denote the probability that A n replies correctly upon a random set of indices Q of size n, as specified in the protocol. Then, given that A n replied correctly to the initial query set Q, the expected number of rewinding attempts with independent Q made by Sim equals 1 /p. Since these rewinding attempts are only made if A n replied correctly to the initial query set Q, we have that the expected number of attempts overall equals p 1/p = 1. This completes the proof. MALICIOUS SENDERS AND DEFENSIBLE RECEIVERS In this section, we reverse the oblivious transfer protocol of Protocol 4.1 to obtain a protocol that is secure in the presence of a malicious sender and private for random inputs in the presence of a defensible receiver. We use the construction of [31] for reversing Protocol 4.1. The protocol is as follows: Protocol 5.1. (reversing oblivious transfer): Inputs: The sender S has a pair of bits (s 0 , s 1 ) for input and the receiver R has a bit r. The protocol: 1. The sender and receiver run an oblivious transfer protocol that is secure in the presence of a malicious receiver and private in the presence of a defensible sender: (a) The sender S, playing the receiver in , inputs ~ r = s 0 s 1 (b) The receiver R, playing the sender in , chooses a random bit R {0, 1} and inputs ~s 0 = and ~ s 1 = r. Denote S's output from by a. 2. S sends R the bit = s 0 a. 3. R outputs s r = . The security of Protocol 5.1 can be easily proven as an information-theoretic reduction, or when the original oblivious transfer protocol is fully secure. In contrast, it is far more subtle in the setting where only privacy in the presence of a defensible sender is assumed. Nevertheless, we do obtain the following claim: Claim 5.2. If is a non-trivial oblivious transfer protocol that is secure in the presence of a malicious receiver and private in the presence of a defensible sender, then Protocol 5.1 is a non-trivial oblivious transfer protocol that is secure in the presence of a malicious sender and private for random inputs in the presence of a defensible receiver. 107 FULLY-SECURE BIT OT In this section, we use the construction of Protocol 4.1 again in order to boost the security of Protocol 5.1 so that it is secure in the presence of both a malicious sender and a malicious receiver; we call such a protocol fully secure to stress that it is secure in the face of any corruption. By Claim 4.2, we have that Protocol 4.1 boosts the security of any oblivious transfer protocol that is private for defensible receivers into one that is secure in the presence of malicious receivers. We can therefore use Protocol 4.1 to boost the security of Protocol 5.1 so that the result is a protocol that is secure in the presence of malicious receivers. This does not suffice, however, because we must show that if the subprotocol used in Protocol 4.1 is secure in the presence of malicious senders, then the result is still secure in the presence of malicious senders. (Claim 4.1 considers only privacy for defensible senders.) This is actually easy to show, and is omitted here due to lack of space. Theorem 6.1. Assume that there exists a non-trivial bit oblivious transfer protocol that is secure in the presence of malicious senders and private for random inputs in the presence of defensible receivers. Then, Protocol 4.1 that is in-stantiated using this , is a non-trivial bit oblivious transfer protocol that is secure in the presence of malicious receivers and senders. Black-box construction of oblivious transfer. Noting that perfectly-binding commitment schemes (as used in Protocol 3.3) can be constructed using black-box access to homomorphic encryption or enhanced trapdoor permutations, and combining Protocols 3.1 and 3.3 with Protocol 4.1, followed by Protocol 5.1 and the construction in Theorem 6.1, we obtain secure bit oblivious transfer with black-box access to a homomorphic encryption scheme or a family of enhanced trapdoor permutations. BLACK-BOX SECURE COMPUTATION Kilian [18] showed that any function can be securely computed given black-box access to a bit oblivious transfer functionality . We therefore have the following theorem, that constitutes our main result: Theorem 7.1. Assume that there exist homomorphic encryption schemes with errorless decryption or families of enhanced trapdoor permutations. Then, for any probabilis-tic polynomial-time functionality f there exists a protocol that uses only black-box access to a homomorphic encryption scheme or to a family of enhanced trapdoor permutations , and securely computes f with any number of corrupted parties and in the presence of a static malicious adversary. We remark that as is standard for the setting of no honest majority, the security guarantee achieved here is that of "se-curity with abort"; see [13, Chapter 7] for formal definitions. REFERENCES [1] W. Aiello, Y. Ishai and O. Reingold. Priced Oblivious Transfer: How to Sell Digital Goods. In EUROCRYPT 2001, Springer-Verlag (LNCS 2045), pages 119135, 2001. [2] M. Ben-Or, S. Goldwasser and A. Wigderson. Completeness Theorems for Non-Cryptographic Fault-Tolerant Distributed Computation. In 20th STOC, pages 110, 1988. [3] M. Blum. Coin Flipping by Phone. In IEEE Spring COMPCOM, pages 133137, 1982. [4] R. Canetti. Security and Composition of Multiparty Cryptographic Protocols. Journal of Cryptology, 13(1):143202, 2000. [5] D. Chaum, C. Cr epeau and I. Damg ard. Multi-party Uncond-itionally Secure Protocols. In 20th STOC, pages 1119, 1988. [6] I. Damg ard and Y. Ishai. Constant-Round Multiparty Computation Using a Black-Box Pseudorandom Generator. In CRYPTO 2005, Springer-Verlag (LNCS 3621), pages 378394, 2005. [7] D. Dolev, C. Dwork and M. Naor. Non-Malleable Cryptography. SIAM Journal on Computing, 30(2):391437, 2000. [8] S. Even, O. Goldreich and A. Lempel. A Randomized Protocol for Signing Contracts. In Communications of the ACM, 28(6):637647, 1985. [9] R. Gennaro, Y. Lindell and T. Malkin. Enhanced versus Plain Trapdoor Permutations for Non-Interactive Zero-Knowledge and Oblivious Transfer. Manuscript in preparation, 2006. [10] R. Gennaro and L. Trevisan. Lower Bounds on the Efficiency of Generic Cryptographic Constructions. In 41st FOCS, pages 305314, 2000. [11] Y. Gertner, S. Kannan, T. Malkin, O. Reingold and M. Viswanathan. The Relationship between Public Key Encryption and Oblivious Transfer. In 41st FOCS, pages 325334, 2000. [12] Y. Gertner, T. Malkin and O. Reingold. On the Impossibility of Basing Trapdoor Functions on Trapdoor Predicates. In 42nd FOCS, pages 126135, 2001. [13] O. Goldreich. Foundations of Cryptography: Volume 2 Basic Applications. Cambridge University Press, 2004. [14] O. Goldreich, S. Micali and A. Wigderson. Proofs that Yield Nothing but their Validity or All Languages in NP Have Zero-Knowledge Proof Systems. Journal of the ACM, 38(1):691729, 1991. [15] O. Goldreich, S. Micali and A. Wigderson. How to Play any Mental Game A Completeness Theorem for Protocols with Honest Majority. In 19th STOC, pages 218229, 1987. [16] R. Impagliazzo and S. Rudich. Limits on the Provable Consequences of One-way Permutations. In CRYPTO'88, Springer-Verlag (LNCS 403), pages 826, 1988. [17] Y.T. Kalai. Smooth Projective Hashing and Two-Message Oblivious Transfer. In EUROCRYPT 2005, Springer-Verlag (LNCS 3494) pages 7895, 2005. [18] J. Kilian. Founding Cryptograph on Oblivious Transfer. In 20th STOC, pages 2031, 1988. [19] J. Kilian. Uses of Randomness In Algorithms and Protocols. MIT Press, 1990. [20] J. Kilian. Improved Efficient Arguments. In CRYPTO'95, Springer-Verlag (LNCS 963), pages 311324, 1995. [21] J.H. Kim, D.R. Simon and P. Tetali. Limits on the Efficiency of One-Way Permutation-Based Hash Functions. In 40th FOCS, pages 535542, 1999. [22] E. Kushilevitz and R. Ostrovsky. Replication Is Not Needed: Single Database, Computationally-Private Information Retrieval. In 38th FOCS, pages 364373, 1997. [23] Y. Lindell. A Simpler Construction of CCA2-Secure Public-Key Encryption Under General Assumptions. In EUROCRYPT 2003, Springer-Verlag (LNCS 2656), pages 241254, 2003. [24] T. Malkin and O. Reingold. Personal communication, 2006. [25] S. Micali. Computationally Sound Proofs. SIAM Journal on Computing, 30(4):12531298, 2000. [26] M. Naor and K. Nissim. Communication Preserving Protocols for Secure Function Evaluation. In 33rd STOC, pages 590599, 2001. [27] M. Naor and B. Pinkas. Efficient Oblivious Transfer Protocols. In 12th SODA, pages 458457, 2001. [28] M. Rabin. How to Exchange Secrets by Oblivious Transfer. Tech. Memo TR-81, Harvard University, 1981. [29] O. Reingold, L. Trevisan, and S. Vadhan. Notions of Reducibility between Cryptographic Primitives. In 1st TCC, pages 120, 2004. [30] A. Sahai. Non-Malleable Non-Interactive Zero-Knowledge and Adaptive Chosen-Ciphertext Security. In 40th FOCS, pages 543553, 1999. [31] S. Wolf and J. Wullschleger. Oblivious Transfer Is Symmetric. To appear in EUROCRYPT 2006. Appears at Cryptology ePrint Archive, Report 2004/336, 2004. [32] A. Yao. How to Generate and Exchange Secrets. In 27th FOCS, pages 162167, 1986. 108

oblivious transfer;encryption scheme;oblivious transfer protocol;secure computation;nonblack-box;malicious adversary;black-box;Theory of cryptography;cryptographic;black-box reductions;trapdoor permutation

45

Bluetooth Dynamic Scheduling and Interference Mitigation

Bluetooth is a cable replacement technology for Wireless Personal Area Networks. It is designed to support a wide variety of applications such as voice, streamed audio and video, web browsing, printing, and file sharing, each imposing a number of quality of service constraints including packet loss, latency, delay variation, and throughput. In addition to QOS support, another challenge for Bluetooth stems from having to share the 2.4 GHz ISM band with other wireless devices such as IEEE 802.11. The main goal of this paper is to investigate the use of a dynamic scheduling algorithm that guarantees QoS while reducing the impact of interference. We propose a mapping between some common QoS parameters such as latency and bit rate and the parameters used in the algorithm. We study the algorithm's performance and obtain simulation results for selected scenarios and configurations of interest.

Introduction Today most radio technologies considered by Wireless Personal Area Network (WPAN) industry consortia and standard groups including the Bluetooth Special Interest Group [1], HomeRF, and the IEEE 802.15, employ the 2.4 GHz ISM frequency band. This same frequency band is already in use by microwave ovens and the popular Wireless Local Area Network (WLAN) devices implementing the IEEE 802.11 standard specifications [8]. However, instead of competing with WLANs for spectrum and applications, WPANs are intented to augment many of the usage scenarios and operate in conjunction with WLANs, i.e., come together in the same laptop, or operate in proximity in an office or conference room environment. For example, Bluetooth can be used to connect a headset, or PDA to a desk-top computer, that, in turn, may be using WLAN to connect to an Access Point placed several meters away. Thus, an issue of growing concern is the coexistence of WLAN and WPAN in the same environment. Several techniques and algorithms aimed at reducing the impact of interference have been considered. These techniques range from collaborative schemes intended for Bluetooth and IEEE 802.11 protocols to be implemented in the same device to fully independent solutions that rely on interference detection and estimation. In particular: Collaborative mechanisms. Mechanisms for collaborative schemes have been proposed to the IEEE 802.15 Coexistence Task Group and are based on a Time Division Multiple Access (TDMA) solution that alternates the transmission of Bluetooth and WLAN packets (assuming both protocols are implemented in the same device and use a common transmitter) [9]. A priority of access is given to Bluetooth for transmitting voice packets, while WLAN is given priority for transmitting data. Non-collaborative mechanisms. The non-collaborative mechanisms range from adaptive frequency hopping [11] to packet scheduling and traffic control [4]. They all use similar techniques for detecting the presence of other devices in the band such as measuring the bit or frame error rate, the signal strength or the signal to interference ratio (often implemented as the Received Signal Indicator Strength (RSSI)). Frequency hopping devices may be able to detect that some frequencies are used by other devices and thus modify their frequency hopping pattern. They can also choose not to transmit on "bad" frequencies. The first technique is known as adaptive frequency hopping, while the second technique is known as MAC scheduling. The main advantage of scheduling is that it does not require changes to the Bluetooth specifications. In this paper we present a Bluetooth Interference Aware Scheduling (BIAS) algorithm to deal with coexistence. This algorithm takes advantage of the fact that devices in the same piconet will not be subject to the same levels of interference on all channels of the band. The basic idea is to utilize the Bluetooth frequency hopping pattern and distribute channels to devices such that to maximize their throughput while ensuring fairness of access among users. In this paper, we propose several extensions to a preliminary discussion of the algorithm [4] in order to address (1) priority scheduling, (2) dynamic changes in the environment, and (3) asymmetric scenarios where packet lengths and data rates are chosen differently in the upstream (slave to master transmission) and downstream (master to slave transmission) directions. In addition, we describe how to map commonly used QOS parameters, namely bit rate, and jitter and the parameters used in BIAS. Simulation results for scenarios and configurations of interest are presented and performance is measured in terms of packet loss and mean access delay. The remainder of this paper is organized as follows. In section 2, we give some general insights on the Bluetooth interference environment. In section 3, we describe BIAS and discuss the mapping of QOS parameters. In section 4, 22 N. GOLMIE we present simulation results and offer concluding remarks in section 5. Interference environment Since Bluetooth operates in the 2.4 GHz band along with other wireless technologies such as 802.11, high and low rate WPAN (802.15.3 and 4), the resulting mutual interference leads to significant performance degradation. In this paper, we assume that interference is caused by an 802.11 spread spectrum network operating in proximity of the Bluetooth piconet. This represents the worst case interference for Bluetooth. Golmie et al. [6] 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. The results indicate that Bluetooth performance may be severely impacted by interference with packet loss of 8% and 18% for voice and data traffic, respectively. In [6], the authors investigate the effect of several factors, such as transmitted power, offered load, packet size, hop rate, and error correction on performance. First, they note that power control may have limited benefits in an interference environment. Increasing the Bluetooth transmission power even ten times is not sufficient to reduce the Bluetooth packet loss. Second, using a shorter packet size leads to less packet loss for Bluetooth at the cost of causing more interference on WLAN. Overall, the results exhibit a strong dependence on the type and characteristics of the traffic distribution used. Additional analytical [5,10] and experimentation [3,7] results confirm these findings. Bluetooth Interference Aware Scheduling In this section, we present a Bluetooth Interference Aware Scheduling (BIAS) algorithm that consists of several components , namely, (i) dynamic channel estimation, (ii) credit computation, and (iii) access priority. A preliminary discussion of BIAS appeared in [4]. In this sequel, we assume that traffic from slave S i to the master (upstream) is characterized by a data rate, i up , equal to (N i peak l iup )/p i where N i peak is the number of packets sent back-to-back within a poll interval, p i , and l iup is the packet length (1, 3, or 5 slots depending on the packet type). Similarly , the data rate in the downstream (from the master to slave S i ) is characterized by i dn equal to (N i peak l i dn )/p i . Note that N i peak and p i are the same in the upstream and downstream, since every packet in the upstream corresponds to one in the downstream. In addition, we assume the following transmission rules for the master and slave. Master The master polls S i every p i slots in order to guarantee i up in the upstream direction. A poll message can be either a data or POLL packet. A data packet is sent if there is a packet in the queue destined for S i . This packet contains the ACK of the previous packet received from S i . In case there is no data to transmit and the master needs to ACK a previous slave transmission, it sends a NULL packet. Slave S i Upon receipt of a packet from the master, the slave can transmit a data packet. This data packet contains the ACK information of the master to slave packet transmission. In case the slave does not have any data to send, it sends a NULL packet in order to ACK the previous packet reception from the master. No ACK is required for a NULL message from the master. In a nutshell, we propose a method that allows the master device, which controls all data transmissions in the piconet, to avoid data transmission to a slave experiencing a "bad" frequency. Furthermore, since a slave transmission always follows a master transmission, using the same principle, the master avoids receiving data on a "bad" frequency, by avoiding a transmission on a frequency preceding a "bad" one in the hopping pattern. This simple scheduling scheme illustrated in figure 1 needs only be implemented in the master device and translates into the following transmission rule. The master transmits in a slot after it verifies that both the slave's receiving frequency, f s , Figure 1. Interference Aware Scheduling. BLUETOOTH DYNAMIC SCHEDULING AND INTERFERENCE MITIGATION 23 Figure 2. Master packet transmission flow diagram. and its own receiving frequency, f m , are "good". Otherwise, the master skips the current transmission slot and repeats the procedure over again in the next transmission opportunity. Figure 2 describes the master's transmission flow diagram. In addition, to checking the slave's and the master's receiving frequencies pair, (f s , f m ), the algorithm incorporates bandwidth requirements, and quality of service guarantees for each master/slave connection in the piconet. This bandwidth allocation is combined with the channel state information and mapped into transmission priorities given to each direction in the master/slave communication. It is shown in the "choose slave" routine in the flow diagram. Note that the master invokes the "choose" routine after serving the retransmission ACK queue for packets sent by the master requiring retransmission . In the remainder of this section, we discuss (a) a dynamic channel estimation procedure, (b) a credit allocation function, and (c) a service priority routine that schedules packet transmissions to devices according to their service requirements and the state of the channel. 3.1. Dynamic channel estimation Estimation is mainly based on measurements conducted on each frequency or channel in order to determine the presence of interference. Several methods are available ranging from BER, RSSI, packet loss rate, and negative ACKs. In this discussion, the estimation is based on negative ACKs, which belongs to the class of implicit methods that do not require messages to be exchanged between the master and the slave devices. First, we define two phases in the channel estimate procedure as illustrated in figure 3. During the Estimation Window, EW, packets are sent on all frequencies regardless of their classification. Note that in case no data traffic is available for transmission , POLL/NULL packets could be exchanged between the master and the slave in order to probe the channel and collect measurements. This POLL/NULL exchanged is designed in most implementations to keep the connection alive and check the status of the slave. It comes at the expense of causing more interference on other systems. EW takes place at the Figure 3. Implicit estimation. beginning of every Estimation Interval, EI, and is followed by an Online phase where the master uses only "good" frequencies to selectively send data and POLL packets to slaves in the piconet. Next, we give a lower bound on the EW and describe how to adjust EI based on the environment's dynamics . Estimation Window. The time to perform the channel estimation depends on the frequency hopping rate since the methods used to perform the classification depend on packet loss measurements per frequency visited. A lower bound calculation is as follows. First, we assume a hop rate of 1600 hops/s given single slot packets. For each receiver the hopping rate is 1600/2 hops/s, or 800 hops/s since nodes receive on every other "frequency" or "hop" in the sequence. Next, we consider the Bluetooth frequency hopping algorithm. In a window of 32 frequencies, every frequency is selected once, then the window is advanced by 16 frequencies, and the process is repeated. Therefore, it takes 5 windows of 32 frequencies in order to visit each of the 79 frequencies twice. In other words, 160 hops visit each frequency twice. The time to visit each frequency four times per receiver is 160/800 2 = 0.4 seconds or 400 ms. In fact, 400 ms constitutes a lower bound assuming full load and single-slot packets. In order to avoid having to fix the EW, or compute it manu-ally , we propose a simple technique to dynamically adjusts the window based on the number of times, N f , each frequency in the band should be visited. For example, if N f is equal to 2, then each receiving frequency in the band is visited at least twice, i.e., the estimation phase ends only when the last frequency in the band has been used twice for each device in the piconet. Note that, avoiding "bad" frequencies can start before EW ends, or as soon as frequency status information becomes available. Estimation Interval. How often to update the channel estimation depends on the application and the dynamics of the scenario used. We propose an adaptive procedure to adjust EI, which is the interval between two consecutive estimation windows. First, we let , be the percentage of frequencies that change classification status (from "good" to "bad" or vice versa) during the previous estimation phase. More formally, let S(f, t) be the status of frequency f at time t. S(f, t) = 1 if f is "good", 0 otherwise. (1) Using the exclusive bit "OR" operation between S(f, t) and S(f, t +1) represents the change of status of frequency f from 24 N. GOLMIE time t to t + 1. A change of status leads to a logic "1" while a no change yields a logic "0". Summing over all frequencies and dividing by the number of frequencies available, which is 79 in this case, is then equal to . t +1 = 1 79 79 f S(f, t) S(f, t + 1) . (2) Initially, EI is set to EI min . Then, EI is updated every interval , t, according to the rationale that if a change were to happen it is likely to happen again in the near future and therefore EI is set to EI min . Otherwise, the window is doubled. EI t +1 = max(2 EI t , EI max ) if t +1 0.1, EI min otherwise. (3) 3.2. Credit allocation The credit system controls the bandwidth allocated to each device in order to ensure that no device gets more than its fair share of the available bandwidth. Thus, devices with a positive credit counter, c i , are allowed to send data. Since the rate in the upstream can be different from the rate in the downstream, we define c i up and c i dn for both the upstream and downstream credits. Credits can be computed according to the upstream and downstream rates negotiated as follows: c i up = i up N, c i dn = i dn N, (4) where N is the number of slots considered in the allocation and i up/down = l i up/down N i peak /p i . Credits are decremented by the number of slots used in each data packet transmission. The transmission of POLL and NULL packets does not affect the credit count based on the rationale that credits are not required for the transmission of POLL and NULL messages . An interesting question is how to compute or derive it from application QOS parameters such as delay, peak bandwidth, and jitter. Let d (seconds), r (bits/s), (seconds) represent delay, peak bandwidth, and jitter, respectively. r is part of the L2CAP QOS parameters and for some applications is negotiated between the master and the slave at connection setup. r is equal to (N peak E l 8)/(p 625 10 -6 ) and = (r l 625 10 -6 )/(E l 8). Note that E l is the number of information bytes contained in a packet of length l. Table 1 gives E l corresponding to the various DH formats. The choice of l depends on the L2CAP packet size, k. When k E 5 , N peak = 1 and l is such that: l = 1 if 0 &lt; k 27, 3 if 27 &lt; k 183, 5 if 183 &lt; k 339. (5) Table 1 Packet encapsulation rate for DH packets. Packet type l E l (bytes) DH1 1 27 DH3 3 183 DH5 5 339 However, when k &gt; E 5 , higher layer packets (L2CAP) are segmented into N peak packets. The aim is to find N peak equal to N peak = k E l (6) such as to minimize N peak l, or the total number of slots needed. Furthermore, since master and slave transmission alternate , the end-to-end delay of a packet accounts for the segmentation and the transmission of packets in both directions. Therefore, the choice of l up and l dn are loosely constrained by the delay requirements as follows: N peak (l up + l dn ) d 625 10 -6 , (7) where 625 10 -6 is the length of a slot in seconds. Finally, the choice of p is determined by as follows: 2 p 625 10 -6 , (8) where 2 is the minimum value for the poll interval since every other slot is dedicated to a master (or slave) transmission. In case r, d, and cannot be determined from the application QOS, can be set to 1 i , the leftover bandwidth after having calculated for all other applications with known service rates ( ). 3.3. Service priority The third component of the algorithm is to give an access priority to devices based on their channel conditions and their allocated credits. We let u i be the probability that a pair of master/slave transmission slots are "good". Thus, u i represents the available spectrum to slave S i , and we write: u i = min 1 - 1 79 , P ( slave i has a good receiving frequency) P (master has a good receiving frequency) , (9) where P ( device i has a good receiving frequency) = Number of good channels i Total number of channels . (10) We use a two-tier system with high and low priorities, denoted by A, and B, respectively. Priority A is used to support delay constrained applications such as voice, MP3, and video. On the other hand, priority B, is used to support best effort connections such as ftp, http, print, email. The scheduling routine services priority A devices first, and priority B devices second. Also, among same tier connections, we choose to give devices with fewer number of good channels the right of way over other devices that have more channels available. The priority access is determined according to a weight factor , w, that is the product of the credits and the probability of BLUETOOTH DYNAMIC SCHEDULING AND INTERFERENCE MITIGATION 25 Table 2 Definition of parameters used in the scheduling algorithm. Parameters Definition i up,dn Rate allocated for device i in the upstream anddownstream w i up,dn Weight for device i c i up,dn Credit for device i N Number of slots considered in the allocation u i Available frequency usage for device i experiencing a bad frequency. w i up and w i dn are computed as follows: w i up = c i up (1 - u i ), w i dn = c i dn (1 - u i ). (11) The master schedules a data transmission for slave i such as to maximize the product of the weights in the up and downstreams : i = max f S w i up w i dn . (12) To transmit a POLL packet, the master looks only at the weight function in the upstream: i = max f S w i up . (13) The selection of a slave is restricted over the set of slaves S that can receive on the master's current transmission frequency , f . Thus, any slave that experiences a "bad" channel on the current transmission frequency is not considered. Four sets of slaves are formed, A f data , A f poll , B f data , and B f poll . A data and A poll represent the set of high priority connections requiring data and POLL packet transmissions, respectively. Similarly, B data and B poll represent low priority connections. First, the algorithm tries to schedule a packet to high priority slaves in group A, then a POLL packet, before it moves to group B. The credit counters and weights are updated accordingly after every master's transmission. Table 2 summarizes the parameters used in the algorithm and their definition. The algorithm's pseudocode is given in table 11. Performance evaluation In this section, we present simulation results to evaluate the performance of BIAS. The experiments illustrate the algorithm's responsiveness to changes in the environment and the support of QOS. The results obtained are compared with Round Robin (RR) scheduling. Our simulation environment is based on a detailed MAC, PHY and channel models for Bluetooth and IEEE 802.11 (WLAN) as described in [6]. The parameters used in the setup vary according to the experiment . The common simulation parameters are summarized in table 3. The simulations are run for 900 seconds of simulated time unless specified otherwise. We run 10 trials using a different random seed for each trial. In addition, to plotting the mean value, we verify that that the statistical variation around the mean values are very small (less than 1%). The performance metrics include the packet loss, the mean access delay, and the channel estimation transient time. The Table 3 Common simulation parameters. Bluetooth parameters Values ACL baseband packet encapsulation DH5 Transmitted power 1 mW WLAN parameters Values Packet interarrival time 2.172 ms Offered load 60% of channel capacity Transmitted power 25 mW Data rate 11 Mbit/s PLCP header 192 bits Packet header 224 bits Payload size 12000 bits packet loss is the percentage of packets dropped due to interference over the total number of packets received. The access delay measures the time it takes to transmit a packet from the time it is passed to the MAC layer until it is suc-cessfully received at the destination. The delay is measured at the L2CAP layer. The estimation transient time measures the time it takes a Bluetooth device to detect the presence of a "bad" frequency, i.e., from the time a packet loss occurs until the frequency is classified "bad". This average is provided on a per frequency basis. 4.1. Experiment 1: base case This experiment includes Bluetooth performance results for the reference scenario when no interference is present. It represents a base case since the effects of BIAS are quantified and compared against the reference scenario. It also covers different levels of interference caused by WLAN systems operating in close proximity. Thus, we examine Bluetooth's performance when 1, 2, and 3 WLAN interfering systems are operational and compare that to the ideal performance when no interference is present. Note that, the maximum number of non-overlapping channels for WLAN systems is 3, i.e., there could be up to 3 WLAN networks operating simultaneously using different non-overlapping channels. In each case, results are obtained with BIAS and RR scheduling. The benefits of using BIAS are discussed in terms of packet loss and access delay. Topology. We use the topology illustrated in figure 4 that consists of 3 WLAN systems (sourcesink pairs), and one Bluetooth piconet with one master and one slave device. In a first step, we record the results of Bluetooth when no WLAN system is present. Then, we add one WLAN system at a time starting with WLAN (Source/Sink) 1, followed by WLAN (Source/Sink) 2, and 3. Traffic. For Bluetooth, a generic source that generates DH5 packets is considered. The packet interarrival mean time in seconds, t B , is exponentially distributed and is computed according to t B = 2 l 0.000625 1 - 1 , (14) 26 N. GOLMIE where l is the packet length in slots and is the offered load. We assume that WLAN is operating in the Direct Sequence Spread Spectrum (DSSS) mode. The WLAN source is transmitting data packets to the sink which is responding with ACKs. The WLAN packet payload is set to 12000 bits transmitted at 11 Mbit/s, while the PLCP header of 192 bits is transmitted at 1 Mbit/s. The packet interarrival time in seconds , t W , is exponentially distributed and its mean is computed according to t W = 192 1000000 + 12224 11000000 1 . (15) Results. Figure 5 gives the packet loss (a) and the mean access delay (b) measured at the slave for a variable Bluetooth offered load (580%). Observe that when no WLAN system is present, the packet loss is zero and the access delay remains flat at around 4 ms. This represents a reference measure for the Bluetooth performance when there is no interference. Each WLAN system addition an increase of 15% in packet loss as shown in figure 5(a). The packet loss is around 15%, 30% and 45% when one, two, and three WLAN systems are present, respectively. Repeating the same experiments using BIAS, brings the packet loss down to zero for any number of WLAN systems. The delay trends captured in figure 5(b) Figure 4. Topology for experiments 1 and 2. are consistent with the packet loss results. Using BIAS yields lower delays than when RR is used. When one WLAN system is present, the delay curve with BIAS is flat at 5 ms (a 1 ms increase compared to the reference case when no interference is present). When 2 WLAN systems are present, the delay curve takes off at 35% with RR, while the curve remains flat until 60% with BIAS. When 3 WLAN systems are present, the delay curve takes off sharply at 15% with RR, while the knee of the curve remains lower with BIAS (shifted to the right). 4.2. Experiment 2: dynamic behavior In this experiment, we focus on BIAS's responsiveness to transient effects and sudden changes in the environment. We measure the channel estimation transient time per frequency and over the entire spectrum. We design an experiment where the WLAN traffic is turned on and off several times during each simulation run (about 30 times). Topology. We use the topology of figure 4 with one WLAN system (Source/Sink 1) and the Bluetooth master/slave pair. Traffic. The traffic is based on bulk data. The offered load for Bluetooth is varied between 10 and 100%, while for WLAN the offered load is set to 60%. For Bluetooth, both DH1 (1 slot) and DH5 (5 slots) packets are used in order to compare the difference in transient times. The time the WLAN connection is ON, T ON , is exponentially distributed with a mean equal to 10 seconds, while the time the WLAN connection is OFF, T OFF , is also exponentially distributed with mean equal to 20 seconds. Each simulation is run for 900 seconds. Unless specified otherwise, we set EI min = 2 seconds, EI max = 100 seconds, N f = 1. Results. Figures 6(a) and 6(b) give the packet loss and access delay, respectively, measured at the Bluetooth slave de (a) (b) Figure 5. Experiment 1. Variable number of WLAN interfering systems. (a) Probability of packet loss. (b) Mean access delay. BLUETOOTH DYNAMIC SCHEDULING AND INTERFERENCE MITIGATION 27 (a) (b) Figure 6. Experiment 2. Variable Bluetooth offered load. (a) Probability of packet loss. (b) Mean access delay. vice. The packet loss obtained with BIAS is negligible (less than 2%) for both DH1 and DH5 packets. On the other hand the packet loss with Round Robin (RR) is close to 10%. The access delay obtained with BIAS for DH1 packets is lower than the delay for DH5 packets for offered loads under 70% (it is around 1.5 ms for DH1 packets, and 4 ms for DH5 packets ). The knee of the curve for DH5 packets is located around 80% of the offered load while it is at 60% for DH1 packets . Observe that BIAS gives lower access delays than RR for DH5 packets (between 40% and 80% offered load). However, the same does not apply to DH1 packets, in which we observe a slight increase in access delay (0.5 ms) with BIAS compared to RR. For short packets (DH1) retransmissions due to packet loss (RR), and delay in transmission due to "bad" frequency avoidance (BIAS), yields comparable delays. Furthermore, given that the probability of packet loss (and retransmission) is small for short packets, RR gives lower access delays on average . Figure 7 gives the time it takes to estimate a "bad" frequency using DH1 and DH5 packets. The use of DH5 packets leads to a higher round trip transmission time, and therefore increases the transient time, up to 1.5 ms while it is around 0 s for DH1 packets. 4.3. Experiment 3: QOS support This experiment highlights the support of QOS in an environment where devices experience different levels of interference and connections have a range of service requirements. Topology. We use the topology illustrated in figure 8. Slaves 1 and 2 experience the same level of interference, while slave 3 does not experience any interference. The y-coordinate of the WLAN FTP server is varied along the y-axis in order to vary the level of interference on the Bluetooth piconet . Figure 7. Experiment 2. Variable Bluetooth offered load. Time to estimate a "bad" channel. Figure 8. Topology for experiment 3. 28 N. GOLMIE Traffic. For Bluetooth, we consider three application profiles , namely, Print, Video, and Email. We use print, video, and email traffic between slaves 1, 2, 3 and the master, respectively . Note that the master is the client process in all three connections. The profile parameters are given in table 4. The WLAN uses the FTP profile described in table 5. Since the video application generates roughly around 93 and 58 packets in the upstream and downstream directions, respectively, and since it is often difficult to predict the exact traffic distributions, the rate is divided evenly between both directions. Thus, we set 2 up 2 dn = 0.25. The two other appli-Table 4 Bluetooth application profile parameters. Parameters Distribution Value Email Send interarrival time (sec) Exponential 120 Send group Constant 3 Receive interarrival time (sec) Exponential 60 Receive group Constant 3 Email size (bytes) Exponential 1024 Print Print requests interarrival time (sec) Exponential 30 File size Normal (30 K, 9 M) Video Frame rate Constant 1 frame/s Frame size (bytes) Constant 17280 (128 120 pixels) Table 5 WLAN application profile parameters. Parameters Distribution Value FTP File interarrival time (sec) Exponential 5 File size (bytes) Exponential 5 M Percentage of get 100% cations, share the leftover bandwidth ( 1,3 up,dn = (1 - 0.5)/4 = 0.125). Results. Figure 9 depicts the results when the WLAN y-coordinate is varied between 0 and 10 meters. In figure 9(a), the packet loss with BIAS is below 0.1% for all three slaves and the master. With RR, slave 1 (Print) and slave 2 (Video) vary between 15% and 3% of packet loss between 0 and 10 meters, respectively. While the packet loss for the master is above 20%. Slave 3 (Email) has a low packet loss with both BIAS and RR since it is far from the WLAN server. The access delay for slave 2 (Video) in figure 9(b) is 0.3 seconds with BIAS, while it is almost double with RR (0.6 seconds). For Print, delays with BIAS are half the delays with RR (0.01 seconds as opposed to 0.02 seconds). The delays for Email are also reduced by half with BIAS. 4.4. Experiment 4: WLAN and multi-Bluetooth piconets interference When two or more Bluetooth piconets are proximally located, one expects few collisions when the packets happen to be transmitted on the same frequency. However, the probability of such collisions is low as discussed in [2] since each piconet has a unique frequency sequence. Given that these packet collisions are random in nature and are already miti-gated by frequency hopping, we do not expect significant performance improvements when BIAS is used since the packet loss is already very low. Furthermore, the fact that frequencies are eliminated due to other Bluetooth piconet interference may even cause delay increases. We illustrate this particular issue using the following scenario. Topology. We use the topology illustrated in figure 10 representing a conference hall environment. It consists of one WLAN AP located at (0, 15) meters, and one WLAN mobile at (0, 0) meters. The WLAN mobile is the server device, (a) (b) Figure 9. Experiment 3. Variable distance. (a) Probability of packet loss. (b) Access delay. BLUETOOTH DYNAMIC SCHEDULING AND INTERFERENCE MITIGATION 29 Figure 10. Topology for experiment 4. Table 6 Profile parameters. Parameters Distribution Value Bluetooth FTP Percentage of put/get 100% Inter-request time (sec) Exponential 5 File size (bytes) Exponential 250 K HTTP Page interarrival time (sec) Exponential 30 Number of objects per page Constant 2 Object 1 size (bytes) Constant 1 K Object 2 size (bytes) Uniform (2 K, 100 K) while the AP is the client. The distance between the WLAN AP and mobile is d W = 15 meters. There 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 meters. We define d B as the distance between a Bluetooth master and slave pair. d B = 1 meter for half of the master and slave pairs, while d B = 2 meters for the other half of the master and slave pairs. Traffic. We run four experiments with different combinations of WLAN and Bluetooth applications, namely, HTTP and FTP. We use the application profiles available in the OP-NET library and configure the parameters according to table 6. The WLAN FTP profile parameters are given in table 5. Results. The results for the Bluetooth packet loss and access delay are given in tables 7 and 8, respectively. The results are grouped by application category (FTP, HTTP), and d B , for each of the WLAN profiles. Overall, the packet loss results with BIAS are comparable to the packet loss obtained with RR. In some instances, the packet loss with BIAS is slightly lower than with RR, however the difference remains less than 2%. The access delays for Bluetooth is given in table 8. The results with BIAS and RR are comparable. However, there are no significant advantages in using BIAS. Tables 9 and 10 give the packet loss and the access delay respectively for the WLAN FTP and HTTP profiles. Ob-Table 7 Bluetooth packet loss probability for experiment 4. BT traffic WLAN traffic FTP HTTP BIAS RR BIAS RR FTP d B = 1 m 0.0103 0.0158 0.0064 0.0356 d B = 2 m 0.1079 0.1210 0.0379 0.0393 HTTP d B = 1 m 0.0012 0.0034 0.0003 0.0002 d B = 2 m 0.0425 0.0614 0.0265 0.0071 Table 8 Bluetooth MAC delay (sec) for experiment 4. BT traffic WLAN traffic FTP HTTP BIAS RR BIAS RR FTP d B = 1 m 0.1805 0.1749 0.1912 0.1739 d B = 2 m 0.3753 0.4574 0.2444 0.2378 HTTP d B = 1 m 0.0840 0.0861 0.0836 0.0835 d B = 2 m 0.0945 0.1121 0.0963 0.0952 Table 9 WLAN probability of packet loss for experiment 4. BT traffic WLAN traffic FTP HTTP BIAS RR BIAS RR FTP 0.1534 0.303 0.2510 0.3481 HTTP 0.0192 0.0961 0.0721 0.1534 Table 10 WLAN MAC delay (sec) for experiment 4. BT traffic WLAN traffic FTP HTTP BIAS RR BIAS RR FTP 0.0017 0.0022 0.0010 0.0011 HTTP 0.0011 0.0018 0.0009 0.0012 serve a significant reduction in packet loss with BIAS for both WLAN applications, in which the packet loss drops from 30% and 34% to 15% and 25% for the FTP and HTTP application, respectively. The access delay shown in table 10 is consistent with the packet loss results and shows slight improvements with BIAS. In summary, the use of BIAS in a multi-Bluetooth and WLAN environment leads to performance improvements for WLAN, while it has little benefits on the Bluetooth performance Concluding remarks In this paper we propose a scheduling technique, BIAS, aimed at eliminating interference on WLAN and alleviating the impact of interference on the Bluetooth performance. This work addresses the need to adjust to changes in the environment, support asymmetric traffic in the upstream and downstream, in addition to the use of different scheduling priorities. 30 N. GOLMIE Table 11 BIAS pseudocode. 1: Every N slots 2: estimate_channel(); 3: compute_credits(); 4: Every even TS f // Master transmission slot 5: if TS f + l dn is clear // Master can receive in next slot 6: { 7: A f data = {set of high priority slaves s.t. ((f "good") and (qsize &gt; 0) and (c dn &gt; 0)} 8: A f poll = {set of high priority slaves s.t. ((f "good") and (c up &gt; 0))} 9: B f data = {set of low priority slaves s.t. ((f "good") and (qsize &gt; 0))} 10: B f poll = {set of low priority slaves s.t. ((f "good") and (c up c dn &gt; 0))} 11: // Service high priority slaves first 12: if (A f data = ) // transmit data packets 13: { 14: i = max A f data (w i up w i dn ) // select device i with the largest weight 15: transmit data packet of size l dn to slave i 16: c i dn,up = c i dn,up - l idn,up ; // decrement credit counter 17: w i dn,up = (1 - u i ) c i dn,up ; // update weights 18: } 19: else if (A f poll = ) // transmit polls 20: { 21: i = max A f poll (w i up ) // select device i with the largest weight 22: transmit poll to slave i 23: c i up = c i up - l iup ; // decrement credit counter 24: w i up = (1 - u i ) c i up ; // update weights 25: } 26: // Then service low priority slaves 27: else if (B f data = ) 28: { 29: i = max B f data (w i up w i dn ) // select device i with the largest weight 30: transmit data packet of size l dn to slave i 31: if (c i dn &gt; 0) c i dn = c i dn - l i dn ; // decrement credit counter 32: else c i up = c i up - l i dn ; // decrement credit counter 33: w i dn,up = (1 - u i ) c i dn,up ; // update weights 34: } 35: else if (B f poll = ) // transmit polls 36: { 37: i = max B f poll (w i up ) // select device i with the largest weight 38: transmit poll to slave i 39: if (c up &gt; 0) c i up = c i up - l iup ; // decrement credit counter 40: else c i dn = c i dn - l iup ; // decrement credit counter 41: w i dn,up = (1 - u i ) c i dn,up ; // update weights 42: } 43: } The performance results obtained are summarized as follows . First, BIAS eliminates packet loss even in the worst interference case when more than 3/4 of the spectrum are occupied by other devices. Delay is slightly increased over the reference scenario (when no interference is present). This increase varies between 1 to 5 ms on average. Furthermore, BIAS is able to rapidly adjusts to changes in the channel. The channel estimation transient time can be as low as 1.5 ms and 250 s for DH5 and DH1 packets, respectively. In addition , BIAS supports QOS and maintains a low access delay for delay-sensitive traffic such as video applications. Finally, we observe that the use of BIAS is not as effective to mitigate interference caused by other Bluetooth piconets. In this case, we note no improvements in access delay and packet loss results , which are comparable to results obtained with Round Robin (RR). An immediate next step for our work consists of developing a channel estimation procedure that is able to differentiate between different types of interference, namely, WLAN and Bluetooth interference. Our preliminary results indicate that this may be helpful in a multi-Bluetooth and WLAN environment . BLUETOOTH DYNAMIC SCHEDULING AND INTERFERENCE MITIGATION 31 Acknowledgements The author would like to thank O. Rebala and A. Tonnerre for their help in developing the simulation models and compiling the results. References [1] Bluetooth Special Interest Group, Specifications of the Bluetooth System , Vol. 1, v. 1.0B "Core" and Vol. 2, v. 1.0B "Profiles" (December 1999). [2] A. El-Hoiydi, Interference between Bluetooth networks upper bound on the packet error rate, IEEE Communications Letters 5 (June 2001) 245247. [3] D. Fumolari, Link performance of an embedded Bluetooth personal area network, in: Proceedings of IEEE ICC'01, Vol. 8, Helsinki, Finland (June 2001) pp. 25732577. [4] N. Golmie, N. Chevrollier and I. Elbakkouri, Interference aware Bluetooth packet scheduling, in: Proceedings of GLOBECOM'01, Vol. 5, San Antonio, TX (November 2001) pp. 28572863. [5] 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, Vol. 8, Helsinki, Finland (June 2001) pp. 25402545. [6] 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) pp. 1118. Extended version appeared in ACM Wireless Networks 9(3) (2003) 201211. [7] I. Howitt, V. Mitter and J. Gutierrez, Empirical study for IEEE 802.11 and Bluetooth interoperability, in: Proceedings of IEEE Vehicular Technology Conference (VTC), Vol. 2 (Spring 2001) pp. 11031113. [8] IEEE Std. 802-11, IEEE Standard for Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specification (June 1997). [9] J. Lansford, R. Nevo and E. Zehavi, MEHTA: A method for coexistence between co-located 802.11b and Bluetooth systems, IEEE P802.11 Working Group Contribution, IEEE P802.15-00/360r0 (November 2000). [10] 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, Seattle, WA (May 2000). [11] B. Treister, A. Batra, K.C. Chen and O. Eliezer, Adapative frequency hopping: A non-collaborative coexistence mechanism, IEEE P802.11 Working Group Contribution, IEEE P802.15-01/252r0, Orlando, FL (May 2001). Nada Golmie received the M.S.E. degree in computer engineering from Syracuse University, Syracuse , NY, and the Ph.D. degree in computer science from the University of Maryland, College Park, MD. Since 1993, she has been a member of the Advanced Network Technologies Division of 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 consortia. 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

WLAN;BIAS;QoS;inteference;dynamic scheduling;Bluetooth;scheduling priorities;interference;coexistence;MAC scheduling;WPANs;WPAN

46

Breadth-First Search Crawling Yields High-Quality Pages

This paper examines the average page quality over time of pages downloaded during a web crawl of 328 million unique pages. We use the connectivity-based metric PageRank to measure the quality of a page. We show that traversing the web graph in breadth-first search order is a good crawling strategy, as it tends to discover high-quality pages early on in the crawl.

INTRODUCTION According to a study released in October 2000, the directly accessible "surface web" consists of about 2.5 billion pages, while the "deep web" (dynamically generated web pages) consists of about 550 billion pages, 95% of which are publicly accessible [9]. By comparison, the Google index released in June 2000 contained 560 million full-text-indexed pages [5]. In other words, Google -- which, according to a recent measurement [6], has the greatest coverage of all search engines -covers only about 0.1% of the publicly accessible web, and the other major search engines do even worse. Increasing the coverage of existing search engines by three orders of magnitude would pose a number of technical challenges , both with respect to their ability to discover, download , and index web pages, as well as their ability to serve queries against an index of that size. (For query engines based on inverted lists, the cost of serving a query is linear to the size of the index.) Therefore, search engines should attempt to download the best pages and include (only) them in their index. Cho, Garcia-Molina, and Page [4] suggested using connectivity -based document quality metrics to direct a crawler towards high-quality pages. They performed a series of crawls over 179,000 pages in the stanford.edu domain and used Copyright is held by the author/owner. WWW10, May 1-5, 2001, Hong Kong. ACM 1-58113-348-0/01/0005. different ordering metrics -- breadth-first, backlink count, PageRank [2], and random -- to direct the different crawls. Under the breath-first ordering, pages are crawled in the order they are discovered. Under the backlink ordering, the pages with the highest number of known links to them are crawled first. Under the PageRank ordering, pages with the highest PageRank (a page quality metric described below) are crawled first. Under the random ordering, the crawler selects the next page to download at random from the set of uncrawled pages. (For repeatability, these crawls were "virtual"; that is, they were performed over a cached copy of these 179,000 pages.) Cho et al. evaluated the effectiveness of each ordering metric by examining how fast it led the crawler to all the "hot" pages. In this context, a "hot" page is a page with either a high number of links pointing to it, or a page with a high PageRank. They found that using the PageRank metric to direct a crawler works extremely well. However, they also discovered that performing the crawl in breadth-first order works almost as well, in particular if "hot" pages are defined to be pages with high PageRank. This paper extends the results of Cho et al. regarding the effectiveness of crawling in breadth-first search order, using a much larger and more diverse data set. While Cho's work was based on a crawl of 179,000 pages from the stanford.edu domain, we performed a crawl of 328 million pages over the entire web, covering more than 7 million distinct hosts. We use connectivity-based page quality metrics, namely Brin and Page's PageRank and variations of it, to measure the quality of downloaded pages over the life of the crawl. We find that not only does breadth-first search download the hot pages first, but also that the average quality of the pages decreased over the duration of the crawl. We also suggest that our crawler's modifications to strict breadth-first search -- made to increase the overall download rate and to avoid overloading any given web server -- enhance its likeliness of retrieving important pages first. The remainder of this paper is structured as follows: Section 2 reviews the PageRank metric we used to evaluate the effectiveness of crawling in breadth-first search order. Section 3 describes the tools we used to conduct our experiments . Section 4 describes the experiments we performed, and the results we obtained. Finally, section 5 offers concluding remarks. PAGERANK There are many conceivable metrics for judging the quality of a web page: by analyzing its content, by measuring 114 its popularity (that is, how often it is viewed), or by examining its connectivity (that is, by determining which other pages link to this page, and vice versa). Metrics based on connectivity have the advantages that they do not require information that is not easily accessible (such as page popularity data), and that they are easy to compute, so they scale well to even very large page collections. They also require retrieving only the links on each page, not the full page contents. Storing the full page contents requires several kilobytes per page, one to two orders of magnitude more than just storing the links. PageRank is the connectivity-based page quality measure suggested by Brin and Page [2]. It is a static measure; it is designed to rank pages in the absence of any queries. That is, PageRank computes the "global worth" of each page. Intuitively, the PageRank measure of a page is similar to its in-degree, which is a possible measure of the importance of a page. The PageRank of a page is high if many pages with a high PageRank contain links to it, and a page containing few outgoing links contributes more weight to the pages it links to than a page containing many outgoing links. The PageRank of a page is expressed mathematically as follows. Suppose there are T total pages on the web. We choose a parameter d (explained below) such that 0 &lt; d &lt; 1; a typical value of d might lie in the range 0.1 &lt; d &lt; 0.15. Let pages p 1 , p 2 , . . . , p k link to page p. Let R(p i ) be the PageRank of p i and C(p i ) be the number of links out of p i . Then the PageRank R(p) of page p is defined to satisfy: R(p) = d T + (1 - d) k X i=1 R(p i ) C(p i ) This equation defines R(p) uniquely, modulo a constant scaling factor. If we scale R(p) so that the PageRanks of all pages sum to 1, R(p) can be thought of as a probability distribution over pages. The PageRank distribution has a simple interpretation in terms of a random walk. Imagine a web surfer who wanders the web. If the surfer visits page p, the random walk is in state p. At each step, the web surfer either jumps to a page on the web chosen uniformly at random, or the web surfer follows a link chosen uniformly at random from those on the current page. The former occurs with probability d, the latter with probability 1 - d. The equilibrium probability that such a surfer is at page p is simply R(p). An alternative way to say this is that the average fraction of the steps that a walk spends at page p is R(p) over sufficiently long walks. This means that pages with high PageRank are more likely to be visited than pages with low PageRank. In our experiments, we set d = 1 7 = 0.14. We also modified PageRank slightly so that pages with no outgoing links contribute their weight equally to all pages. That is, the random surfer is equally likely to jump to any page from a page with no outgoing links. We ran experiments using both the original PageRank algorithm, which does not distinguish between links to pages on the same versus different hosts, and a variant of PageRank which only considers links to different hosts. TOOLS We used two tools in conducting this research: Mercator and the Connectivity Server 2, both developed at our lab. We used Mercator to crawl the web, and the Connectivity Server 2 to provide fast access to the link information downloaded from the crawl. Mercator is an extensible, multithreaded, high-performance web crawler [7, 10]. It is written in Java and is highly configurable. Its default download strategy is to perform a breadth-first search of the web, with the following three modifications: 1. It downloads multiple pages (typically 500) in parallel. This modification allows us to download about 10 million pages a day; without it, we would download well under 100,000 pages per day. 2. Only a single HTTP connection is opened to any given web server at any given time. This modification is necessary due to the prevalence of relative URLs on the web (about 80% of the links on an average web page refer to the same host), which leads to a high degree of host locality in the crawler's download queue. If we were to download many pages from the same host in parallel, we would overload or even crash that web server. 3. If it took t seconds to download a document from a given web server, then Mercator will wait for 10t seconds before contacting that web server again. This modification is not strictly necessary, but it further eases the load our crawler places on individual servers on the web. We found that this policy reduces the rate of complaints we receive while crawling. For the experiments described below, we configured Mercator to extract all the links from each downloaded page and save them to disk; for disk space reasons, we did not retain the pages themselves. We conducted a crawl that attempted to download 532 million pages over the course of 58 days (which we refer to as days 1 to 58 throughout the paper ). Of all those download attempts, 328 million returned valid, unique HTML pages; the others resulted in TCP- and DNS-errors, non-200 HTTP return codes, non-HTML documents , or duplicates. Mercator's download rate decreased over the course of the crawl, due to increasing access times to one of its disk-based data structures that keeps track of which URLs have already been seen. The median download day was 22; the mean download day was 24.5. The extracted links data was then loaded into the Connectivity Server 2 (CS2) [11], a database for URLs and links. A build of CS2 takes a web crawl as input and creates a database representation of the web graph induced by the pages in the crawl. A CS2 database consists of all URLs that were crawled, extended with all URLs referenced at least five times by the crawled pages. (Incorporating uncrawled URLs with multiple links pointing to them ensured that we did not ignore any popular URLs. Setting the threshold at five incoming links reduced the set of uncrawled URLs by over 90%, which enabled us to fit the database within the 16 GB of RAM available to us.) The CS2 database also contains all links among those URLs and host information for each URL. It maps each URL to all of its outgoing and its incoming links. It is possible to get all the incoming links for a given URL, or just the links from different hosts. CS2 stores links in both directions in, on average, 2.4 bytes per link (as compared to 8 bytes per link in the original connectivity server (CS1) described in [1]). Like CS1, 115 0 5 10 15 20 25 30 35 40 45 50 55 Day of crawl 0 2 4 6 8 Average PageRank Figure 1: Average PageRank score by day of crawl CS2 is designed to give high-performance access when run on a machine with enough RAM to store the database in memory. On the 667 MHz Compaq AlphaServer ES40 with 16 GB of RAM used in our experiments, it takes 70-80 ms to convert a URL into an internal id or vice versa, and 0.1 ms/link to retrieve each incoming or outgoing link as an internal id. The database for our crawl of 328 million pages contained 351 million URLs and 6.1 billion links. Therefore, one iteration of PageRank ran in about 15 minutes. AVERAGE PAGE QUALITY OVER A LONG CRAWL In this section, we report on our experiments. We implemented PageRank and its variants over the CS2 interface, and ran each algorithm for 100 iterations on the 6.1 billion link database. (In all our experiments, the PageRank computation converged within less than 100 iterations.) Although the PageRank scores are conventionally normalized to sum to 1 (making it easier to think of them as a probability distribution), we normalized them to sum to the number of nodes in the graph (351 million). This way, the average page has a PageRank of 1, independent of the number of pages. Figure 1 shows the average PageRank of all pages downloaded on each day of the crawl. The average score for pages crawled on the first day is 7.04, more than three times the average score of 2.07 for pages crawled on the second day. The average score tapers from there down to 1.08 after the first week, 0.84 after the second week, and 0.59 after the fourth week. Clearly, we downloaded more high quality pages, i.e., pages with high PageRank, early in the crawl than later on. We then decided to examine specifically when we had crawled the highest ranked pages. We examined the pages with the top N PageRanks, for increasing values of N from 1 to 328 million (all of the pages downloaded). Figure 2 graphs the average day on which we crawled the pages with the highest N scores. Note that the horizontal axis shows the values of N on a log scale. All of the top 10 and 91 of the top 100 pages were crawled on the first day. There are some anomalies in the graph between N equals 100 and 300, where the average day fluctuates between 2 and 3 (the second and third days of the crawl). These anomalies are caused by 24 pages in the top 300 (8%) that were downloaded after the first week. Most of those pages had a lot of local links (links from pages on the same host), but not many remote links. In other words, the 1 10 100 1000 10000 100000 1e+06 1e+07 1e+08 top N 5 10 15 20 25 Average day top N pages were crawled Figure 2: Average day on which the top N pages were crawled pages on the same host "endorse" each other, but few other hosts endorse them. We address this phenomenon later in the last experiment, shown in Figure 4. After N equals 400, the curve steadily increases to day 24.5, the mean download day of the entire crawl. Our next experiment checks that pages with high PageRank are not ranked high only because they were crawled early. For example, a page whose outgoing links all point to pages with links back to it might have an artificially high PageRank if all of its outgoing links have been crawled, but not too many other pages. For this experiment we ran the PageRank algorithm on the graph induced by only the first 28 days of the crawl. This graph contains 217 million URLs and 3.8 billion links between them. We then compared the top ranked pages between the two data sets. We found that of the top 1 million scoring pages, 96% were downloaded during the first 4 weeks, and 76% of them were ranked in the top 1 million pages in the 28 day data set. That is, it was clear that those pages were important even before the crawl had finished. Figure 3 generalizes these statistics: for each value of N, we plot the percentage of overlap between the top N scoring pages in the 28 day crawl versus the 58 day crawl. Although the top few pages are different, by the top 20 ranked pages there is an 80% overlap. The overlap continues in the 60-80% range through the extent of the entire 28 day data set. This figure suggests that breadth-first search crawling is fairly immune to the type of self-endorsement described above: although the size of the graph induced by the full crawl is about 60% larger than the graph induced by the 28 day crawl, the longer crawl replaced only about 25% of the "hot" pages discovered during the first 28 days, irrespective of the size of the "hot" set. Some connectivity-based metrics, such as Kleinberg's algorithm [8], consider only remote links, that is, links between pages on different hosts. We noticed that some anomalies in Figure 2 were due to a lot of local links, and decided to experiment with a variant of the PageRank algorithm that only propagates weights along remote links. This modification of PageRank counts only links from different hosts as proper endorsements of a page; links from the same host are viewed as improper self-endorsement and therefore not counted. Figure 4 shows our results: the average PageRank for pages downloaded on the first day is even higher than when all links are considered. The average PageRank for the first day is 12.1, while it's 1.8 on the second day and 1.0 on the 116 1 10 100 1000 10000 100000 1e+06 1e+07 1e+08 top N pages 0 20 40 60 80 100 % overlap between data sets Figure 3: The percent overlap between the top N ranked pages in the first 28 vs all 58 days of the crawl fourth day. The average PageRank then declines gradually down to 0.6 on the last day. Notice that the average PageRank on the first day of crawling is higher than in Figure 1, and that the curve falls more sharply. This drop indicates that our crawling strategy is not biased toward self-endorsing hosts, as a crawler using the standard version of PageRank would be. We believe that this lack of bias is due in part to our crawler's politeness policies, which impose a rate limit on its accesses to any particular host. There are some flaws with a metric based only on remote links. For example, http://www.yahoo.com/ has a very high PageRank score. However, it only has local outlinks, so its weight gets evenly distributed over all pages in the graph, rather than just to the other pages in Yahoo! to which it points. Transitively, the pages on other hosts to which Yahoo! links do not benefit from the high score of http://www.yahoo.com/. In the future work section below, we outline some ideas for remedying this problem. CONCLUSIONS The experiments described in this paper demonstrate that a crawler that downloads pages in breadth-first search order discovers the highest quality pages during the early stages of the crawl. As the crawl progresses, the quality of the downloaded pages deteriorates. We speculate that breadth-first search is a good crawling strategy because the most important pages have many links to them from numerous hosts, and those links will be found early, regardless of on which host or page the crawl originates. Discovering high-quality pages early on in a crawl is desirable for public web search engines such as AltaVista or Google, given that none of these search engines is able to crawl and index more than a fraction of the web. Our results have practical implications to search engine companies. Although breadth-first search crawling seems to be a very natural crawling strategy, not all of the crawlers we are familiar with employ it. For example, the Internet Archive crawler described in [3] does not perform a breadth-first search of the entire web; instead, it picks 64 hosts at a time and crawls these hosts in parallel. Each host is crawled exhaustively; links that point to other hosts are saved to seed subsequent crawls of those hosts. This crawling strategy has no bias towards high-quality pages; if the hosts to be crawled are picked in random order, the quality of downloaded pages will be uniform throughout the crawl. 0 5 10 15 20 25 30 35 40 45 50 55 Day of crawl 0 2 4 6 8 10 12 Average PageRank (remote links only) Figure 4: Average PageRank when only remote links are considered Similarly, the Scooter web crawler used until recently by AltaVista downloaded pages in essentially random order. (At this point, AltaVista is using Mercator.) This approach made it easier to provide politeness guarantees -- essentially, it spread the load imposed by the crawler evenly over all web servers -- but as a result, the quality of the discovered pages is uniform over the life of the crawl. We cannot make any statements about other large-scale web crawlers. Most search engine companies treat their crawling strategy as a trade secret, and have not described it in the literature. Cho et al. [4] showed that using a connectivity-based ordering metric for downloads, such as PageRank, will steer the crawler towards even higher-quality pages than using breadth-first search. However, computing PageRank values for several hundred million or more pages is an extremely expensive computation. It took us over a day to compute the PageRanks of our graph of 351 million pages, despite the fact that we had the hardware resources to hold the entire graph in memory! Using PageRank to steer a crawler would require multiple such computations over larger and larger graphs, in order to take newly discovered pages into account, and is essentially infeasible in real time. On the other hand, crawling in breadth-first search order provides a fairly good bias towards high quality pages without the computational cost. We believe that crawling in breadth-first search order provides the better tradeoff. FUTURE WORK There are two directions in which we would like to extend this work. One direction is to try a variant of PageRank which weighs links to pages on remote hosts differently than links to other pages on the same host. From the experiment that generated Figure 4 above, we learned that remote links should count more than local links, but that weights should be propagated along local links as well (e.g., to distribute the weight of http://www.yahoo.com/ to the pages that Yahoo! recommends). We suspect that some search engines already use different weights for links, but there has been no formal study of how to divide the weights among the links or even whether the division should be static (e.g., remote links get 80% of the total weight) or proportional to the number of total links (e.g., each remote link gets four times the weight of each local link). The other direction is to try different connectivity-based 117 metrics. While PageRank is the only connectivity measure we know aimed at ranking all of the pages on the world wide web, Kleinberg's algorithm [8] is another well-known connectivity analysis algorithm targeted towards computing quality scores for pages. The algorithm computes two scores for each document: a hub score and an authority score. Pages with high authority scores are expected to have high-quality content; the authority scores are similar in intent to PageRanks . Kleinberg's algorithm is designed to rank the results of a query to a search engine, and only considers a small set of pages when it computes authority scores. However, we believe that we can extend the algorithm to consider the entire graph of the web. REFERENCES [1] K. Bharat, A. Broder, M. Henzinger, P. Kumar, and S. Venkatasubramanian. The connectivity server: Fast access to linkage information on the web. In Proceedings of the 7th International World Wide Web Conference, pages 469477, Brisbane, Australia, April 1998. Elsevier Science. [2] S. Brin and L. Page. The anatomy of a large-scale hypertextual web search engine. In Proceedings of the 7th International World Wide Web Conference, pages 107117, Brisbane, Australia, April 1998. Elsevier Science. [3] M. Burner. Crawling towards eternity: Building an archive of the world wide web. Web Techniques Magazine, 2(5):3740, May 1997. [4] J. Cho, H. Garcia-Molina, and L. Page. Efficient crawling through URL ordering. In Proceedings of the 7th International World Wide Web Conference, pages 161172, Brisbane, Australia, April 1998. Elsevier Science. [5] Google Inc. Press release: "Google launches world's largest search engine." June 26, 2000. Available at http://www.google.com/press/pressrel/pressrelease26.html [6] M. Henzinger, A. Heydon, M. Mitzenmacher, and M. Najork. On near-uniform URL sampling. In Proceedings of the 9th International World Wide Web Conference, pages 295308, Amsterdam, Netherlands, May 2000. Elsevier Science. [7] A. Heydon and M. Najork. Mercator: A scalable, extensible web crawler. World Wide Web, 2(4):219229, Dec. 1999. [8] J. Kleinberg. Authoritative sources in a hyperlinked environment. In Proceedings of the 9th ACM-SIAM Symposium on Discrete Algorithms, pages 668677, San Francisco, CA, Jan. 1998. [9] P. Lyman, H. Varian, J. Dunn, A. Strygin, and K. Swearingen. How much information? School of Information Management and Systems, Univ. of California at Berkeley, 2000. Available at http://www.sims.berkeley.edu/how-much-info [10] Mercator Home Page. http://www.research.digital.com/SRC/mercator [11] J. L. Wiener, R. Wickremesinghe, M. Burrows, K. Randall, and R. Stata. Better link compression. Manuscript in progress. Compaq Systems Research Center, 2001. VITAE Marc Najork is a senior member of the research staff at Compaq Computer Corporation's Systems Research Center. His current research focuses on high-performance web crawling and web characterization. He was a principal contributor to Mercator, the web crawler used by AltaVista. In the past, he has worked on tools for web surfing , 3D animation, information visualization , algorithm animation, and visual programming languages. He received a Ph.D. in Computer Science from the University of Illinois at Urbana-Champaign in 1994. Janet L. Wiener is a member of the research staff at Compaq Computer Corporation's Systems Research Center. She currently focuses on developing algorithms to characterize the web and tools (such as the Connectivity Server) to support those algorithms . Prior to joining Compaq in 1998, she was a research scientist at Stanford University working on data warehousing, heterogeneous data integration , and semi-structured data. She received a Ph.D. from the University of Wisconsin-Madison in 1995, and a B.A. from Williams College in 1989. 118

;high quality pages;breadth first search;crawl order;ordering metrics;Crawling;crawling;PageRank;page quality metric;breadth-first search;connectivity-based metrics

47

Broadcasting Information via Display Names in Instant Messaging

Many instant messenger (IM) clients let a person specify the identifying name that appears in another person's contact list. We have noticed that many people add extra information to this name as a way to broadcast information to their contacts. Twelve IM contact lists comprising 444 individuals were monitored over three weeks to observe how these individuals used and altered their display names. Almost half of them changed their display names at varying frequencies, where the new information fell into seventeen different categories of communication supplied to others. Three themes encompass these categories: Identification ("who am I"?), Information About Self ("this is what is going on with me") and Broadcast Message ("I am directing information to the community"). The design implication is that systems supporting person to person casual interaction, such as IM, should explicitly include facilities that allow people to broadcast these types of information to their community of contacts.

INTRODUCTION Millions of people use instant messenger (IM) clients daily to communicate with friends, relatives, co-workers and even online dating contacts. With this explosion of use, researchers have taken to studying instant messaging and its impact. Much of the research regarding IM has been focused on its primary uses: maintaining awareness of a contact's presence and availability, how people (usually dyads) converse via text chat, and how they exploit other features such as file sharing and receipt of notifications. For example, studies of IM use in the workplace expose how it supports collaboration, communication and project activities [3, 10, 13], as well as its negative effects [15] such as disruption [4]. In more social contexts, researchers found a positive relationship between the amount of IM use and verbal, affective and social intimacy [9]. IM also proves important in the life of teens, where it helps support the maintenance of their social relationships [8]. Other computer-mediated communication tools, such as MUDs (Multi-User Domains or Multi-User Dungeons), IRC (Internet Relay Chat), and broadcast messaging tools also allow spontaneous real-time (or synchronous) communication with other users. However, there are significant differences between them. IM is predominately used between people who are known to each other outside of cyberspace, e.g., friends and associates. IM conversations are also private, and tend to be between pairs of people. They are also person centered and not group centered: while a contact list may collect one's `buddies', these lists are not shared across contacts. In contrast, MUDs and IRC are public channels, where any conversation is heard by all people currently in the MUD or IRC. Most tend to be used by `strangers', i.e., those who are unknown to each other in real space, and usually involve more than two individuals. Indeed, the norm is for participants to protect their anonymity by displaying a pseudonym rather than their real names. Any personal messages that are posted are usually in relation to their virtual identity. However, a few experimental MUD-like systems do focus on teams, where they provide its members with rich awareness information of one another and more power in their collaboration tools, e.g., Sideshow [2], Notification Collage [7], or Community Bar [12]. Broadcast messaging tools [11] sit in the middle, where real-time messages usually comprising notifications and announcements (not conversations) are sent to large groups of people who are somehow associated with one another, e.g., Tickertape [6]. The big `win' of IM is that it provides one's ad hoc set of contacts with awareness of one's online state, which in turns serves as an estimate of one's availability for conversation. While not completely accurate [13], even this minimal information suffices to create opportunities for lightweight text-based conversations and to reduce the equivalent of `telephone tag'. While many research systems go far beyond IM in the rich awareness information they give to others [e.g., 2, 7, 12, 16], questions remain about privacy implications of distributing this information. IM contacts are identified by the system through e-mail addresses. While unique, these email addresses may be cryptic 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. GROUP'05, November 69, 2005, Sanibel Island, Florida, USA. Copyright 2005 ACM 1-59593-223-2/05/0011...$5.00. 89 to a human viewer e.g., a person may not be able to infer that 12gorwan@yahoo.com is really Gregor McEwan. Consequently, the designers of most IM clients include a feature that lets a person create and/or change their display name at any time; this name is shown to others instead of the email address. For example, in MSN Messenger (Figure 1) a person can raise the `Personal Settings...' dialog by selecting the drop-down menu attached to their name (i.e., `Stephanie' in the figure), and edit the `Display name' text field which we also call the display field within it. All contacts immediately see this new name. Because we are heavy IM users, we noticed that many of our contacts change their display field to do more than simply identify or label themselves. Figure 1 illustrates this, where we see that various people have used this feature to publicly broadcast what they are doing (e.g., `anitsirK Marking') or an event in their life (e.g., `Employed'), or a personal state of mind (e.g., `Chasing Insanity'). Other examples we saw include using the display field as a way to publicize personal status, specify location, post comments, ask questions, and even post popular culture references. These obviously augment IM's preset availability messages (i.e. away, busy, be right back) in far richer ways than the system was explicitly designed to support. We believed that people's appropriation of the IM display name feature into a public broadcast facility is a phenomenon worth exploring. Why was this space being appropriated for messages broadcast to an entire contact list? What were users trying to communicate to others and how is this information different than that in a normal IM conversation? How often do these messages or alternate communications occur? To answer these and other questions, we conducted a three week study, where we monitored changes in each person's display field within contact lists held by various users of MSN Messenger. We tracked how often contacts changed their display name, and what these changes were. We also categorized these changes into communication purposes. After briefly summarizing some related work, we describe the methodology used to acquire display name usage data. This is followed by our results, a discussion of the findings, implications of the work, and recommendations for future work. RELATED WORK There are a variety of articles describing how people identify themselves on the Internet, usually in MUDs and IRC. Yet most of these stress how identity is formed through pseudo-anonymity [1,17], i.e., where a person creates a virtual identity to project a different persona of who they are while protecting their real identity. People's choices of names and/or avatars are usually one part of identity creation. This work is not particularly applicable to IM, as people on a contact list are typically known to one another. Grinter & Palen's [8] study of teen use of IM is far more relevant, and partially reflects our own interests. While their work broadly considers IM as an emerging feature of teen life, they do mention that teens found the preset availability messages to be too impersonal. To combat feelings of exclusion or to avoid being rude, teens would personalize the display name area to include a message which explains their unavailability, changes in their local environment (i.e., `Going quiet because Mom just arrived'), and for justifying their lack of presence on the system (i.e., `Out for dinner'). The use of IM names to broadcast messages is an everyday world phenomenon, and has been anecdotally noticed by non-scientists . For example, one reporter noted in a newspaper article that changes to her display name are her main form of IM communication rather than actual chat conversations [14]. Social scientists talk more generally about computer mediated communications and how they can be used to build communities. Etzioni and Etzioni [5] argue that in order to form and sustain bonds, a community of connected individuals needs what they call "interactive broadcasting". This is composed of two major elements: the ability to broadcast messages to many people within the community simultaneously, and the ability for those addressed by the message to provide feedback, not just to the message originator, but to other message recipients as well. In this context, a broadcast message can be considered a request for interaction from some (or all) members of a group [11]. A variety of designers have implemented this broadcast capability into their systems. For example, IRC, Notification Collage [7], Community Bar [12] and Tickertape [6] are all tools that implement interactive broadcasting. A message (which may include multimedia information) can be posted and broadcast to the group, and it is possible for everyone to view the information without directly contributing to the conversation. Those who want to respond can do so, in full view of all users. All these systems allow for communal feedback, i.e., where everyone sees the response. Unlike IM, however, these systems include a strong notion of a common group by providing a public space for interaction. In summary, there are discussions of how broadcasting information contributes to community building, and there are systems that are based on public dissemination of information within a group. However, excepting a few discussions of this Figure 1: MSN Messenger; modified display names are visible 90 phenomenon [8,14], there has been no real analysis of how people have appropriated the display name feature of IM. Given the importance and widespread of IM, we believe this analysis is critical if we are to understand how we can improve IM systems. METHODOLOGY This study investigates how people use the display name feature in IM clients to broadcast information other than one's name. We do this by capturing changes in each person's display field as they appear in contact lists over time and over everyday use, by asking people to explain what these changes meant, and by counting, categorizing and analyzing these changes. 3.1 Research questions We wanted to identify three main behavioural patterns within our captured data: 1. At what frequency do users change the information in their display field when using an IM client such as MSN Messenger? 2. What are the main communication categories that represent the information held by these display field changes? 3. What is the frequency distribution of these categories? A fourth interesting but secondary question was: 4. Are changes to the display name related to the demographics of age or sex? 3.2 Participants We had two classes of participants. Our primary participants were those who made their contact list available to us. Our secondary participants were those who comprised the people on the contact lists. Twelve participants were recruited as primary participants, all Computer Science graduate students or faculty at the University of Calgary. They ranged in age from 23 to 50, and were regular users of MSN Messenger. These participants provided access to their IM contact lists. They were also willing to annotate the collected data. While the number of contacts on each person's list varied somewhat, this variance was irrelevant to our study. Our secondary participants were the 444 contacts found on the contact lists of the 12 primary participants. These contacts covered a broad range of demographics and social relationships, i.e., fellow students, workmates, friends, family members and other relatives. While the display names used by these 444 people were collected as data, they were not contacted directly. 3.3 Materials and Data Capture Each participant (whether primary or secondary) used their own pre-existing and unaltered MSN Messenger client on their own computer (running Windows) for everyday purposes. We wrote a logging program to collect all contact list data from each primary participant. It monitored every person's display field as it appeared in the contact list. The software worked by tapping into the programming API of MSN Messenger (regardless of its version) to monitor activities within it. This logging program was given only to the 12 primary participants. No special software was needed by the 444 secondary participants, as their data was captured via the logging software on the primary participant's computer. The 12 primary participants installed our software on whatever computers they wished. When installed, it worked in tandem with MSN Messenger to collect data on everyday IM usage in the field. The program monitored whether the participant was logged in to MSN Messenger. If logged in, it recorded the initial set of display names and any display name changes of the secondary participants on the contact list. The initial set of display names were needed to notice if a change occurred since the primary participant's last login. As part of our analysis, we used the standard features of Microsoft Excel 2003 to sort and consolidate the data files. Relevant data was then transferred to Minitab v.14 to tally distributions, calculate any statistics and create visual representations of the data. Further analysis of the categories of communication used in the display field was conducted using paper cut-outs and post-it notes to create an affinity diagram; this is detailed later. 3.4 Method Once primary participants agreed to participate in the study, we gave them instructions on how to install the logging program on their computer. We did not have to be present for this, so people could install it on whichever computers they regularly used, be it at work or at home. The program then ran automatically; the only indication of its operation was a small red notebook icon appearing in the participant's system tray. This icon allowed a participant to abort the collection process if they wished, but none chose this option. Data was collected for approximately three weeks, but did require the person to be logged onto MSN Messenger. If a primary participant was not logged on, no data about their contacts was recorded. This meant that some display field changes of secondary participants could have been missed. 3.5 Analysis At the end of three weeks, the primary participants were instructed to open the data file with Excel, and indicate the sex and approximate age of each listed contact member in a pre-designated column. For each display name change, they were also asked to categorize the type of information the contact was trying to broadcast to others. We did not predefine any categories. Participants created their own category labels and categorized names into them however they wished. We chose this approach because we felt that participants would have a far better understanding of the true meaning of a person's display field changes than someone unfamiliar with the contact; we also felt that as recipients of this information, their interpretation was important. We also believed that they would generate a greater and therefore richer breadth of categories. Once the categorizations were completed, the data files were transferred to the primary investigator. The investigator consolidated all of the data files into one master file, and removed any duplicate entries. These duplicate entries occurred for two reasons. More than one person had a particular contact on their list. Each time a participant logged in, their entire contact list was recorded in the data file. If a contact had not changed their name while the participant was offline, a duplicate entry was created. 91 When duplicate entries occurred, all but the earliest occurrence of the display name change was removed. A category list was created for each primary participant based on his or her individual categorizations of display name changes. Because these category names could differ between participants, we needed to re-categorize these names into a master category list. To do this, all categories were printed on separate slips of paper for easy sorting. We then created an affinity diagram to resort these categories, where entries from all the lists were sorted into groups based on similarity. These groups then formed a master category (see Figure 2). A master category title was then chosen that best represented the theme for the grouping. After this master list was established, the entries in the consolidated file were then re-categorized based on these new divisions; this would allow us to create a distribution profile. We should mention that many entries into the display field contained more than one textual element, each of which could be categorized differently. When this happened, we treated the display field as holding multiple entries. An example of this is shown here, where the contact's display field contains two elements; `Johnboy' could be categorized as a Name Variation, while `yonge&eglington' (a street junction in Toronto) is categorized as an indicator of Location. In this case, this display field entry would be split into two text fragments, where each fragment would be counted in the category that best fit. As we will see, these types of dual entry usually occurred because people tend to keep their names (or an identifying variation thereof) visible to others in order to identify themselves. Occasionally a display field would contain two elements where neither were identifiers. For example, the text shown here is categorized as two elements: `packing' is an Activity, and `sad to be leaving' is a Mood. Only rarely did display field entries contain more than two elements. RESULTS Our first research question was: At what frequency do users change the information in their display field when using an IM client such as MSN Messenger? Before answering this question, recall that the recording of display field changes of a secondary participant on a contact list only happened when the primary participant was logged on to MSN Messenger. If the primary participant was logged out, no display field changes to their contacts were recorded. While a single change would be noted by comparing the last recorded version of the contact's display field to the one recorded when the primary participant logs on, multiple changes during the logout period would be lost. This means we cannot calculate the exact display name change distribution across all contacts. Still, our numbers should be a good estimate of what happens. At the very least they represent a lower bound that somewhat underestimates how often display fields changes occur. The data certainly suffices to indicate the range of activities and individual differences across 444 people. Figure 3 illustrates the distribution of contacts according to how often they changed the contents of the display field. Our results show that 58% of our 444 contacts (258 people) never changed the contents of the display field during the three week period. For the remaining 42% of contacts (186 people), we counted a total of 1968 display name changes, or an average of 11 display name changes per person over the three week period, or up to 4 times a week. Never 58.1% Rarely 12.2% Weekly 5.6% Several times a day 4.5% Daily 3.4% Several times a week 16.2% Figure 3: Distribution of contacts according to how often they change the display field contents Figure 2: Example affinity diagram used to group participant categorizations into master categories. 92 However, this average is misleading, for we also found that people change their display names at different frequencies. We created six rate change categories. Based on a contact's data, we placed each contact into the category that best estimated that contact's change rate. Figure 3 displays this distribution of contacts among the six rate change categories. We see that the 42% of contacts who change their display name do so at different rates. About 8% (4.5% + 3.4%) of contacts change their names from once to several times a day. About 22% of them change their names less often, from once to several times a week (16.2% + 5.6%). The final 12% change it rarely, i.e., once or twice over the three week period. The person who had the highest display field change rate is worth added discussion, as it suggests what happens with contacts who used this feature heavily. This person changed her display field early in the morning, and notified contacts when she arrived at school. Around 4 pm the changes started again, continuing until approximately 11 pm when she went to bed. Her changes would incorporate details on what was occupying her time. Changes would state particulars: when she was studying, babysitting, or watching TV, and her emotional reactions to these events. If she found something entertaining or interesting on TV, she would post quotes. If she was bored, she would put out a request for someone/anyone to call. In essence, this person used her display field as a web log, where she recorded and disseminated information to her community. Even though we had no further knowledge of this person, a sense of who she was and what her life was garnered through all the changes that she made to her display name. 4.2 Communication categories Our second research question was: What are the main communication categories that represent the information held by these display field changes? After analyzing the categories created by our primary participants through the affinity diagramming process, we identified seventeen master categories. These are listed below in alphabetic order. A description of each category is given along with illustrative examples taken from our data. Many examples contain more than one textual element, usually an identifier, as we present them as they appeared in the display field. To protect confidentiality, name information has been changed. Activities include things or activities that a person has done in the past, is currently involved in, or is about to participate in the future. It also includes countdowns to an upcoming event. Examples include: Amy - House hunting! Joe was drunk on a Tuesday...shameful. Braced: 60% done my portfolio! Adverts include advertisements or promotions for items or events, and things that people have for sale. Easton Synergy Grip 100 Flex Iginla Blade Left (Brand spanking new): $225 headachey -- Tim Stuart Tribute and Fundraiser November 6th @ 8PM -- ask for details Comments are personal comments, expressions of an individual's opinion and general statements on how they view things in the world around them. Jan[et] - Airlines are Evil Bee - undocumented code should be illegal Nancy: you don't need English to live in Vancouver Default contains only the default unaltered entries to the display field. After installation, the IM client displays a person's e-mail address in the field. These may or may not actually contain a person's name as part of the email address. johnsmith@hotmail.com Jyn2l@hotmail.com Directions contain entries where a reader is being directed to a web site or link. Examples are: Bee-http://java.sun.com/features/1999/05/duke_gallery.html jessie {http://littlemisskool.blogspot.com} CHECK THIS====&gt; http://www.blitzkreigentt.com/.... Constructed &lt;==== Fun contains entries that contain puns, inside jokes, humorous statements, and items placed for the amusement of others. Melanie. me: "come see, its a lunar eclipse"; kate: "where?" what do you call a fish with no eyes: f sh Huffy - Home is where you hang your @ Joe Like a Vermin, trapped for the very first time Handles contains those display name entries that hold a person's handle. A handle is like a well known nickname: it is a consistent title or name that people give themselves to represent their identity on the internet. As we will see later, IM handles are not used for the pseudo-anonymity purposes as found in IRC public forums. hunnybear Iceman spidermax Location contains information about a person's current location or future destination. It can also contain travel information. Many times this location information is permanently attached to the display name when localized at a particular computer, as in "home" or "work". This label can indicate to others the type of communications that are appropriate. Mat Singh...going home in 10 days! In the dominican republic Dan James [Office] mike -&gt; lab meeting Messages contain information of significance directed at an individual on a person's contact list or to the group as a whole. darren~thanks nate for the halo hookup SirMe - Happy Birthday, Angie! Melanie. Nick, ill be on the 3 30 or whatever bus at the college. &lt;&lt;school&gt;&gt; Mood contains entries that give indications of a person's mood, feelings, health or state of being. i give up Adam feels rejected britney - disoriented haze Joe - as if shot in the head, yet still charging blindly forward. Bee - double espresso whee!!! Maggs - Not Feeling Well. Name contains entries of a person's given name. This category contains no nicknames, handles or variations on the name. Rebecca Fred Jones 93 Notices contains entries that give notice of a particular event, share news or display announcements. DBosch [ We're home owners! ] Tracey... down 24.2 lbs Jennifer - party is cancelled NaKuL - new msn login Gretchen -- Holy Cole's coming to vancouver!! Presence contains items which provide more detailed information about a person's online presence or availability beyond the standard status indicators. Bee - really am busy, only msg me in emergency Melanie. &gt;&gt; off for family time&lt;&lt; mike - reading at my desk/disregard (Away) status Flickerin: be Back at 630ish Questions contain rhetorical questions and questions that are posed to stimulate response. This category also contains questions that are requests for assistance, similar to those that appear in company broadcast messaging systems when a person is searching for an expert in a given area [6]. Luke -- Anyone took CS322? I need some help with cilog! Joe - who keeps messing with my chair?? Shri- Needin' a Physics Toolkit w/Dynamics + Collisions + Fields, any ideas? Melanie. Anyone have a working printer? Quotes contains quotations taken from movies, tv, books, plays or lyrics from music. It also contains references to pop culture. Dusit - If you can dodge a wrench, you can dodge anything! b33qZ -- king jeremy the wicked... oh, rules his world... Andrea - so long and thanks for all the fish Unknown contains all the entries in which the meaning of the text is too cryptic that it could not be categorized by either the primary participant or the investigator. It is assumed that once deciphered that each of these entries could be placed in one of the other sixteen categories. b33qZ [nts:perri] Andy ~ Ah ' Black_Venom (In 432) »~-jd-~«--&gt; SkRoNk &lt;-- yeh social ppl Variations contain entries where the identifier is a variation on the person's name. This can include an abbreviated version of the full name, a nickname in which the given name is still identifiable, or a variation in the way the letters of the name are printed or ordered. DiAnNe kev Maggs timbob Einahpets 4.3 Category Distribution Frequency Our third research question was: What is the frequency distribution of these categories? First, the 2226 logged display fields were analyzed to reveal a total of 3603 elements (recall that some display fields could have more than one information element in it). Second, each element was then located in a single communications category that best represented its meaning. Figure 4 shows these category counts in two sections. The top part plots the Name, Variations and Handle categories. We separated these `Identification' categories from the other categories because the information they contain satisfy the original purpose of the display field i.e., to hold identifying information. The frequency distribution of the remaining 14 categories are then listed. The bar representing the counts of the number of elements within each of these categories are further distinguished into three groups. The lightest section of each bar represents the group of category elements that were the only element contained by the display field. The medium coloured section shows the number of category elements whose text coexisted with another element found in one of the three `Identification' categories in the display field. The darkest section of the bar groups category elements whose text coexisted in the display field with another element found in any category other than the three `Identification' categories. The figure shows that approximately 49%, or 1766/3603 of the categorized elements, were in one of the three `Identification' categories, i.e., Name (32.4%), Variations (10%) or Handle (6.4%). This makes sense, for meaningful self-identification is the expected use of the IM display name feature. The darkly colored regions of their bars also reveal that identification elements in total coexist with other pieces of information in the display field over 67% (1186/1766) of the time. For example, the Name was included with other elements 825/1168 (71%) of the time. Similarly, Variations and Handles was included Figure 4: Bar chart displaying category distribution 94 205/359 (43%) and 156/239 (65%) in conjunction with other elements. Note that there are no medium coloured regions in these bars. This is because elements within the Name, Variations and Handle categories never co-existed with each other. They only occurred in conjunction with elements in the other 14 category types. The other 14 categories of communication identify information unrelated to identification. Collectively, these categories comprise the other 51% of the total number of elements (1837 of 3603 total). Within these 1837 elements, we see that the most frequent categories of communication used are Mood at 19.4% (357/1837), Comments at 17.8% (327/1837) Activities at 16.6% (305/1837), Location at 12.5% (230/1837), Messages at 8.3% (152/1837), followed by Quotes, Notices and Fun. The other categories occur less often, but still at a significant level. The modest size of the lightly coloured section of all these categories suggest that this information often appeared in tandem with other categories. Most of time, this was one of the Name, Variations, or Handle elements, as represented by the medium-coloured section in each bar. Still, the presence of the darkly coloured bar sections showed that two non-identifier category elements may coexist in a display field. 4.4 Demographics of People Who Change Their Display Names Our final research question was: Are changes to the display name related to the demographics of age or sex? The 444 contacts comprised somewhat more males than females. The primary participants reported 232 males, 189 females, and 1 male/female (the account was known to be used by a couple). The sex of the remaining 22 contacts was not reported. The dominant age range of the 444 contacts was between 21-30 years old. Table 1 summarizes the age demographics of the 444 contacts, as reported by our 12 primary participants. Since the exact age of each contact was sometimes uncertain, we used age group categories to capture their estimated ages. We then analyzed whether age or sex of a person was related to the number of changes that person made. First, we removed records for those contacts whose sex was not reported. We then performed a chi-square analysis on the remaining 421 contacts to determine whether there was a relationship between sex and the rate that a person changed their display field. Sex and display name change rate were found to be independent, 2 (5, N = 421) = 7.54, p = 0.183. That is, no relationship exists between the sex of a person and how often a person changes the display name. We performed a similar chi-square analysis for age and display name change rates, where unreported people were excluded. Age groups were collapsed into three age ranges: &lt;20, 21 to 30, and 31+. This was done for analytic reasons, since several cells in the chi-square analysis would have contained counts of less than one with the original divisions. Age range and name change rates were found to be not independent, 2 (10, N = 413) = 20.507, p = 0.025. That is, a relationship exists between the age of a person and their likelihood of changing their display name. This result will be examined further in the discussion. DISCUSSION The most important thing revealed by our study is that a good number of people persistently used the display name feature to publicly broadcast information about themselves to their friends, and that this happened outside of individual chat sessions. They did this in spite of the fact that IM display fields are not explicitly designed to be a public broadcast system. This suggests that systems should be designed to better support this kind of broadcast activity. Details are discussed below. 5.1 Interpreting the results People change the information in their display field. From this study we have learned that the changing of the information in an IM display field is not an oddity or something done occasionally by certain individuals. Rather, it is a popular behaviour: 42% of users in our study changed their display name, and 25% did so several times a week or more. This behaviour happens in spite of the fact that the Instant Messenger client we studied does not make changing the display name immediately accessible (e.g., through direct manipulation): people had to raise menus, dialog boxes, and form fill the text. People use the display field for identification, to give information about self, and to broadcast messages. People used the limited text that could be displayed in the display field in rich ways. Seventeen different categories were needed to describe the various communications placed in the display field. Stepping back, three themes encompass these categories. The first theme is Identification: "who am I"? The second theme is Information About Self: "this is what is going on with me". The third theme is Broadcast Messages: "I am directing information to the community". These are described separately in the following three sections. Identification is fundamental. Identifying oneself to personal contacts by typing one's own name in the display field is the original purpose of this feature; the name replaces the default email address as a way to uniquely identify a person. This proved necessary because e-mail addresses are a poor substitute for a name; some email services enforce cryptic email addresses, and others are so oversubscribed that all but the rarest names are already taken. Table 1: Age distribution of contact group Age Group Count Percent &lt;15 7 1.69 16-20 24 5.81 21-25 179 43.34 26-30 126 30.51 31-35 36 8.72 36-40 18 4.36 40+ 23 5.57 N = 413, Unreported = 31 95 While people identified themselves in several ways, inserting one's real Name or a recognizable Variation of it (e.g., initials or nicknames) proved the two most common communication categories. Handles was also popular (a constant representative name that superficially resembles nicknames in IRC or discussion groups on the Internet [1, 17]). Regardless of the differences between these categories, in all cases the names, variations or handles presented are not used to maintain pseudo-anonymity or complete anonymity as in IRC or MUDs. Rather, the identifier is something that the contact group uses to recognize a known individual. Another indicator of the importance of the Identification categories is that many users keep their name visible even when they add extra information to the display field (the black bars in the three identification categories, and the grey bars in the other 14 categories in Figure 4). People do this in spite of the limited display space: in a normally sized IM window about 30-50 display field characters are viewable. As well, the usual order of this information is a name followed by the extra information. A typical example is illustrated in Figure 5. This inclusion of identity is likely done as a courtesy behaviour so that others can distinguish between contacts without resorting to deciphering the e-mail address. Extra information is usually about self. Of the remaining 14 categories, the majority of them provide information about `self'. Elements in these `about self' categories dominate the frequency count (~85% of the non-name elements), with the top four categories providing information about Mood, Comments, Activities, and Location. These top categories all present information about the person at a moment in time: they annotate how they are feeling, what they are doing, or where they are. Similarly, the lesser used Presence category indicates if they are available, thus augmenting the preset status indicators, while Quotes and Fun are indirect indicators of state of mind and personality traits. Obviously, these people want to disclose an additional level of information revealing personal state and action to their community of friends, close contacts and collaborators. The regular association of this kind of information with one's name means that this information is truly about self; this is in sharp contrast to the personas found in chat systems, where people construct an artificial pseudonym identity through avatars or nicknames [1, 17]. People want to be able to broadcast information without involving conversation. Most of the remaining categories (about 14% of the non-name elements) contain communicative messages intended for the group. In particular, Messages, Notices, Questions and Directions are categories that either provide information thought to be of interest to the group or are posted to stimulate a response. Most of these are undirected e.g., `Does anyone know...'. Occasionally, a message may be specifically directed to an individual, yet this is done in a forum public to the community of contacts. Clearly, people are adapting the IM display field into a form of public broadcast communication facility; they are thus fulfilling one element of the broadcasting system described by Etzioni and Etzioni [5]. Since each user's contact list contains a different set of names, a responder (who may change their display name to respond to another's broadcast message) is likely not sending that response to the same community of people. This hampering of responses suggests that display names are less effective for creating the running dialogs common to IRC, MUDs and other public broadcast systems [6, 11, 17]. Asynchronous messaging. In MSN Messenger, the direct chat facility is session based. That is, direct chat cannot be used by one person to leave information for a currently `Offline' participant to read later. In contrast, the display name persists across sessions, meaning that asynchronous communication to offline participants is possible. For example, consider the message `SirMe - Happy Birthday, Angie!' that was found in the Messages category. By including this in his display name, SirMe is leaving an asynchronous message that Angie (and others) can see when they come on line. Younger users may change their display names more frequently than older users; sex does not make a difference. The demographics of our study suggest some demographic trends, which are described below. However, we caution that, due to the way we collected data, the demographic findings and how they relate to display name changes are at best tentative. First, the age ranges of our secondary contacts (as being 14 65 years old) were likely heavily influenced by the fact that these contacts were culled from the lists of only 12 primary participants (from 22 50 years old), most of whom were within the 21-30 age group, weighing the data with a similar age range. Second, our data is incomplete as display field change data for secondary contacts was not collected when their associated primary contact was off line. Third, ages of secondary participants were estimated, which affects the analysis we could do. In spite of this tentative flavour, we include our results as they suggest trends and future areas of study. We saw a fairly balanced number of males and females in our sample: 55% were male, 45% were female. The chi-square analysis for sex and display field change rates indicated that the two variables are independent, i.e., the sex of the participant does not suggest how often that person would change their display name. However, the chi-square analysis for participant age and display field changes suggests that they are related 1 . We subsequently examined the chi-square table data to compare the observed count with the expected count for each cell of age group crossed with rate. Discrepancies between the observed and expected counts indicate a pattern where younger users are more apt to frequently change their display name when compared to older users. This trend may reflect a "computer generation" gap where younger users would be more apt to change their display name. It could also reflect a culture gap, where younger users are using it for social reasons [8], while older users are using it for workplace purposes [13]. 1 While the chi-square test determined that the two variables are not independent, it does not provide details on how the two variables are related. If true values of age and average change rates were available instead of our estimated categories (a subject of a future study), other statistical analyses could be used to reveal this detail. Figure 5: A typical display field showing how people retain identity (Name), followed by other information (Activity) 96 5.2 Implications for practitioners People persistently use the display field not only to identify themselves to their community of contacts, but to reveal personal information about self and to broadcast messages. They do this in spite of the fact that the display field facility was designed for other purposes; the IM community co-opted this feature to fill their real desires and needs. The first major implication is that IM and similar facilities need first-class interface features that let people broadcast identifying information, information about self, and public messages. Because some people change this information fairly often, this information should be easy to create and alter, e.g., through direct manipulation. Some of these capabilities are only now being supplied by a few major IM vendors. For example, the new version of MSN Messenger (v. 7.0), released shortly after our study was performed), includes a dedicated space for adding and editing a personal message (Figure 6, top). A person can directly alter this text by clicking within it: no menus or dialog boxes have to be navigated or raised. Other people see this personal information as visually distinguished text, e.g., the italicized text within the contact list (Figure 6, bottom). The personal information message is also proprietary to the machine, similar to the display picture. Thus people can set unique location labels to various computers if desired, i.e. home or work. The Community Bar (CB) [12] is a multimedia groupware system being developed by collaborators in our laboratory. Elements of its design are partially influenced by our study results. People within an ad hoc group inhabit places, and all see the equivalent of a contact list within a place. For example, Figure 7 shows a place called `IM Paper' and three participants within it. To support `Identification', each participant is represented by a `Presence item', which shows running video (or photo) of them, their name. To support `Information about self', the Presence item also includes optional personal information (which may wrap across multiple lines) that persists across login sessions. A person can quickly modify this personal information by a popup raised whenever he or she moves their mouse over their item (Figure 7, right side). To support `Broadcast Messages', it also lets people broadcast and respond to public messages to all people in the group. This public broadcast is not available in MSN Messenger 7, For example, Figure 7 (bottom) illustrates a public text chat dialog that lets anyone in the group post messages; all see its contents and all can post responses. Not shown is a sticky note facility, where a person can post a persistent message to all. Finally, certain categories of information are supported. For example, `Directions' are satisfied by letting people post a `web item' (not illustrated): a thumbnail of a web page of interest that others can navigate to via a single button press. Another implication of our study is that people use many different categories of information especially when describing self which in turn suggests that people are trying to provide others with a rich picture of themselves. Yet most systems, even the current ones shown above, only let people set one attribute of themselves in their personal message space (although they may combine these in a text fragment). Perhaps future systems will let people construct an `avatar' of sorts with multiple attributes that distinguish these categories, so that (say) mood, location and activity are treated independently rather than compete for a small space. While these (and likely other) systems suggest point design solutions to our implications, what is important is that our study has placed this work on a solid intellectual footing. It provides details of what people have done, and has identified the categories of information that people supply. For example, we suspect that MSN Messenger's inclusion of a personal information field arose because its designers noticed that people were moulding the technology to suit their needs, and they wanted to "fix the interface" to better fulfill these needs. In contrast, our study helps designers understand why Figure 6: MSN Messenger v7.0 separates editing and display of names and personal messages. Figure 7: Snapshot of Community Bar displaying personal message space within presence item 97 appropriation occurred in the first place. Looking at the 17 categories of communication that are used in messages found in the display name space, we saw that most are personal, or about the self. In taking over this space, users are not `hacking' to make IM do totally different things. Rather, they are adding richness to their identity beyond their simple name label. They are expressing identity, and they own this expression by using a text field that only they can alter. We also saw that there is some use of the display field for public broadcasting of messages. This suggests that there is a problem with the way we compartmentalize systems: IM systems with no real notion of groups or public broadcast, versus IRC and similar systems where public broadcasts dominate. The real solution likely amalgamates both: a system that somehow supports both public and private discussions between ad hoc (and perhaps non-overlapping ) groups. To our knowledge, only very few systems (such as the Community Bar above [12]) are trying to tackle this fusion of genres. CONCLUSION Most studies of communication using instant messenger clients have been focused on the activities within the main chat window. In contrast, this study examined how contacts appropriate IM technology to publicly broadcast information by adding extra text to their display name. We exposed patterns of behaviour, where we saw that almost half of the contacts we monitored change their display names with varying frequencies. We established a set of seventeen communication categories for the different types of personal messages added to the display field. We saw that people did want to identify themselves (the Name, Variations and Handles category), and that these were true identities that contacts would recognize versus anonymous pseudonyms not known by others within the social group. We also saw that the most popular communications were those that added personal information about self: a person's psycho-physiological status, one's current activities, details of their location, and expressions of personal comments and opinions. We also saw that people occasionally used it to broadcast messages to the group, a facility not otherwise available in IM. These findings suggest that personal information and public broadcast of messages, currently supported through this creative appropriation by users, should be provided as a first class interface feature in IM design. This is just the first of a set of studies that could be done. Much has been discovered, although these results should be verified and refined further. For example, modest refinements of our study protocol would allow us to more precisely capture the frequency of changes within the display field and their distribution within the different communication categories. However, we suspect that the actual categories of communication will not change dramatically. We would also like to consider the author's intentions of a display name change along with the recipient's opinion. More importantly, we intend to study behaviour and communication patterns within systems that provide explicit support for personal information supply (such as MSN v7.0) and public broadcast (such as the Community Bar). ACKNOWLEDGMENTS Many thanks to all those who participated in this project and took precious time to make communication category evaluations for their many contacts' display name changes; their input was invaluable. Special thanks to Gregor McEwan and Michael Boyle for their advice both on intellectual content and on our recording software. This work was funded in part by the NSERC Nectar Research Networks program. REFERENCES [1] Bechar-Israeli, H. (1995). From &quot;Bonehead&quot; to &quot;cLoNehEAd&quot;: Nicknames, play and identity on internet relay chat. J. Computer-Mediated Communication, 1 (2). [2] Cadiz J. J., Venolia G. D., Jancke G. & Gupta A. Designing and Deploying an Information Awareness Interface. Proc ACM CSCW (2002). 314-323. [3] Cameron, A. F. & Webster, J. (2005). Unintended consequences of emerging communication technologies: Instant Messaging in the workplace. Computers in Human Behaviour, 21, 85-103. [4] Cutrell E. B., Czerwinski M. & Horvitz E. Effects of instant messaging interruptions on computing tasks. In Proc ACM CHI Extended Abstracts (2002). 99-100. [5] Etzioni, A. & Etzioni, O. (1999). Face-to-face and computer-mediated communities, a comparative analysis. The Information Society, 15, 241-248. [6] Fitzpatrick, G., Parsowith, S., Segall, B., & Kaplan, S. (1998). Tickertape: Awareness in a single line. Proc ACM CHI (1998). 281-282. [7] Greenberg, S. & Rounding, M. (2001). The Notification Collage: Posting Information to Public and Personal Displays. Proc ACM CHI (2001). 514-521. [8] Grinter, R. E. & Palen, L. (2002). Instant messaging in teen life. Proc ACM CSCW (2002). 21-30. [9] Hu, Y., Wood, J. F., Smith, V., & Westbrook, N. (2004). Friendships through IM: Examining the relationship between instant messaging and intimacy. J Computer-Mediated Communication, 10 (1). [10] Isaacs E., Walendowski A., Whittaker S., Schiano D. J. & Kamm C. (2002) The character, functions, and styles of instant messaging in the workplace, Proc ACM CSCW (2002). 11-20. [11] Jania, F. (2003). Broadcast Messaging: Messaging to the masses. Queue, 1 (8), 38-43. [12] McEwan, G. and Greenberg, S. Supporting Social Worlds with the Community Bar. To appear in Proceedings of ACM Group 2005, Sanibel Island, Florida, Nov 6-9. [13] Nardi, B. A., Whittaker, S., & Bradner, E. Interaction and outeraction: instant messaging in action. In Proc ACM CSCW (2000), 79-88. [14] Piepmeyer, A. I've been replaced by a screen name. The Daily Utah Chronicle, October 31, 2003. www.dailyutahchronicle.com/news/2003/10/31/Opinion/ Ive-Been.Replaced.By.A.Screen.Name-545565.shtml [15] Rennecker J. & Godwin L. Theorizing the Unintended Consequences of Instant Messaging for Worker Productivity, Sprouts: Working Papers on Information Environments, Systems and Organizations, 3 (Summer). Retrieved Dec 3, 2004 //weatherhead.cwru.edu/sprouts/ 2003/030307.pdf [16] Tang, J. C. & Begole, J. (2003). Beyond instant messaging. Queue, 1 (8), 28-37. [17] Turkle, S. (1997). Life on the Screen: Identity in the Age of the Internet. New York: Simon & Schuster Inc. 98

communication;Communication Catogories;Name Variation Handles;Identification Is Fundamental;Related IM Research;Distribution Frequency Of Various Catogories;Display Names;Instant messenger;awareness;MSN messager;Broadcast Information;Catorgorisation Of Display Names;Instant Messaging;display name

48

Building a Research Library for the History of the Web

This paper describes the building of a research library for studying the Web, especially research on how the structure and content of the Web change over time. The library is particularly aimed at supporting social scientists for whom the Web is both a fascinating social phenomenon and a mirror on society. The library is built on the collections of the Internet Archive, which has been preserving a crawl of the Web every two months since 1996. The technical challenges in organizing this data for research fall into two categories: high-performance computing to transfer and manage the very large amounts of data, and human-computer interfaces that empower research by non-computer specialists.

1. BACKGROUND 1.1 Research in the History of the Web The Web is one of the most interesting artifacts of our time. For social scientists, it is a subject of study both for itself and for the manner in which it illuminates contemporary social phenomena. Yet a researcher who wishes to study the Web is faced with major difficulties. An obvious problem is that the Web is huge. Any study of the Web as a whole must be prepared to analyze billions of pages and hundreds of terabytes of data. Furthermore, the Web changes continually. It is never possible to repeat a study on the actual Web with quite the same data. Any snapshot of the whole Web requires a crawl that will take several weeks to gather data. Because the size and boundaries of the Web are ill defined, basic parameters are hard to come by and it is almost impossible to generate random samples for statistical purposes. But the biggest problem that social scientists face in carrying out Web research is historical: the desire to track activities across time. The Web of today can be studied by direct Web crawling, or via tools such as the Google Web API1 , while Amazon has recently made its older Alexa corpus commercially available for the development of searching and related services2 . However, the only collection that can be used for more general research into the history of the Web is the Web collection of the Internet Archive . 1.2 The Internet Archive Everybody with an interest in the history of the Web must be grateful to Brewster Kahle for his foresight in preserving the content of the Web for future generations, through the not-for-profit Internet Archive and through Alexa Internet, Inc., which he also founded. 1 The Google Web Search API allows a client to submit a limited number of search requests, using the SOAP and WSDL standards. See: http://www.google.com/apis/. 2 See http://websearch.alexa.com/welcome.html for the Alexa corpus made available by Amazon. This site also has a description of the relationship between Alexa Internet and the Internet Archive. 3 The Internet Archive's Web site is http://www.archive.org/. Copyright is held by the author/owner(s). JCDL'06, June 11-15, 2006, Chapel Hill, North Carolina, USA. ACM 1-5959 3-354-9/06/0006. 95 Vannevar Bush Best Paper Candidate The Internet Archive began to collect and preserve the Web in 1996. With a few gaps in the early years, the collection has added a full crawl of the Web every two months since then. Most but not all of this data comes from the Alexa crawls. Statistics of the sizes of the separate crawls are complicated by the fact that a single crawl may contain several variants of the same URL, but in August 2005 the total volume of data was 544 Terabytes (TB). This is the size of the compressed data. As discussed below, the overall compression ratio is about 10:1, so that the total size of the collection is approximately 5 to 6 Petabytes uncompressed. Table 1 gives estimates of the size of the individual crawls for each year. Table 1. Estimates of crawl sizes (compressed) Year Web pages (TB per crawl) Metadata (TB per crawl) 1996 1 0.2 1997 2 0.4 1998 3 0.6 1999 4 0.8 2000 10 1.2 2001 15 2 2002 25 3 2003 30 4 2004 45 6 2005 60 10 We are working with the Internet Archive to build a research library based on this collection. In summer 2005, we began work on the system that is being used to transfer a major subset of the data to Cornell and to organize it for researchers, with a particular emphasis on supporting social science research. This paper describes the technical design, performance testing, and progress in implementation. The overall goals of the library and plans for its use in research are described in a separate paper [1]. 1.3 User Studies In building any library, the objective is to organize the collections and services so that they provide the greatest range of opportunities for users, both now and in the future. Inevitably the design is a trade-off between predictions of what users will find helpful and the practicalities of building and maintaining the library. This trade-off is particularly important for a library of the whole Web because of the computing challenges of managing very large amounts of data. Therefore, the design of the library began with interviews of potential users to identify how the collections might be organized to be most valuable to them. Two users studies were carried out, with sociologists and with computer science researchers. 1.3.1 Sociology In fall 2005, Cornell received support from the National Science Foundation's Next Generation Cybertools program for a project that combines sociology research with continuing development of the Web library 4 . In this project, the specific areas of research are diffusion of ideas, including polarization of opinions and the spread of urban legends. Conventionally, sociologists have studied such phenomena by analysis of small surveys with hand-coded data. One aim of the project is to develop a new methodology for such research built around very large-scale collections of Web data, with automated tools used to extract, encode and analyze the data. Social science researchers identified a number of specific studies that they would like to carry out using historical Web data. Many of the studies have the same general structure: (a) extract a subset of the Web for detailed analysis, (b) encode selected attributes of the pages in that subset, (c) repeat for the corresponding subsets at several different dates, (d) analyze the changes over time. The criteria by which a portion of the Web is chosen for analysis are extremely varied. Some desirable criteria are impossible with today's computing, e.g., they require understanding of the content of a page. However, simple criteria such as domain names provide a good starting point for many purposes, particularly when combined with focused Web crawling to refine the subsets for analysis. Once a subset has been extracted, social science researchers want to analyze the text, for which full text indexes are important. They also wish to analyze the structure of links between pages for the social relationship that they represent. Such research requires interdisciplinary efforts by computer scientists and social scientists. Some of the analysis tools already exist, e.g., using full text indexes of Web pages to trace the movement of individuals. Others tools are themselves subjects of computer science research in natural language processing and machine learning, e.g., to analyze the text of Web pages for sentiments, opinions, and other features of interest to social scientists. 1.3.2 Computer Science Ten computer scientists who carry out research on the Web contributed to the user studies. Their research areas include the structure and evolution of the Web, data mining, digital libraries, machine learning, and natural language processing. Most of their interest focuses on the textual content of Web pages and the structure of the Web as revealed by the graph of links between pages. Several of the researchers commented that they expend ninety percent of their effort gathering test data; even then they have difficulty in determining how robust the results are across time. A fundamental tool for such research is the Web graph of links between pages. Studies of the graph are very important in understanding the structure of the Web, and the graph is the basis of practical tools such as PageRank [3] or Hubs and Authorities [9]. Despite its importance, there have been few studies that have looked at changes in the Web graph over time. Many of the classical studies of the Web graph were based on early AltaVista crawls and have never been repeated. Algorithmic research needs graphs of at least one billion pages, preferably stored in the main memory of a single computer. For textual research on the Web there are two additional requirements. The first is snapshots that are repeated across time 4 Michael Macy (principal investigator), et al., &quot;Very Large Semi-Structured Datasets for Social Science Research&quot;. NSF grant SES-0537606. http://www.infosci.cornell.edu/SIN/cybertools/ 96 that can be used for burst analysis, and other time based research. The second is full text indexes of substantial numbers of Web pages. Focused Web crawling is of particular importance in digital libraries research. Part of the original motivation for developing this library was an interest in automatic selection of library materials from the Web [2, 10]. Using the actual Web for research in focused crawling is technically difficult and the results are often hard to interpret since no experiment can ever be repeated with exactly the same data. ARCHITECTURE The Internet Archive uses highly compressed file formats developed in conjunction with Alexa Internet. Compressed Web pages are packed together in large files using the ARC format [4]. The pages in each ARC file are in essentially random order, usually the sequence in which the Web crawler originally captured them. Every ARC file has an associated DAT file, which contains metadata for the pages including URL, IP address, crawl date and time, and hyperlinks from the page. The files are compressed with gzip. Ten years ago the decision was made that ARC files should be approximately 100 MB, which seemed big at the time, but this size is now too small for efficiency and will need to be increased. The sizes of the DAT files depend on the number of pages in the associated ARC files, but average about 15 MB. The compression ratios also vary widely. The ratio is more than 20:1 for text files but close to 1:1 for files that are already compressed efficiently, such as videos. The overall ratio for ARC files is about 10:1. 2.1.1 The Database The Cornell Web library uses a relational database to store metadata about the Web pages and a separate Page Store to store the actual pages. In addition, the unprocessed ARC and DAT files received from the Internet Archive are copied to a tape archive. In choosing a relational database, we considered but rejected two approaches that have been successful in related applications. The first option was to use a modern digital library repository with support for rich data models, such as XML and RDF, and search services that support semi-structured data, such as XQuery. Such capabilities are appealing, but we know of no repository system that can manage tens of billions of objects. The scale of the Web precludes such an approach. The second option was to follow the model of production services for the Web, such as Google [7] and Yahoo. They provide low cost processing and data storage by spreading their systems across very large numbers of small, commodity computers used as servers. This is the approach used by the Internet Archive to store its collections and for its very popular Wayback Machine 5 . We rejected this architecture for a research library for two principal reasons: (a) there are many algorithmic computations on large datasets where a single large computer is intrinsically more efficient than a distributed cluster of smaller machines, and (b) even when the research can be done effectively, clusters of computers are more difficult to program by researchers who are carrying out Web-scale research. As an example, each server at the Internet Archive has an index of the files stored on it, but there is only a very limited central index. The Wayback Machine allows a user to retrieve all the pages in the 5 The Wayback Machine is accessible at http://www.archive.org/. entire collection that have a given URL. It relies on a protocol in which an identifier is broadcast and each server responds with a list of matches. This is very efficient for this specific purpose, but it would be extremely difficult to extract the flexible subsets required by social science researchers with this organization of data. A relational database has many advantages for the Web library and one major disadvantage. Foremost among the advantages is scalability. Commercial relational database systems are highly optimized for storing, loading, indexing, extracting, backing-up, and restoring huge volumes of data. Usability is another important advantage. A relational schema provides a single image of the collection, expressed in a manner that is familiar to many researchers. The disadvantage is a loss of flexibility. The design and implementation of the database attempt to reconcile the expected uses that will be made of the library against scalability constraints, but it will be difficult to make major changes to the schema without rebuilding the entire database. The actual Web pages are stored in a separate Page Store. At the Internet Archive, if two Web pages are identical they are stored twice. With the new Page Store, duplicate pages are stored only once. Rather surprisingly, there is as yet very little data about how many pages remain unchanged between crawls, but we expect that elimination of duplicates will save significant online storage, especially with large audio and video files. The Page Store is implemented as a set of compressed files, one file for each page received in the ARC files. Since many pages on the Web do not change between crawls, the Preload subsystem checks for content duplicates using an MD5 check sum of the content. Thus, a copy of the content is stored only once however many pages have that content. In order to guarantee fast access to the stored content, each page's content is compressed individually. The architecture of the Page Store allows decisions to be made about which pages to store online at any given time. For example, the library might decide not to store large audio and video files online. While all metadata will be online at all times, an individual Web page could be accessed from the online Page Store, the off-line tape archive, or over the Internet from the Internet Archive. 2.2 Equipment The library is housed at the Cornell Theory Center, which is the university's high-performance computing center. The choice of equipment and the use of a relational database were closely related decisions. The Theory Center has expertise in distributed cluster computing, but because of the very high data rates, a symmetric multi-processor configuration was chosen instead. Figure 1 shows part of the configuration of the central computer. Figure 1. Configuration of the main computer system 97 The system is shared with another data-intensive program of research, the analysis of data from the Arecibo radio telescope in Puerto Rico. Each group has use of a dedicated Unisys ES7000/430 server, with 16 Itanium2 processors running at 1.5 Gigahertz. The memory can be shared between the servers, but in practice each project has sole use of 32 GB. Each server has a RAID disk subsystem attached via a dual-ported fiber-channel. The operating system is Microsoft Windows Server 2003. For the Web library the disk subsystem provides an initial 45 TB of disk storage. We plan to extend the capacity to 240 TB by 2007. There are no technical barriers to adding additional fiber channels and disk capacity. In the longer-term, disk prices are falling faster than the growth in the size of Web crawls, which gives confidence that the library will be able to keep up with the growth of the Web. By using a symmetric multi-processor configuration with a high performance disk subsystem, we are able to balance processing and disk access requirements. Since the data sets are local to the system on which the database is located, the system can perform bulk-loading tasks without incurring any networking penalties. The large real memory is an added attraction of this configuration. It allows researchers to carry out substantial computation entirely in memory. For instance, it is possible to process a Web graph of one billion pages within memory. 2.3 The Human Interface The design of the human interface is perhaps the most challenging aspect of developing the library. The social science research groups that we are working with have the technical skills to write scripts and simple programs. Many are experts in statistical calculations. But they should not be expected to write large or complex computer programs. The current design supports three categories of users. The Basic Access Service provides a Web Services API that allows a client to access pages in the collection by any metadata that is indexed in the database, e.g., by URL and date. The Retro Browser, which is described below, uses this API to allow a user to browse the collection as it was at a certain date. The Subset Extraction Service supports users who wish to download sets of partially analyzed data to their own computers for further analysis. A Web form is provided to define a subset of the data (e.g., by date, URL, domain, etc.), extract subsets of the collection, and store them as virtual views in the database. Sets of analysis tools, many of which are already under development, can be applied to the subset and the results downloaded to a client computer. Technically advanced users can be authorized to run their own programs on the central computer. To support the Basic Access Service and the Subset Extraction Service, we provide a dedicated Web server, which is housed next to the main system. SCALABILITY EXPERIMENTS Although the library has been generously funded by the National Science Foundation, we do not yet have sufficient capacity to download and mount online the entire Web collection of the Internet Archive. This is the long term goal, but during the initial phase, care has been taken to balance the several parts of the system: online storage, network bandwidth, processing of the incoming data, database, performance, and the need to archive, back-up, and restore the data. In spring 2005, several undergraduate and masters students carried out independent projects to estimate sizes and processing requirements 6 . To test database performance before large amounts of actual data were available, we used the R-MAT algorithm to generate a synthetic graph with properties similar to the Web graph [5]. This test graph has one billion nodes with more than seven billion links, and domain names generated according to their distribution on the real Web [11]. Based on these benchmarks, the decision was made to install a 100 Mb/sec network connection to the Internet Archive and to load data at a sustained rate of 250 GB/day, beginning January 2006. This rate will enable the library to acquire and mount online by the end of 2007 a complete crawl of the Web for each year since 1996. This phase will require approximately 240 TB of disk storage. Note that the disk requirement differs from the estimates of raw data shown in Table 1. The database with its indexes is less highly compressed than the raw data, but savings are made in the storage of duplicate data, both in the database and the Page Store. During fall 2005, first generation software was written for (a) the data flow system that brings data to the library, and (b) the user API and tool sets. They are described in the next two sections. DATA FLOW Figure 2 shows the flow of data into the library. When ARC and DAT files are received from the Internet Archive, the first step is to store them in the tape archive. The Preload system then unpacks the raw data, extracts metadata, and prepares batch files for loading into the database and Page Store. Figure 2. Flow of data into the library Figure 2 does not show the data tracking system. This is a major subsystem that manages the data transfers, monitors all errors, and tracks the tens of millions of files within the library. 4.1 Networking Internet2 is used to transfer data from the Internet Archive in San Francisco, California to Cornell University in Ithaca, New York. For this purpose, a 100Mbit/sec link has been established from the Internet Archive to Internet2. Both Cornell and the Internet Archive have internal networks with 1 Gbit/sec or greater performance. In the future, the National LambdaRail and the TeraGrid are intriguing possibilities. These new networks have the capacity to go 6 These student reports are available at http://www.infosci.cornell.edu/SIN/WebLib/papers.html. User tools Preload system Database Page store Tape archive Internet 2 98 beyond bulk data transfer and support genuine distributed processing between the Web library and the Internet Archive. For example, if large audio and video files are not stored online, an application could use the TeraGrid to retrieve individual large files from the Internet Archive on demand. At the end of December 2005, a series of experiments were run to measure the sustained throughput of multi-threaded FTP transfers over Internet2, using Transport Layer Security. These measurements showed transfer rates of 280 GB per day before system tuning, or rather better than 30 percent of the theoretical maximum throughput of the link to the Internet Archive. This is sufficient for the planned rate of 250 GB per day. If greater bandwidth proves necessary, the link from the Internet Archive to Internet2 can be upgraded to 500Mbps inexpensively, while the Cornell Theory Center will soon have enormous bandwidth available via the TeraGrid. 4.2 Preload Subsystem The Preload subsystem takes incoming ARC and DAT files, uncompresses them, parses them to extract metadata, and generates two types of output files: metadata for loading into the database and the actual content of the Web pages to be stored in the Page Store. Metadata for loading into the database is output in the form of 40GB text files, a separate file for every database table. To satisfy the speed and flexibility requirements, the Preload system is designed to run as a set of independent single-thread processes, avoiding all inter-process communication and locking over input or output data. Likewise, each process writes its own output files. This design allows for easy configuration of the system to run a required number of processes on a given number of processors, on one or more machines. To determine each process's input, input files are partitioned by the first k bits of the MD5 hash sum of the filename, where 2k is the total number of processes in the system. The design of the subsystem does not require the corresponding ARC and DAT files to be processed together. A series of experiments were run to test the performance of the Preload system, using the metadata from the synthetic Web graph. The experiments used 1, 2, 4 and 8 processors, with the data partitioned into 16 parts, according to the first 4 bits of the hash sum. Separate experiments were made for ARC and DAT files. Figure 3 shows results for the ARC files. The x-axis shows the number of CPUs used, and the y-axis shows the throughput in KB/sec. The white bar shows throughput per processor and the shaded bar shows total throughput. Adding more processors slightly decreases the throughput per processor due to contention for random disk accesses. The total throughput increases steadily up to four processors. After that, disk contention becomes too high, and the throughput actually declines. The results for DAT files are similar, with the total throughput flattening after four processors. From these experiments, we conclude that four processors are optimal. The corresponding throughputs are 73 MB/sec (about 6 TB/day) for ARC files, and 12 MB/sec (about 1 TB/day) for DAT files. When metadata from the DAT files is uncompressed its size increases by a factor of about 11:1. Fortunately, much of this data is duplicated. For example, a given URL may occur many times in a crawl and be replicated in many crawls. Therefore duplicate elimination has been a major consideration in refining the database design. Figure 3. Performance of Preload system (ARC files) Note that during the above experiments no other processes were running on the system. The overall performance will be lower when the Preload system is run in production mode at the same time as other subsystems. Also, during the preliminary phase, only text and HTML files were fully analyzed to extract links and anchor text. Processing of other file formats will be added in the future. Some of these formats will be computationally intensive, e.g., PDF files. 4.3 Database Design The relational database uses Microsoft SQL Server 2000. Three important goals when designing the database schema and deciding how to load the data to the database were: (a) minimize the storage requirements, (b) maximize the load throughput, and (c) support efficient logging, backup, and restore. Conceptually, for each crawl, the database stores metadata about each page (e.g., information about the content of the page, URL of the page) and about the links between them (including anchor text and text surrounding the anchor text). However, to avoid storing redundant data, the schema is denormalized. The denormalized schema is shown below in Figure 4. Information about URL and domain names is stored in a look-up table, since the same URLs and domain names appear many times in a single crawl and across crawls. For similar reasons, anchor text, text surrounding the anchor text, and information about page content are stored in the look-up tables Dictionary and Page Content respectively, as shown in the schema in Figure 4. To make the loading of the data faster, separate tables for each of Page, Page Content and Link are created for each crawl while the other tables (e.g., the URL table) are shared among crawls. 80 70 60 50 40 30 20 10 Throughput (MB/s) 1 2 4 8 Number of CPUs 99 Figure 4. Database design: the de-normalized schema The Preload subsystem outputs separate files conforming to the schema described above and these files are bulk loaded into the database. There are many parameters that affect the bulk load performance. These parameters include: batch size, file size, degree of parallelism, and interaction with the logging, backup and recovery system. The synthetic Web data was used to understand how these parameters affect the loading performance and to tune them. The first sets of experiments were used to determine the optimal file size for bulk loading and the number of CPUs used. In these experiments, default, recovery and backup mechanisms were used [6]. The results of the first set of experiments indicate that it is optimal to load each file as a single batch; a file size of 40GB and 4 CPUs gave the best performance, and around 800GB could be loaded in one day. However, the experiments to determine the file size and degree of parallelism showed significant variability. Using the default logging provided by MS SQL Server 2000, checkpointing and nightly backups were consuming enormous resources. They interfered with the bulk loading process and are a probable cause of the variance seen in the performance benchmarks. Two observations were made to overcome the performance penalty. First, data is append-only while being bulk loaded and is read-only after the bulk load is complete; logging, recovery and backup mechanisms can be customized to increase the performance and decrease the variance in loading times. Second, tables in the schema can reside on different disks and thus can be written in parallel. Following these two observations, in the current design each table can be put onto a separate disk as shown in Figure 5. Moreover, Page, Page Content and Links information for each crawl are put into separate files. This partitioning according to crawls is easy in MS SQL Server as separate tables for each of Page, Page Content and Link are created for each crawl. The database load subsystem is divided into two programs: a high-level program that organizes the processes and a low level program that runs separate loads in parallel. The workflow for loading each table in each crawl consists of five major steps: (a) Get the files produced by the Preload subsystem for the current table in the current crawl and write the relevant log information to an administrative database; commit the transaction writing the log information. (b) Write files of the table to the disk corresponding to the current table via the low level program; files corresponding to different tables can be written in parallel. (c) Create the necessary indexes. (d) Back-up the newly written data. (e) Write to the log the relevant information to indicate that processing of the files for the current table in the current crawl is complete and commit the transaction writing the log information. In MS SQL Server 2000, backups and index creation are all atomic. Figure 5. Database design: organization of file system This new design is being implemented and tested, as of January 2006. First indications are that the performance will comfortably meet the required performance goals. Extensive benchmarking is required to tune many parameters, such as batch size, file size, degree of parallelism, and the index management. 4.4 Archiving Archiving and back-up are expensive operations with complex trade-offs. Without care, the networking bandwidth and disk throughput used in logging, back-up, and writing to the tape library could have a major impact on the system throughput. As described above, the database design allows the database files to be backed up incrementally. This provides two options for restoring the database, by reprocessing the raw data or from the back-up. The Page Store is not backed-up. If parts of it were ever corrupted, they would have to be restored by reprocessing the raw data. A current design project is to reorganize the Page Store to permit efficient restoration. The library uses a robotic tape library with LTO3 tape drives. This is shared with other systems at the center. All unprocessed ARC and DAT files are copied to the tape library to be stored indefinitely. This preserves another copy of the Internet Archive's data for the long-term. This data is unique and could never be replaced. The industry standard life of these tapes is thirty years but our expectation is that six years is a more probable time before the tape library is replaced and all the data will have to be copied onto fresh media. SUPPORT FOR THE USER Figure 6 shows the architecture of the interface that the library offers to users. This is a three-tier architecture. The data tier consists of the relational database of metadata and the Page Store; the middleware tier provides services to access the data, tools to analyze it, and a choice of Web Services APIs; clients interact with the middleware tools either through the APIs, or directly. 5.1 Clients The user tools system was designed to be extensible and scalable. Specifically, it supports two categories of users: (a) users who analyze the data remotely from their own computers, perhaps at another institution or even in another country, and (b) Page Page Content Link - Destination URL URL Path Dictionary Domain Page Link Page Content Custom Log Everything Else 100 computationally intensive users, who may wish to run very heavy analyses of the data using the Cornell Theory Center computing environment. Corresponding to these two categories of users the architecture supports two types of clients: Web services clients and clients that execute their own programs on the library's servers. The first category requires little technical sophistication from the user, while the second category trades complexity against the flexibility of being able to write custom programs and to use the full processing power available. Figure 6. Architecture of the interfaces provided for users of the library Web services clients are intended to be used remotely, with moderate demands on their access to the data. Examples of these clients include running queries using a full text index, or fetching specific pages using existing indexes. Web service clients may also be used to start, control, and retrieve results from experiments run by high-performance clients. Web services are implemented by using Microsoft's ATL server libraries. They run on a dedicated Web server. The clients themselves can be implemented in any language or environment that supports Web services standards. Users of these clients do not need to know how the data is stored. They are provided with forms that are automatically converted to SQL commands by the middleware. The high-performance clients will for the most part run within the Cornell Theory Center. They will usually have a high bandwidth connection to the database. Clients of this form may carry out research that need lots of computation power, e.g., experiments that process very large subsets or analyze how the structure of the Web changes over time. These clients are implemented by linking against dynamic link libraries (DLLs) provided by the application server tier. 5.2 Access to the Database The application server tier accesses the database using Microsoft database tools. Two main areas of functionality have been implemented in this tier: Basic Access Services (BAS), and the Subset Extraction Services (SES). Each consists of two parts: a set of services, implemented as a series of dynamic link libraries (DLLs) written in C++, and a Web Services API. Basic Access Services are for clients that interact directly with the database. They allow a client to fetch pages given a combination of URL and date of crawl. They also allow a client to check within which crawls a given page is available. For example, a focused Web crawler can retrieve pages from a specified crawl, using a simple client script that interfaces to the BAS Web services API. Subset Extraction Services allow a client to select a part of the data as a subset. Once created, this subset is stored in the database as a view. Such subsets are useful for running experiments over a smaller, perhaps random, sample of the Web, as well as selecting relevant pages for a particular experiments, such as those from a given domain. For example, a researcher studying government Web sites might extract textual pages from the .gov domain for a selected range of dates. 5.2.1 Users Beyond Cornell This digital library is intended for the use of all academic researchers, not only those based at Cornell. Technically, this is straightforward. The library is connected to the Internet, including Internet2, and will soon be available via the TeraGrid. We are currently developing a code of use policies for researchers. Mining this data has potential for abuses of privacy. Cornell researchers need to follow the university's procedures for such research and we need to find convenient ways to extend the code of practice to all users of the library. 5.3 User Tools 5.3.1 The Retro Browser The Retro Browser is an example of a Web services client that is already implemented and running in a test environment [12]. To a user, it appears to be a regular Web browser, except that it browses an historical snapshot of the Web. The Retro Browser is designed with the non-technical user in mind. The design assumes that the user may not be technically proficient and should not be expected to install new software or run special scripts. After the user has made an initial choice of a date in Web history, the Retro Bowser behaves like any other Web browser. The user uses a standard browser to carry out all the standard Web tasks, such as download an applet, run a script, submit a forms, etc. The only difference is that every URL is resolved to the record in the Web library for the specified date. The major component of the Retro Browser is a Web server configured to be used as a proxy server. To obtain the data from the database, the proxy server utilizes the Basic Access Web Service API. The Retro Browser client interacts with the Retro Browser proxy in a standard HTTP client-server fashion. To fetch the appropriate page from the database requires a URL and the date of the crawl, which is represented by a crawl ID. The proxy server expects a session cookie specifying a crawl ID with every request. If such a cookie is not found with the request, the user is asked to specify the crawl ID. Further requests may or may not contain a cookie since cookies are tied to a domain. However, the Retro Browser proxy ensures that the cookie is replicated for all domains using a series of redirects. In this manner, the architecture ensures that the user is asked to specify the crawl date for the first request only. 5.3.2 Analysis of the Web Graph A set of analysis tools is under development that will provide more complex access and analysis functions. These tools are part of the application server tier. They are applied to subsets of the data and accessed either directly or through the Subset Extraction Services API. Data Tier Application Server Tier Client Tier Page Store Metadata Basic Access Services Subset Extraction Services BAS Web Service SES Web Service Clients High Performance Clients Retro-Browser 101 One group of tools operates on the Web graph of a subset. Hyperlinks from each Web page are stored in the database. Representation of the graph is by its adjacency matrix using a compressed sparse row representation. Preliminary software has been written to read all the links from a given subset of the data and construct the adjacency matrix. The matrix is then stored in the file system in a compressed form, which allows performing the basic operations, such as matrix addition and multiplication. The Cuthill-McKee algorithm is used to reorder the nodes to create dense blocks within the matrix to increase the compression ratio and allow in-memory processing [8]. 5.3.3 Full Text Indexes Full text indexes are a vital tool for many researchers. For instance a social science researcher may wish to track the Web pages that refer to a named individual or may identify trends by burst analysis of terms used on the Web. Our initial approach is to provide an indexing service for data subsets, using the Nutch search engine. It is straightforward to extract a subset, which is represented by a database view, and create a full text index of all textual pages in the subset. The only problem is the processing necessary to index a very large subset. We are in discussions with Cutting, the principal developer of the Lucene and Nutch search engines 7 . He has been working with the Internet Archive to build indexes of very large collections of Web pages in ARC format. For this purpose, they are developing a modified version of Nutch, known as Nutch WAX (Web Archive eXtensions). Rather than duplicate this effort, we are exploring the possibility of providing access to these indexes through the Basic Access Service. ACKNOWLEDGEMENTS We wish to thank Brewster Kahle, Tracey Jaquith, John Berry and their colleagues at the Internet Archive for their support of this work. The following Cornell students have contributed to the development of the library described in this paper: Mayank Gandhi, Nicholas Gerner, Min-Daou Gu, Wei Guo, Parul Jain, Karthik Jeyabalan, Jerrin Kallukalam, Serena Kohli, Ari Rabkin, Patrick Colin Reilly, Lipi Sanghi, Shantanu Shah, Dmitriy Shtokman, Chris Sosa, Samuel Benzaquen Stern, Jimmy Yanbo Sun, Harsh Tiwari, Nurwati Widodo, Yu Richard Wang. This work has been funded in part by the National Science Foundation, grants CNS-0403340, DUE-0127308, and SES-0537606 , with equipment support from Unisys and by an E-Science grant and a gift from Microsoft Corporation. 7 The Lucene search engine is described at http://lucene.apache.org/. Nutch is described at http://lucene.apache.org/nutch/. REFERENCES [1] Arms, W., Aya, S., Dmitriev, P., Kot, B., Mitchell, R., Walle, L., A Research Library for the Web based on the Historical Collections of the Internet Archive. D-Lib Magazine. February 2006. http://www.dlib.org/dlib/february06/arms/02arms.html [2] Bergmark, D., Collection synthesis. ACM/IEEE-CS Joint Conference on Digital Libraries, 2002. [3] Brin, S., and Page. L., The anatomy of a large-scale hypertextual Web search engine. Seventh International World Wide Web Conference. Brisbane, Australia, 1998. [4] Burner, M., and Kahle, B., Internet Archive ARC File Format, 1996. http://archive.org/web/researcher/ArcFileFormat.php [5] Chakrabarti, D., Zhan, Y., and Faloutsos, C., R-MAT: recursive model for graph mining. SIAM International Conference on Data Mining, 2004. [6] Gerner, N., Sosa, C., Fall 2005 Semester Report for Web Lab Database Load Group. M.Eng. report, Computer Science Department, Cornell University, 2005. http://www.infosci.cornell.edu/SIN/WebLib/papers/Gerner200 5.doc. [7] Ghemawat, S., Gobioff, H. and Leung, S., The Google File System. 19th ACM Symposium on Operating Systems Principles, October 2003. [8] Jeyabalan, K., Kallukalam, J., Representation of Web Graph for in Memory Computation. M.Eng. report, Computer Science Department, Cornell University, 2005. http://www.infosci.cornell.edu/SIN/WebLib/papers/Jeyabalan Kallukalam2005.doc. [9] J. Kleinberg. Authoritative sources in a hyperlinked environment. Ninth ACM-SIAM Symposium on Discrete Algorithms, 1998. [10] Mitchell, S., Mooney, M., Mason, J., Paynter, G., Ruscheinski, J., Kedzierski, A., Humphreys, K., iVia Open Source Virtual Library System. D-Lib Magazine, 9 (1), January 2003. http://www.dlib.org/dlib/january03/mitchell/01mitchell.html [11] Shah, S., Generating a web graph. M.Eng. report, Computer Science Department, Cornell University, 2005. http://www.infosci.cornell.edu/SIN/WebLib/papers/Shah2005a .doc. [12] Shah, S., Retro Browser. M.Eng. report, Computer Science Department, Cornell University, 2005. http://www.infosci.cornell.edu/SIN/WebLib/papers/Shah2005b .pdf. 102

User Interface;Dataflow;Internet Archive. 1. BACKGROUND 1.1 Research in the History of the Web The Web is one of the most interesting artifacts of our time. For social scientists;history of the Web;basic parameters are hard to come by and it is almost impossible to generate random samples for statistical purposes. But the biggest problem that social scientists face in carrying out Web research is historical;or via tools such as the Google Web API;Database Management;it is a subject of study both for itself and for the manner in which it illuminates contemporary social phenomena. Yet a researcher who wishes to study the Web is faced with major difficulties. An obvious problem is that the Web is huge. Any study of the Web as a whole must be prepared to analyze billions of pages and hundreds of terabytes of data. Furthermore;Storage;Flexible Preload System;Internet Archive;digital libraries;Scalability;the Web changes continually. It is never possible to repeat a study on the actual Web with quite the same data. Any snapshot of the whole Web requires a crawl that will take several weeks to gather data. Because the size and boundaries of the Web are ill defined;Database Access;User Support;computational social science;Full Text Indexes;the desire to track activities across time. The Web of today can be studied by direct Web crawling

49

Building a Sense of History: Narratives and Pathways of Women Computing Educators

This working group laid the groundwork for the collection and analysis of oral histories of women computing educators. This endeavor will eventually create a body of narratives to serve as role models to attract students, in particular women, to computing; it will also serve to preserve the history of the female pioneers in computing education. Pre-conference work included administration of a survey to assess topical interest. The working group produced aids for conducting interviews, including an opening script, an outline of topics to be covered, guidelines for conducting interviews, and a set of probing questions to ensure consistency in the interviews. The group explored issues such as copyright and archival that confront the large-scale implementation of the project and suggested extensions to this research. This report includes an annotated bibliography of resources. The next steps will include training colleagues in how to conduct interviews and establishing guidelines for archival and use of the interviews.

INTRODUCTION During the SIGCSE Technical Symposium held in Reno, NV in February 2003, a significant number of events focused on under-representation of women in the computing curriculum and as computing educators. Eric Roberts' keynote talk, "Expanding the Audience for Computer Science" [21], was a moving discussion of inclusiveness and a lament about the consequences of non-inclusion . At the Friday luncheon, Jane Margolis and Allan Fisher discussed results from their groundbreaking work, Unlocking the Clubhouse [15]. Several private discussions begun at the conference and continuing for some time afterward led to a November 2004, proposal for this Working Group. In this report, we document the results from a Working Group of computer science educators at the 2005 ITiCSE conference held in Lisbon, Portugal. We were drawn together by our shared concern about women's under-representation among computing educators. We wished to honor women who had persevered in the early days of this field and to make their stories available as a resource for those following after. ITiCSE working groups are convened for the purpose of intensive collaborative work on a topic of common interest among the participants, prior to and during the conference, generally 174 completing the Working Group's task by conference end. In contrast, this group was convened to lay the groundwork for a project that we hope will continue for some time to come. The Working Group leaders spent the preceding 18 months formulating the charter for the Working Group: to collect oral histories from pioneering women in computing education. The goal of the Working Group meetings at ITiCSE in Lisbon was to formulate a plan that could bring the charter to fruition. We envision that the result of this project will be a large oral history collection of broad scope with potential value to researchers and others engaged in a variety of different projects. Because this project could result in a large quantity of data, it cannot be stored by one person in her private file space. The data must be maintained and administered by an agency or institution prepared for such a task. We write this report for multiple audiences: 1. Those who want a concise account of what the group accomplished in Lisbon. 2. Those whose work will proceed from and build on that done in Lisbon for the oral history project. 3. Those who want insights into the evolution and dynamic of a working group. 4. Those seeking historical information about the beginnings of the oral history project. PREPARING FOR THE PROJECT This section outlines key steps and insights developed prior to the ITiCSE conference. 2.1 Building a Background The initial vision of this project was to collect stories, or narratives, from successful computing educators, in particular from women. We were particularly interested in the various paths these individuals had followed through their careers. We considered resources related to women and computing education, in particular factors that seemed to lead to success in the field. We found that the area of inquiry known as oral history includes techniques conducive to the type of data-gathering we visualized. Key resources for our project include a set of oral history evaluation guidelines [20], an Oral History Association [17], and a tutorial for conducting oral history from the Oral History Institute at Baylor University [2]. We discuss these resources further in Section 4 and in the annotated bibliography. 2.2 Project Vision While it was clear from its inception that the primary focus for this Working Group would be women computing educators, we recognized that this is potentially the first phase of a longer-term project. The techniques developed in this first phase could be used in later phases, eventually developing into a broader project covering the history of computing education as a profession. This longer-term project should lead to a collection of oral histories from both men and women in the field as well as other artifacts. While we expect that future investigators will analyze the materials collected during each phase of the project, analysis of the materials is not the driving factor at this time. We feel it is vital to create an accessible repository of the data to support future investigations. 2.3 Survey to Gather Ideas In order to gather ideas about the project from a broad community of individuals, we designed a survey to request ideas from colleagues. In recognition of the longer-range potential of this work, the survey solicited information for the full field of computing education, rather than restricting responses to the narrower focus of women computing educators. The full survey can be found in Appendix D. We targeted two on-line communities with vested interest in the topic of this Working Group: Systers [23] and SIGCSE.members [22]. By the end of the conference, the survey had resulted in responses from 24 different individuals. Respondents offered ideas for questions, thoughts about how to recruit additional subjects for the interviews, and advice for how to proceed. The respondents suggested 60 educators as potential interviewees, of whom 34 are women. Several respondents also indicated interest in becoming involved with the project as planners, interviewers, or subjects. FORMULATING THE PROJECT GOALS At the heart of this project is the recognition that women are under-represented in the computing field [12]. In particular, Working Group members had a variety of ideas for how to address the lack of women in computing education. Among the ideas: providing role models capturing stories of women of different ages to provide a history of women in computing education exploring the history of early women computing educators to learn about and honor the stories of these women, who often faced difficult circumstances recording difficulties that women educators encountered during their careers, and in some cases overcame, as a source of inspiration and support Considering the challenges faced by women in early computing education also brought up questions about how they managed those challenges: What internal reserves and external resources did they draw on? How did they sustain their confidence in their own capabilities, often as the only woman in what was at times a hostile environment? This led the group to consider self-efficacy beliefs, which Bandura ([2], p. 391) defines as "people's judgments of their capabilities to organize and execute courses of action required to attain designated types of performances". A person's self-efficacy beliefs can play a significant role in her capacity to manage difficulties: if she believes she can actualize her intentions, obstacles presented by the environment impose less drag. 3.2 Focus for the Short Term This section addresses a number of key points that the group must consider for the near term in order to coordinate work by a distributed set of volunteers. 3.2.1 Protocol for Collecting Stories A key task of the Working Group was to establish a protocol to be followed by all volunteers on this project. The resources related to 175 collecting oral histories provided a rich source of information for defining the protocol to be used over the life of the project. A clear protocol will ensure consistency in the quality and general content of the interviews, especially for interviewers with little experience in collecting oral histories. We discuss the protocol further in Section 5. 3.2.2 Identifying Potential Subjects The primary focus for this phase of the project is women computing educators who are late in their careers. The project will seek an international sample in order to ensure a more complete picture; during the conference in Lisbon, many non-U.S. educators showed interest in the project. It is urgent to capture narratives from older and more senior educators while these pioneers are still able to participate in the interview process. The Working Group has created a list of potential subjects, including ideas drawn from the results of the survey described in Section 2.3. 3.2.3 Legal Issues: Consent, Access, Ownership A paramount concern for the project is the set of legal issues associated with this form of academic inquiry. Pressing concerns that must be resolved include the following: obtaining permissions to ensure that the materials are openly accessible for use in future studies and analyses; determining who will have access to the materials collected in this project; determining ownership of the materials; and designing appropriate copyright and permission forms. 3.2.4 Storage and Transcription of the Interviews When an interview is complete, the recording(s), notes, and any supplemental materials must be prepared for later use and analysis. As a temporary measure, a copy on CD will provide secure storage of the materials until incorporated into a formal repository. To make the interviews more accessible to future users such as researchers and historians, common practice is to develop a transcript. Besides being easier to scan quickly, a good quality transcript makes it easier to create notes and cross-reference parts of the interview. There are two main approaches to creating transcripts for an interview: listening to the taped interview and capturing the dialog manually, a tedious and exacting process, or using voice recognition software to automatically create a transcript in a computer file, a process that tends to be very error-prone. Once the transcription is complete, careful editing can make the work clearer and more accessible. However, editing requires a deft touch, using pre-determined guidelines specific to the project. The editing process may consider issues such as how and whether to correct errors (for example, should transcriber errors be fixed during editing? Should errors of fact acknowledged by the interview subject be set right?) and whether to clear out irrelevant information (for example, deleting [presumably meaningless] false starts to make the transcript's meaning clearer). Editing may also introduce paragraphs and subheadings to help highlight topics and make it easier for a future reader to traverse the transcribed interview. 3.3 Archival of the Project Materials For security and availability of the collected materials, it will be vital to identify a means for the long-term storage of the interviews and other artifacts. In addition, the repository can be used to maintain a bibliography of results related to the overall project. While it is premature to determine where this work might eventually reside, an excellent example of the appropriate style of storage and availability is the repository related to the history of computing maintained by the Charles Babbage Institute [4]. BACKGROUND To set the context for the Working Group's project, this section considers four background areas: the area of inquiry known as oral history, resources related to the history of computing, resources on the history of women in computing, and work related to the history of computing education. 4.1 What is Oral History? Oral history is a method of inquiry with a rich tradition and specific guidelines. While folklore and storytelling are examples of oral history through the ages, modern techniques have improved the reliability of the data one can gather in an oral history project. The Wikipedia article on oral history [25] explains: "Oral history is an account of something passed down by word of mouth from one generation to another. Oral history is considered by some historians to be an unreliable source for the study of history. However, oral history is a valid means for preserving and transmitting history." The Oral History Association [17] has published guidelines that address several aspects of conducting oral history, including responsibility to subjects, the public, and the profession; interview content and conduct; storage and preservation of media and interviews; and an excellent bibliography. The Oral History Primer from the Department of History at California State University, Long Beach [6] offers an overview of many of the aspects of conducting an oral history project, such as how to design the study, how to conduct and process the interview, and how to use the completed interview. This resource offers a sample outline, a sample transcript, and a sample agreement form. As the Working Group prepared for the meetings in Lisbon, a number of oral history projects helped us formulate ideas about how the materials from such a project can be planned for and archived. For example, the London Voices project [14] gathered oral histories from a variety of individuals and has made these stories available via a Museum of London website. The Oral History Directory from Alexander Street Press [18] is an ambitious effort to index the major oral history collections in English throughout the world. During our working group presentation at the ITiCSE conference, we learned of another project in Brazil, O Museu da Pessoa (Museum of the Person) [16], which can provide additional ideas. The annotated bibliography in Appendix E lists relevant projects we have discovered thus far. 176 One of the Working Group members in Lisbon, William Aspray, is a historian of computing who has conducted over 200 interviews eliciting oral histories. The materials related to these interviews are in the repositories of the Charles Babbage Institute for History of Computing [4], which we discuss further in the next section. Aspray's participation in the Working Group provided key inputs and examples as the group developed the guidelines and planning reflected in this report. 4.2 History of Computing Resources Interest in the history of computing is broad-based. A variety of historical projects focus on areas as diverse as artifacts (e.g., punched cards, old computers), the timeline of events and developments in computing, and the people involved in driving the field forward. This section highlights a few computing history projects that seem particularly relevant in the context of this Working Group's project. The Charles Babbage Institute (CBI) [4] was started in 1978 and by 1989 became an historical archives and research center of the University of Minnesota. CBI preserves relevant historical documentation in a variety of media, conducts and fosters research in history and archival methods, and sponsors scholarly meetings and publications related to the history of computing. The resources on this site include a set of more than 300 oral histories, of which no more than 5% appear to be from women. The IEEE Annals of the History of Computing [8], a quarterly publication started in 1979, features scholarly articles by leading computer scientists and historians, as well as firsthand accounts by computer pioneers. The Annals is intended to serve as a focal point for collecting and disseminating information on historical projects and organizations, oral history activities, and international conferences. The IFIP Working Group 9.7 on the History of Computing [9], established in 1992, focuses on the history of computing and informatics with a view to providing the impetus to preserve the records and artifacts of information processing inventions, practices, and activities throughout the world under the auspices of IFIP and its constituent organizations. Among the goals of this group are to encourage the development of national archives, to identify pioneers worthy of appreciation and distinction, to develop publication plans for histories of Information Processing, and to promote the inclusion of historical modules in appropriate curricula. The Virtual Museum of Computing (VMoC) [24], maintained by Jonathan Bowen of London South Bank University, is a collecting point that leads to many different sites across the web. Sections currently featured on the VMoC site include corporate history and overviews, history of computing organizations, and general historical information. The History of Computing project [7], started by Cornelis Robat in the late 1980s, is now supported by a non-profit foundation founded in April, 2000. This project is based in the Netherlands and has partners from throughout the world, including the Ukraine, Poland, and Mexico. The project seems focused on gathering artifacts into an enormous database to ensure that important historical information remains available. 4.3 Resources on the History of Women in Computing Especially relevant to this Working Group's efforts are projects to collect oral histories of women in computing. Janet Abbate [1] is conducting a research project to develop a history of women in computing in the United States and Britain since World War II. Her project draws on oral history interviews with more than fifty women who were active in computer science departments and the software industry. A project that apparently never came to fruition is mentioned on a history of computing site created by J.A.N. Lee [13]. This project was called "Women in (the) Computing History" (with the acronym "witch"). The description of this project states: "In keeping with the tradition of documenting women's history through oral histories, the Women in (the) Computing History mailing list hopes to augment traditional resources of women's and histories of computing by being a repository for women's own stories throughout the history of computing. All in computing, too, not just those of us formally schooled in the computing sciences." Unfortunately, it appears that this project has disappeared from view, as we have thus far been unable to establish contact with anyone associated with the project. The IFIP Working Group 9.8 on Women and Information Technology [10] was established in 2001. Aspects of this group's charge include the exchange of women's experiences as scholars and professionals in information technology, integration of feminist perspectives into computer science, and developing an understanding of the gendered aspects in design, realization, and implementation of information systems. The aims that seem especially relevant for this project are analyzing the role of gender in computing education and educational strategies to support and retain girls and women. 4.4 History of Computing Education Resources Considered separately from resources related to the History of Computing, few resources address the history of computing education. In 1982, the Mathematical Association of America published a perspective on the field of Computer Science. The first chapter is an in-depth exploration of the development of Computer Science, with emphasis on the educational underpinnings of this field [19]. In August, 2004, when the IFIP 18th World Computer Congress was held in Toulouse, France, one component of the Congress was a History of Computing in Education conference. A book published in 2004 derives from contributions made at this conference. This book [11] considers two aspects: the impact of computing on education over the past forty years and as a pedagogical tool in computing education. Various articles consider how organizations have used computers to enhance teaching and learning, spanning experiences in elementary education through university studies in several countries. 177 WORK DURING ITiCSE Once the Working Group convened in Lisbon, the face-to-face meeting time was spent primarily on four activities: refining the purpose of the project, discussing and demonstrating the relevant techniques, developing a protocol to guide the process of planning and conducting interviews, and training members in how to use the interviewing techniques and materials. Each of these aspects are covered below. 5.1 Refining Purpose During the Working Group meetings, we refined the purpose and methods of the project. We realized the need to differentiate between the purpose of the interviews (how they are structured and the kind of information they elicit) and the purpose of the project as a whole (how the interviews will be used). We also came to realize that our original notion of interviewing "successful" women computing educators constrained the project in two ways: 1) defining what we meant by "success" and 2) losing the stories and lessons of those who did not continue in computing education. 5.2 Demonstration During the two days of meetings before the ITiCSE conference began, a key aspect of the Working Group's efforts was to explore the theory and techniques guiding this project. To this end, the group discussed general techniques for how to use oral histories. Two of the group members, Aspray and Barker, have social science backgrounds: Aspray is a historian of computing, while Barker is a social scientist whose work focuses on women in computing. Because most group members had little experience with conducting this type of inquiry, Aspray overviewed the purposes of oral history and methods for conducting interviews. To make the techniques tangible, Aspray conducted a demonstration interview with Working Group leader Barbara Owens as the subject. In preparation for the interview, Aspray and the remaining Working Group members formulated a set of topics and prompts to include in the interview. The demonstration interview was recorded on several digital devices, both to test the devices and to avoid the possible loss of information due to technical difficulties. After the demonstration interview was completed, Aspray and Barker led the group in deconstructing 1 the interview. During this session, the group reflected on what went well, what could be improved, and what to change in the future. 5.3 The Protocol A major product of our Working Group was a protocol for this project. After much discussion, we concluded that having a common set of materials would be vital for achieving consistent results in interview sessions conducted by a wide variety of volunteers. The protocol materials that will be used to support the interview process include an opening script, an outline of topics, a set of sample probing questions or prompts, and guidelines for conducting interviews. We discuss each of these items in the remainder of this section. 1 Deconstructing an interview is different than analyzing the results; the former focuses on process, while the latter considers content. 5.3.1 The Opening Script The opening script is used by the interviewer to set the scene before beginning the session. For example, the interviewer should caution the subject that it is common for sensitive topics to come up during the course of a session and that the subject should feel free to ask that the recording be turned off. As the session gets underway and the interviewer starts the recording device(s), specific opening information should be read onto the recording in order to provide a full context for this session. The interviewer could state, for example: "This is an interview with (interview subject's name) from (name of institution), conducted by (interviewer's name). This interview is being recorded on (date) at (city, country). It is part of the (computing education oral history series / formalized name yet to be determined). "Did we give and pronounce your name correctly?" 2 After this, the interviewer can begin giving prompts, such as "Tell us about your parents, for example what they did for a living." In this example statement, using the pronoun "us", rather than "me", can help the subject remember that her story is being told for a wider audience than just the interviewer at hand. 5.3.2 Outline of Topics The Working Group developed an outline of relevant topics to be used in guiding the interviews. The outline can also assist the interviewer in preparing for the interview, with the goal of making the face-to-face time with the subject as effective as possible. The Outline of Topics that the Working Group developed appears in Appendix A. 5.3.3 Sample Probing Questions Prompts are follow-up questions designed to elicit more detailed answers or follow up a thread introduced in an earlier answer. Because an interviewer must feel free to pursue topics that emerge as the session progresses, the prompts set provides examples for how the interview can proceed, rather than a strict step-by-step recipe. The Working Group developed a list of example prompts, which appears in Appendix B. 5.3.4 Guidelines for Conducting Interviews This oral history project will require many interviewers in order to increase the number of stories that can be collected within a limited timeframe and across a wide geographical area. Guidelines will help coordinate the efforts across volunteers in order to achieve a level of consistency across the results. Guidelines will also help the volunteers prepare for and conduct sessions. The guidelines can assist an interviewer in establishing the proper setting, maintaining an appropriate flow, and helping the subject focus on the issues at hand. To prepare for the session, the interviewer should study relevant background materials such as the subject's resume or vita, their professional publications, and anything written about the subject 2 This final sentence is relevant primarily for names that are unusual or difficult to pronounce. 178 in secondary literature. This information can help the interviewer plan and prioritize the specific prompts, as well as the order of prompts, to be used during the interview session. At the same time, the interviewer should not use the outline of topics as "tick off" items. An effective interview will be interactive in nature, with the specific choice and ordering of prompts based on previous answers. Because the duration of a session must be limited to no more than an hour or two, the time must be used effectively. This makes it essential for the interviewer to come to the session as well prepared as possible. The face-to-face time during the session should be used to explore tacit knowledge and the reasons for certain behaviors and outcomes, providing insights into the motivations behind events in the subject's life. To use the time well, the interviewer must avoid spending precious time during the session pursuing information that can be gleaned from the subject's vita or other readily available materials. The Working Group's guidelines for conducting interviews appear in Appendix C. 5.4 Training On the second day of the ITiCSE Working Group meetings, the group divided into two sub-groups, each of which included three computing educators and a "consultant" (either Aspray or Barker). In these sub-groups, the computing educators tested the tentative protocol to conduct practice interviews with one another. Each interview session lasted about 15 minutes, with one computing educator interviewing a second using the list of topic areas, while the third member watched and listened from the side. During the practice interviews, both sub-groups explored technologies that can be used to record the interviews and transcribe the audio recordings, testing multiple devices and in one group using a headset to capture the answers for automatic transcription. The "consultants" observed during the interviews, then helped the group deconstruct the interviews and critique the methods. REFLECTIONS AFTER ITiCSE Working Group members have offered the following as the most positive outcomes of the time in Lisbon (given in no particular order): 1. Learning techniques of oral history and observing an experienced interviewer using the techniques during a demonstration. 2. Hearing diverse ideas about project goals and reaching consensus. 3. Fleshing out the protocol for conducting interviews, thus making clear what should be asked during a session. The protocol includes a detailed set of guidelines for conducting an oral history interview (Appendix C), an opening script (Section 5.3.1), a topic outline (Appendix A), and sample prompts (Appendix B). 4. Being trained in interview techniques, which allowed the group to experiment with the equipment and pilot the process of gathering histories. Several members expressed the desire for additional training and for the opportunity to review recorded interviews conducted by more experienced individuals. 5. Understanding that the operative dynamic in an interview/oral history differs from that in conversation, although the similarities make it tricky to balance the exchange. Members felt very positive about the experience of the practice interviews. One member reported that she often found herself caught up in the stories the subjects were telling, leading her to realize that it takes effort to learn to stick to the list of topics that the interviewer wants to cover. 6. Seeing the importance of privacy considerations, as well as the need to obtain permission and plan for storage and access. 7. Getting to know the other group members and hearing significant parts some of their stories. Even this small sample gave group members a feeling for the wide variety of paths taken and challenges overcome. 8. Discovering that the individual paths were an interplay between a recitation of facts (dates and places) and the deeply felt emotional life that often motivated a person's actions. This underscored the need for a respectful and reasonably well-trained approach by each interviewer. The group encountered a number of difficulties with the software and equipment. It became clear that the equipment is the weakest link in performing an interview. The interviewer cannot be certain that the equipment is functioning as expected until he or she takes a break to review the recording. Based on experimentation during the Working Group meetings, the Working group can make the following observations: The group used several different models of the Olympus DVR (Digital Voice Recorder) and was able to get each model to work properly. Direct recording to the computer worked well through the Olympus VN-240PC digital recorder. Transferring the recordings to CD was simple and seemed like an excellent way to create a temporary archive. While the Dragon Naturally Speaking Preferred speech recognition software [5] may be helpful, it will require further experimentation to use it effectively. While the group's experiences with the i-River recording devices were not successful, one member has been pleased with the performance of this device in the past. In general the digital recorders worked well. However, in every session at least one of the recorders failed, generally due to inexperience, human error, or time limitations. A key conclusion is that equipment redundancy is imperative. We decided it is safest to use at least three recording devices during each interview in order to ensure the best possible quality of recording. Group members were surprised at the difficulty of transcribing recorded interviews. Some members had hoped there would be useful tricks or slight-of-hand for doing transcription. Unfortunately, creating a good quality transcription is simply a lengthy and intense process. A group member who transcribed her own interview found that it took nearly five times as long as the interview duration to complete the transcription of the session! During early planning for the Working Group, the co-leaders had hoped to "... conduct initial analysis of pilot interview data, and identify emergent themes". In the end, the group spent no time 179 with formal analysis of the practice interviews. Instead, the group used the time to hone interview techniques and understand how to move the project forward. During the first day of the ITiCSE conference (after the Working Groups had each met for two full days), each Working Group presented their group's mission and progress for conference attendees. The main impressions that Working Group members brought away from this presentation were very positive, with many attendees showing strong interest in the project and offering encouragement as well as suggestions for potential subjects. WHAT COMES NEXT? While the experience during the ITiCSE conference was valuable, the time in Lisbon was too short and the expectations too high for the group to be able to complete everything it had hoped to accomplish. By the end of the conference and the completion of this report, the Working Group had prepared an annotated bibliography, learned about oral histories, piloted hardware and software for recording, and set the stage for ongoing collection of histories, including a protocol to follow in planning for and conducting interviews. While the Working Group did consider legal and ethical issues during their discussions, a great deal must be resolved before the process of active interviewing can begin. In particular, access and ownership issues must be resolved before we can begin collecting interviews. The Working Group has an excellent start in recruiting volunteers to help in carrying out all aspects of the project. However, the work of the volunteers must be coordinated in order to produce coherent results. In addition, volunteers who conduct interviews must be trained in the techniques. Various Working Group members have agreed to propose workshops and other training opportunities at a variety of venues and events. A challenge will be to select the set of subjects from the many suggestions we have received. For the current stage of work, we will include only women computing educators who are retired or in the latter stages of their careers. The entire project has an underlying sense of urgency because many of the pioneers are in poor health or have already passed away. We have seen clear interest in eventually expanding the project to include the stories of women in earlier parts of their careers and men at any stage of their careers. Obtaining one or more sources of funding will be essential to achieving the full vision of the project. Funding can support aspects such as transcription and review, travel to conduct training or to meet with subjects, and setting up permanent archival facilities. While finding a permanent home for the oral histories is not essential during the early phases of the project, it is important if the collected stories are to be useful and usable. In addition to providing for archival of the recordings and transcriptions, the eventual home should allow for including contextual materials, such as course and curriculum artifacts. The archival capability must include sophisticated support for indexing and searching in order to support future visitors in browsing the collection and analyzing the interview transcriptions and other artifacts. Ultimately, whether this project will succeed or fail depends on the level of engagement we can generate for all phases of the project. To start, we must involve the computing education community in collecting stories from women computing educators who have retired or are about to retire. At the same time, we must create and maintain a sense of excitement about the potential of the project. If there are sufficiently many interested volunteers, the full-blown project to collect stories from men and from women earlier in their careers could certainly get underway in parallel with the current efforts. ACKNOWLEDGMENTS The individuals who met in Lisbon enjoyed the unique opportunity to learn these techniques and plan for what we hope will be a productive long-term project. The group was fortunate to have additional individuals involved in the pre-conference discussions, several of whom made key contributions to the preparations. In particular, the Working Group is grateful to Bettina Bair for her enthusiastic support. Bettina set up the group wiki and provided feedback as well as many ideas for resources. We are also grateful to the others who participated in the pre-conference discussions, including Anne Applin and Amardeep Kahlon. We thank the individuals who responded to our survey and offered suggestions of future subjects and possible questions. Comments from the anonymous reviewers allowed us to refine the purpose of the report and improve the presentation. Late discussions with Susan Gerhart provided additional ideas and inspiration for future work. REFERENCES The references given here are used directly in the text. In Appendix E we provide an annotated reference list, which repeats several of these references supplemented with our annotations. [17] Abbate, J., Finding our Place in History: Six Decades of Women in Computing, Grace Hopper Celebration of Women in Computing. October 6-9, 2004. Chicago, IL. online: gracehopper.org/Proceedings/PDF/wpp_Abbate.pdf ; last modified 10 January 2005, accessed 17 June 2005. [18] Baylor University Institute for Oral History, Oral History Workshop on the Web, online: www.baylor.edu/Oral_History/ , last modified 25 April 2005, accessed 17 June 2005. [19] Bandura, A. Social foundations of thought and action: A social cognitive theory. Englewood Cliffs, NJ: Prentice Hall, 1986. [20] Charles Babbage Institute, Center for the History of Information Technology, University of Minnesota. online: www.cbi.umn.edu/index.html , accessed 17 June 2005. [21] Dragon Naturally Speaking Preferred speech recognition software, 1st-Dragon information page: www.1st-dragon .com/dragnatspeak1.html , accessed 17 July 2005. [22] Gluck, S. B. An Oral History Primer, Department of History, California State University, Long Beach, online: www.csulb.edu/depts/history/relprm/oralprimer/ OHprimer.html . last updated 6 March 2001, accessed 28 July 2005. [23] The History of Computing project. online: www.thocp.net/ . accessed 20 July 2005. [24] IEEE Annals of Computing. online: www.computer.org/annals/. accessed 28 July 2005. 180 [25] IFIP Working Group 9.7 on the History of Computing. online: http://www.comphist.org/ . last modified 12 July 2005; accessed 20 July 2005. [26] IFIP Working Group 9.8 on Women and Information Technology. online: www.informatik.uni-bremen .de/~oechteri/IFIP/ . last modified 24 February 2004; accessed 20 July 2005. [27] Implagliazzo, J., & J. A. N. Lee, History of computing in education, Boston: Kluwer, 2004. [28] Lazowska, E. Pale and male: 19th century design in a 21st century world. Women in computing history, SIGCSE Bulletin inroads. 34(2) (June 2002). pp. 11-12. [29] Lee, J.A.N. History of Computing. online: ei.cs.vt.edu/~history/ . last updated 6 December 2002. accessed 20 July 2005. [30] London Voices, Museum of London, online: www.museumoflondon.org.uk/MOLsite/londonsvoice s/ , last modified 8 July 2004, accessed 17 June 2005. [31] Margolis, J. & Fisher, A. Unlocking the Clubhouse: The Carnegie Mellon Experience, SIGCSE Bulletin inroads, 34(2), June 2002. pp. 79-83. [32] Museu da Pessoa [translation: Museum of the Person]. online: www.museudapessoa.net/ . accessed 20 July 2005. [33] Oral History Association. online: omega.dickinson.edu/organizations/oha/ . accessed 20 July 2005. [34] Oral History Directory, Alexander Street Press, online: www.alexanderstreet2.com/oralhist/ , last modified 18 March 2005, accessed 19 June 2005. [35] Pollack, S. V. The Development of Computer Science. In S. V. Pollack (Ed.) Studies in Computer Science. MAA Studies in Mathematics, Volume 22. The Mathematical Association of America, 1982. [36] Ritchie, D. A. Oral History Evaluation Guidelines. Pamphlet Number 3. Oral History Association. Oral History Association. online: www.dickinson.edu/oha/pub_eg.html , last modified 4 May 2004, accessed 17 June 2005. [37] Roberts, E. Expanding the audience for computer science. PowerPoint version of keynote talk presented at 2003 SIGCSE Technical Symposium, Reno, Nevada. [38] SIGCSE.members, the members-only mailing list of the ACM Special Interest Group for Computer Science Education. Subscription information online at www.sigcse.org/ . [39] Systers, an online community for technical women in computing. Subscription information online at www.mecca.org/ . [40] Virtual Museum of Computing (VMoC), online: http://vmoc.museophile.org/ ; last modified 4 January 2005, accessed 20 July 2005. [41] Wikipedia, Oral History. online: en.wikipedia.org/wiki/Oral_history/ . last modified 17 June 2005, accessed 20 July 2005.

Oral History;Computing Education History

5

A Dependability Perspective on Emerging Technologies

Emerging technologies are set to provide further provisions for computing in times when the limits of current technology of microelectronics become an ever closer presence. A technology roadmap document lists biologically-inspired computing and quantum computing as two emerging technology vectors for novel computing architectures <A href="5.html#12">[43]. But the potential benefits that will come from entering the nanoelectronics era and from exploring novel nanotechnologies are foreseen to come at the cost of increased sensitivity to influences from the surrounding environment. This paper elaborates on a dependability perspective over these two emerging technology vectors from a designer's standpoint. Maintaining or increasing the dependability of unconventional computational processes is discussed in two different contexts: one of a bio-inspired computing architecture (the Embryonics project) and another of a quantum computational architecture (the QUERIST project).

INTRODUCTION High-end computing has reached nearly every corner of our present day life, in a variety of forms taylored to accommodate either general purpose or specialized applications. Computers may be considerred as fine exponents of the present days' technological wave if not their finest, they certainly do count as solid, indispensable support for the finest. From the very beginning of the computing advent, the main target was squeezing out any additional performance. The inception period was not always trouble-free, accurate computation results being required at an ever faster pace on a road that has become manifold: some applications do require computational speed as a top priority; others are set for the highest possible dependability, while still delivering sufficient performance levels. Several definitions for dependability have been proposed: "the ability of a system to avoid service failures that are more frequent or more severe than is acceptable" <A href="5.html#11">[2], or "the property of a computer system such that reliance can justifiably be placed on the service it delivers" <A href="5.html#11">[9]<A href="5.html#12">[45]. Dependability is therefore a synthetic term specifying a qualitative system descriptor that can generally be quantified through a list of attributes including reliability, fault tolerance, availability, and others. In real world, a dependable system would have to operate normally over extended periods of time before experiencing any fail (reliability, availability) and to recover quickly from errors (fault tolerance, self-test and self-repair). The term "acceptable" has an essential meaning within the dependability's definition, setting the upper limits of the damages that can be supported by the system while still remaining functional or computationally accurate. A dependability analysis should take into consideration if not quantitative figures for the acceptable damage limit, at least a qualitative parameter representation for its attributes. Dependable systems are therefore crucial for applications that prohibit or limit human interventions, such as long-term exposure to aggressive (or even hostile) environments. The best examples are long term operating machines as required by managing deep-underwater/nuclear activities and outer space exploration. There are three main concerns that should be posed through a system's design in order to achieve high dependability <A href="5.html#12">[42]: 1. Specifying the dependability requirements: selecting the dependability requirements that have to be pursued in building the computing system, based on known or assumed goals for the part of the world that is directly affected by the computing system; 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. 2. Designing and implementing the computing system so as to achieve the dependability required. However, this step is hard to implement since the system reliability cannot be satisfied CF'06, May 35, 2006, Ischia, Italy. Copyright 2006 ACM 1-59593-302-6/06/0005...$5.00. 187 simply from careful design. Some techniques can be used to help to achieve this goal, such as using fault injection to evaluate the design process. 3. Validating a system: gaining confidence that a certain dependability requirement/goal has been attained. This paper will address these main concerns through an attempt to provide an in-depth view over modern computing directions and paradigms, which we consider to be representative for the efforts involved in improving overall dependability. 1.1 Motivations We have listed some of the applications of dependable computing systems as linked to activities that take place in special environments, such as deep underwater or outer space. At a very first sight, these applications would appear specific enough to not encourage a specific design for dependability approach in computing. However, evidence suggest this is hardly the case; on the contrary, it is difficult to imagine a domain left unconquered by computer systems during times when industrial, transport, financial services and others do rely heavily on accurate computer operation at any given moment. If computer innacuracies could be more easily overlooked at home, professional environments cannot accept such missbehaviors. Yet the recent history of computing provides evidence that dependability is not a sine qua non feature. During their life cycle, electronic devices constantly suffer a number of influences that manifest predominantly over transient regimes, which in turn introduce a variety of errors unified in the literature under the name of transient faults, soft errors or single event upsets (SEUs). The rate electronic devices are affected with is known under the term of soft error rate or simply SER and is measured in fails per unit time. Because it relies on transient phenomena due to changing states and logical values, digital electronics makes up for a special category that is also affected by soft errors. No matter the name they are referred under, these errors affect the computing processes and are due to electromagnetic noise and/or <A href="5.html#12">external radiations rather than design or manufacturing flaws [28]. One cause at the origin of soft fails affecting digital devices is known to be due to radioactive decay processes. Radioactive isotopes, widely used for a range of purposes, might contaminate semiconductor materials leading to soft errors; evidence is available throughout the literature, both by empirical observations and experimental results <A href="5.html#12">[20]. Consequently, cosmic rays, containing a broad range of energized atomic/subatomic particles may lead to the appearance of soft fails. Computers therefore are susceptive to soft errors, an issue that will potentially become essential with the advent of emerging technologies. As acknowledged by the International Technology Roadmap for Semiconductors (ITRS), issued at the end of 2004 <A href="5.html#12">[43], the microelectronics industry faces a challenging task in going to and beyond 45nm scale in order to address "beyond CMOS" applications. Scaling down the technology will enable an extremely large number of devices to be integrated onto the same chip. However, the great challenge will be to ensure the new <A href="5.html#11">devices will be operational at this scale [6], since they will exhibit a sensitive behavior to soft fails. In order to address the negative effects brought by technology scaling, it is to be expected that significant control resources will need to be implem<A href="5.html#11">ented [3]. Another challenging aspect concerning emerging technologies is to match the newly developed device technologies with new system architectures, a synergistic/collaborative development of the two being seen as likely to be very rewarding. The potential of biologically-inspired and quantum computing architectures is acknowledged by the ITRS report on emerging technologies <A href="5.html#12">[43] (see <A href="5.html#2">Figure 1). This paper will investigate the relevance of soft fails and attempt to provide means of harnessing their negative effects on modern computing in the context of biologically-inspired and quantum computing architectures. Figure 1: Bio-inspired and quantum computing are acknowledged as architectural technology vectors in emerging technologies <A href="5.html#12">[43] 1.2 Paper Outline This paper is structured as follows. Section <A href="5.html#3">2 will address the first main concern, that is, specifying and selecting dependability requirements that will have to be pursued when building a computational platform. Parameters that describe and quantify dependability attributes, such as reliability, will be introduced, with a highlight on their accepted models and their issues. A particular consideration will be given to the failure rate parameter, which is the basis of all reliability analyses. Section <A href="5.html#7">3 will approach some of the means for design for dependability; it will therefore elaborate upon two emerging technology vectors, as seen by the ITRS report <A href="5.html#12">[43], which define two novel architectures, namely biologically-inspired (or bio-inspired ) and quantum computing. We will introduce two projects and their corresponding architectures, called Embryonics (as a biologically-inspired computing platform) and QUERIST (as a quantum computing platform designed to allow and study error injection). These two architectures are representative for the coming age of nano-computing, where computational processes take place as encoded at the very inner core level of matter, be it semiconductor material (for nanoelectronics, targetted here by the Embryonics project) or atomic scale dynamics (for quantum computing, targetted here by the QUERIST project). This section will then introduce dependability aspects within bio-inspired computing (the Embryonics project being investigated in SubSection <A href="5.html#7">3.1) and within quantum computing (the QUERIST <A href="5.html#10">project being investigated in SubSection 3.2). Finally<A href="5.html#10">, Section 4 will present the conclusions and prospects for designing emerging technology dependable computing systems, as we see them. 188 DEPENDABILITY ATTRIBUTES An important dependability attribute for any given system lies in its capacity to operate reliably for a given time interval, knowing that normal operation was delivered at initial time <A href="5.html#11">[8]. Reliability functions are modelled as exponential functions of parameter , which is the failure rate. The reliability of a system is the consequence of the reliability of all of its subsystems. The heterogeneity of the system leads to a difficult quantitative assessment of its overall reliability; moreover, estimating the reliability functions is further made difficult because formal rigour is not commercially available, this being kept under military mandate <A href="5.html#12">[44]. The failure rate for a given system can be modelled as a function of the failure rates of its individual subsystems, suggestions being present in the MIL-HDBC-217 document, which is publicly available <A href="5.html#12">[44]. However, this document has been strongly criticized for its failure rate estimations based on the Arrhenius model, which relates the failure rate to the operating temperature: B E K T Ke = (1) where K is a constant, K B is Boltzmann's constant, T is the absolute temperature and E is the "activation energy" for the process <A href="5.html#11">. Quantitative values for failure rates show significant differences between those predicted using MIL-HDBC-217 and those from testing real devices (see ). There are two conclusions that can be drawn from this: B <A href="5.html#11">[18] <A href="5.html#3">Figure 2 1. quantitative estimations for failure rate values are strongly dependant on the quality of information used; unfortunately, current reliable information about electronic devices is known to be lacking <A href="5.html#12">[44]; 2. despite differences between predicted and real values, the MIL-HDBC-217 methodology can be useful for qualitative analyses in order to take decisions regarding sub-system parts that should benefit from improved designs. Figure 2. Predicted vs real failure rates plotted against temperature <A href="5.html#11">[18] So far the failure rate of digital devices has been considerred as due to internal causes. However, this is not always the case, soft fails being equally present due to the aggressive influences of the external environment, which <A href="5.html#12">also have to be modelled [22]. The external envirnment features highly dynamic changes in its parameters, which will eventually affect the normal operation of digital devices that lack sufficient protection or ability to adapt. Ideally, computing devices would behave in a consistent and accurate manner regardless of fluctuations in environmental parameters. This is either a consequence of soft error mitigation techniques or due to flexible hardware/software functionality that allow the system as a whole to adapt to environamental changes and tolerate induced faults. While certain soft error mitigation techniques are available, the technology scaling towards nanoelectronics affects their efficiency by integrating a larger number of devices per chip (which requires a larger amount of redundant/control logic or other measures), which feature, at the same time, smaller dimensions (which renders an electronic device much more senzitive to the influence of stray energetic particles that reach it as part of cosmic rays). Both aspects are involved in the development of the two emerging technology vectors mentioned <A href="5.html#2">in SubSection 1.1, although having slightly different motivations: while the nature of the quantum environment prohibits precise computation in the absence of fault tolerance techniques, such techniques are targetted by bio-inspired computing as means of improving the dependability of a computing platform. 2.1 Bio-Inspired Computing If living beings may be considered to fulfill computational tasks, then Nature is the ultimate engineer: each of the living beings exhibit solutions that were successfully tested and refined in such ways human engineers will never afford. One reason is time: the testing period coinciding with the very existence of life itself. Another reason is variety and complexity: Nature has found and adapted a variety of solutions to address complex survivability issues in a dynamically changing environment. No matter how Nature approached the process of evolution, engineering could perhaps benefit most from drawing inspiration from its mechanisms rather from trying to develop particular techniques. Bio-inspired computing is not a new idea. John von Neumann was preoccupied to design a machine that could replicate itself and was quite interested in the study of how the behavior of the human brain could be implemented by a computer <A href="5.html#11">[13][14]. He also pioneered the field of dependable computing by studying the possibility of building reliable machines out of unreliable components <A href="5.html#11">[15]. Unfortunately, the dream of implementing his self-reproducing automata could not become true until the 1990s, when massively programmable logic opened the new era of reconfigurable computing. But when trying to adapt nature's mechanisms in digital devices, it becomes most evident that biological organisms are rightfully the most intricate structures known to man. They continuously demonstrate a highly complex behavior due to massive, parallel cooperation between huge numbers of relatively simple elements, the cells. And considering uncountable variety of living beings, with a life span up to several hundreds (for the animal regnum) or even thousands (for the vegetal regnum) of years, it seems nature is the closest spring of inspiration for designing dependable, fault tolerant systems. Investigating the particularities of natural systems, a taxonomy of three categories of processes<A href="5.html#12"> can be identified [32]: 1. Phylogenetic processes constitute the first level of organization of the living matter. They are concerned with the temporal evolution of the genetic heritage of all individuals, 189 therefore mastering the evolution of all species. The phylogenetic processes rely on mechanisms such as recombination and mutation, which are essentially nondeterministic; the error rate ensures here nature's diversity. 2. Ontogenetic processes represent the second level of organization of the living matter. They are also concerned with the temporal evolution of the genetic heritage of, in this case, a single, multicellular individual, therefore mastering an individual's development from the stage of a single cell, the zygote, through succesive cellular division and specialization, to the adult stage. These processes rely on deterministic mechanisms; any error at this level results in malformations. 3. Epigenetic processes represent the third level of organization of the living matter. They are concerned with the integration of interactions with the surrounding environment therefore resulting in what we call learning systems. This taxonomy is important in that it provides a model called POE (from Phylogeny, Ontogeny and Epigenesis) that inspires the combination of processes in order to create novel bio-inspired <A href="5.html#4">hardware (see Figure 3). We believe this is also important from a dependability engineering perspective, for the following reasons: 1. Phylogenetic processes were assimilated by modern computing as evolutionary computation, including genetic algorithms and genetic programming. The essence of any genetic algorithm is the derivation of a solution space based on recombination, crossover and mutation processes that spawn a population of individuals, each encoding a possible solution. One may consider that each such step, with the exception of discovering the solution, is equivalent to a process of error injection, which in turn leads to wandering from the optimal solution (or class of solutions). However, genetic algorithms prove to be successful despite this error injection, the fitness function being responsible for the successful quantification of the significance of the "error". Therefore genetic computation is intrinsicaly resilient to faults and errors, largely due to the fact that they are part of the very process that generates the solutions. 2. Ontogenetic processes have been implemented in digital hardware with modular and uniform architectures. Such an architecture enables the implementation of mechanisms similar to the cellular division and cellular differentiation that take place in living beings<A href="5.html#12"> [31]. These mechanisms bring the advantage of distributed and hierarchical fault tolerance strategies: the uniformity of the architecture also makes any module to be universal, that is, to be able to take over the role of any other damaged module. 3. Epigenetic processes were assimilated by modern computing mainly as artificial neural networks (or ANNs) as inspired by the nervous system, and much less as inspired by the immune or endocrine systems from superior multicellular living beings. ANNs are known to have a generalization capacity, that is, to respond well even if the input patterns are not part of the patterns used during the learning phase. This means that ANNs possess a certain ability to tolerante faults, whether they manifest at the inputs or inside their intenal architecture. With the advent of field programmable logic (of which the most salient representative are the FPGAs) it is now possible to change hardware functionality through software, thus allowing information to govern matter in digital electronics. This is not dissimilar to what happens in nature: information coded in DNA affects the development of an organism. A special kind of such digital devices that change dynamically their behavior are known as evolvable or adaptive hardware; they are bio-inspired computing systems whose behaviors may change according to computational targets, or, if harsh or unknown environments are to be explored, for the purpose of maximizing dependability. Figure 3. The POE model of bio-inspired systems<A href="5.html#12"> [32] 2.2 Quantum Computing Error detection and correction techniques are vital in quantum computing due to the destructive effect of the environment, which therefore acts as an omnipresent error generator. Error detection and correction must provide a safe recovery process within quantum computing processes through keeping error propagation under control. Without such dependability techniques there could be no realistic prospect of an operational quantum computational device <A href="5.html#11">[19]. There are two main sources of errors: the first is due to the erroneous behavior of the quantum gate, producing the so-called processing errors; the second is due to the macroscopic environment that interacts with the quantum state, producing the storing and transmitting errors. The consistency of any quantum computation process can be destroyed by innacuracies and errors if the error probability in the basic components (qubits, quantum gates) excedes an accuracy threshold. This is a critical aspect since the microscopic quantum states are prone to frequent errors. The main error source is the decoherence<A href="5.html#11"> effect [16]. The environment is constantly attempting to measure the sensitive quantum superposition state, a phenomenon that cannot be avoided technologically since it is not (yet) possible to isolate them perfectly. The superposition state will decay through measuring and will therefore become a projection of the state vector onto a basis vector (or eigenstate). The most insidious error, however, appears when decoherence affects the quantum amplitudes without destroying them; this is similar to small analog errors. Issues stated above are solved, on one hand, through intrinsic fault tolerance by technological implementation <A href="5.html#11">(topological interactions [1]) and, on the other hand, by error correcting techniques at the unitary (gate network) level. We will focus on the error detecting and correcting techniques, which are difficult to approach due to quantum constraints: the useful state 190 can neither be observed (otherwise it will decohere), nor can it be cloned. 2.2.1 Background <A href="5.html#11">As expressed in bra-ket notation [16], the qubit is a normalized vector in some Hilbert space , 2 H { } 0 , 1 being the orthonormal basis: 0 1 0 1 a a = + ( are the so-called quantum amplitudes, representing the square root of the associated measurement probabilities for the eigenstates 0 1 , a a C 0 and 1 respectively, with 0 1 2 2 1 a a + = ). Therefore, the qubit can be affected by 3 types of errors: Bit flip errors are somewhat similar to classical bit flip errors. For a single qubit things are exactly the same as in classical computation: 0 1 , 1 0 . For 2 or more qubits, flip errors affecting the state may modify it or leave it unchanged. For instance, if we consider the so-called cat state ( 1 00 11 2 Cat = + ) <A href="5.html#11">[19], and the first qubit is affected by a bit flip error, the resulting state will be ( ) 1 10 01 2 Cat + . But, if both qubits are affected by bit flips, there will be no change in the state: ( ) 1 11 00 2 Cat Cat + = . Phase errors affect the phase of one of the qubit's amplitudes and is expressed as 0 0 , 1 - 1 . This type of error is very dangerous, due to its propagation behavior but it only makes sense when dealing with superposition states. If we consider an equally weighted qubit superposition state and inject a phase error, this results in ( ) ( ) 1 1 0 1 0 1 2 2 + . There is a strict correspondence between bit flip and phase error types due to the way they map onto Hilbert spaces with the same dimension but different basis. The bit flip is an error from the space with basis 2 H { } 0 , 1 , whereas the phase error appears in the same space with basis ( ) ( ) 1 1 0 1 , 0 1 2 2 + or { } , + 1 a . The space basis conversion, in this case, is made by applying the Hadamard transform; <A href="5.html#5">Figure 4 shows an example of transforming a bit flip error into a phase error (A, and vice versa (B. Small amplitude errors: amplitudes of the quantum bit can be affected by small errors, similar to analog errors. Even if such an error does not destroy the superposition and conserves the value of the superposed states, small amplitude errors could accumulate over time, eventually ruining the computation. In order to avoid this situation, specific methodologies for digitizing small errors are used to reduce them to a non-fault or a bit-flip 0 and a <A href="5.html#11">[19]. Due to the quantum physics laws, fault tolerance techniques have to comply with the following computational constraints: The observation destroys the state. Since observation is equivalent to measurement, this leads to destroying the useful state superposition. Information copying is impossible. Quantum physics renders the cloning of a quantum state impossible, meaning that a quantum state cannot be copied correctly. Therefore quantum error correction must address the following problems: Non-destructive measurement. Despite the first constraint it is necessary to find a way to measure the encoded information without destroying it. Because the encoded state cannot be measured directly, one needs to properly prepare some scratch (ancilla) qubits, which can then be measured. Fault-tolerant recovery. Due to the high error rate in quantum computational devices, it is likely that the error recovery itself will be affected by errors. If the recovery process is not fault-tolerant , then any error coding becomes useless. Phase error backward propagation. If we consider the XOR gate from <A href="5.html#5">Figure 5(A, a flip error affecting the target qubit (b) will propagate backwards and also affect the source qubit. This is due to the gate network equivalence from <A href="5.html#5">Figure 5(B and the basis transformation described by<A href="5.html#5"> Figure 4. Figure 4. Correspondence between bit flip and phase errors Figure 5. (A The backward propagation of a phase error for the XOR gate; (B Gate network equivalence In order to deal with the problems described the next strategies have to be followed: Digitizing small errors. The presence of small errors is not a major concern, as they can be digitized using a special technique based on measuring auxiliary<A href="5.html#11"> (ancilla) qubits [19]. Ancilla usage. Since qubit cloning is impossible, a majority voting strategy is difficult to implement. However, by using ancilla qubits, the eigenstate information can be duplicated inside the existing superposition, resulting in the entanglement of the ancilla with the useful data. Because any measurement performed on the ancilla could have repercussions on the useful qubits, the appropriate strategy will employ special coding for both data qubits and ancilla (data errors only will be copied onto the ancilla), followed by the computation of an error syndrome, which has to be obtained through measuring the ancilla (see <A href="5.html#6">Figure 6). Avoiding massive spreading of phase errors. As shown previously, a phase error on the target qubit will propagate on all source qubits. The solution is to use more ancilla qubits as targets, so that no ancilla qubit is used more than once. 191 Figure 6. Fault-tolerant procedure with ancilla qubits Ancilla and syndrome accuracy. Setting the ancilla code to some known quantum state could be an erroneous process. Computing the syndrome is also prone to errors. Hence, on one hand, one has to make sure that the ancilla qubits are in the right state by verifying and recovering them if needed; on the other hand, in order to have a reliable syndrome, it must be computed repeatedly. Error recovery. As the small errors can be digitized (therefore, they are either corrected or transformed into bit flip errors), the recovery must deal only with bit flip and phase errors. A state that needs to be recovered is described by: 0 1 1 0 0 1 0 1 1 0 0 1 if no error occurs 0 1 for a flip error 0 1 0 1 for a phase error 0 1 for both flip and phase errors error a a a a a a a a a a + + + . Correcting a bit flip error means applying the negation unitary transformation to the affected qubit. To correct phase and combined errors, the following unitary operators will have to be applied respectively: . 0 1 1 0 N x U = = 1 0 0 , 0 1 0 Z Y N Z i U U U U i = = = 2.2.2 Quantum Error Correcting Codes Quantum error coding and correcting (QECC) is performed with special coding techniques inspired from the classic Hamming codes. The classical error coding is adapted so that it becomes suitable for the quantum strategy, allowing only the ancilla qubits to be measured. The state-of-the-art in QECC is represented by the stabilizer encoding, a particular case being the Steane codes (the Shor codes may<A href="5.html#12"> also be used [29]). Steane's 7-qubit code is a single error correcting code inspired from classical Hamming coding and can be adapted for ancilla coding as well. Therefore it cannot recover from two identical qubit faults, but it can recover from a bit flip a phase flip. The Steane 7-qubit coding of 0 and 1 consists of an equally weighted superposition of all the valid Hamming 7-bit words with an even and odd number of 1s, respectively: ( ) , 0 1 2 3 0 1 2 0 1 2 3 0 1 2 32 32 1 0 2 1 0000000 0010111 0101110 0111001 2 1001011 1011100 1100101 1110010 u u u u c c c S even u u u u c c c = = + + + + + + + ( ) , 0 1 2 3 0 1 2 0 1 2 3 0 1 2 32 32 1 1 2 1 1111111 1101000 1010001 1000110 2 0110100 0100011 0011010 0001101 u u u u c c c S odd u u u u c c c = = + + + + + + + + (3) Applying the Steane coding on an arbitrary given quantum state 0 1 0 a a = + 1 transforms it into 0 1 0 1 S S a a = + S . This code was designed to correct bit-flip errors, but by changing the basis (through a Hadamard transform) the phase error transforms into a bit flip error, which can then be corrected: ( ) ( ) 1 0 0 0 1 2 1 1 1 0 1 2 S S S S S S S H H = = + = = S (4) 2.2.3 Fault Tolerance Methodologies Quantum error-correcting codes exist for r errors, , 1 r r N . Therefore a non-correctable error occurs if a number of 1 r + errors occur simultaneously before the recovery process. If the probability of a quantum gate error or storage error in the time unit is of order , then the probability of an error affecting the processed data block becomes of order 1 r + , which is negligible if r is sufficiently large. However, by increasing r the safe recovery also becomes more complex and hence prone to errors: it is possible that 1 r + errors accumulate in the block before the recovery is performed. Considering the relationship between r and the number of computational steps required for computing the syndrome is polynomial of the order p r . It was proven that in order to reduce as much as possible the error probability r must be chosen so that 1 1 ~ p r e <A href="5.html#11">[7][19]. By consequence, if attempting to execute N cycles of error correction without any r+1 errors accumulating before the recovery ends, then 1 ~ exp p N . Therefore the accuracy degree will be of the form ( ) ~ log p N , which is better than the accuracy degree corresponding to the no-coding case, 1 ~ N . However, there exists a so that if then non-correctable error becomes likely, which limits the length of the recovery process. Given the extremely large number of gates employed by a quantum algorithm implementation, also has to be very large; for Shor's algorithm must be higher than max N max N N &gt; max N max N 9 3 10 <A href="5.html#12">[30]. <A href="5.html#7">As shown in Figure 7, the required accuracy degree approaches today's technological limits (tipically 10 -3 for p=4) after N=10 5 . For a fault tolerant encoding solution for Shor algorithm implementation this should have happened after N=10 9 <A href="5.html#11"> [19][34]<A href="5.html#12">. + (2) Additional fault tolerance must be employed in order to preserve reliable quantum computation over an arbitrary number of computational steps. Concatenated coding represents one such technique, which improves the reliability by shaping the size of 192 the code blocks and the number of code levels. It is also resource <A href="5.html#11">demanding and vulnerable to correlated errors [19][37]. Another approach, replacing the concatenated codes, is based on Reconfigurable Quantum Gate Arrays (RQGAs) <A href="5.html#12">[34][37], which are used for configuring ECC circuits based on stabilizer codes <A href="5.html#11">[7][33]. By using a quantum configuration register for the RQGA (i.e. a superposition of classical configurations), the reconfigurable circuit is brought to a state where it represents a simultaneous superposition of distinct ECC circuits. After measuring the configuration register, only one ECC circuit is selected and used; if k distinct ECC circuits were superposed and the gate error rate is , then the overall gate error probability becomes k (s<A href="5.html#7">ee Figure 8). As a result, the accuracy threshold value for the RQGA solution clearly dominates the technological accuracy limit, as s<A href="5.html#7">hown in Figure 9 <A href="5.html#12">[37]. Figure 7. Accuracy plots: p=3 for xi 1 , p=4 for xi 2 , p=5 for xi 3 ; xi 4 for no-coding, ref is the reference accuracy (i.e. the accuracy allowed by today's state of the art technology) Figure 8. A quantum configuration register acts as a superposition of k distinct circuits sharing the same input state and the same output qubits DEPENDABLE SYSTEM DESIGN In order to model the erroneous behavior of a device of system it is necessary to understand the causality of phenomena concerned. A defect affecting a device from a physical point of view is called a fault, or a fail. Faults may be put in evidence through logical misbehavior, in which case they transform into errors. Finally, errors accumulating can lead to system<A href="5.html#11"> failure [8]. The fault-error -failure causal chain is essential to developping techniques that reduce the risk of error occurrence, even in the presence of faults, in order to minimize the probability of a system failure, and can be architecture specific. We will elaborate next on techniques used by a bio-inspired and by a quantum computing platform. Figure 9. Evolution of accuracy threshold value for RQHW stabilizer codes (xir); the technological accuracy limit (lim) is also provided for a relevant comparison 3.1 The Embryonics Approach Several years before his untimely death John von Neumann began developping a theory of automata, which was to contain a systematic theory of mixed mathematical and logical forms, aimed to a better understanding of both natural systems and computers <A href="5.html#11">[14]. The essence of von Neumann's message appears to entail the formula "genotype + ribotype = phenotype". He provided the foundations of a self-replicating machine (the phenotype), consisting of its complete description (the genotype), which is interpreted by a ribosome (the ribotype). Embryonics (a contraction for embryonic electronics) is a long term research project launched by the Logic Systems Laboratory at the Swiss Federal Institute of Technology, Lausanne, Switzerland. Its aim is to explore the potential of biologically-inspired mechanisms by borrowing and adapting them from nature into digital devices for the purpose of endowing them with the remarkable robustness present in biological <A href="5.html#12">entities [39]. Though perhaps fuzzy at a first glance, analogies between biology and electronics are pres<A href="5.html#8">ented in Table 1 <A href="5.html#11">[12][31]. But if we consider that the function of a living cell is determined by the genome, and that a computer's functionality is determined by the operating program, then the two worlds may be regarded as sharing a certain degree of similarity. Three fundamental features shared by living entities are required to be targetted by Embryonics in order to embody the formula "genotype + ribotype = phenotype" into digital hardware: multicellular organisms are made of a finite number of cells, which in turn are made of a finite number of chemically bonded molecules; each cell (beginning with the original cell, the zygote) may generate one or several daughter cell(s) through a process called cellular division; both the parent and the daughter cell(s) share the same genetic information, called the genome; different types of cells may exist due to cellular differentiation, a process through which only a part of the genome is executed. These fundamental features led the Embryonics project to settle for an architectural hierarchy of four levels (see <A href="5.html#8">Figure 10). We will not delve very deep inside the Embryonics'phylosophy, as such details were broadly covered by<A href="5.html#11"> literature [12][20]<A href="5.html#12">[23][24] [25][40]; we will, however, introduce each of the four levels in 193 order to be able to see how this bio-inspired platform fits modern design for dependability efforts. Table 1. Analogies present in Embryonics <A href="5.html#11">[12] Biology Electronics Multicellular organism Parallel computer systems Cell Processor Molecule FPGA Element Figure 10. Structural hierarchy in Embryonics<A href="5.html#11"> [12] The upmost level in Embryonics, bearing a certain similarity to what is found in nature, is the population level, composed of a number of organisms. One level down the hierarchy constitutes the organismic level, and corresponds to individual entities in a variety of functionalities and sizes. Each of the organisms may be further decomposed into smaller, simpler parts, called cells, which in turn may be decomposed in molecules. According to Embryonics, a biological organism corresponds in the world of digital systems to a complete computer, a biological cell is equivalent to a processor, and the smallest part in biology, the molecule, may be seen as the smallest, programmable element in digital electronics (s<A href="5.html#8">ee Table 1). An extremely valuable consequence of the Embryonics architecture is that each cell is &quot;universal&quot;, containing a copy of the whole of the organism's genetic material, the genome. This enables very flexible redundancy strategies, the living organisms being capable of self-repair (healing) or self-replication (cloning) <A href="5.html#11">[12]. Self-replication may be of great interest in the nanoelectronics era, where extremely large areas of programmable logic will probably render any centralized control very inefficient. Instead, the self-replication mechanism implemented in Embryonics will allow the initial colonization of the entire programmable array in a decentralized and distributed manner. <A href="5.html#8">Figure 11 presents an example of such colonization. At initial time the configuration bitstream (containing the genome) enters the bottom left corner of a programmable array and, at each clock cycle, the genome is pushed through and partitions the programmable space accordingly. From a dependability standpoint, the Embryonics hierarchical architecture offers incentives for an also hierarchical self-repair strategy. Because the target applications are those in which the failure frequency must be very low to be "acceptable", two levels of self-repair are offered: at the molecular level (programmable logic is susceptible to soft fail occurrences) and at the cellular level (soft fails manifest at this level as soft errors). Let us consider an example of a simple cell made of 3 lines and 3 columns of molecules, of which one column contains spare molecules. If a fault occurs inside an active cell, it can be repaired through transferring its functionality toward the appropriate spare molecule, which will become active (see <A href="5.html#8">Figure 12). Figure 11. Space colonization in Embryonics <A href="5.html#11">[11] Figure 12. Self-repair at the molecular level: faulty molecule E is replaced by spare molecule H, which becomes active <A href="5.html#12">[39] The self-repair process at molecular level ensures the fault recovery as long as there are spare molecules left for repair. However, it is possible for a cell to experience a multiple error, in which case the self-repair mechanism at the molecular level can no longer reconfigure the inside of the cell successfully. If such a situation arises, then a second self-repair strategy is trigerred at a higher level. The cell will "die", therefore trigerring the self-repair at the cellular level, the entire column containing the faulty cell (cell C in this example) being deactivated, its role being taken by the nearest spare column to the right (see <A href="5.html#9">Figure 13). A critique that could be addressed to the current Embryonics design would be its strategy of self-repair at the higher, cellular level: in case of a faulty cell, an entire column containing that cell will be deactivated, its role being transferred to the first available column of spares to the right (s<A href="5.html#9">ee Figure 13). There are two points in which this strategy could benefit: 194 1. Instead of deactivating a whole column of cells, it would be more efficient to only deactivate the faulty cell only (see <A href="5.html#9">Figure 14). The resources affected by the role transfer would be greatly reduced (one cell versus an entire column), coupled with the fact that particle flux generating soft fails is unlikely to be homogeneous and isotrope. This means regions assimilable more likely to cells rather than entire column of cells would be more affected by soft fails, not to mention that during genetic data transfer (required by taking over the role of the faulty cell) there is a greater risk of enduring a new soft fail (moving data is much more sensitive to soft fails than static data) <A href="5.html#11">[5][10]. 2. Such a strategy would be consistent with that used for the self-repair at the molecular level, which would simplify a thorough reliability analysis. Concatenated coding would also seem easier to be implemented and the strategy consistency would mean that concatenated coding would not be limited to a two-level hierarchy <A href="5.html#12">[20][21]. Figure 13. Molecular self-repair failure: the cell "dies" (bottom), triggering the cellular self-repair (top) <A href="5.html#12">[39] We consider a cell of M lines and N columns, being composed of modules of M lines and n+s columns (for instance, the cell <A href="5.html#8">presented in Figure 12 consists of a single such module of two active columns and one spare column), of which s are spares. In order to meet certain reliability criteria, it is necessary to know what is the number s of spare columns of molecules that correspond to n columns of active molecules, that is, the horizontal dimensions for such a module. We will not provide a thorough reliability analysis, as this has been done previously <A href="5.html#11">[4][17]<A href="5.html#12">[20][21]; instead, we will analyze the influences of the proposed consistent self-repair strategy at both molecular and cellular levels through the use of logic molecules. Therefore Equation (5) holds: ( ) { }( ) { }( ) ( ) 1 k ModRow i R t =Prob no fails t Prob i fails t N k n s = + = + (5) where ( ) ModRow R t represents the reliability function for a row within a module. Then, considering the failure rate for one molecule , the probability of all molecules (both active and spare) to operate normally in a module's row becomes: { }( ) ( ) n s t Prob no fails t e + = (6) The probability of a row enduring i fails in the active molecules part is the conditional probability of having n-i active molecules operating normally, while a number of s-i spare molecules are ready to activate (that is, they are not affected by errors themselves): { }( ) { }( ) { }( Prob i fails t Prob i fails active t Prob i spares ok t ) = (7) { }( ) ( ) ( ) ( ) 1 n i t n i t n Prob i fails active t e e i = (8) { }( ) ( ) ( ) 1 it k i t k Prob i spares ok t e e i = (9) Then the reliability function for an entire cell is the cummulated reliability functions for the total number of modules: ( ) ( ) MN n s Cell ModRow R t R t + = (10) Figure 14. Proposed reconfiguration strategy at the cellular level A self-repair strategy that conserves the consistency between the molecular and the cellular level would allow for a more straightforward reliability analysis. Basically, it would be sufficient to substitute dimension parameters in Equations (5) (10) with those approapriate to the analysis of an organism instead of a cell. To illustrate this, we will consider an organism of M * lines and N * columns, being composed of modules of M * lines and n * +s * columns, of which s * are spares; we will also use the organism partitioning into modules, similar to the partitioning of cells used before. Therefore Equation (5) transforms into Equation (11): ( ) { }( ) { }( ) ( ) 1 * * * * k * * CellMR i R t =Prob no fails t Prob i fails t N k n s = + = + (11) where ( ) CellMR R t represents the reliability function for a row of cells within an organism module. In this case, the significance of the terms will be as follows: { }( ) ( ) * * n s * Cell Prob no fails t R t + = (12) 195 While Equation (7) continues to hold under the form of Equation (13), the significance of its terms will change according to the dimensions at the cellular level: { }( ) { }( ) { }( * * * Prob i fails t Prob i fails active t Prob i spares ok t = ) i k i (13) { }( ) ( ) ( ) ( ) * * 1 * n i Cell Cell n Prob i fails active t R t R t i = (14) { }( ) ( ) ( ) ( ) * * 1 * i Cell Cell k Prob i spares ok t R t R t i = (15) Finally, similar to Equation (10), the reliability function for an entire organism is the cummulated reliability functions for the total number of its modules: ( ) ( ) * * * * M N n s Org CellMR R t R t + = (16) Equations (5) to (16) provide the basics for a thorough reliability analysis for the proposed, uniform strategy of hierarchical reconfiguration, as opposed to the analysis provided by<A href="5.html#12"> [21], which specifically targetted the current Embryonics architecture. Despite having settled the reliability model, both analyses are incomplete, in that the failure rate parameter is missing, which makes a precise, quantitative dependability target difficult to meet. However, a reliability analysis is still valuable from a qualitative point of view, allowing a direct comparison of different systems. 3.2 The QUERIST Approach In order to deal with errors induced by the constant influence of the external environment upon computational processes, the following assumptions were made: errors appear randomly, are uncorrelated (neither in space, nor in time), there are no storage <A href="5.html#11">errors, and there are no leakage phenomena involved [19]. Classical HDL-based fault injection methodologies can be mapped to simulating quantum circuits without intervention provided that the new error and fault models are taken into account <A href="5.html#12">[35]. Of course, efficiency criteria require that they be adapted to one of the available efficient simulation frameworks <A href="5.html#12">[36][38][41]. QUERIST (from QUantum ERror Injection Simulation Tool) is the name of such a project, fostering simulated fault injection techniques in quantum <A href="5.html#12">circuits [34]. Similar to classical computation, simulated fault injection is used in order to evaluate the employed FTAMS (Fault Tolerance Algorithms and Methodologies) <A href="5.html#12">[26][27]. An overview of the QUERIST<A href="5.html#11"> project is presented in Figure 15. The three cycles of initialization, simulation, and data computation are common to both classical and quantum approaches. The first cycle takes the quantum circuit HDL description as an input. Two abstract inputs are considered, the HDL model and the assumed error model; the first influences how the HDL description is presented, while the second one dictates the test scenario by defining the start/stop simulation states (since qubits are equally prone to error, all the signals must be observed). HDL modeling of quantum circuits in order to attain efficient simulation is discussed in <A href="5.html#12">[34][35][36][38]. The outputs of the first cycle, which are also inputs for the simulation cycle, consist of a test scenario and an executable HDL model with the corresponding entanglement analysis, dictated by <A href="5.html#12">the bubble-bit encoded quantum states [36][38]. The output of the second cycle consists of time diagrams for all qubits, from the start to the stop states. Useful information, extracted from the raw, bubble-bit-represented, qubit traces are compared to correct qubit values, the result being the probabilistic accuracy threshold value, in the third cycle. The initialization and simulation cycles depend on specific aspects of quantum circuit <A href="5.html#12">simulation [35]. The data processing cycle is independent from the specific simulation framework and is aimed at determining the accuracy threshold as the main reliability measure that also defines the feasibility of the quantum circuit implementations. Suppose that, at simulation time t we observe signals { } 0 1 , ,..., n s s s . In our analysis, s i is the state observed during non-faulty simulation, so for the same state in a faulty environment we will have the state * i s . For validation of the quantum FTAMs, we need to compare i s with * i s . This can be done by using operator ( ) * , i i dif s s . This means that the total number of overall state errors at simulation time t is . The error rate on the overall observed states at moments ( 1 * 0 , t n i i i e dif s s = = ) 0 1 1 , ,..., m t t t will be given by 1 0 1 m sim j j t e m = = . The used FTAMs are only valid if the relationship between the experimental sim and the assumed singular error rate is of the order 2 ~ sim <A href="5.html#11">[19]. CONCLUSIONS This paper presented arguments in favor of two novel computing architectures for the purpose of addressing the challenges raised by the forthcoming nanoelectronics era. Distributed self-testing and self-repairing will probably become a must in the next years as centralized control logic is expected to become unable to harness the extremely large number of devices, all equally prone to errors, that will be integrated onto the same chip. Bio-inspired computing brings valuable techniques that explore the potential of massively parallel, distributed computation and fault-tolerance that will likely provide an essential help to jumpstart new nanoelectronic architectures. As one of the representatives of bio-inspired computing, the Embryonics project presents a hierarchical architecture that achieves fault tolerance through implementing an also hierarchical reconfiguration. A similar approach for maximizing fault tolerance is present in quantum computing, the QUERIST project; even if bio-inspired and quantum computing may seem dissimilar at a first glance, they both achieve fault tolerance by adapting the same techniques from classical computing and using essentially the same error model. Nanoelectronics will potentially change the way computing systems are designed, not only because of the sheer number of devices that will coexist onto the same chip, but also because of the sensitivity of these devices. 196 Figure 15. An overview of the QUERIST project Therefore, if nanoelectronics is to be employed to build dependable computing machines (a certain contradiction notwithstanding), valuable expertise in design can be drawn from natural sciences. While biology provides countless examples of successfully implemented fault tolerance strategies, physics offers theoretical foundations, both of which were found to share common ground. It is perhaps a coincidence worth exploring in digital computing. REFERENCES [1] Aharonov, D., Ben-Or, M. Fault Tolerant Quantum Computation with Constant Error. Proc. ACM 29th Ann. Symposium on Theory of Computing, El Paso, Texas, May 1997, pp. 176-188. [2] Avizienis, A., Laprie, J.C., Randell, B., Landwehr, C. Basic Concepts and Taxonomy of Dependable and Secure Computing. IEEE Transactions on Dependable and Secure Computing, 1, 1 (Jan-Mar 2004), 11-33. [3] Butts, M., DeHon, A., Golstein, S.C. Molecular Electronics: Devices, Systems and Tools for Gigagate, Gigabit Chips. Proc. Intl. Conference on CAD (ICCAD'02), 2002, pp. 433-440 . [4] Canham, R., Tyrrell, A. An Embryonic Array with Improved Efficiency and Fault Tolerance. Proc. IEEE NASA/DoD Conference on Evolvable Hardware, Chicago Il, 2003, 275-282 . [5] Gaisler, J. Evaluation of a 32-Bit Microprocessor with Built-In Concurrent Error Detection. Proc. 27th Annual Intl. Symposium on Fault-Tolerant Computing (FTCS-27), 1997, pp. 42-46. [6] Goldstein, S.C. The Challenges and Opportunities of Nanoelectronics. Proc. Government Microcircuit Applications and Critical Technology Conference (GOMAC Tech 04 ), Monterey, CA, March 2004. [7] Gottesman, D. Class of quantum error-correcting codes saturating the quantum Hamming bound. Phys. Rev. A 54, 1996, pp. 1862-1868. [8] Johnson, B.W. Design and Analysis of Fault-Tolerant Digital Systems. Addison-Wesley, 1989. [9] Laprie, J.-C. (Ed.). Dependability: Basic Concepts and Terminology. Dependable Computing and Fault-Tolerant Systems Series, Vol. 5, Springer-Verlag, Vienna, 1992. [10] Liden, P., Dahlgren, P., Johansson, R., Karlsson, J. On Latching Probability of Particle Induced Transients in Combinational Networks. Proc. Intl. Symposium on Fault-Tolerant Computing (FTCS-24), 1994, pp.340-349. [11] Mange, D., Sipper, M., Stauffer, A., Tempesti, G. Toward Robust Integrated Circuits: The Embryonics Approach. Proc. of the IEEE, vol. 88, No. 4, April 2000, pp. 516-541. [12] Mange, D. and Tomassini, M. eds. Bio-Inspired Computing Machines: Towards Novel Computational Architectures. Presses Polytechniques et Universitaires Romandes, Lausanne, Switzerland, 1998. [13] Von Neumann, J. The Computer and the Brain (2 nd edition). Physical Science, 2000. [14] Von Neumann, J. The Theory of Self-Reproducing Automata. A. W. Burks, ed. University of Illinois Press, Urbana, IL, 1966. [15] Von Neumann, J. Probabilistic Logic and the Synthesis of Reliable Organisms from Unreliable Components. In C.E. Shannon, J. McCarthy (eds.) Automata Studies, Annals of Mathematical Studies 34, Princeton University Press, 1956, 43-98. [16] Nielsen, M.A., Chuang, I.L. Quantum Computation and Quantum Information. Cambridge University Press, 2000. [17] Ortega, C., Tyrrell, A. Reliability Analysis in Self-Repairing Embryonic Systems. Proc. 1st NASA/DoD Workshop on Evolvable Hardware, Pasadena CA, 1999, 120-128. [18] O'Connor, P.D.T. Practical Reliability Engineering. John Wiley & Sons, 4 th edition, 2002. [19] Preskill, J. Fault Tolerant Quantum Computation. In H.K. Lo, S. Popescu and T.P. Spiller, eds. Introduction to Quantum Computation, World Scientific Publishing Co., 1998. 197 [20] Prodan, L. Self-Repairing Memory Arrays Inspired by Biological Processes. Ph.D. Thesis, "Politehnica" University of Timisoara, Romania, October 14, 2005. [21] Prodan, L., Udrescu, M., Vladutiu, M. Survivability Analysis in Embryonics: A New Perspective. Proc. IEEE NASA/DoD Conference on Evolvable Hardware, Washington DC, 2005, 280-289. [22] Prodan, L., Udrescu, M., Vladutiu, M. Self-Repairing Embryonic Memory Arrays. Proc. IEEE NASA/DoD Conference on Evolvable Hardware, Seattle WA, 2004, 130-137 . [23] Prodan, L., Tempesti, G., Mange, D., and Stauffer, A. Embryonics: Electronic Stem Cells. Proc. Artificial Life VIII, The MIT Press, Cambridge MA, 2003, 101-105. [24] Prodan, L., Tempesti, G., Mange, D., and Stauffer, A. Embryonics: Artificial Cells Driven by Artificial DNA. Proc. 4th International Conference on Evolvable Systems (ICES2001), Tokyo, Japan, LNCS vol. 2210, Springer, Berlin, 2001, 100-111. [25] Prodan, L., Tempesti, G., Mange, D., and Stauffer, A. Biology Meets Electronics: The Path to a Bio-Inspired FPGA. In Proc. 3rd International Conference on Evolvable Systems (ICES2000), Edinburgh, Scotland, LNCS 1801, Springer, Berlin, 2000, 187-196. [26] Rimen, M., Ohlsson, J., Karlsson, J., Jenn, E., Arlat, J. Validation of fault tolerance by fault injection in VHDL simulation models. Rapport LAAS No.92469, December 1992. [27] Rimen, M., Ohlsson, J., Karlsson, J., Jenn, E., Arlat, J. Design guidelines of a VHDL-based simulation tool for the validation of fault tolerance. Rapport LAAS No93170, Esprit Basic Research Action No.6362, May 1993. [28] Shivakumar, P., Kistler, M., Keckler, S.W., Burger, D., Alvisi, L. Modelling the Effect of Technology Trends on the Soft Error Rate of Combinational Logic. Proc. Intl. Conference on Dependable Systems and Networks (DSN), June 2002, pp. 389-398. [29] Shor, P. Fault-tolerant quantum computation. arXiv.org:quant-ph/9605011, 1996. [30] Shor, P. Algorithms for Quantum Computation: Discrete Logarithms and Factoring. Proc. 35th Symp. on Foundations of Computer Science, 1994, pp.124-134. [31] Sipper, M., Mange, D., Stauffer, A. Ontogenetic Hardware. BioSystems, 44, 3, 1997, 193-207. [32] Sipper, M., Sanchez, E., Mange, D., Tomassini, M., Perez-Uribe , A., Stauffer, A. A Phylogenetic, Ontogenetic and Epigenetic View of Bio-Inspired Hardware Systems. IEEE Transactions on Evolutionary Computation, 1, 1, April 1997, 83-97. [33] Steane, A. Multiple Particle Interference and Quantum Error Correction. Proc. Roy. Soc. Lond. A 452, 1996, pp. 2551. [34] Udrescu, M. Quantum Circuits Engineering: Efficient Simulation and Reconfigurable Quantum Hardware. Ph.D. Thesis, "Politehnica" University of Timisoara, Romania, November 25, 2005. [35] Udrescu, M., Prodan, L., Vladutiu, M. Simulated Fault Injection in Quantum Circuits with the Bubble Bit Technique. Proc. International Conference &quot;Adaptive and Natural Computing Algorithms&quot;, pp. 276-279. [36] Udrescu, M., Prodan, L., Vladutiu, M. The Bubble Bit Technique as Improvement of HDL-Based Quantum Circuits Simulation. IEEE 38th Annual Simulation Symposium, San Diego CA, USA, 2005, pp. 217-224. [37] Udrescu, M., Prodan, L., Vladutiu, M. Improving Quantum Circuit Dependability with Reconfigurable Quantum Gate Arrays. 2nd ACM International Conference on Computing Frontiers, Ischia, Italy, 2005, pp. 133-144. [38] Udrescu, M., Prodan, L., Vladutiu, M. Using HDLs for describing quantum circuits: a framework for efficient quantum algorithm simulation. Proc. 1st ACM Conference on Computing Frontiers, Ischia, Italy, 2004, 96-110. [39] Tempesti, G. A Self-Repairing Multiplexer-Based FPGA Inspired by Biological Processes. Ph.D. Thesis No. 1827, Logic Systems Laboratory, The Swiss Federal Institute of Technology, Lausanne, 1998. [40] Tempesti, G., Mange, D., Petraglio, E., Stauffer, A., Thoma Y. Developmental Processes in Silicon: An Engineering Perspective. Proc. IEEE NASA/DoD Conference on Evolvable Hardware, Chicago Il, 2003, 265-274. [41] Viamontes, G., Markov, I., Hayes, J.P. High-performance QuIDD-based Simulation of Quantum Circuits. Proc. Design Autom. and Test in Europe (DATE), Paris, France, 2004, pp. 1354-1359. [42] Yu, Y., Johnson, B.W. A Perspective on the State of Research on Fault Injection Techniques. Technical Report UVA-CSCS-FIT-001, University of Virginia, May 20, 2002. [43] ***. ITRS International Technology Roadmap for Semiconductors , Emerging Research Devices, 2004, http://www. itrs.net/Common/2004Update/2004_05_ERD.pdf [44] ***. Society of Reliability Engineers, http://www.sre.org/ pubs/ [45] ***. http://www.dependability.org/wg10.4/ 198

emerging technologies;Self replication;Embryonics;Computing technology;Error detection;Fault tolerance;Digital devices;Computing architecture;environment;Soft errors;Dependable system;Computing system;System design;Correction techniques;bio-inspired digital design;Bio-inspired computing;Reliability;Dependability;Nano computing;Failure rate;Emerging technologies;Nanoelectronics;bio-inspired computing;Self repair;evolvable hardware;Computer system;quantum computing;fault-tolerance assessment;QUERIST;Bio-computing;reliability;Quantum computing

50

Building Sustainable Community Information Systems: Lessons from a Digital Government Project

This paper introduces a rationale for and approach to the study of sustainability in computerized community information systems. It begins by presenting a theoretical framework for posing questions about sustainability predicated upon assumptions from social construction of technology and adaptive structuration theories. Based in part on the literature and in part on our own experiences in developing a community information system, we introduce and consider three issues related to sustainability: stakeholder involvement, commitment from key players, and the development of critical mass.

INTRODUCTION New technologies make it feasible and in many cases practical for individuals, groups, and organizations to collaborate in the development of joint information systems. In fact, over the last three decades of evolution, few applications of information technology have stimulated so much interest on the part of so many. Collaborative information systems are attractive to users because they make it possible to find information from diverse sources in an easy and efficient way. Such systems make good sense for information providers because it becomes possible to attract a larger audience than a solitary effort might otherwise be able to command and to pool resources to achieve certain economies in scale and technology expense. The advantages of collaborative computerized information systems have been widely recognized, but this has been particularly the case for those with the goal of making community information more available, accessible, and oriented toward community development. Computerized community information systems are diverse in form and, over time, have come to be known by many different names, including community bulletin boards, civic networks, community networks, community information networks, televillages, smart communities, and Free-Nets. They have been initiated by many different sponsors, including government organizations at the federal, state, and local levels, academic organizations, libraries, and ad hoc groups of citizens that may or may not later transform their enterprises into not-for-profit organizations [7]. With respect to longevity, these projects have come and gone, only to be replaced by newer and more sophisticated manifestations of the same basic information sharing capabilities. Consistent with the evolution of technology over the last thirty years, Kubicek and Wagner [14] analyze the historical trajectory of community networks to understand how these applications have evolved over time based upon their animating ideas, the zeitgeist of the time, the state of technology access, and the kinds of services such applications make available. Their analysis makes it possible to see that there has never been a standard for design or operation when it comes to community information systems. Instead, each such project has been very much a social experiment, born of a cluster of varied ideas related to the general theme of using technology to promote the development of vibrant geographically-based communities. Since there has been no standard to follow, each instance of computerized community information system can be seen as an experiment in accommodating the tensions between access to hardware/software infrastructure, design of the particular application or system, user needs, and the initiating and ongoing resources that support these efforts. These projects can be resource intensive; thus, a variety of institutional actors have lent their financial support particularly over the past decade. The successive rounds of funding for community technology projects by the Department of Commerce's National Telecommunications and Information Administration (now called the Technology Opportunities Program) is a case in point. The Digital Government Program of the National Science Foundation has Copyright held by the author 145 provided support for such ventures, as have many private foundations and technology corporations. From the perspective of funding organizations, the nature of the experiment at the heart of CCINs is essentially this: how to build applications that achieve their civic goals, that provide services perceived as valuable by their users, and that can command continuing support from the community beyond the horizon of initial funding. From a purely academic perspective, the more general question centers on, as Venkatesh [28] has put it, the "lifecycle" of community information systems. More specifically, we wish to know how such systems "originate, stabilize, and change in their sociohistorical context" (p. 339). We do not have extensive knowledge about the extent to which community information systems achieve their goals, endure over time, or the conditions that facilitate effectiveness and sustainability. However, based on what we do know, it is apparent that such enterprises are fragile. Perhaps the closest we have come to a standard or model is the relatively extensive set of experiments in community networking in the 1990s called Free-Nets , which were fashioned after the public broadcasting system and intended to serve their localities by providing access to wide-area computer networks and information about the community. Founded in 1989, the National Public Telecomputing Network, an umbrella organization for Free-Nets, went bankrupt in 1996. After successive decreases in the cost of computing equipment and Internet access, and the development of the World Wide Web, many Free-Nets went out of business [14]. Studies of community networks funded by the federal and state governments also suggest that community information systems have difficulty enduring beyond their initial funding [26] [21]. In this paper, we introduce and consider conditions that facilitate the sustainability of computerized community information systems. We base our discussion in part on our own efforts to develop a community information system called Connected Kids in Troy, New York, a project that began in a formal sense in 1999 and continues today. We begin by presenting a theoretical framework for posing questions about sustainability based on the social construction of technology and adaptive structuration theory. Drawing on the literature as well as our experiences, we introduce and discuss three issues we believe to be critically related to sustainability: stakeholder involvement, commitment from key players, and critical mass. THEORETICAL FRAMEWORK All computerized community information systems are designed, although whether researchers and participants understand the significance of design and its relevance to sustainability varies from context to context. In some cases, where community networks have originated as the indigenous creation of technology-savvy citizens, it may appear to researchers that the design of the information system is a natural expression of community development unfettered by theoretical considerations. However, in other cases, design is taken more seriously and treated as an element that can be purposefully controlled in order to achieve particular kinds of effects. In either case, we argue that the material form, functionalities, conceptual configuration, and impact of technology is shaped by the uses, goals, interests, and ideologies of those who participate in its development and others who use it following development. In the literature addressing the social construction of technology, this argument is frequently illustrated by showing how users appropriate new technologies for their own purposes, which may be contrary to those of designers (see e.g. [15] [25]). However, we take this position one step further by suggesting that community information systems, and information and communication technologies more broadly, reflect the interests, orientations, and indeed the nave social theories of their designers, as well as being shaped ex post facto by their users [8]. On this basis, we have argued that academic researchers need to become involved in the process of technology design as a way of exploring how to improve the design of technology and as a way to test social theory, including communication, information, and democratic theory. However, our position suggests that users of information systems must also be included in their initial conceptualization and design in order to develop systems that reflect users' needs, goals, and values. This leads quite naturally to creating interdisciplinary (e.g. computer scientists, information scientists, social scientists) application design teams that provide for participation by community members; it is in such collaborative arrangements that sustainable community information systems may be designed. The social construction of technology argues that technologies are shaped by both designers and users and suggests that information system design be undertaken collaboratively by those implicated in both the technical and social conceptualization of the system. However, issues of sustainability ultimately focus on reproduction of the system. Once designed, information systems must be deployed, and once deployed, they must be re-enacted on a routine basis by their users to be sustained. Adaptive structuration theory is one of the most fully developed theoretical perspectives for understanding how new technologies come to reproduce social structures or to generate structural change in particular social contexts. DeSanctis and Poole [4] base their work on structuration processes originally described by Giddens [5]. Giddens [5] suggests that technologies in organizations either reproduce existing social structure or change social structure by virtue of the kinds of structures that are instantiated when social actors use technologies. Structure consists of rules and resources that actors draw upon to produce social behavior. For DeSanctis and Poole [4], social structures are physically incorporated in new technologies in two complementary ways. First, technologies embody rules and resources embedded in the form of particular material capabilities, functionalities, and features that comprise a variety of behavioral options to be used in constructing social action. Second, the "spirit" of a technology, also considered to be a property of the technology, expresses the values and goals that are brought to bear upon the tasks the technology was originally intended to accomplish. Together the features and spirit of a technology comprise its "structural potential," or the range of possible actions that users can draw upon to constitute or reproduce social structures in technology use. Orlikowski [17] disputes a portion of this conceptualization, noting that, according to Giddens [5], structure has a "virtual" rather than material existence, and thus can never be physically incorporated into technology. Instead, &quot;[w]hile a technology may be seen to embody particular symbol and material properties, it does not embody structures because those are only instantiated in practice&quot; ([17], p. 206) and, if reproduced, are systematically repeated over time. 146 Orlikowski's [17] point is that users may draw upon only some of a technology's features, and may do that in ways that depart substantially from the original conceptualizations of designers. In essence, users "enact" technology in their collective, systematic, and routine use of a technology, reproducing some of the technology and some of its associated structures through practice. Orlikowski's [17] term "technologies-in-practice" references the idea that as users engage selectively with particular technological features, particular structures, or sets of rules and resources associated with the technology, are selectively reconstituted. Thus, a technology-in-practice is a &quot;repeatedly experienced, personally ordered and edited version of the technological artifact, being experienced differently by different individuals and differently by the same individuals depending on the time or circumstances&quot; ([17], p. 408). Applied to sustainability, our questions center on the conditions under which users "appropriate" the system. For community information systems, there are generally two kinds of users-information providers and information consumers--and, of course, the same individuals may play both user roles. Thus, our questions become: Under what conditions do users collectively and routinely draw upon and apply particular features of a community information system? When do they reference the way their system &quot;should&quot; work in order to construct a shared perspective about community action? Through regular and routine enactments of technology in regular use, users reproduce the rules and resources or structures of community life that are instantiated in technology use. This is not to say that "unfaithful" appropriations, or those that are out-of-line with the spirit of the technology, cannot occur; but it is to say that it is unlikely they will sustain a community information system. FACTORS RELATED TO SUSTAINABLE COMMUNITY INFORMATION SYSTEMS We begin our discussion of factors related to sustainability by distinguishing between the effectiveness and the sustainability of computerized community information systems. Community information systems are designed and advocated with many goals in mind, some of which focus on traditional issues of community development, such as decreasing unemployment, stimulating economic growth, improving health and social welfare, and others focus on building social capital, or enhancing interest and participation in government decision making processes. The issue of effectiveness addresses whether such systems are achieving the goals for which they were designed. Sustainability, on the other hand, addresses whether the information system is able to endure past its initial launching phase, whether it is used and reproduced by its intended audience, and whether it can continue to attract resources beyond those obtained for initial development and deployment. Clearly these two concepts are not irrelevant to each other, but neither are they the same. It is possible that questions of sustainability logically precede those of effectiveness, but there may also be important relationships between effectiveness and sustainability. Sustainability has long been a consideration in the development of information systems. Indeed, the failure rate of new IT applications in the public sector has motivated significant interest in addressing the issue of sustainability and speculation about the extent to which participation in system development is related ultimately to system adoption and use [10] [11]. More specifically, government services are increasingly out-sourced to not-for-profit organizations that may not be experienced in collaboration [3]. Information technology makes it possible for organizations to collaborate in providing information but whether or not such collaborations actually take place is more than a technical issue. The development of any information system, and particularly collaborative systems, requires organizations to change, in a very real way, some of their routine modes of operation and incorporate new behaviors. Scholl's [22] research finds that stakeholder involvement and the commitment of senior executives to be highly related to the integration of e-government projects into business process change for government organizations. Stakeholder involvement has long been acknowledged as a key element in the construction of community information system, although applied to this context rather than that of traditional hierarchical organizations, the idea bears further scrutiny. We have also seen the commitment of key executives playing a role in our own development work. We discuss each of these two ideas at some length below, and add a third: development of a critical mass of users. 3.1 Stakeholder Involvement Our work was motivated in part by Schuler's [23] invitation to academic researchers to collaborate with communities in building community networking projects. At the time, it was fair to characterize our institution's hometown, Troy, NY, as a "digitally divided" community. Our experiences suggested that new technologies and their potential seemed to be of interest to the members of the community (see [9]). But many community and government organizations lacked access to hardware and network connections as well as the expertise needed for using this equipment. It seemed most likely that we would need to do more to generate interest in the development of a community information service in order to stimulate participation from likely stakeholders in such a project. Connected Kids was conceived in Fall 1999 in the course of discussions among Troy City Government representatives on the topic of how new technologies might usefully be employed to provide services to the community. At the time we learned that the mayor sought to reinvigorate the City's office of youth services and had speculated about whether these technologies could be used to provide one of that office's primary and most popular functions, which was to disseminate information about resources and programs sponsored by not-for-profit organizations as well as those sponsored by Troy's own Department of Recreation. It seemed clear to us the World Wide Web might indeed be used for such purposes. Thus, Connected Kids was conceived as both a digital government project as well as a community information system. We received initial assurances that the City would administer the information system after it had been successfully designed and deployed. Connected Kids began with sensitivity to the need for stakeholder involvement, particularly that of participating organizations that we hoped would be information providers. We were aware that the "best way to kill a community network," was to fail to involve the community in system development [24]. We took seriously Gygi's [6] prediction that the degree of community involvement 147 and the extent to which the project represented community interests and participation would likely affect political and economic outcomes. Thus, although our project began initially as a collaboration between academic researchers and government administrators, we moved quickly to invite community organizations to participate at an orientation meeting in February 2000 and held a series of focus group discussions in October 2000 in which we explored with representatives of participating organizations how such an information system might be conceptualized to best meet their information needs. In Fall 2001 and Winter 2002 we undertook a series of participatory design sessions in which representatives of participating organizations were introduced to portions of a system prototype based on their previous contributions and asked to describe their experiences and suggest improvements. Finally, as we designed interfaces in Summer and Fall 2002, we again consulted with representatives of participating organizations in user testing sessions. By Fall 2003 and Winter 2004, we had demonstrated the system and trained numerous representatives of participating organizations, who reportedly found our interface pages easy to use. However, these same organizations were not spontaneously--or frequently-entering information about their programs or activities for youth. Based on contributions from our collaborating organizations, the design of Connected Kids reflected much of the best wisdom about community information systems: the system could be used to both create and easily update data [2] [13]; we had involved end user groups (kids and their parents) in the design as well [27] [12]; and the system focused principally on information deemed crucial by our participating organizations, information that we expected had the capacity to be integrated into the routine lives of the communities they serve [21]. Further, access to technology lost its urgency as an issue, since it is no longer the case that our participating organizations lack access to networking technology. Thus, we did not attribute our problems with data entry to system attributes. Instead, we considered the suggestion by Scott and Page [18] that "sustainable technologies are processes (authors' emphasis); they are not products." In traditional hierarchical organizations, lower levels of stakeholder involvement may be sufficient for system acceptance. However, a community information system requires that members of the community contribute information and it must be seen to be in their continuing interest to do so. In Fall 2004, we have sought to create a quasi-formal governance body to administer the project, a Connected Kids Advisory Board, recruiting representatives from 10 organizations (from among the most influential) to commit to guiding the short-term future of the project (approximately 1 year) as we transition to system deployment in Spring 2005. Our Board has now met for several months, and it remains to be seen whether this vehicle will foster a sufficient level of system participation, perhaps ownership, to sustain Connected Kids through deployment and beyond. 3.2 Commitment from Key Players Scholl [22] finds that support from key executives is critical to incorporating e-government projects into an organization's business processes, and our experience underscores this finding. In fact, we would expand the range of individuals likely to be considered "key." Not only are senior executives important, but so also are others in the organization that have any significant job-related association with the information system under development. Application development projects take place over potentially long periods of time and involve many different individuals in many different roles. Job occupants in the public sector may be comparatively stable, but they are not permanent. Those who champion an application development project may not be around when it is time to deploy the system. What is generally not recognized when academic researchers undertake technology projects in organizational contexts is that they may need to become the primary advocates for deployment of the project. This is not a typical role for researchers, who may with ample justification see their obligations confined to simply performing the research or developmental work on the project. However, researchers who seek to develop sustainable products may find themselves required to situate the project politically within the organization or group of organizations for which it was originally intended. They may in fact be the only individuals who can play this particular role. In the case of Connected Kids, we secured commitment from both the mayor and deputy mayor of Troy, along with, of course, that of our primary organizational liaison. We continued to work for quite some time, reporting regularly on progress to our liaison, without realizing that this individual was getting progressively involved in turf battles with another technology-oriented actor in city government. As our liaison's influence within city government eroded, so also did support for our project without our awareness. Once we understood what was happening, we acted quickly, and luckily in sufficient time, to re-establish the importance of the project with the mayor and deputy mayor. From that point hence our primary liaison was the deputy mayor. Unfortunately, mayoral administrations come and go, and the administration that was our primary government partner was voted out of office in November 2003. Within six months all the individuals who had any primary working relationship with our project were gone, and we faced the need to re-create commitment with a new mayor and deputy mayor, a process that took considerable time and that has delayed implementation by nearly a year. Of course, this is not something we could have prevented. However, it is interesting to note that our new liaison with city government is an individual who had worked in city government under both administrations. 3.3 Critical Mass of Users Ultimately, for a community information system to endure, it must establish a significant number of regular users, who enact the technology for at least some of the purposes for which it was originally intended, and in so doing, reproduce community structures that are instantiated in the technology. In our case, this means bringing an audience of end users to the system who are interested in information about youth that is disseminated through it. Connected Kids is in many respects similar to an electronic "public good" [20], that is, a product established through the contributions and for the benefit of a set of actors that also has the effect of benefiting other users. In our case, the system was designed by and for youth organizations, which serve as information providers. We have sought to show how these organizations may appropriate the technology and accomplish what Bannon and Griffin [1] suggest, which is to use the technology as a means to "further their own 148 ends" (p. 48) rather than as an end in itself. However, the added value of a collaborative information system is that in bringing an audience to information distributed by one organization, that audience is also available to peruse the information of other organizations. Thus the overall effect is to increase the cumulative size of the audience for all involved. Further, the external user audience benefits from the ease of accessing information from a wide variety of organizations that all provide services for youth. Patterson and Kavanaugh [19] argue that the pro-social benefits of a public good are achieved when the system achieves a critical mass of users. In our case, this would equal the number of users that information providers consider to make it worth their continuing efforts to input information about their activities. As Markus [16] points out, the number of users will depend on the diversity and value of the information available through the system. Thus, sustainability is dependent on the reciprocal interdependence of both information providers and information users. Both information in the system and use must achieve critical mass, and this must happen relatively soon after deployment. Our strategy is to bring both of these activities together in time. We have asked the Connected Kids Advisory Board to develop a marketing campaign that will accompany the deployment of the system and they are currently embarked on this activity. Bolstered by the participation of RPI and the City of Troy, we seek to attract a large external audience to experiment with the system. Advisory Board members recognize that the success of the marketing campaign depends on the presence of considerable amount of high quality information in the system, and have committed to providing it. In this way, we seek to jumpstart a virtuous circle in which sufficient quantities of information and use reciprocally reinforce each other creating a critical mass of information providers and consumers. CONCLUSION Poised for deployment, Connected Kids enables us to test a range of expectations generated by this set of considerations regarding sustainability. Within the next year, we should be able to assess relationships related to stakeholder involvement such as those between factors such as perceptions of involvement in system design and administration; actual participation in system design and testing activities; and perceptions of ownership with outcomes such as the extent of data contributed to the system, perceptions of commitment to the system, and significance of organizational resources devoted to participation. Further, we should be able to assess relationships between the perceived amount and quality of information in the system and end user satisfaction, likelihood of returning to the system, and interest in becoming more involved in system activities. Of course, we will continue to be able to comment anecdotally on what we have learned about the politics of technology diffusion in public sector organizations. REFERENCES [1] Bannon, L.J., and Griffin, J. New technology, communities, and networking: Problems and prospects for orchestrating change. Telematics and Informatics, 18, (2001), 35-49. [2] Cowan, D.D., Mayfield, C. T., Tompa, F. W., and Gasparini, W. New role for community networks. Communications of the ACM, 41, 4, (1998), 61-63. [3] Dawes, S.S., Bloniarz, P. A., and Kelly, K. L. Some Assembly Required: Building a Digital Government for the 21st Century. Albany, NY: Center for Technology in Government, 1999. http://www.ctg.albany.edu/research/workshop/dgfinalreport. pdf [4] DeSanctis, G., and Poole, M.S. Capturing the complexity in advanced technology use: Adaptive structuration theory. Organization Science, 5, (1994), 121-147. [5] Giddens, A. The Constitution of Society. University of California Press, Berkeley, CA, 1984. [6] Gygi, K. Uncovering Best Practices: A Framework for Assessing Outcomes in Community Computer Networking, 1996. http://www.laplaza.org/about lap/archives/cn96/gygi.html [7] Harrison, T., and Stephen, T. Researching and creating community networks. In Doing Internet Research, S. Jones ed. Sage, Newbury Park, CA, 1999, 221-241. [8] Harrison, T., and Zappen, J. Methodological and theoretical frameworks for the design of community information systems. Journal of Computer-Mediated Communication, 8, 3 (2003). http://www.ascus.org/jcmc/vol8/issue3 [9] Harrison, T.M., Zappen, J.P., and Prell, C.L. Transforming new communication Technologies into Community Media. In Community Media in the Information Age: Perspectives and Prospects, N. Jankowski and O. Prehn eds., Hampton, Cresskill, NJ, 2002, 249-269. [10] Heeks, R. Better information age reform: Reducing the risk of information systems failure. In Reinventing government in the information age: International Practice in IT-Enabled Public Sector Reform, R. Heeks ed., Routledge, London, 1999, 75-109. [11] Heeks, R., and Bhatnager, S. Understanding success and failure in information age reform. In Reinventing government in the information age: International Practice in IT-Enabled Public Sector Reform, R. Heeks ed., Routledge, London, 1999, 50-73. [12] Howley, K. Equity, access, and participation in community networks. Social Science Computer Review, 16, (1998), 402-410. [13] Keenan T. P., and Trotter, D. M. The changing role of community networks in providing citizen access to the Internet. Internet Research: Electronic Networking Applications and Policy, 9, 2, (1999) 100-108. [14] Kubicek, H., and Wagner, R.M. Community networks in a generational perspective: The change of an electronic medium within three decades. Information, Communication, and Society, 5, (2002), 291-319. [15] Lievrouw, L., and Livingstone, S. The social shaping and consequences of ICTs. In The Handbook of New Media, L. Lievrouw and S. Livingstone, Eds. Sage, London, 2002, 121 . 149 [16] Markus, L. Toward a "critical mass" theory of interactive media. Communication Research, 14, (1987), 491511. [17] Orlikowski, W. Using technology and constituting structures: A practice lens for studying technology in organizations. Organization Science, 11, (2000), 404-428. [18] Page, M. and Scott, A. Change agency and women's learning. Information, Communication & Society, 4, 4 (2001), 528-559. [19] Patterson, S.J., and Kavanaugh, A.L. Building a sustainable community network: An application of critical mass theory. Electronic Journal of Communication, 2, (2001). http://www.cios.org/www/ejc/v11n201.htm. [20] Rafaeli, S., and LaRose, R. Electronic bulletin boards, and "public goods" explanations of collaborative mass media. Communication Research, 20, 2, (1993), 277-297. [21] Rosenbaum, H. Web-based community networks: A study of information organization and access. ASIS '98: Access in the global information economy. In Proceedings of the 61st Annual Meetings of the American Society for Information Society, 35, 1998, 516-527. [22] Scholl, H.J. Current practices in E-government-induced business process change. Proceedings of the National Conference on Digital Government, Digital Government Research Center, 2004, 99-108. [23] Schuler, D. Community Computer Networks: An Opportunity for Collaboration among Democratic Technology Practitioners and Researchers, 1997. http://www.sn.org/ip/commnet/oslo-197.text [24] Schuler, D. How to Kill Community Networks. Hint: We May Have Already Started, 1996. http://www.scn.org/ip/commnet/kill-commnets.html [25] Sproull, L., and Kiesler, S. Connections: New Ways of Working in the Networked Organization. MIT Press, Cambridge, MA, 1991. [26]

critical mass;construction of technology;key players;community network;computerized community information system;participatory design;sustainability;skateholder involvement;Community networks

51

Can Machine Learning Be Secure?

Machine learning systems offer unparalled flexibility in dealing with evolving input in a variety of applications, such as intrusion detection systems and spam e-mail filtering. However , machine learning algorithms themselves can be a target of attack by a malicious adversary. This paper provides a framework for answering the question, "Can machine learning be secure?" Novel contributions of this paper include a taxonomy of different types of attacks on machine learning techniques and systems, a variety of defenses against those attacks, a discussion of ideas that are important to security for machine learning, an analytical model giving a lower bound on attacker's work function, and a list of open problems.

INTRODUCTION Machine learning techniques are being applied to a growing number of systems and networking problems, particularly those problems where the intention is to detect anomalous system behavior. For instance, network Intrusion Detection Systems (IDS) monitor network traffic to detect abnormal activities, such as attacks against hosts or servers. Machine learning techniques offer the benefit that they can detect novel differences in traffic (which presumably represent attack traffic) by being trained on normal (known good) 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. ASIACCS'06, March 2124, 2006, Taipei, Taiwan Copyright 2006 ACM 1-59593-272-0/06/0003 ... $ 5.00 attack (known bad) traffic. The traditional approach to designing an IDS relied on an expert codifying rules defining normal behavior and intrusions [26]. Because this approach often fails to detect novel intrusions, a variety of researchers have proposed incorporating machine learning techniques into intrusion detection systems [1, 16, 18, 24, 38, 41]. On the other hand, use of machine learning opens the possibility of an adversary who maliciously "mis-trains" a learning system in an IDS. A natural question arises: what techniques (in their attacks) can an adversary use to confuse a learning system? This paper explores the larger question, as posed in the title of this paper, can machine learning be secure? Specific questions that we examine include: Can the adversary manipulate a learning system to permit a specific attack? For example, can an attacker leverage knowledge about the machine learning system used by a spam e-mail filtering system to bypass the filtering? Can an adversary degrade the performance of a learning system to the extent that system administrators are forced to disable the IDS? For example, could the attacker confuse the system and cause valid e-mail to be rejected? What defenses exist against adversaries manipulating (attacking) learning systems? More generally, what is the potential impact from a security standpoint of using machine learning on a system ? Can an attacker exploit properties of the machine learning technique to disrupt the system? The issue of machine learning security goes beyond intrusion detection systems and spam e-mail filters. Machine learning is a powerful technique and has been used in a variety of applications, including web services, online agent systems, virus detection, cluster monitoring, and a variety of applications that must deal with dynamically changing data patterns. Novel contributions of this paper include a taxonomy of different types of attacks on machine learning techniques and systems, a variety of defenses against those attacks, a discussion of ideas that are important to security for machine Invited Talk learning, an analytical model giving a lower bound on attacker's work function, and a list of open problems. The rest of this paper is organized as follows: Section 2 discusses machine learning and how it is typically used in a system, Section 3 develops a taxonomy of attacks, Section 4 introduces potential defenses against attacks and explores their potential costs, Section 5 identifies several of the ideas that are important to security for machine learning, Section 6 presents an analytical model that examines an attack to manipulate a naive learning algorithm, Section 7 discusses related work, potential research directions, and our conclusions REVIEW A machine learning system attempts to find a hypothesis function f that maps events (which we call points below) into different classes. For example, an intrusion detection system would find a hypothesis function f that maps an event point (an instance of network behavior) into one of two results: normal or intrusion. One kind of learning system called supervised learning works by taking a training data set together with labels identifying the class for every point in the training data set. For example, a supervised learning algorithm for an IDS would have a training set consisting of points corresponding to normal behavior and points corresponding to intrusion behavior. The learning algorithm selects the hypothesis function f that best predicts the classification of a point. More complicated learning algorithms can deal with event points that are both labeled and unlabeled and furthermore can deal with continuous streams of unlabeled points so that training is an ongoing process. In this paper, we call these algorithms online learning systems. This remainder of this subsection presents a concise overview of concepts in statistical learning theory. The presentation below is formal and can be skipped on a first reading. For a fuller discussion with motivation, refer to [11, 31]. A predictive learning problem is defined over an input space X, an output space Y, and a loss function : Y Y R. The input to the problem is a training set S, specified as {(x i , y i ) X Y}, and the output is a hypothesis function f : X Y. We choose f from a hypothesis space (or function class) F to minimize the prediction error given by the loss function. In many cases, researchers assume stationarity , that the distribution of data points encountered in the future will be the same as the distribution of the training set. Stationarity allows us to reduce the predictive learning problem to a minimization of the sum of the loss over the training set: f = argmin f F X (x i ,y i ) S (f (x i ), y i ) (1) Loss functions are typically defined to be non-negative over all inputs and zero when f (x i ) = y i . A commonly used loss function is the squared-error loss sq (f (x i ), y) = (f (x i ) -y) 2 . The hypothesis space (or function class) F can be any representation of functions from X to Y, such as linear functions, polynomials, boolean functions, or neural networks. The choice of F involves a tradeoff between expressiveness and ability to generalize. If F is too expressive, it can overfit the training data. The extreme case is a lookup table that maps x i to y i for each instance of the training set but will not generalize to new data. A linear function, on the other hand, will generalize what it learns on the training set to new points, though it may not be sufficiently expressive to describe intricate data sets. We typically use simple function classes, such as linear functions, to avoid overfitting. We can describe a more general learning problem by dropping the requirement that the training examples include all the labels y i . The case where all labels are present is referred to as supervised learning, when no labels are present the problem is unsupervised, and when some labels are present the problem is semi-supervised. In all these cases we can pose the learning problem as the minimization of some measure over the training set: f = argmin f F X (x i ) S L(x i , f ) (2) 2.2 Terminology and Running Example To illustrate some of our contributions, we use a running example throughout this paper: a network Intrusion Detection System (IDS). This IDS receives network events x X and classifies each event x as either f (x) = normal or f (x) = intrusion. The literature describes a number of algorithms for learning f over time, but we wish to consider the impact of malicious input on the learning algorithm. This paper poses the question: can a malicious party send events to the IDS that will cause it to malfunction? Possible types of attacks on the IDS include attacks on the learning algorithm, causing the IDS to create an f that misclassifies events. As we discuss in the next section, this is only one of a number of types of attacks that an adversary can make on an IDS. It is important to be careful about notation here. When we speak of attacks, we mean an attack on the learning system (e.g., the learner in an IDS). Attacks may try to make the learner mis-learn, fail because of denial of service, report information about its internal state, etc. "Attack" should be distinguished from "intrusion." An attack targets a learning system; an intrusion targets a computer system (such as a system protected by an IDS). While many researchers use the word "attack" to include intrusions, in this paper we are careful to use the word "attack" only to mean an attack on a learner. We do not want to restrict ourselves to particular learning algorithms used by intrusion detection systems to choose hypotheses. However, we allow adversaries that have deep understanding of the learning algorithms. Similarly, we do not discuss mechanisms for translating network level events into a form relevant to the learner. We call each unit of data seen by the learner a data point, or simply a point. In the context of the IDS, our discussion encompasses continuous, discrete, or mixed data. We assume that X is a metric space, allowing us to freely discuss distances Integrity Availability Causative: Targeted Permit a specific intrusion Create sufficient errors to make system unusable for one person or service Indiscriminate Permit at least one intrusion Create sufficient errors to make learner unusable Exploratory: Targeted Find a permitted intrusion from a small set of possibilities Find a set of points misclassified by the learner Indiscriminate Find a permitted intrusion Table 1: The attack model. between points. Furthermore, we assume the set of points classified as normal by the IDS forms multiple contiguous subsets in X. The border of this set is called the decision boundary. Below, we consider a variety of scenarios and assumptions. ATTACKS We give relevant properties for analyzing attacks on machine learning systems. Influence Causative - Causative attacks alter the training process through influence over the training data. Exploratory - Exploratory attacks do not alter the training process but use other techniques, such as probing the learner or offline analysis, to discover information. Specificity Targeted - The specificity of an attack is a continuous spectrum. At the targeted end, the focus of the attack is on a particular point or a small set of points. Indiscriminate - At the indiscriminate end, the adversary has a more flexible goal that involves a very general class of points, such as "any false negative." Security violation Integrity - An integrity attack results in intrusion points being classified as normal (false negatives). Availability - An availability attack is a broader class of attack than an integrity attack. An availability attack results in so many classification errors, both false negatives and false positives, that the system becomes effectively unusable. These three axes define a space of attacks; Table 1 provides a concise summary. In causative attacks, the adversary has some measure of control over the training of the learner. An attack that causes the learner to misclassify intrusion points, for example an attack that fools an IDS into not flagging a known exploit as an intrusion, is a causative integrity attack. The distinction between targeted and indiscriminate causative integrity attacks is the difference between choosing one particular exploit or just finding any exploit. A causative availability attack causes the learner's performance to degrade. For example , an adversary might cause an IDS to reject many legitimate HTTP connections. A causative availability attack may be used to force the system administrator to disable the IDS. A targeted attack focuses on a particular service, while an indiscriminate attack has a wider scope. Exploratory attacks do not attempt to influence learning; they instead attempt to discover information about the state of the learner. Exploratory integrity attacks seek to find intrusions that are not recognized by the learner. 3.2 Online Learning A learner can have an explicit training phase or can be continuously trained (online learner). Online learning allows the learner to adapt to changing conditions; the assumption of stationarity is weakened to accommodate long-term changes in the distribution of data seen by the learner. Online learning is more flexible, but potentially simplifies causative attacks. By definition, an online learner changes its prediction function over time, so an adversary has the opportunity to shape this change. Gradual causative attacks may be difficult to detect. DEFENSES In this section we discuss potential defenses against attacks. This section describes speculative work, and the efficacy of these techniques in practice is a topic for future research. 4.1 Robustness To increase robustness against causative attacks we constrain the class of functions (hypotheses) that the learner considers. The constraint we consider is the statistical technique of regularization. Regularization extends the basic learning optimization in Equation (1) by adding a term J(f ) that penalizes complex hypotheses: f = argmin f F 8 &lt; : X (x i ,y i ) S (f (x i ), y i ) + J(f ) 9 = ; (3) Here adjusts the trade-off. The penalty term J(f ) can be as simple as the sum of squares of the parameters of f . Regularization is used in statistics to restrict or bias the choice of hypothesis when the problem suffers from lack of Integrity Availability Causative: Targeted Regularization Randomization Regularization Randomization Indiscriminate Regularization Regularization Exploratory: Targeted Information hiding Randomization Information hiding Indiscriminate Information hiding Table 2: Defenses against the attacks in Table 1. data or noisy data. It can also be interpreted as encoding a prior distribution on the parameters, penalizing parameter choices that are less likely a priori. Regularization and prior distributions can both be viewed as penalty functions in Equation (3) [42]. The constraint added to the learning problem by the penalty term may help our defenses in two ways. First, it has the effect of smoothing the solution, removing complexity that an adversary might exploit in attacks. Second, prior distributions can be a useful way to encode expert knowledge about a domain or use domain structure learned from a preprocess-ing step. In the simplest case, we might have a reasonable guess for the parameters (such as the mean) that we wish to refine; in a more complex situation, we could perform an analysis of a related dataset giving correlation information which informs a multivariate Gaussian prior on the parameters [28]. When the learner has more prior information (or constraints) on which to base the learning, there is less dependence on exact data fitting, so there is less opportunity for the adversary to exert influence over the learning process. 4.2 Detecting Attacks The learner can benefit from the ability to detect attacks even if they are not prevented. Detecting attacks can be difficult even when the adversary is not attempting to conceal them. However, we may be able to detect causative attacks by using a special test set. This test set could include several known intrusions and intrusion variants, as well as some random points that are similar to the intrusions. After the learner has been trained, misclassifying a disproportionately high number of intrusions could indicate compromises. To detect naive exploratory attacks, a separate clustering algorithm could be run against data classified by the learner. The sudden appearance of a large cluster near the decision boundary could indicate systematic probing. This type of defense is akin to port scan detection, which has become an arms race between port scanners and IDS [26]. Detecting an attack gives the learner information about the adversary's capabilities. This information may be used to reformulate defense strategies. As the adversary's control over the data increases, the best strategy for the learner is to ignore potentially tainted data. Otherwise, the adversary can exploit misplaced trust. These ideas have been formalized within the context of deception games [14, 32], which typically assume all players know the extent to which other players may manipulate data. However , if the parties estimate each other's abilities, more sophisticated strategies emerge. 4.3 Disinformation In some circumstances, the learner may be able to alter the data seen by the adversary. This strategy of disinformation has the goal of confusing the adversary's estimate of the learner's state. In the simplest case, the adversary would then be faced with a situation not unlike a learner under an indiscriminate causative availability attack. The goal of the learner is to prevent the adversary from learning the decision boundary. Please note how the roles of adversary and learner have been reversed. A more sophisticated learner could trick the adversary into believing that a particular intrusion was not included in the training set. This apparently permitted "intrusion" would act as a honeypot [27], causing the adversary to reveal itself. An increase in the incidence of that particular attack would be detected, revealing the existence of an adversary. In this case again, roles would reverse, and the adversary would face a situation analogous to a learner subjected to a targeted causative integrity attack. 4.4 Randomization for Targeted Attacks Targeted attacks hinge on the classification of one point or a small set of points. They are more sensitive to variations in the decision boundary than indiscriminate attacks because boundary movement is more likely to change the classification of the relevant points. This suggests randomization as a potential tool against targeted causative attacks. In such an attack, the adversary has to do a particular amount of work to move the decision boundary past the targeted point. If there is some randomization in the placement of the boundary and the adversary has imperfect feedback from the learner, more work is required . 0 4.5 Cost of Countermeasures The more we know about the distribution of training data, the less room there is for an adversary to manipulate the learner. The disadvantage, however, is that the legitimate data has less influence in the learning process. A tension exists between expressivity and constraint: as the learner includes more prior information, it loses flexibility to adapt to the data, but as it incorporates more information from the data, it becomes more vulnerable to attack. Equation (3) makes this tradeoff explicit with . In the adversarial scenario, this tradeoff becomes more relevant because the adversary may have influence over the data. Randomization increases the adversary's work, but it also will increase the learner's base error rate. Determining the right amount of randomization is an open problem. 4.6 Summary of Defenses Table 2 shows how our defenses discussed here relate to attack classes presented in Table 1. (Information hiding is an additional technique discussed in Section 5 below.) DISCUSSION A number of defenses and attacks upon machine learning algorithms hinge upon the types of information available to the adversary. Some of these involve information about the decision boundary. Below we consider factors that influence the security and secrecy of the decision boundary. 5.2 Scale of Training Some machine learning systems are trained by the end user, while others are trained using data from many users or organizations . The choice between these two models is sometimes cast as a tradeoff between the amount of training data and the secrecy of the resulting classifier [3]. This issue also applies to an IDS; if an IDS is trained each time it is deployed then it will have comparatively little data regarding normal network traffic. It will also have no chance to learn about novel intrusions before seeing them in the wild. Conversely, an IDS that uses a global set of rules would be able to adapt to novel intrusion attempts more quickly. Unfortunately, any adversary with access to a public IDS classification function can test to ensure that its intrusion points will be accepted by deployments of the same classification function. These issues are instances of a more general problem. In some cases, it seems reasonable to assume the adversary has little access to information available to the learner. However , unless the adversary has no prior knowledge about the learning problem at hand, we cannot assume all of the information provided in the training set is secret. Therefore, it is unclear how much is gained by attempting to keep the training set, and therefore the state of the classifier, secret. Many systems already attempt to achieve a balance between global and local retraining [3]. Systems that take this approach have the potential to outperform systems that perform training at a single level. However, the relationships between multilevel training, the adversary's domain knowledge , and secrecy are not yet well understood. 5.2.1 Adversary Observations Even without prior knowledge regarding a particular system , an adversary still may deduce the state of the learning algorithm. For example, if the learning system provides feedback to the adversary (e.g., "Request denied"), then a probing attack could be used to map the space of acceptable inputs. If the adversary has no information regarding the type of decision boundary used by the learner, this process could require a number of probes proportional to the size of the space. On the other hand, if the adversary knows which learning algorithm is being used, a few well-chosen probes could give the adversary sufficient knowledge of the learner's state. As a standard security practice, we assume the learning algorithm itself to be common knowledge. Instead of expecting the learning algorithm to be a secret, some systems attempt to prevent the adversary from discovering the set of features the learning algorithm uses. This may be realistic in systems with a small number of deployments . Ideally, we could produce an information theoretic bound on the amount of information an adversary could gain by observing the behavior of a particular algorithm on a particular point. Using these bounds, we could reason about the algorithm's robustness against probing attacks. In this setting, it may also be interesting to distinguish between information gained from normal points drawn from the data's underlying distribution, intrusion points from a third party, and (normal or intrusion) attack points of the adversary's choosing. An adversary with sufficient information regarding training data, classifications of data points, or the internal state of a learner would be able to deduce the learner's decision boundary . This knowledge could simplify other types of attacks. For instance, the adversary could avoid detection by choosing intrusion points that will be misclassified by the learner, or launch an availability attack by manipulating normal points in a way that leads to misclassification. In either case, by increasing the number of points that are in the region that the defender incorrectly classifies, the adversary could increase the error rate. Some algorithms classify points by translating them into an abstract space and performing the actual classification in that space. The mapping between raw data and the abstract space is often difficult to reason about. Therefore, it may be computationally difficult for an adversary to use knowledge of a classifier's decision boundary to generate "interesting" attack points that will be misclassified. One can imagine classes of decision boundaries that are meaningful, yet provably provide an adversary with no information regarding unclassified points. Even with complete knowledge of the state of a learner that uses such a decision boundary, it would be computationally intractable to find 0 one of a few "interesting" points in a sufficiently large search space. In some cases, the decision boundary itself may contain sensitive information. For example, knowledge of the boundary may allow an adversary to infer confidential information about the training set. Alternatively, the way the decision boundary was constructed might be a secret. 5.2.2 Security Properties The performance of different algorithms will likely degrade differently as the adversary controls larger fractions of the training set. A measurement of an algorithm's ability to deal with malicious training errors could help system designers reason about and decide between different learners. A simple approach would be to characterize an algorithm's performance when subjected to a particular type of attack, but this would lead to an arms race as adversaries devise classes of attacks not well represented during the evaluation of the algorithm. Depending on the exact nature of the classification problem , it may be possible to make statements regarding the strength of predictions. For example, after making a classification a learning algorithm could examine the training set for that classification. It could measure the effect of small changes to that training set; if small changes generate large effects, the training set is more vulnerable to manipulation. THEORETICAL RESULTS In this section we present an analytic model that examines a causative attack to manipulate a naive learning algorithm. The model's simplicity yields an optimal policy for the adversary and a bound on the effort required to achieve the adversary's objective. We interpret the resulting bound and discuss possible extensions to this model to capture more realistic settings. We discuss an outlier detection technique. Outlier detection is the task of identifying anomalous data and is a widely used paradigm in fault detection [40], intrusion detection [23], and virus detection [33, 34]. We find the smallest region that contains some fixed percentage of the observed data, which is called the support of the data's distribution. The outlier detector classifies points inside the support as normal and those outside as anomalous. Outlier detection is often used in scenarios where anomalous data is scarce or novel anomalies could arise. 6.1 A Simple Model One simple approach to outlier detection is to estimate the support of the normal data by a multi-dimensional hypersphere . As depicted in Figure 1(a) every point in the hypersphere is classified as normal and those outside the hypersphere are classified as outliers. The training algorithm fixes the radius of the hypersphere and centers it at the mean of the training data. The hypersphere can be fit into the learning framework presented above by a squared loss function, sphere ( X, x i ) = `x i - X 2 , where X is the centroid of the data {x i }. It is easy to show that the parameter that minimizes Equation (1) is the mean of the training data. To make the hypersphere adaptive, the hypersphere is retrained on new data allowing for a repeated attack. To prevent arbitrary data from being introduced, we employ a conservative retraining strategy that only admits new points to the training set if they are classified as normal; we say the classifier bootstraps itself. This learning framework is not meant to represent the state of the art in learning techniques ; instead, it is a illustrative technique that allows for an exact analysis. 6.2 Attack Strategy The attack we analyze involves an adversary determined to alter our detector to include a specific point G by constructing data to shift the hypersphere toward the target as the hypersphere is retrained. We assume the goal G is initially correctly classified as an anomaly by our algorithm. For instance , in the IDS domain, the adversary has an intrusion packet that our detector currently classifies as anomalous. The adversary wants to change the state of our detector to misclassify the packet as normal. This scenario is a causative targeted integrity attack. Before the attack, the hypersphere is centered at X 0 and it has a fixed radius R. The attack is iterated over the course of T &gt; 1 training iterations. At the i-th iteration the mean of the hypersphere is denoted by X i . We give the adversary complete control: the adversary knows the algorithm, its feature set, and its current state, and all points are attack points. At each iteration, the bootstrapping policy retrains on all points that were classified as normal in a previous iteration. Under this policy, the adversary's optimal strategy is straightforward -- as depicted in Figure 1(b) the adversary places points at the location where the line between the mean and G intersects with the boundary . This reduces the attack to a single dimension along this line. Suppose that in the i-th iteration, the adversary strategically places i points at the i-th optimal location achieving optimal displacement of the mean toward the adversary's goal, G. The effort of the adversary is measured by M defined as P T i=1 i . Placing all attack points in the first iteration is not optimal. It achieves a finite shift while optimal strategies achieve unbounded gains. As we discuss below, the attack strategy must be balanced. The more points placed during an iteration , the further the hypersphere is displaced on that iteration . However, the points placed early in the attack effectively weigh down the hypersphere making it more difficult to move. The adversary must balance current gain against future gain. Another tradeoff is the number of rounds of iteration versus the total effort. 6.3 Optimal Attack Displacement We calculate the displacement caused by a sequence { i } of attack points. For T iterations and M total attack points, the function D R,T ( { i }) denotes the relative displacement caused by the attack sequence. The relative displacement is the total displacement over the radius of the hypersphere, X T - X 0 R . Let M i be defined as P i j=1 j , the cumulative mass. Using these terms, the relative distance is D R,T ( {M i }) = T T X i=2 M i -1 M i (4) (a) Hypersphere Outlier Detection (b) Attack on a Hypersphere Outlier Detector Figure 1: Depictions of the concept of hypersphere outlier detection and the vulnerability of naive approaches. In Figure 1(a) a bounding hypersphere centered at X of fixed radius R is used to estimate the empirical support of a distribution excluding outliers. Samples from the "normal" distribution being modeled are indicated by with three outliers indicated by . Meanwhile, Figure 1(b) depicts how an adversary with knowledge of the state of the outlier detector could shift the outlier detector toward a first goal G. It could take several iterations of attacks to shift the hypersphere further to include the second goal G . where we constrain M 1 = 1 and M T = M [25]. By finding an upper bound to Equation (4), we can bound the minimal effort M of the adversary. For a particular M , we desire an optimal sequence {M i } that achieves the maximum relative displacement, D R,T (M ). If the adversary has no time constraint, the solution is M i = i, which corresponds to placing a single point at each iteration. However, if the adversary expedites the attack to T &lt; M iterations, the optimal strategy is given by M i = M i -1 T -1 . This value is not always an integer, so we have: D R,T (M ) T - (T - 1) M -1 T -1 T (5) 6.4 Bounding the Adversary's Effort From these results we find a bound on the adversary's effort M . Since M 1 and T &gt; 1, Equation (5) is monotonically increasing in M . If the desired relative displacement to the goal is D R , the bound in Equation (5) can be inverted to bound the minimal effort M required to achieve the goal. Since D R &lt; T , this bound is given by: M ,, T-1 T - D R T -1 (6) The bound in Equation (6) gives us a worst-case bound on the adversary's capability when the adversary has complete control of the learner's training. For large relative displacements D R &gt; 1, the bound decreases exponentially as the number of iterations is increased. The bound has a limiting value of M e D R -1 . The adversary must tradeoff between using a large number of attack points or extending the attack over many iterations. A tightly-fit hypersphere with small radius will be more robust since our displacement is relative to its radius. An apparent deficiency of this analysis is the weak bound of M where 0 &lt; 1 that occurs when D R 1. This an important range since the adversary's goal may be near the boundary. The deficiency comes directly from our assumption of complete adversarial control. The lack of initial non-adversarial data allows our adversary to ensure a first step of one radius regardless of M . Therefore, the adversary can reach the objective of D R 1 with any M 1 in a single iteration. A more complex model could allow for initial data. By considering an initial N training points that support the hypersphere before the attack, we can obtain a stronger bound: M N he D R - 1 i (7) This stronger bound ensures that even for small D R , the adversary's effort is a multiple of N that increases exponentially in the desired displacement [25]. We could extend the model by adding non-adversarial data at every training iteration, for this corresponds to scenarios where the adversary only controls part of the data. CONCLUSIONS The earliest theoretical work we know of that approaches learning in the presence of an adversary was done by Kearns and Li [15]. They worked in the context of Valiant's Probably Approximately Correct (PAC) learning framework [35, 36], extending it to prove bounds for maliciously chosen errors in the training data. Specifically, they proved that if the learner is to perform correctly, in general the fraction of training points controlled by the adversary must be less than /(1 + ), where is the desired bound on classification errors by the learner [4, 6, 30]. Results from game theory may be relevant to adversarial learning systems. In particular, deception games involve players that have partial information and influence the information seen by other players. Some of these games involve continuous variables generated by various probability distributions [5, 9, 17, 29, 32], while others apply to scenarios with discrete states [14]. This work and adversarial learning both ask many of the same questions, and they both address the same underlying issues. Integration of game theoretic concepts is a promising direction for work in this area. Dalvi et al. examine the learn-adapt-relearn cycle from a game-theoretic point of view [8]. In their model, the learner has a cost for measuring each feature of the data and the adversary has a cost for changing each feature in attack points. If the adversary and learner have complete information about each other and we accept some other assumptions , they find an optimal strategy for the learner to defend against the adversary's adaptations. Research has also begun to examine the vulnerability of learners to reverse engineering. Lowd and Meek introduce a novel learning problem for adversarial classifier reverse engineering in which an adversary conducts an attack that minimizes a cost function [21]. Under their framework, Lowd and Meek construct algorithms for reverse engineering linear classifiers. Moreover, they build an attack to reverse engineer spam filters [22]. Although they are not machine learning systems, publicly verifiable digital watermarks also must deal with sensitivity (probing) attacks. An information theoretic analysis of the sensitivity attack quantifies the amount of information revealed per probe. Randomization of thresholds within the watermark verification algorithm increase the number of probes necessary to remove a digital watermark [19]. An interesting junction of learning and game theory has dealt with combining advice from a set of experts to predict a sequence with the goal of doing at least as well as the best expert in all possible sequences [7, 13, 37]. In this domain, adaptive weighting schemes are used to combine the experts , each accessed by how well it performs compared to the best expert for an adversarially chosen sequence. Amongst these schemes are the Aggregating Algorithm [37] and the Weighted Majority Algorithm [20]. There has also been work on attacking statistical spam filters . Wittel and Wu [39] discuss the possibility of crafting attacks designed to take advantage of the statistical nature of such spam filters, and they implement a simple attack. John Graham-Cumming describes implementing an attack he calls "Bayes vs. Bayes," in which the adversary trains a second statistical spam filter based on feedback from the filter under attack and then uses the second filter to find words that make spam messages undetectable by the original filter [10]. Methods exist to perform exact learning of a concept using answers to a series of queries. These queries return a coun-terexample when a "no" response is generated. In many scenarios, it has been shown that learning is possible even in the worst case [2]. Control theory has been proposed as an alternative to game theory and search oriented expert-systems for military command and control systems [12]. The motivation behind this proposal is the difficulty associated with modeling (or even predicting) the goals of a military adversary. 7.2 Research Directions Can machine learning be secure? Does adding machine learning to a system introduce vulnerability? This paper proposes a framework for understanding these questions. We present a model for describing attacks against learning algorithms, and we analyze a simple attack in detail. We discuss potential defenses against attacks and speculate about their effectiveness. Here we lay out the directions for research that we see as most promising. To evaluate and ensure the security of machine learning, these are among the most important areas that must be addressed: Information How crucial is it to keep information secret from an adversary? If an adversary has full knowledge of the system, are all the exploratory attacks trivial? If the adversary has no knowledge about the system, which attacks are still possible? Arms race Can we avoid arms races in online learning systems? Arms races have occurred in spam filters. Can game theory suggest a strategy for secure re-training? Quantitative measurement Can we measure the effects of attacks? Such information would allow comparison of the security performance of learning algorithms. We could calculate risk based on probability and damage assessments of attacks. Security proofs Can we bound the amount of information leaked by the learner? If so, we can bound the accuracy of the adversary's approximation of the learner's current state. Detecting adversaries Attacks introduce potentially detectable side effects such as drift, unusual patterns in the data observed by the learner, etc. These attacks are more pronounced in online learning. When do these side effects reveal the adversary's attack? ACKNOWLEDGMENTS Thanks to Michael I. Jordan, Peter Bartlett and David Mol-nar for their insightful discussions and comments regarding this work. We gratefully acknowledge support from the National Science Foundation and the Homeland Security Advanced Research Projects Agency. The views expressed here are solely those of the authors and do not necessarily reflect the views of the funding agencies or any agency of the U.S. government. REFERENCES [1] I. Androutsopoulos, J. Koutsias, K. V. Chandrinos, G. Paliouras, and C. D. Spyropolous. An evaluation of naive Bayesian anti-spam filtering. Proceedings of the Workshop on Machine Learning in the New Information Age, pages 917, 2000. [2] D. Angluin. Queries and concept learning. Machine Learning, 2(4):319342, Apr. 1988. [3] Apache, http://spamassassin.apache.org/. SpamAssassin. [4] P. Auer. Learning nested differences in the presence of malicious noise. Theoretical Computer Science, 185(1):159175, 1997. [5] V. J. Baston and F. Bostock. Deception games. International Journal of Game Theory, 17(2):129134, 1988. [6] N. H. Bshouty, N. Eiron, and E. Kushilevitz. PAC learning with nasty noise. Theoretical Computer Science, 288(2):255275, 2002. [7] N. Cesa-Bianchi, Y. Freund, D. P. Helmbold, D. Haussler, R. E. Schapire, and M. K. Warmuth. How to use expert advice. Journal of the ACM, 44(3):427485, May 1997. [8] N. Dalvi, P. Domingos, Mausam, S. Sanghai, and D. Verma. Adversarial classification. In Proceedings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 99108, Seattle, WA, 2004. ACM Press. [9] B. Fristedt. The deceptive number changing game in the absence of symmetry. International Journal of Game Theory, 26:183191, 1997. [10] J. Graham-Cumming. How to beat an adaptive spam filter. Presentation at the MIT Spam Conference, Jan. 2004. [11] T. Hastie, R. Tibshirani, and J. Friedman. The Elements of Statistical Learning: Data Mining, Inference and Prediction. Springer, 2003. [12] S. A. Heise and H. S. Morse. The DARPA JFACC program: Modeling and control of military operations. In Proceedings of the 39th IEEE Conference on Decision and Control, pages 25512555. IEEE, 2000. [13] M. Herbster and M. K. Warmuth. Tracking the best expert. Machine Learning, 32(2):151178, Aug. 1998. [14] J. P. Hespanha, Y. S. Ateskan, and H. H. Kizilocak. Deception in non-cooperative games with partial information. In Proceedings of the 2nd DARPA-JFACC Symposium on Advances in Enterprise Control, 2000. [15] M. Kearns and M. Li. Learning in the presence of malicious errors. SIAM Journal on Computing, 22:807837, 1993. [16] A. Lazarevic, L. Ert oz, V. Kumar, A. Ozgur, and J. Srivastava. A comparative study of anomaly detection schemes in network intrusion detection. In D. Barbar a and C. Kamath, editors, Proceedings of the Third SIAM International Conference on Data Mining, May 2003. [17] K.-T. Lee. On a deception game with three boxes. International Journal of Game Theory, 22:8995, 1993. [18] Y. Liao and V. R. Vemuri. Using text categorization techniques for intrusion detection. In Proceedings of the 11th USENIX Security Symposium, pages 5159, Aug. 2002. [19] J.-P. M. Linnartz and M. van Dijk. Analysis of the sensitivity attack against electronic watermarks in images. In D. Aucsmith, editor, Information Hiding '98, pages 258272. Springer-Verlag, 1998. [20] N. Littlestone and M. K. Warmuth. The weighted majority algorithm. Information and Computation, 108(2):212261, 1994. [21] D. Lowd and C. Meek. Adversarial learning. In Proceedings of the Eleventh ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 641647, 2005. [22] D. Lowd and C. Meek. Good word attacks on statistical spam filters. In Proceedings of the Second Conference on Email and Anti-Spam (CEAS), 2005. [23] M. V. Mahoney and P. K. Chan. Learning nonstationary models of normal network traffic for detecting novel attacks. In Proceedings of the Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 376385, 2002. [24] S. Mukkamala, G. Janoski, and A. Sung. Intrusion detection using neural networks and support vector machines. In Proceedings of the International Joint Conference on Neural Networks (IJCNN'02), pages 17021707, 2002. [25] B. Nelson. Designing, Implementing, and Analyzing a System for Virus Detection. Master's thesis, University of California at Berkeley, Dec. 2005. [26] V. Paxson. Bro: A system for detecting network intruders in real-time. Computer Networks, 31(23):24352463, Dec. 1999. [27] N. Provos. A virtual honeypot framework. In Proceedings of the 13th USENIX Security Symposium, 2004. [28] R. Raina, A. Y. Ng, and D. Koller. Transfer learning by constructing informative priors. In Neural Information Processing Systems Workshop on Inductive Transfer: 10 Years Later, 2005. [29] M. Sakaguchi. Effect of correlation in a simple deception game. Mathematica Japonica, 35(3):527536, 1990. [30] R. A. Servedio. Smooth boosting and learning with malicious noise. Journal of Machine Learning Research (JMLR), 4:633648, Sept. 2003. [31] J. Shawe-Taylor and N. Cristianini. Kernel Methods for Pattern Analysis. Cambridge University Press, 2004. [32] J. Spencer. A deception game. American Math Monthly, 80:416417, 1973. [33] S. J. Stolfo, S. Hershkop, K. Wang, O. Nimeskern, and C. W. Hu. A behavior-based approach to secure email systems. In Mathematical Methods, Models and Architectures for Computer Networks Security, 2003. [34] S. J. Stolfo, W. J. Li, S. Hershkop, K. Wang, C. W. Hu, and O. Nimeskern. Detecting viral propagations using email behavior profiles. In ACM Transactions on Internet Technology, 2004. [35] L. G. Valiant. A theory of the learnable. Communications of the ACM, 27(11):11341142, Nov. 1984. [36] L. G. Valiant. Learning disjunctions of conjunctions. In Proceedings of the 9th International Joint Conference on Artificial Intelligence, pages 560566, 1985. [37] V. Vovk. Aggregating strategies. In M. Fulk and J. Case, editors, Proceedings of the 7th Annual Workshop on Computational Learning Theory, pages 371383, San Mateo, CA, 1990. Morgan-Kaufmann. [38] L. Wehenkel. Machine learning approaches to power system security assessment. IEEE Intelligent Systems and Their Applications, 12(5):6072, Sept.Oct. 1997. [39] G. L. Wittel and S. F. Wu. On attacking statistical spam filters. In Proceedings of the First Conference on Email and Anti-Spam (CEAS), 2004. [40] W. Xu, P. Bodik, and D. Patterson. A flexible architecture for statistical learning and data mining from system log streams. In Temporal Data Mining: Algorithms, Theory and Applications, Brighton, UK, Nov. 2004. The Fourth IEEE International Conference on Data Mining. [41] D.-Y. Yeung and C. Chow. Parzen-window network intrusion detectors. In Proceedings of the Sixteenth International Conference on Pattern Recognition, pages 385388, Aug. 2002. [42] K. Yu and V. Tresp. Learning to learn and collaborative filtering. In Neural Information Processing Systems Workshop on Inductive Transfer: 10 Years Later, 2005.

Indiscriminate attack;Targeted attack;Statistical Learning;Exploratory attack;Machine learning;Security Metrics;Integrity;Availability;Spam Filters;Causative attack;Intrusion Detection;Machine Learning;Intrusion detection system;Computer Security;Game Theory;Adversarial Learning;Learning algorithms;Computer Networks;Security

52

Catenaccio: Interactive Information Retrieval System through Drawing

The Catenaccio system integrates information retrieval with sketch manipulations. The system is designed especially for pen-based computing and allows users to retrieve information by simple pen manipulations such as drawing a picture. When a user draws a circle and writes a keyword, information nodes related to the keyword are collected automatically inside the circle. In addition, the user can create a Venn diagram by repeatedly drawing circles and keywords to form more complex queries. Thus, the user can retrieve information both interactively and visually without complex manipulations. Moreover, the sketch interaction is so simple that it is possible to combine it with other types of data such as images and real-world information for information retrieval. In this paper, we describe our Catenaccio system and how it can be effectively applied.

INTRODUCTION Pen-based computers, such as personal digital assistants (PDA) and tablet PCs, have been developed. These computers are characterized by simple sketch interfaces similar to drawing a picture on paper in the real world. This drawing manipulation is not especially useful for communicating details, but is effective for general use. It is especially useful for creative activities, so there have been a number of research reports on improving sketch manipulation [1, 2, 3]. In addition, some game devices (e.g., Nintendo DS [4]) support such kinds of interactions and provide many types of game content. In these systems, a user can use the entire system window as a workspace and create 3D CG from 2D drawings. However, as the original applications may not support information retrieval, the user has to use conventional retrieval applications along with pen-based input styles. Considerable research has been done to support the use of information visualization for retrieving information [5]. Technical visualization methods such as zooming and scaling can be used to effectively display huge amounts of data [6, 7, 8]. However, existing visualization systems focus on mouse manipulation (e.g., click and drag), so they are not effectively designed for pen-based interactions such as a drawing. The most popular method of retrieving information is no doubt keyword searching. Search engines via the Web (e.g., Google and Yahoo) have been generally used for keyword searching [12, 13], and people feel that they cannot live without such search engines. Generally, keyword searching requires users to input one or more keywords. In these systems, users can retrieve information related to the keywords with Boolean operations (e.g., AND, OR and NOT). However, the systems are based on conventional input methods. Users of pen-based computers have to write a query into a fixed dialog box with a stylus or pen. Therefore, we have been developing an information retrieval system based on simple sketch manipulations. Our goal is to devise an effective and simple information retrieval system that works on pen-based computers, so we integrated a keyword searching that is one of the most usual methods with sketch manipulation that is one of the simple interactions. In our system, users retrieve information by drawing a Venn diagram instead of inputting keywords to a dialog box. Because the Venn diagram can be used to display Boolean operations (e.g., AND, OR, and NOT) visually and create some relationships at the same time, users can recognize the relationships at a glance. Moreover, the system allows users to use other types of data as elements in a Venn diagram (Fig. 1). In this paper, we describe our Catenaccio system that integrates information retrieval with sketch manipulations, and explain how it can be effectively applied for information retrieval. 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. AVI '06, May 2326, 2006, Venezia, Italy. Copyright 2006 ACM 1-59593-353-0/06/0005...$5.00. 79 Figure 1: Venn diagram: Venn diagram can be used to display Boolean operations and some relationships at the same time (top). The user can create an original Venn diagram (bottom). Figure 2. Basic manipulation: A user of the Catenaccio system draws a circle and writes a keyword inside that circle. Information nodes related to the keyword are then collected within the circled area. RELATED WORK A wide variety of information visualization systems are used for information retrieval [5]. Treating information as visualized nodes (e.g., images and simple shapes) allows users to interact with the information space visually. Moreover, several techniques (e.g., scaling, zooming, focus and context) are used to display a huge amount of information more effectively [6, 7, 8, 19]. Especially, spring model [11] provides useful ways to recognize the relationships between nodes. In these systems, related nodes move when the user clicks and drags a node. That is, node positions are dynamically changed through the user manipulations, so users retrieve information interactively. InfoCrystal [9] is also a visual tool focused on information retrieval. The system uses Venn diagrams to treat huge amounts of information effectively. However, conventional systems are designed for mouse interactions (e.g., click and drag) and their layouts are predefined, so they are not suitable for Pen-based computing, especially drawing or writing by hand. There are also several sketch interfaces focusing on pen-based computing [1, 2, 3]. Most of them enhance drawing manipulations and focus on 3D creations performed with 2D manipulation. Characteristically, the manipulations required for these systems are simple and are similar to drawing a stroke on a piece of paper with a pen. Sketch [1] users can draw 3D curves by performing 2D manipulations. This system calculates a 3D curve by combining a 2D stroke and a shadow stroke. Users of Harold [2] and Tolba [3] can create flat models in a 3D space by using sketch-based manipulation, effectively creating a 2.5D scene in a 3D space. Figure 3. Drawing a Venn diagram: By repeatedly drawing circles and keywords, users create Venn diagrams, and can then retrieve information by forming complex queries. Information nodes related with both "CG" and "3D" are collected. SYSTEM OVERVIEW The Catenaccio system is focused on pen-based computing and provides an interactive and visual information retrieval environment of using drawing manipulations. 3.1 Drawing Circles and Writing Keywords A user of the Catenaccio system draws a circle and writes a keyword inside that circle. The system automatically recognizes both the circle area and the keyword. Information nodes related to the keyword are then collected within the circled area. By making a continuous series of simple drawings, the user can create a Venn diagram to form a more complex query. Since the entire window is both a search area and a drawing canvas, the user can use the workspace freely. Since all the manipulations required for information retrieval are based on sketch manipulations, users can design an original Venn diagram related their interests. Thus, users can freely exploit the whole application window as both an input and a search area and retrieve information without complex GUIs. The circle provides an area where information nodes related to the keyword will be collected, and the keyword provides a query to search for related information nodes from a database. By continuing to use simple drawings, users can form more complex queries. The related information nodes are moved with a force that depends on the distance between the node position and the center of the circle (Fig. 2). Figure 3 shows an example of creating a Venn diagram by continuing to draw circles and keywords. In the example, when a user retrieves information that has two keywords "CG" and "3D", the user first draws a circle and writes "CG" (Fig. 3 (1)), and then draws another circle and writes "3D" (Fig. 3 (2)). Information nodes related to both keywords appear in the shared area of the Venn diagram (Fig. 3 (3 and 4)). Moreover, in the Venn diagram, the user can view four areas at a glance (Fig. 3 (3)). 80 Figure 4: Venn diagram of a drawing and an image, and Venn diagram of a drawing and an image that contains character information. Figure 5: Combination with real-world information: Capturing real-world information as a picture (1, 2) and drawing a circle around the keyword brings up related information nodes (3, 4). 3.2 Combination with Other Types of Data Users can now easily take pictures using digital cameras and cell phones that contain CCD cameras. As a result, they may have a huge amount of original image data in their computers. These data include some information such as name, time, or place, so we considered using them for information retrieval. We have developed prototype applications to explore the potential of Catenaccio. A Venn diagram is basically constructed by combining keywords and areas, so it is possible to combine that diagram with other types of data such as images and real-world information. Images contain name, time, or place information, and that information becomes a good trigger for retrieving other information and it can be used for queries. In addition, the image data has a rectangular shape that is useful for setting an area by controlling its position and size. The example in Figure 4 shows how users can use images to create Venn diagrams by combining drawings with image data. In the example, a file named "Mr. Tobita" becomes a query for the Venn diagram, so information related to "VR" AND "Mr. Tobita" is collected. In this case, even if users have forgotten someone's name, they can still retrieve related information through image contents. Figure 5 shows an example a Venn diagram with real-world information. Generally, using captured data for an interaction trigger is a common technique in AR systems [20]. We also use it for elements of Venn diagrams. Users first capture real-world information through digital cameras attached to their computers (Fig. 5 (1, 2)), and then draw circles around keywords on the captured data and another circle to collect information nodes (Fig. 5 (3)). As the system recognizes the keyword inside the first circle, related information nodes appear inside the second curve (Fig. 5 (4)). Figure 6: Recognition of user drawings: The system labels the user drawing area (1, 2). The system recognizes written keywords by using an OCR library (3) Previously, we have proposed a similar information retrieval system [10]. That system provides natural interactions, however, it completely depends on real-world objects. However, Catenaccio provides not only real-world information, but also drawing manipulations. Thus, the user can retrieve information even if there are not enough real-world objects. IMPLEMENTATION The entire workspace is bitmapped as in conventional 2D paint systems, and user drawing manipulations are reflected in the bitmap. The system supports two types of drawing, writing keywords and area drawings. The keyword writing is displayed as green, and the area drawing is displayed as blue. To set an area, the system labels the inside of a green area and knows the size and position of the area (Fig. 6 (1)). Then, the system divides the area into four layers for node animations (Fig. 6(2)). For keyword writing, the system sends the result to an OCR library to search for its meaning (Fig. 6 (3)). Catenaccio is a prototype system now, and the relationship between keywords and information nodes are predefined in a temporary database. The database contains three types of data: node names, keywords, and relationship levels. After the image recognition processes, nodes related to the keyword are selected and start moving until they are in the area appropriate to their relationship. The force is calculated by spring model [11]. For example, the node with the strongest relationship with the keyword receives a force that takes it to the deepest area. DISSCUSSION We have had some opportunities to demonstrate our system. Here, we discuss user interactions with Catenaccio based on comments made by visitors to our demonstrations. Also we consider the limitations of the system and our plans for future work. From our demonstrations, the visitors quickly understood the concepts of our system that integrates information retrieval with 81 drawing manipulations. Using Venn diagrams makes recognizing the relationships between information nodes and keywords easy. Most visitors could create simple Venn diagrams and set related nodes into the diagrams after watching a simple demonstration. We observed that some users drew interesting Venn diagrams that resembled pictures. The system facilitates creative activities, so we expect users will be able to create more original, and increasingly effective drawings for information retrieval. Especially, we received good reactions from users regarding the combination of drawing and an image to create a Venn diagram. By exploiting such combinations, the system augments keyword searching, and it is different from conventional search engines [12, 13]. Moreover, combining keywords and user-drawn pictures to create Venn diagrams is possible. Our system focuses on information retrieval for pen-based input. However, information retrieval using Venn diagrams is quite rough. We plan to combine our system with other types of sketch-based systems such as VelvetPath [16] to support more detailed interaction. With such a combination, users can use Catenaccio for general retrieval of information and then use VelvetPath to examine the information in more detail. In this case, all the manipulations would still be based on drawing or handwriting, so a user can handle a large amount of data in a natural way. Drawing manipulation is also useful for finger gestures. Many AR systems support the use of finger gestures as an input method [17, 18]. As the system recognizes user finger gestures, users can create Venn diagrams by manipulating real-world objects, drawing circles, and writing keywords. CONCLUSION We described the Catenaccio system that is focused on Pen-based computing and allows users to retrieve information by drawing Venn diagrams. The system recognizes user writing and drawings (keywords and circles) and places information related to the keywords inside the circles. Using this input, the system provides an interactive and visual information retrieval method. We described some examples of retrieving information through simple drawings. We also provided several examples of unique Venn diagrams created by combining drawings with images and real-world information. ACKNOWLEDGMENTS We thank Tetsuji Takada and Sinji Daigo for the visible and useful suggestions on this work. REFERENCES [1] R. C. Zeleznik, K. P. Herndon, and J. F. Hughes. An Interface for Sketching 3D Curves. In Proceedings of ACM SIGGRAPH '96, pp. 163-170, 1996. [2] J. M. Cohen, J. F. Hughes, and R. C. Zeleznik. Harold: A World Made of Drawings. In Proceedings of NPAR2000 (Symposium on Non-Photorealistic Animation and Rendering), pp. 83-90, 2000. [3] O. Tolba, J. Doresey, and L. McMillan. Sketching with Projective 2D Strokes. In Proceedings of ACM UIST '99, pp. 149-157, 1999. [4] Nintendo DS: http://www.nintendo.co.jp/ds/ [5] S. K. Card, J. D. MacKinlay, and B. Shneiderman. Readings in Information Visualization: Using Vision to Think. Morgan Kaufmann, 1999. [6] H. Koike. Fractal views: a fractal-based method for controlling information display. In Proceedings of ACM Transactions on Information Systems, Vol. 13, No. 3, pp. 305-323, July 1995. [7] G. W. Furnas. Generalized fisheye views. In Proceedings of the ACM Transactions on Computer-Human Interaction, Vol. 1, No. 2, pp. 126-160, 1994. [8] B. B. Bederson, J. D. Hollan, K. Perlin, J. Meyer, D. Bacon, and G. Furnas. Pad++: A Zoomable Graphical Sketchpad for Exploring Alternate Interface Physics. Journal of Visual Languages and Computing, Vol. 7, No. 1, pp. 3-31, 1996. [9] A. Spoerri. Visual tools for information retrieval. In Proceedings of VL'93, pp. 160-168, 1993. [10] H. Koike, Y. Sato, Y. Kobayashi, H. Tobita and M. Kobayashi. Interactive Textbook and Interactive Venn Diagram. In Proceedings of ACM CHI2000, pp. 121-128 , 2000. [11] R. Davidson and D. Harel. Drawing Graphics Nicely Using Simulated Annealing. In Proceedings of ACM Transactions on Graphics, Vol. 15, No. 4, pp. 301-331, 1996. [12] Google: http://www.google.com [13] Yahoo: http://www.yahoo.com [14] T. Calishain and R. Dornfest. GoogleHack: 100 Industrial-Strength Tips & Tricks . O'RELLY, 2003. [15] P. Bausch. AmazonHack: 100 Industrial-Strength Tips & Tools. O'RELLY, 2003. [16] H. Tobita. VelvetPath: Layout Design System with Sketch and Paint Manipulations, In Proceedings of EUROGRAPHICS2003 Short Presentations, pp. 137-144 , 2003. [17] J. Rekimoto. SmartSkin: An Infrastructure for Freehand Manipulation on Interactive Surfaces, In Proceedings of ACM CHI2002, 113-120, 2002. [18] X. Chen, H. Koike, Y. Nakanishi, K. Oka, and Y. Sato. Two-handed drawing on augmented desk system, In Proceedings of AVI 2002, 2002. [19] E. Orimo and H. Koike. ZASH: A browsing system for multi-dimensional data. In Proceedings of IEEE VL '99, pp. 266-286, 1999. [20] J. Rekimoto and K. Nagao. The world through the computer: Computer augmented interaction with real world environments. In Proceedings of ACM UIST'95, pp. 29-36, 1995. 82

Sketch manipulation;Information node;Venn diagram;Visual information retrieval;Pen-based computing;Image data;Information retrieval;keyword searching;2D system;sketch manipulations;interactive system;Catnaccio system;Interactive information retrieval system

53

Compression of Inverted Indexes For Fast Query Evaluation

Compression reduces both the size of indexes and the time needed to evaluate queries. In this paper, we revisit the compression of inverted lists of document postings that store the position and frequency of indexed terms, considering two approaches to improving retrieval efficiency:better implementation and better choice of integer compression schemes. First, we propose several simple optimisations to well-known integer compression schemes, and show experimentally that these lead to significant reductions in time. Second, we explore the impact of choice of compression scheme on retrieval efficiency. In experiments on large collections of data, we show two surprising results:use of simple byte-aligned codes halves the query evaluation time compared to the most compact Golomb-Rice bitwise compression schemes; and, even when an index fits entirely in memory, byte-aligned codes result in faster query evaluation than does an uncompressed index, emphasising that the cost of transferring data from memory to the CPU cache is less for an appropriately compressed index than for an uncompressed index. Moreover, byte-aligned schemes have only a modest space overhead:the most compact schemes result in indexes that are around 10% of the size of the collection, while a byte-aligned scheme is around 13%. We conclude that fast byte-aligned codes should be used to store integers in inverted lists.

INTRODUCTION Search engines have demanding performance requirements. Users expect fast answers to queries, many queries must be processed per second, and the quantity of data that must be searched in response to each query is staggering. The demands continue to grow:the Google search engine, for example, indexed around one billion documents a year ago and now manages more than double that figure 1 . Moreover, the increasing availability and affordability of large storage devices suggests that the amount of data stored online will continue to grow. Inverted indexes are used to evaluate queries in all practical search engines [14]. Compression of these indexes has three major benefits for performance. First, a compressed index requires less storage space. Second, compressed data makes better use of the available communication bandwidth; more information can be transfered per second than when the data is uncompressed. For fast decompression schemes, the total time cost of transfering compressed data and sub-sequently decompressing is potentially much less than the cost of transferring uncompressed data. Third, compression increases the likelihood that the part of the index required to evaluate a query is already cached in memory, thus entirely avoiding a disk access. Thus index compression can reduce costs in retrieval systems. We have found that an uncompressed inverted index that stores the location of the indexed words in web documents typically consumes more than 30% of the space required to store the uncompressed collection of documents. (Web documents often include a great deal of information that is not indexed, such as HTML tags; in the TREC web data, which we use in our experiments, on average around half of each document is indexable text.) When the index is compressed, the index size is reduced to between 10%15% of that required to store the uncompressed collection; this size includes document numbers, in-document frequencies, and word positions within documents. If the index is too large to fit entirely within main memory, then querying the uncompressed index is slower:as we show later, it is up to twice as slow as the fastest compressed scheme. In this paper, we revisit compression schemes for the in-1 See http://www.google.com/ 222 verted list component of inverted indexes. We also propose a new method for decoding lists. There have been a great many reports of experiments on compression of indexes with bitwise compression schemes [6, 8, 12, 14, 15], which use an integral number of bits to represent each integer, usually with no restriction on the alignment of the integers to byte or machine-word boundaries. We consider several aspects of these schemes:how to decode bitwise representations of integers efficiently; how to minimise the operations required for the most compact scheme, Golomb coding; and the relative performance of Elias gamma coding, Elias delta coding, Golomb coding, and Rice coding for storing indexes. We question whether bitwise compression schemes are the best choice for storing lists of integers. As an alternative, we consider bytewise integer compression schemes, which require that each integer is stored in an integral number of blocks, where each block is eight bits. The length of each stored integer can therefore be measured in an exact number of bytes. An additional restriction is to require that these eight-bit blocks must align to machine-word or byte boundaries . We propose and experimentally investigate several variations of bytewise schemes. We investigate the performance of different index compression schemes through experiments on large query sets and collections of Web documents. We report two surprising results. For a 20 gigabyte collection, where the index is several times larger than main memory, optimised bytewise schemes more than halve the average decoding time compared to the fastest bitwise approach. For a much smaller collection, where the index fits in main memory, a bytewise compressed index can still be processed faster than an uncompressed index. These results show that effective use of communication bandwidths is important for not only disk-to-memory transfers but also memory-to-cache transfers. The only disadvantage of bytewise compressed indexes is that they are up to 30% larger than bitwise compressed indexes; the smallest bitwise index is around 10% of the uncompressed collection size, while the bytewise index is around 13%. INVERTED INDEXES An inverted index consists of two major components:the vocabulary of terms--for example the words--from the collection , and inverted lists, which are vectors that contain information about the occurrence of the terms [14]. In a basic implementation, for each term t there is an inverted list that contains postings &lt; f d,t , d &gt; where f d,t is the frequency f of term t in the ordinal document d. One posting is stored in the list for each document that contains the term t. Inverted lists of this form--along with additional statistics such as the document length l d , and f t , the number of documents that contain the term t--are sufficient to support ranked and Boolean query modes. To support phrase querying or proximity querying, additional information must be kept in the inverted lists. Thus inverted list postings should be of the form &lt; f d,t , d, [o 0,d,t . . . o f d,t ,d,t ] &gt; The additional information is the list of offsets o; one offset is stored for each of the f d,t occurrences of term t in document d. Postings in inverted lists are usually ordered by increasing d, and the offsets likewise ordered within the postings by increasing o. This has the benefit that differences between values--rather than the raw values--can be stored, improving the compressibility of the lists. Other arrangements of the postings in lists are useful when lists are not necessarily completely processed in response to a query. For example, in frequency-sorted indexes [9, 10] postings are ordered by f d,t , and in impact-ordered indexes the postings are ordered by quantised weights [1]. These approaches also rely on compression to help achieve efficiency gains, and the improvements to compression performance we describe in this paper are as applicable to these methods as they are to the simple index representations we use as a testbed for our compression methods. Consider an example inverted list with offsets for the term "Matthew": &lt; 3, 7, [6, 51, 117] &gt;&lt; 1, 44, [12] &gt;&lt; 2, 117, [14, 1077] &gt; In this index, the terms are words, the offsets are word positions within the documents, and the lists are ordered by d. This inverted list states that the term "Matthew" occurs 3 times in document 7, at offsets 6, 51, and 117. It also occurs once in document 44 at offset 12, and twice in document 117, at offsets 14 and 1077. Ranked queries can be answered using the inverted index as follows. First, the terms in the user's query are located in the inverted index vocabulary. Second, the corresponding inverted lists for each term are retrieved from disk, and then processed by decreasing f t . Third, for each posting in each inverted list, an accumulator weight A d is increased; the magnitude of the increase is dependent on the similarity measure used, and can consider the weight w q,t of term t in the query q, the weight w d,t of the term t in the document d, and other factors. Fourth, after processing part [1, 6] or all of the lists, the accumulator scores are partially sorted to identify the most similar documents. Last, for a typical search engine, document summaries of the top ten documents are generated or retrieved and shown to the user. The offsets stored in each inverted list posting are not used in ranked query processing. Phrase queries require offsets and that a given sequence of words be contiguous in a matching document. For example, consider a combined ranked and phrase query: "Matthew Richardson" Richmond To evaluate such a query, the same first two steps as for ranked querying are applied. Then, instead of accumulating weights, it is necessary to construct a temporary inverted list for the phrase, by fetching the inverted list of each of the individual terms and combining them. If the inverted list for "Matthew" is as above and the inverted list for "Richardson" is &lt; 1, 7, [52] &gt; &lt; 2, 12, [1, 4] &gt; &lt; 1, 44, [83] &gt; then both words occur in document 7 and as an ordered pair. Only the word "Richardson" is in document 12, both words occur in document 44 but not as a pair, and only "Matthew" occurs in document 117. The list for "Matthew Richardson" is therefore &lt; 1, 7, [51] &gt; 223 After this, the ranking process is continued from the third step, where the list for the term "Richmond" and the newly created list are used to adjust accumulator weights. Phrase queries can involve more than two words. COMPRESSING INVERTED INDEXES Special-purpose integer compression schemes offer both fast decoding and compact storage of inverted lists [13, 14]. In this section, we consider how inverted lists are compressed and stored on disk. We limit our discussions here to the special-purpose integer compression techniques that have previously been shown to be suitable for index compression, and focus on their use in increasing the speed of retrieval systems. Without compression, the time cost of retrieving inverted lists is the sum of the time taken to seek for and then retrieve the inverted lists from disk into memory, and the time taken to transfer the lists from memory into the CPU cache before they are processed. The speed of access to compressed inverted lists is determined by two factors:first, the com-putational requirements for decoding the compressed data and, second, the time required to seek for and retrieve the compressed data from disk and to transfer it to the CPU cache before it is decoded. For a compression scheme to allow faster access to inverted lists, the total retrieval time and CPU processing costs should be less than the retrieval time of the uncompressed representation. However, a third factor makes compression attractive even if CPU processing costs exceed the saving in disk transfer time:compressing inverted lists increases the number of lists that can be cached in memory between queries, so that in the context of a stream of queries use of compression reduces the number of disk accesses. It is therefore important that a compression scheme be efficient in both decompression CPU costs and space requirements. There are two general classes of compression scheme that are appropriate for storing inverted lists. Variable-bit or bitwise schemes store integers in an integral number of bits. Well-known bitwise schemes include Elias gamma and delta coding [3] and Golomb-Rice coding [4]. Bytewise schemes store an integer in an integral number of blocks, where a block is eight bits in size; we distinguish between blocks and bytes here, since there is no implied restriction that a block must align to a physical byte-boundary. A simple bytewise scheme is variable-byte coding [2, 13]; uncompressed integers are also stored in an integral number of blocks, but we do not define them as bytewise schemes since, on most architectures , an integer has a fixed-size representation of four bytes. In detail, these schemes are as follows. Elias coding [3] is a non-parameterised bitwise method of coding integers. (Non-parameterised methods use static or fixed codes to store integers.) The Elias gamma code represents a positive integer k by 1 + log 2 k stored as a unary code, followed by the binary representation of k without its most significant bit. Using Elias gamma coding, small integers are compactly represented; in particular, the integer 1 is represented as a single 1-bit. Gamma coding is relatively inefficient for storing integers larger than 15 [13]. Elias delta codes are suited to coding larger integers, but are inefficient for small values. For an integer k, a delta code stores the gamma code representation of 1 + log 2 k , and then the binary representation of k without its most significant bit. Golomb-Rice bitwise coding [4] has been shown to offer more compact storage of integers and faster retrieval than the Elias codes [13]; indeed, it is bitwise optimal under the assumption that the set of documents with a given term is random. The codes are adapted to per-term likelihoods via a parameter that is used to determine the code emitted for an integer. In many cases, this parameter must be stored separately using, for example, an Elias code. For coding of inverted lists, a single parameter is used for all document numbers in a postings list, but each posting requires a parameter for its offsets. The parameters can be calculated as the lists are decoded using statistics stored in memory and in the lists, as we discuss later. Coding of an integer k using Golomb codes with respect to a parameter b is as follows. The code that is emitted is in two parts:first, the unary code of a quotient q is emitted, where q = (k - 1)/b + 1; second, a binary code is emitted for the remainder r, where r = k - q b - 1. The number of bits required to store the remainder r is either log 2 b or log 2 b . To retrieve the remainder, the value of the "toggle point" t = 1 ((log 2 k)+1))-b is required, where indicates a left-shift operation. After retrieving log 2 b bits of the remainder r, the remainder is compared to t. If r &gt; t, then one additional bit of the remainder must be retrieved. It is generally thought that caching calculated values of log 2 b is necessary for fast decoding, with a main-memory penalty of having to store the values. However, as we show later, when the standard log library function is replaced with a fast bit-shifting version, caching is unnecessary. Rice coding is a variant of Golomb coding where the value of b is restricted to be a power of 2. The advantage of this restriction is that there is no "toggle point" calculation required , that is, the remainder is always stored in exactly log 2 b bits. The disadvantage of this scheme is that the choice of value for b is restricted and, therefore, the compression is slightly less effective than that of Golomb coding. For compression of inverted lists, a value of b is required. Witten et al. [14] report that for cases where the probability of any particular integer value occurring is small--which is the usual case for document numbers d and offsets o--then b can be calculated as: b = 0.69 mean(k) For each inverted list, the mean value of document numbers d can be approximated as k = N/f t where N is the number of documents in the collection and f t is the number of postings in the inverted list for term t [14]. This approach can also be extended to offsets:the mean value of offsets o for an inverted list posting can be approximated as k = l d /f d,t where l d is the length of document d and f d,t is the number of offsets of term t within that document. As the statistics N, f t , and l are often available in memory, or in a simple auxiliary structure on disk, storage of b values is not required for decoding; approximate values of l can be stored in memory for compactness [7], but use of approximate values has little effect on compression effectiveness as it leads to only small relative errors in computation of b. In bytewise coding an integer is stored in an integral number of eight-bit blocks. For variable-byte codes, seven bits in each block are used to store a binary representation of the integer k. The remaining bit is used to indicate whether the current block is the final block for k, or whether an additional block follows. Consider an example of an integer k 224 in the range of 2 7 = 128 to 2 14 = 16, 384. Two blocks are required to represent this integer:the first block contains the seven least-significant bits of the integer and the eighth bit is used to flag that another block follows; the second block contains the remaining most-significant bits and the eighth bit flags that no further blocks follow. We use the convention that the flag bit is set to 1 in the final block and 0 otherwise. Compressing an inverted index, then, involves choosing compression schemes for the three kinds of data that are stored in a posting:a document number d, an in-document frequency f d,t , and a sequence of offsets o. A standard choice is to use Golomb codes for document numbers, gamma codes for frequencies, and delta codes for offsets [14]. (We explore the properties of this choice later.) In this paper, we describe such a choice as a GolD-GamF-DelO index. 3.1 Fast Decoding We experiment with compression of inverted lists of postings that contain frequencies f d,t , documents numbers d, and offsets o. For fast decompression of these postings, there are two important considerations:first, the choice of compression scheme for each component of the posting; and, second, modifications to each compression scheme so that it is both fast and compatible with the schemes used for the other components. In this section, we outline the optimisations we use for fast decompression. Our code is publically available and distributed under the GNU public licence. 2 Bitwise Compression We have experimented with a range of variations of bitwise decompression schemes. Williams and Zobel [13] reported results for several efficient schemes, where vectors that contain compressed integers are retrieved from disk and subse-quently decoded. 3 In their approach, vector decoding uses bitwise shift operations, bit masks, multiplication, subtraction , and function calls to retrieve sequences of bits that span byte boundaries. In our experiments on Intel Pentium-based servers running the Linux operating system, we have found that bitwise shift operations are usually faster than bit masks, and that the function calls are slow. By opti-mising our code to use bitwise shifts and to remove nested function calls, we have found that the overall time to decode vectors--regardless of the compression scheme used--is on average around 60% of that using the code of Williams and Zobel. Other optimisations that are specific to Golomb-Rice coding are also of value. Golomb-Rice decoding requires that log 2 b is calculated to determine the number of remainder bits to be retrieved. It is practicable to explicitly cache values of log 2 b in a hash table as they are calculated, or to pre-calculate all likely-to-be-used values as the retrieval query engine is initialised. This saves recalculation of logarithms when a value of b is reused in later processing, with the penalty of additional memory requirements for storing the lookup table. We measured the performance of Golomb coding with and 2 The search engine used in these experiments and our integer compression code is available from http://www.seg.rmit.edu.au/ 3 The code used by Williams and Zobel in their experiments is available from http://www.cs.rmit.edu.au/ ~hugh/software/ without caching. Timings are average elapsed query evaluation cost to process index information for 25,000 queries on a 9.75 gigabyte (Gb) collection of Web data [5] using our prototype retrieval engine on a GolD-GamF-GolO index (that is, Golomb document numbers, gamma frequencies, Golomb offsets); we discuss collection statistics and experimental design further in Section 4. The cache lookup table size is unrestricted. We found that, without caching of log 2 b values, the average query evaluation time is 0.961 seconds. Caching of log 2 b values as they are calculated during query processing roughly halves the average query evaluation time, to 0.494 seconds. Pre-calculating and storing the values offers almost no benefit over caching during query processing, reducing the time to 0.491 seconds; this reflects that only limited b values are required during query evaluation. Caching of toggle points yields 0.492 seconds. As toggle points are calculated using bitwise shifts, addition, and subtraction, this is further evidence that bitwise shifts are inexpensive on our hardware. An alternative approach to managing log computations is to replace the standard library log function with a loop that determines log 2 b using bitwise shifts and equality tests; the logarithm value can be determined by locating the position of the most-significant 1-bit in b. We found that this led to slight additional improvements in the speed of decoding Golomb codes, outperforming explicit caching. All Golomb-Rice coding results reported in this paper are computed in this way. Bytewise Compression We have experimented with improvements to variable-byte coding. Unlike in bitwise coding, we have found that masking and shifting are equally as fast because of the large number of shifts required. We use shifts in our experiments. Perhaps the most obvious way to increase the speed of variable-byte decoding is to align the eight-bit blocks to byte boundaries. Alignment with byte boundaries limits the decoding to only one option:the flag bit indicating if this is the last byte in the integer is always the most significant bit, and the remaining seven bits contain the value. Without byte alignment, additional conditional tests and operations are required to extract the flag bit, and the seven-bit value can span byte boundaries. We would expect that byte alignment would improve the speed of decoding variable-byte integers. Figure 1 shows the effect of byte alignment of variable-byte integers. In this experiment, variable-byte coding is used to store the offsets o in each inverted list posting. The optimised Golomb coding scheme described in the previous section is used to code document numbers d and Elias gamma coding is used to store the frequencies f d,t . We refer to this as a GolD-GamF-VbyO index. The graph at the left of Figure 1 shows total index size as a percentage of the uncompressed collection being indexed . The first bar shows that, without byte alignment, the GolD-GamF-VbyO index requires almost 13% of the space required by the collection. The second bar shows that padding to byte alignment after storing the Gamma-coded f d,t values increases the space requirement to just over 13.5% of the collection size. We discuss the other schemes in this figure later in this section. The graph at the right of Figure 1 shows elapsed query evaluation times using different index designs. Timings are 225 Original Original with byte boundary Signature block Signature block with byte boundary 0 5 10 15 20 Size (% of Collection) Original Original with byte boundary Signature block Signature block with byte boundary Scanning Scanning with byte boundary Scanning with signature block Scanning with signature block and byte boundary 0.0 0.2 0.4 0.6 0.8 1.0 Average Query Time (Seconds) Figure 1: Variable-byte schemes for compressing offsets in inverted lists in a GolD-GamF-VbyOindex. Four different compression schemes are shown and, for each, both original and scanning decoding are shown. Scanning decoding can be used when offsets are not needed for query resolution. the average elapsed query evaluation cost to process the inverted lists for 25,000 queries on a 20 Gb collection of Web [5] data using our prototype retrieval engine. Queries are processed as conjunctive Boolean queries. The first bar shows that the average time is around 0.7 seconds for the GolD-GamF-VbyO index without byte alignment. The second bar shows that the effect of byte alignment is a 25% reduction in average query time. Therefore, despite the small additional space requirement, byte-alignment is beneficial when storing variable-byte integers. A second optimisation to variable-byte coding is to consider the query mode when processing the index. For querying that does not use offsets--such as ranked and Boolean querying--decoding of the offsets in each posting is unnecessary . Rather, all that is required are the document numbers d and document frequencies f d,t . An optimisation is therefore to only examine the flag bit of each block and to ignore the remaining seven bits that contain the value. The value of f d,t indicates the number of offsets o stored in the posting . By examining flag bits until f d,t 1-bits are processed, it is possible to bypass the offsets with minimal processing. We call this approach scanning. Scanning can also be used in query modes that do require offset decoding. As we discussed earlier, phrase querying requires that all terms are present in a matching document. After processing the inverted list for the first term that is evaluated in a phrase query, a temporary inverted list of postings is created. This temporary list has a set D of documents that contain the first term. When processing the second term in the query, a second set of document numbers D are processed. Offsets for the posting associated with document d D can be scanned, that is, passed over without decoding, if d is not a member of D. (At the same time, document numbers in D that are not in D are discarded .) We show the performance of scanning in Figure 1. The fifth and sixth bars show how scanning affects query evaluation time for variable-bytes that are either unaligned and aligned to byte boundaries in the GolD-GamF-VbyO index. Scanning removes the processing of seven-bit values. This reduces the cost of retrieving unaligned variable-bytes to less than that of the aligned variable-byte schemes; the small speed advantage is due to the retrieval of smaller lists in the unaligned version. Scanning has little effect on byte-aligned variable bytes, reflecting that the processing of seven-bit values using shift operations has a low cost. Overall, however, byte-alignment is preferred since the decoding cost of offsets is expensive in an unaligned scheme. A third optimisation is an approach we call signature blocks, which are a variant of skipping. Skipping is the approach of storing additional integers in inverted lists that indicate how much data can be skipped without any processing [14]. Skipping has the disadvantage of an additional storage space requirement, but has been shown to offer substantial speed improvements [14]. A signature block is an eight-bit block that stores the flag bits of up to eight blocks that follow. For example, a signature block with the bit-string 11100101 represents that five integers are stored in the eight following eight-bit blocks:the string 111 represents that the first three blocks store one integer each; the string 001 represents that the fourth integer is stored over three blocks; and, the string 01 represents that the final integer is stored over two blocks. As all flag bits are stored in the signature block, the following blocks use all eight bits to store values, rather the seven-bit scheme in the standard variable-byte integer representation. The primary use of signature blocks is skipping. To skip offsets, f d,t offset values must be retrieved but not processed. By counting the number of 1-bits in a signature block, the number of integers stored in the next eight blocks can be determined. If the value of f d,t exceeds this, then a second or subsequent signature block is processed until f d,t offsets have been skipped. The last signature block is, on average, half full. We have found that bitwise shifts are faster than a lookup table for processing of signature blocks. The speed and space requirements are also shown in Figure 1. Not surprisingly, the signature block scheme requires more space than the previous variable-byte schemes. This space requirement is further increased if byte alignment of blocks is enforced. In terms of speed, the third and fourth bars in the right-hand histogram show that signature blocks are slower than the original variable-byte schemes when offsets are processed in the GolD-GamF-VbyO index. These results are not surprising:signature blocks are slow to process when they are unaligned, and the byte-aligned version is slow because processing costs are no less than the original variable-byte schemes and longer disk reads are required. As shown by the seventh bar, when offsets are skipped the unaligned signature block scheme is slower than the original 226 variable-byte scheme. The savings of skipping with signature blocks are negated by more complex processing when blocks are not byte-aligned. In contrast, the right-most bar shows that the byte-aligned signature block scheme with skipping is slightly faster on average than all other schemes. However, we conclude--given the compactness of the index and good overall performance--that the best all-round scheme is the original variable-byte scheme with byte alignment . Therefore, all variable-byte results reported in the Section 4 use the original byte-aligned variable-byte scheme with scanning. Customised Compression Combinations of bitwise and bytewise compression schemes are also possible. The aim of such approaches is to combine the fast decoding of bytewise schemes with the compact storage of bitwise schemes. For example, a simple and efficient custom scheme is to store a single bit that indicates which of two compression schemes is used, and then to store the integer using the designated compression scheme. We have experimented with several approaches for storing offsets . The simplest and most efficient approach we tested is as follows:when f d,t = 1, we store a single bit indicating whether the following offset is stored as a bitwise Elias delta code or as a bytewise eight-bit binary representation. When storing values, we use Elias delta coding if the value is greater than 256 and the binary scheme otherwise. This scheme has the potential to reduce space because in the median posting f d,t is 1 and the average offset is around 200. Selective use of a fixed-width representation can save storage of the 6-bit prefix used to indicate magnitude in the corresponding delta code. We report the results with this scheme, which we call custom , in the next section. This was the fastest custom scheme we tested. Other approaches we tried included switching between variable-byte and bitwise schemes, using the custom scheme when f d,t is either 1 or 2, and other simple variations . We omit results for these less successful approaches. RESULTS All experiments described in this paper are carried out on an Intel Pentium III based machine with 512 Mb of main-memory running the Linux operating system. Other processes and disk activity was minimised during timing experiments , that is, the machine was under light-load. A theme throughout these experiments and greatly impacting on the results is the importance of caching. On a modern machine, caching takes place at two levels. One level is the caching of recently-accessed disk blocks in memory, a process that is managed by the operating system. When the size of the index significantly exceeds memory capacity, to make space to fetch a new inverted list, the blocks containing material that has not been accessed for a while must be discarded. One of the main benefits of compression is that a much greater volume of index information can be cached in memory. For this reason, we test our compression schemes with streams of 10,000 or 25,000 queries extracted from a query log [11], where the frequency distribution of query terms leads to beneficial use of caching. Again, queries are processed as conjunctive Boolean queries. The other level at which caching takes place is the retention in the CPU cache of small blocks of data, typically of 128 bytes, recently accessed from memory. CPU caching DelD-GamF-GolO GolD-GamF-GamO GolD-GamF-DelO GolD-GamF-GolO GolD-GamF-RicO GolD-GamF-VbyO GolD-VbyF-VbyO RicD-GamF-RicO RicD-VbyF-VbyO VbyD-VbyF-VbyO Custom No compression 0 10 20 30 40 Index Size (% of Collection) DelD-GamF-GolO GolD-GamF-GamO GolD-GamF-DelO GolD-GamF-GolO GolD-GamF-RicO GolD-GamF-VbyO GolD-VbyF-VbyO RicD-GamF-RicO RicD-VbyF-VbyO VbyD-VbyF-VbyO Custom No compression 0 1 2 3 4 Average Query Time (Seconds x 10^-2) Figure 2: Performance of integer compression schemes for offsets in inverted lists, in an index with Golomb document numbers and gamma frequencies. In this experiment, the index fits in main memory. A 500 Mb collection is used, and results are averaged over 10,000 queries. is managed in hardware. In current desktop computers, as many as 150 instruction cycles are required to fetch a single machine-word into the CPU. At a coarser level, compression of postings lists means that the number of fetches from memory to cache during decompression is halved. Small collection Figure 2 shows the relative performance of the integer compression schemes we have described for storing offsets, on a 500 Mb collection of 94,802 Web documents drawn from the TREC Web track data [5]; timing results are averaged over 10,000 queries drawn from an Excite search engine query log [11]. The index contains 703, 518 terms. These results show the effect of varying the coding scheme used for document numbers d, frequencies f d,t , and offsets o. In all cases where both bitwise and variable-byte codes are used, the bitwise codes are padded to a byte boundary before a variable-byte code is emitted; thus, for example, in a GolD-GamF -VbyO index, there is padding between the gamma 227 frequency and the sequence of variable-byte offsets. Not all code combinations are shown; for example, given that the speed advantage of using variable-byte document numbers is small, we have not reported results for index types such as VbyD-GamF-RicD, and due to the use of padding a choice such as VbyD-GamF-VbyD. Given the highly skew distribution of f d,t values, Golomb or Rice are not suitable coding methods, so these have not been tried. In the "no compression" case, fixed-width fields are used to store postings. Document numbers are stored in 32 bits, frequencies in 16 bits, and offsets in 24 bits; these were the smallest multiples of bytes that would not overflow for reasonable assumptions about data properties. The relative performance of Elias delta and gamma, Rice, and Golomb coding is as expected. The non-parameterised Elias coding schemes result in larger indexes than the param-terised Golomb-Rice schemes that, in turn, result in slower query evaluation. The average difference between offsets is greater than 15, making Elias delta coding more appropriate overall than gamma coding; the latter is both slower and less space-efficient. On the lower graph in Figure 2, comparing the fourth and fifth columns and comparing the fifth and eighth columns, it can be seen that choice of Golomb or Rice codes for either offsets or document numbers has virtually no impact on index size. Comparing the fifth and eighth columns on the upper graph, the schemes yield similar decoding times for document numbers. However, Rice codes are markedly faster for decoding offsets, because no toggle point calculation is required . Among the bitwise schemes, we conclude that Rice coding should be used in preference to other schemes for coding document numbers and offsets. The most surprising result is the effect of using the optimised byte-boundary variable-byte scheme for coding offsets . Despite the variable-byte index being 26% larger than the corresponding Rice-coded index, the overall query evaluation time is 62% less. Further speed gains are given by coding all values in variable-byte codes. Indeed, variable-byte decoding is faster even than processing uncompressed lists. This result is remarkable:the cost of transfering variable-byte coded lists from memory to the CPU cache and then decoding the lists is less than the cost of transferring uncompressed lists. To our knowledge, this is the first practical illustration that compression improves the efficiency of an in-memory retrieval system. We conclude from this that variable-byte coding should be used to store offsets to reduce both disk retrieval and memory retrieval costs. In experiments with integers, Williams and Zobel found that variable-byte coding is faster than the bitwise schemes for storing large integers of the magnitude stored in inverted lists [13]. Our result confirms this observation for retrieval systems, while also showing that the effect extends to fast retrieval from memory and that improvements to variable-byte coding can considerably increase decoding speed. The custom scheme uses both Elias delta and a binary bytewise scheme, reducing query evaluation to around 58% of the time for the Elias delta scheme. However, the custom scheme is almost twice as slow as the variable-byte scheme and, therefore, has little benefit in practice. Large collection Figure 3 shows the results of a larger experiment with an index that does not fit within the main-memory of our ma-DelD -GamF-GolO GolD-GamF-GamO GolD-GamF-DelO GolD-GamF-GolO GolD-GamF-VbyO RicD-GamF-RicO RicD-VbyF-VbyO VbyD-VbyF-VbyO No compression 0 10 20 30 40 Index Size (% of Collection) DelD-GamF-GolO GolD-GamF-GamO GolD-GamF-DelO GolD-GamF-GolO GolD-GamF-VbyO RicD-GamF-RicO RicD-VbyF-VbyO VbyD-VbyF-VbyO No compression 0.0 0.5 1.0 1.5 2.0 Average Query Time (Seconds) Figure 3: The performance of integer compression schemes for compressing offsets in inverted lists, with Golomb-coded document numbers and gamma-coded offsets. In this experiment, the index is several times larger than main memory. A 20 Gb collection is used, and results are averaged over 25,000 queries. chine. Exactly the same index types are tried as for the experiment above. A 20 Gb collection of 4,014,894 Web documents drawn from the TREC Web track data [5] is used and timing results are averaged over 25,000 Boolean queries drawn from an Excite search engine query log [11]. The index contains 9,574,703 terms. We include only selected schemes in our results. We again note that we have not used heuristics to reduce query evaluation costs such as frequency-ordering or early termination. Indeed, we have not even used stopping; with stopwords removed, query times are greatly impoved. Our aim in this research is to measure the impact on index decoding time of different choices of compression method, not to establish new benchmarks for query evaluation time. Our improvements to compression techniques could, however, be used in conjunction with the other heuristics, in all likelihood further reducing query evaluation time compared to the best times reported previously. 228 The relative speeds of the bitwise Golomb, Elias delta, and variable-byte coded offset schemes are similar to that of our experiments with the 500 Mb collection. Again, variable-byte coding results in the fastest query evaluation. Perhaps unsurprisingly given the results described above, an uncompressed index that does not fit in main-memory is relatively much slower than the variable-byte scheme; the disk transfer costs are a larger fraction of the overall query cost when the index does not fit in memory, and less use can be made of the memory cache. Indexes with variable-byte offsets are twice as fast as indexes with Golomb, delta, or gamma offsets , and one-and-a-half times as fast as indexes with Rice offsets. VbyD-VbyF-VbyO indexes are twice as fast as any index type with non-variable-byte offsets. In separate experiments we have observed that the gains demonstrated by compression continue to increase with collection size, as the proportion of the index that can be held in memory declines. Despite the loss in compression with variable-byte coding, indexes are still less than one-seventh of the size of the indexed data, and the efficiency gains are huge. CONCLUSIONS Compression of inverted lists can significantly improve the performance of retrieval systems. We have shown that an efficiently implemented variable-byte bytewise scheme results in query evaluation that is twice as fast as more compact bitwise schemes. Moreover, we have demonstrated that the cost of transferring data from memory to the CPU cache can also be reduced by compression:when an index fits in main memory, the transfer of compressed data from memory to the cache and subsequent decoding is less than that of transferring uncompressed data. Using byte-aligned coding , we have shown that queries can be run more than twice as fast as with bitwise codes, at a small loss of compression efficiency. These are dramatic gains. Modern computer architectures create opportunities for compression to yield performance advantages. Once, the main benefits of compression were to save scarce disk space and computer-to-computer transmission costs. An equally important benefit now is to make use of the fact that the CPU is largely idle. Fetching a single byte from memory involves a delay of 12 to 150 CPU cycles; a fetch from disk involves a delay of 10,000,000 cycles. Compression can greatly reduce the number of such accesses, while CPU time that would otherwise be unused can be spent on decoding. With fast decoding, overall costs are much reduced, greatly increasing query evaluation speed. In current computers such architecture considerations are increasingly important to development of new algorithms for query processing. Poor caching has been a crucial shortcoming of existing algorithms investigated in this research. There are several possible extensions to this work. We plan to investigate nibble-coding, a variant of variable-byte coding where two flag bits are used in each variable-byte block. It is likely that this approach may improve the performance of signature blocks. We will also experiment with phrase querying in practice and to explore the average query evaluation speed when partial scanning is possible. REFERENCES [1] V. Anh, O. de Kretser, and A. Moffat. Vector-Space ranking with effective early termination. In W. Croft, D. Harper, D. Kraft, and J. Zobel, editors, Proc. ACM-SIGIR International Conference on Research and Development in Information Retrieval, pages 3542, New York, Sept. 2001. [2] E. de Moura, G. Navarro, N. Ziviani, and R. Baeza-Yates. Fast and flexible word searching on compressed text. ACM Transactions on Information Systems, 18(2):113139, 2000. [3] P. Elias. Universal codeword sets and representations of the integers. IEEE Transactions on Information Theory, IT-21(2):194203, Mar. 1975. [4] S. Golomb. Run-length encodings. IEEE Transactions on Information Theory, IT12(3):399401, July 1966. [5] D. Hawking, N. Creswell, and P. Thistlewaite. Overview of TREC-7 very large collection track. In E. Voorhees and D. Harman, editors, Proc. Text Retrieval Conference (TREC), pages 91104, Washington, 1999. National Institute of Standards and Technology Special Publication 500-242. [6] A. Moffat and J. Zobel. Self-indexing inverted files for fast text retrieval. ACM Transactions on Information Systems, 14(4):349379, Oct. 1996. [7] A. Moffat, J. Zobel, and R. Sacks-Davis. Memory-efficient ranking. Information Processing & Management, 30(6):733744, 1994. [8] G. Navarro, E. de Moura, M. Neubert, N. Ziviani, and R. Baeza-Yates. Adding compression to block addressing inverted indexes. Information Retrieval, 3(1):4977, 2000. [9] M. Persin. Document filtering for fast ranking. In W. Croft and C. van Rijsbergen, editors, Proc. ACM-SIGIR International Conference on Research and Development in Information Retrieval, pages 339348, Dublin, Ireland, 1994. [10] M. Persin, J. Zobel, and R. Sacks-Davis. Filtered document retrieval with frequency-sorted indexes. Journal of the American Society for Information Science, 47(10):749764, 1996. [11] A. Spink, D. Wolfram, B. J. Jansen, and T. Saracevic. Searching the web:The public and their queries. Journal of the American Society for Information Science, 52(3):226234, 2001. [12] A. Vo and A. Moffat. Compressed inverted files with reduced decoding overheads. In R. Wilkinson, B. Croft, K. van Rijsbergen, A. Moffat, and J. Zobel, editors, Proc. ACM-SIGIR International Conference on Research and Development in Information Retrieval, pages 290297, Melbourne, Australia, July 1998. [13] H. Williams and J. Zobel. Compressing integers for fast file access. Computer Journal, 42(3):193201, 1999. [14] I. Witten, A. Moffat, and T. Bell. Managing Gigabytes: Compressing and Indexing Documents and Images. Morgan Kaufmann Publishers, Los Altos, CA 94022, USA, second edition, 1999. [15] N. Ziviani, E. de Moura, G. Navarro, and R. Baeza-Yates. Compression:A key for next-generation text retrieval systems. IEEE Computer, 33(11):3744, Nov. 2000. 229

Variable byte;Decoding;Efficiency;integer coding;Bytewise compression;Search engine;retrieval efficiency;Integer Compression;Inverted indexes;index compression;Optimisation;Compression;Inverted index;Document retrieval

54

Computing Consistent Query Answers using Conflict Hypergraphs

A consistent query answer in a possibly inconsistent database is an answer which is true in every (minimal) repair of the database. We present here a practical framework for computing consistent query answers for large, possibly inconsistent relational databases. We consider relational algebra queries without projection , and denial constraints. Because our framework handles union queries, we can effectively (and efficiently) extract indefinite disjunctive information from an inconsistent database. We describe a number of novel optimization techniques applicable in this context and summarize experimental results that validate our approach.

INTRODUCTION Traditionally, the main role of integrity constraints in databases was to enforce consistency. The occurrence of integrity violations was prevented by DBMS software. However , while integrity constraints continue to express important semantic properties of data, enforcing the constraints has become problematic in current database applications. For example, in data integration systems integrity violations may be due to the presence of multiple autonomous data sources. The sources may separately satisfy the constraints , but when they are integrated the constraints may not hold. Moreover, because the sources are autonomous, the violations cannot be simply fixed by removing the data involved in the violations. Example 1. Let Student be a relation schema with the attributes Name and Address and the key functional dependency N ame Address. Consider the following instance of Student: The first two tuples may come from different data sources, so it may be impossible or impractical to resolve the inconsistency between them. However, there is clearly a difference between the first two tuples and the third one. We don't know whether Jeremy Burford lives in Los Angeles or New York, but we do know that Linda Kenner lives in Chicago. An approach to query answering that ignores inconsistencies will be unable to make this distinction the distinction between reliable and unreliable data. On the other hand, any approach that simply masks out inconsistent data (the first two tuples in this example) will lose indefinite information present in inconsistent databases. In this example, we know that there is a student named Jeremy Burford (existential information) and that Jeremy Burford lives in Los Angeles or New York (disjunctive information). The above example illustrates the need to modify the standard notion of query answer in the context of inconsistent databases. We need to be able to talk about query answers that are unaffected by integrity violations. In [2], the notion of consistent query answer was proposed to achieve that ob jective . [2] introduced the notion of repair: a database that satisfies the integrity constraints and is minimally different from the original database. A consistent answer to a query, in this framework, is an answer present in the result of the query in every repair. Example 2. In Example 1, there are two repairs corresponding to two different ways of restoring the consistency: either the first or the second tuple is deleted. If a query asks for all the information about students, only the tuple (Linda Kenner,Chicago) is returned as a consistent answer because it is the only tuple that is present in both repairs. On the other hand, if a query asks for the names of students living in Los Angeles or New York, then Jeremy Burford is a consistent answer. The framework of [2] has served as a foundation for most of the subsequent work in the area of querying inconsistent databases [3, 5, 11, 12, 13, 15, 17, 19, 23] (see [7] for a survey and an in-depth discussion). The work presented here addresses the issue of computing consistent query answers for projection-free queries and denial integrity constraints. It is shown in [13] that this task can be done in polynomial time, using the notion of conflict hypergraph that succinctly 417 represents all the integrity violations in a given database. This line research is pursued further in the present paper. The main contributions of this paper are as follows: A complete, scalable framework for computing consistent answers to projection-free relational algebra queries in the presence of denial constraints. Our approach uses a relational DBMS as a backend and scales up to large databases. Novel optimization techniques to eliminate redundant DBMS queries. Encouraging experimental results that compare our approach with an approach based on query rewriting and estimate the overhead of computing consistent query answers. No comprehensive results of this kind exist in the literature. Because our query language includes union, our approach can extract indefinite disjunctive information present in an inconsistent database (see Example 1). Moreover, consistent query answers are computed in polynomial time. Other existing approaches are either unable to handle disjunction in queries [2, 12, 17] or cannot guarantee polynomial time com-putability of consistent query answers [3, 5, 11, 15, 19, 23]. The latter is due to the fact that those approaches rely on the computation of answers sets of logic programs with disjunction and negation a p 2 -complete problem. Only the approach of [2, 12] (which uses query rewriting) and the approach presented here scale up to large databases. Related research is further discussed in Section 6. The plan of the paper is as follows. In Section 2, we introduce basic concepts. In Section 3, we present our approach to computing consistent answers to projection-free queries and describe its implementation in a system called Hippo. In Section 4, we describe several techniques for eliminating redundant DBMS queries, that we have implemented in Hippo. In Section 5, we discuss a number of experiments we have conducted with Hippo and query rewriting. In Section 6, we briefly discuss related work. Section 7 contains conclusions and a discussion of possible future research directions. BASIC NOTIONS AND FACTS In this paper we work in the relational model of data. We recall that a database schema S is a set of relation names with attribute names and types. An instance of a database is a function that assigns a finite set of tuples to each relation name. For the purposes of this paper we consider only two fixed database domains N (natural numbers) and D (unin-terpreted constants). We also use the natural interpretation over N of binary relational symbols =, =, &lt;, &gt;, and we assume that two constants are equal only if they have the same name. We also view I as a structure for the first-order language over the vocabulary consisting of symbols of S, and standard built-in predicates over N (=, =, &lt;, &gt;). In this article, we use projection-free (-free) relational algebra expressions, defined using the following grammar: E :: R | (E) | E E | E E | E \ E. |R| is the arity of the relation symbol R and (unless specified otherwise) for the sake of simplicity we assume that attribute names are consecutive natural numbers. We extend this to expressions, i.e. |E| is the arity of the expression, and E.i is the reference to the i-th column resulting from the expression E (used in conditions for subexpressions). Morover, t[i] is the value on the i-th position of t, t[i, j] is an abbreviation for a tuple (t[i], . . . , t[j]), and with |t| we denote the length of the tuple t. We say that a tuple t is compatible with an expression E if the length of the tuple is equal to the arity of the expression, i.e. |t| = |E|. For a given expression E, QA E (I) is the result of evaluating E in the database instance I. In this paper we use only the set semantics of relational algebra expressions. We also use relational calculus queries consisting of quantifier-free first-order formulas which may be open (having free variables) or ground. In fact, our approach can handle relational algebra queries that require projection, as long as they can be translated to quantifier-free relational calculus queries. That's why we can deal with the relational algebra query corresponding to the query Student (X, LosAngeles ) Student(X, NewYork ) in Example 1.We also occasionally use SQL. 2.2 Repairs and consistent query answers An integrity constraint is a consistent closed first-order formula. In this paper we consider only the class of denial integrity constraints of the form: x 1 , . . . , x k . [R i 1 ( x 1 ) . . . R i k ( x k ) (x 1 , . . . , x k )] , (1) where is a boolean expression consisting of atomic formulas referring to built-in predicates. The number k is called the arity of a constraint. Note that, for example, functional dependencies and exclusion constraints are of the above form. Below we give another example. Example 3. Consider the relation Emp with attributes Name, Salary, and Manager, with Name being the primary key. The constraint that no employee can have a salary greater that that of her manager is a denial constraint: n, s, m, s , m . [Emp(n, s, m) Emp(m, s , m ) s &gt; s ]. Definition 1 (Consistent database). A database instance I is consistent with a set of integrity constraints C if I |= C (i.e., C is true in I); inconsistent otherwise. Definition 2. For a given database instance I of schema S, its set of facts (I) is the set of all positive facts that hold in this database: (I) = {R(t)|R S t I(R)}. Definition 3 (Database distance). Given two instances I 1 and I 2 of the same database, the distance between those instances (I 1 , I 2 ) is the symmetric difference between sets of facts of those instances: (I 1 , I 2 ) = ((I 1 ) \ (I 2 )) ((I 2 ) \ (I 1 )). Definition 4 (Proximity relation). Given three instances I, I 1 , I 2 , the instance I 1 is closer to I than the instance I 2 if the distance between I 1 and I is contained in the distance between I 2 and I, i.e. I 1 I I 2 (I, I 1 ) (I, I 2 ). 418 Definition 5 (Database repair). For a given instance I and set of integrity constraints C, I is a repair of I w.r.t. C if I is the closest instance to I, which is consistent with C , i.e. I |= C and I is I -minimal among the instances that satisfy C. By Rep C (I) we denote the set of all repairs of I with respect to C. The following fact captures an important property of repairs of denial constraints: each repair is a maximal consistent subset of the database. Fact 1. If C consists only of denial constraints, then: I Rep C (I) (I ) (I). Definition 6 (Core instance). For a given instance I, its core w.r.t a set of integrity constraints C is an instance Core I C such that: Core I C (R) = I Rep C (I) I (R). For any relation R and set of integrity constraints C, if there exists a relational algebra expression R C such that that for any instance I: QA R C (I) = Core I C (R), we call R C a core expression of the relation R w.r.t the set of integrity constraints C. Fact 2. If C is a set of denial integrity constraints, then for any R S there exists a core expression R C of R w.r.t C. Example 4. Suppose we have a table P (A, B) with a functional dependency A B. The core expression for P in SQL is: SELECT * FROM P P1 WHERE NOT EXISTS ( SELECT * FROM P P2 WHERE P1.A = P2.A AND P1.B &lt;&gt; P2.B); Having defined repairs, we can define consistent answers to queries. In general, the intuition is that the consistent query answer is an answer to the query in every repair. In this paper we consider consistent answers for two classes of queries. Definition 7 (CQA for ground queries). Given a database instance I and a set of denial integrity constraints C, we say that true (resp. false) is the consistent answer to a ground query w.r.t. C in I , and we write I |= C , if in every repair I Rep C (I), I |= (resp. I |= ). Definition 8 (CQA for relational algebra). Given a database instance I and a set of denial integrity constraints C, the set of consistent answers to a query E w.r.t. C in I is defined as follows: CQA E C (I) = I Rep C (I) QA E (I ). 2.3 Conflict hypergraphs The conflict hypergraph [13] constitutes a compact, space-efficient representation of all repairs of a given database instance . Note that this representation is specifically geared toward denial constraints. Definition 9 (Conflict). For a given integrity constraint c of form (1), a set of facts {R i 1 (t 1 ), . . . , R i k (t k ) }, where t j I(R i j ), is a conflict in a database instance I if (t 1 , . . . , t k ). By E c,I we denote the set of all conflicts generated by the integrity constraint c in I. Definition 10 (Conflict hypergraph). For a given set of integrity constraints C and a database instance I, a conflict hypergraph G C,I is a hypergraph with the set of vertices being the set of facts from the instance I, and the set of hyperedges consisting of all conflicts generated by constraints from C in I, i.e. G C,I = ( V I , E C,I ), where V I = (I), and E C,I = cC E c,I . Definition 11 (Maximal independent set). For a hypergraph G = (V, E), the set of vertices is a maximal independent set if it is a maximal set that contains no hyperedge from E. Fact 3. Let I be a database instance, and C a set of denial constraints, then for any repair I Rep C (I), (I ) is a maximal independent set M in G C,I , and vice versa. As shown in the following example in case of denial constraints the set of conflicts can be defined using a simple query. Example 5. Suppose we have a table P (A, B) with a functional dependency A B. The SQL expression for selecting all conflicts from P generated by the functional constraint is: SELECT * FROM P P1, P P2 WHERE P1.A = P2.A AND P1.B &lt;&gt; P2.B; Definition 12 (Data complexity). The data complexity of consistent answers to ground first-order queries is the complexity of determining the membership in the set D C, = {I|I |= C }, where is a fixed ground first-order query, and C is a fixed finite set of integrity constraints. We note that for a fixed set of integrity constraints, the conflict hypergraph is of polynomial size (in the number of tuples in the database instance). IMPLEMENTATION We review here the algorithm [13] for checking the consistency of ground queries in the presence of denial constraints, and then show how to use it to answer -free queries relational algebra queries, which which correspond to open quantifier-free relational calculus queries. We assume here that we work with a set of integrity constraints consisting only of denial constraints. The input to the algorithm consists of a ground quantifier-free formula , a set of integrity constraints C, and a database instance I. We want the algorithm to answer the question whether I |= C . Theorem 1. [13] The data complexity of consistent answers to quantifier-free ground queries w.r.t a set of denial constraints is in P . 419 The proof of this theorem can be found in [13] together with the corresponding algorithm that we call HProver. This algorithm takes the query in CNF, and a conflict hypergraph G C,I that corresponds to the database instance I in the presence of integrity constraints C. The first step of Input: = 1 . . . k ground input formula in CNF, G C,I = ( V I , E C,I ) conflict hypergraph of I w.r.t. C. 1 for i {1, . . . , k} do 2 let i R i 1 ( t 1 ) . . . R i p ( t p ) R i p+1 ( t p+1 ) . . . R i m ( t m ). 3 for j {p + 1, . . . , m} do 4 if t j I(R i j ) then 5 next i; 6 B {R i p+1 ( t p+1 ) , . . . , R i m ( t m ) } 7 for j {1, . . . , p} do 8 if t j I(R i j ) then 9 choose e j {e E C,I |R i j ( t j ) e} nondeterm. 10 B B (e j \ {R i j ( t j ) }). 11 if B is independent in G C,I then 12 return false; 13 return true; Figure 1: Algorithm HProver the algorithm reduces the task of determining whether true is the consistent answer to the query to answering the same question for every conjunct i . Then each formula i is negated and the rest of the algorithm attempts to find a repair I in which i is true, i.e., in which 1. t j I (R i j ) for (j = p + 1, . . . , m) 2. t j I (R i j ) for (j = 1, . . . , p) Such a repair corresponds to a maximal independent set M in the conflict hypergraph such that: 1 . every of R i p+1 (t p+1 ), . . . , R i m (t m ) is an element of M , 2 . none of R i 1 (t 1 ), . . . , R i p (t p ) is an element of M . If the algorithm succeeds in building an independent set satisfying the properties 1 and 2 , such a set can be extended to a maximal one which also satisfies those properties. That means that there is a repair in which i , and thus also , is true. If the algorithm does not succeed for any i, i = 1, . . . , k, then true is the consistent answer to . The condition 1 is satisfied by simply including the appropriate facts in M . The condition 2 is satisfied by excluding the appropriate facts from M . A fact can be excluded if it is not in (I) or if it belongs to a hyperedge whose remaining elements are already in M . 3.2 Finding an envelope Any relational algebra expression E can be translated to a corresponding first-order formula E ( x) in a standard way. Since we consider only -free algebra expressions, the formula E ( x) is quantifier-free. To be able to use HProver, we have to ground this formula, i.e., find an appropriate set of bindings for the variables in the formula. This will be done by evaluating an envelope query over the database. An envelope query should satisfy two properties: (1) it should return a superset of the set of consistent query answers for every database instance, and (2) it should be easily constructible from the original query. The result of evaluating an envelope query over a given database will be called an envelope. Suppose K E is an envelope query for a query E. We have that CQA E C (I) = { t QA K E (I) | I |= C E ( t)}. If an expression E does not use the difference operator (and thus is a monotonic expression), E itself is an envelope query, as stated by the following lemma: Lemma 1. For any monotonic relational expression E, the following holds: CQA E C (I) QA E (I). However when E is not monotonic, then the set of consistent query answers may contain tuples not contained in QA E (I). That kind of a situation is shown in the example below. Example 6. Suppose we have two relations R(A, B) and S(A, B, C, D), and we have functional dependency over R : A B. In case when I(R) = {(1, 2), (1, 3)}, and I(S) = {(1, 2, 1, 3)}, the set of answers to the query E = S \ (R(A 1 , B 1 ) B 1 =B 2 R(A 2 , B 2 )) is , while the set of consistent query answers is {(1, 2, 1, 3)}. To obtain the expression for an envelope, we define two operators F and G by mutual recursion. The operator F defines the envelope by overestimating the set of consistent answers. The auxiliary operator G underestimates the set of consistent answers. Definition 13. We define the operators F and G recursively : F (R) = R, F (E 1 E 2 ) = F (E 1 ) F (E 2 ), F (E 1 \ E 2 ) = F (E 1 ) \ G(E 2 ), F (E 1 E 2 ) = F (E 1 ) F (E 2 ), F ( (E)) = (F (E)), G(R) = R C , G(E 1 E 2 ) = G(E 1 ) G(E 2 ), G(E 1 \ E 2 ) = G(E 1 ) \ F (E 2 ), G(E 1 E 2 ) = G(E 1 ) G(E 2 ), G( (E)) = (G(E)). Because C consist only of denial constraints, Fact 2 guarantees that the expression R C exists, and therefore the operators are well defined. The pair of operators (F, G) has the following properties: Lemma 2. For any -free relational algebra expression E: QA G(E) (I) QA E (I) QA F (E) (I), and CQA G(E) C (I) CQA E C (I) CQA F (E) C (I). Lemma 3. For any -free relational algebra expression E: I Rep C (I). QA G(E) (I) QA E (I ) QA F (E) (I) With those two lemmas we can prove the following theorem. Theorem 2. If C contains only denial constraints, then for any -free relational algebra expression E the following holds for every database instance I: QA G(E) (I) CQA E C (I) QA F (E) (I). 420 3.3 The system Hippo We have implemented a system called Hippo for finding consistent answers to -free relational algebra queries. The data is stored in an RDBMS (in our case, PostgreSQL). The flow of data in Hippo is shown in Figure 2. The only E : , , \, Estimating F (E) : , , , \ Evaluation Conflict Detection DB Translation E : , , Envelope Conflict Hypergraph Grounding HProver Answer Set IC Figure 2: Data flow in Hippo output of this system is the Answer Set consisting of the consistent answers to the input query E with respect to a set of integrity constraints IC in the database instance DB. Before processing any input query, the system performs Conflict Detection, and creates the Conflict Hypergraph. We assume that the number of conflicts is small enough to allow us to store the hypergraph in main memory. We keep in main memory only the set of hyperedges corresponding to conflicts in database. The set of all the vertices represents the entire contents of the database and thus may be too big to fit in main memory. In this way, we guarantee that our approach is scalable. The processing of a query E consists of Estimating it to an envelope query F (E) that after Evaluation b y an RDBMS gives us the Envelope. Also, the system performs Translation of the input query E to a corresponding first-order logic formula E . Now, for every tuple from the Envelope we perform Grounding of E . Having now a first-order ground query we can check if true is the consistent answer to this query using HProver. Depending on the result of this check we return the tuple or not. It's important to notice here that because the hypergraph is stored in main memory, HProver doesn't need any immediate knowledge of the integrity constraints (no arrow from IC to HProver). This is because in HProver the independence of constructed sets B is being checked only for sets of vertices that are contained in the database, and if such vertices are in any conflict, it is registered in the hypergraph. HProver makes, however, database accesses to check tuple membership in database relations. OPTIMIZATIONS The previous section showed how to build a system for computing consistent query answers. But even though we have decided to store the conflict hypergraph in main memory , we still have to perform tuple membership checks (steps 4 and 8 in the HProver algorithm). To check if a tuple is present in a given table, we execute a simple membership query. For every tuple from the envelope we have to perform several tuple checks (depending on the complexity of the query). Executing any query is usually a costly operation in the database context. Therefore tuple membership checks are a significant factor in the algorithm execution time. In this section we address the problem of eliminating tuple membership checks. We propose two improvements: 1. The first infers information about the tuples present in the database from the current envelope tuple. That makes it possible to answer some tuple checks without interrogating the database. 2. The second supplements the first by extending the envelope expression so that we can find the results of all relevant tuple checks without executing any membership query. 4.1 Knowledge gathering In this section we address the problem of answering tuple checks. Definition 14 (Relevant facts). For a given -free expression E and a tuple t compatible with E, the set TC(E, t) of relevant facts is defined recursively: TC(R, t) = {R(t)}, TC(E 1 E 2 , t) = TC(E 1 , t) TC(E 2 , t), TC(E 1 \ E 2 , t) = TC(E 1 , t) TC(E 2 , t), TC(E 1 E 2 , (t 1 , t 2 )) = TC(E 1 , t 1 ) TC(E 2 , t 2 ), TC( (E), t) = TC(E, t). The set of facts TC(E, t) consists of all facts that HProver may need when working with the query E (t) (we conjecture that the same set of facts will be needed by any practical checker of consistent query answers for quantifier-free queries). In the following example we show that the tuple t itself may carry information that can be used to derive some relevant facts. Example 7. Recall that relation attributes are named by natural numbers. Assume that we have two tables R(1, 2), P (1, 2) and a query E = F (E) = 1=a (R(RP )). Suppose that a tuple t = (a, b, c, d) is the only result of the evaluation of F (E) in a database instance I. The set of relevant facts is TC(E, t) = {R(a, b), R(c, d), P (c, d)}. A natural consequence of the semantics of relational algebra expressions is that t QA 1=a (R(RP )) (I) implies (a, b) I(R). We can use this information to avoid performing some membership queries. At the same time the tuple t itself doesn't carry enough information to decide whether (c, d) belongs to either I(R), I(P ), or both of them. We call the process of inferring the information from result of the evaluation of a query knowledge gathering. Formally, we define the set of derived facts in the following way: Definition 15 (Knowledge gathering). For a given -free expression E and a tuple t compatible with E 421 we define the set KG recursively: KG(R, t) = {R(t)}, KG(E 1 E 2 , t) = KG(E 1 , t) KG(E 2 , t), KG(E 1 \ E 2 , t) = KG(E 1 , t), KG( (E), t) = KG(E, t), KG(E 1 E 2 , (t 1 , t 2 )) = KG(E 1 , t 1 ) KG(E 2 , t 2 ). We note here that the cardinality of the set of facts inferred with KG is linear in the size of the query and doesn't depend on the value of the tuple t. Now we state the main property of KG. Theorem 3 (Soundness of KG). Given a database instance I and a -free expression E t QA F (E) (I).R(t ) TC(E, t). R(t ) KG(E, t) I |= R(t ). Knowledge gathering is also complete in the case of {, }-expressions , i.e. it derives all relevant facts that hold in the database I. Theorem 4 (Completeness of KG for {, }). Given a database I and any {, }-query E. t QA F (E) (I).R(t ) TC(E, t). I |= R(t ) R(t ) KG(E, t). 4.2 Extended knowledge gathering In general, when the expression translates to a disjunctive query we need to extend the query so that the resulting tuple carries some additional information allowing us to derive all relevant facts. The extended approach described in detail below is illustrated first by the following example. Example 8. For the previously considered expression E = 1=a (R (R P )) the extended approach constructs the expression 1=a (R (R P )) 3,4 -- R 3,4 -- P , where is the left outer join operator 1 . Suppose now, I(R) = {(a, b), (e, f)} and I(P ) = {(c, d), (e, f)}. Then the evaluation of the extended envelope expression yields the following: 1=a (R (R P )) 3,4 -- R 3,4 -- P a b a b a b a b c d c d a b e f e f e f Now, consider the tuple (a, b, c, d, , , c, d). We can decompose it into two parts (a, b, c, d) and (, , c, d). The first part is simply the tuple from the envelope F (E), and it can be used to infer the fact R(a, b). The second part allows us to make two other important inferences. Namely, (c, d) I(R) and (c, d) I(P ). Our goal is to minimally extend the expression so that we can derive all relevant facts. In order to find what information is not guaranteed to be gathered from evaluation of the envelope expression, we generalize the definitions of KG and TC to non-ground tuples consisting of distinct variables. 1 For clarity we simplify the notion of the outer join condition . When writing S 3,4 -- T we mean S S.3=T.1S.4=T.2 ----------- T , and we assume the left join operator is left associative Definition 16 (Complementary set). For a given -free expression E, the complementary set (E) is defined as follows: (E) = TC(E, x) \ KG(E, x), where x = (x 1 , . . . , x |E| ). Example 9. Taking again under consideration the expression E = 1=a (R (R P )) and x = (x 1 , . . . , x 4 ) we have: TC(E, x) = {R(x 1 , x 2 ), P (x 3 , x 4 ), R(x 3 , x 4 ) }, KG(E, x) = {R(x 1 , x 2 ) }. R(x 1 , x 2 ) TC(E) means that for any tuple (t 1 , t 2 , t 3 , t 4 ) from the evaluation of the envelope expression for E, HProver may perform the tuple check R(t 1 , t 2 ). We have also R(x 1 , x 2 ) KG(E) and therefore we are able to answer this check using knowledge gathering. On the other hand R(x 3 , x 4 ) TC(E) means that for HProver may perform a tuple check R(t 3 , t 4 ). Since we don't have that R(x 3 , x 4 ) KG(E) we cannot guarantee that we can answer tuple checks R(t 3 , t 4 ) without executing a membership query on the database, even though we are able to answer tuple checks R(t 1 , t 2 ). The complementary set for the discussed expression is: (E) = {R(x 3 , x 4 ), P (x 3 , x 4 ) }. Analogous examples can be used to show that the simple knowledge gathering is not sufficient to avoid membership checks when processing expressions with the difference operator . Next, we extend the envelope expression so that it evaluation provides us with all information sufficient to answer the tuple checks. Definition 17 (Extended envelope expression). For a given -free expression E the extended envelope expression is defined as follows: H(E) = F (E) |R| j=1 E.(i+j-1)=R.j ---------------R (x i ,...,x i+|R|-1 )(E) R. The notation means that we have as many outer joins as there are elements in (E). They can appear in any order. We also define the following auxiliary expression: S(E) = E R(x i ,...,x i+|R|-1 )(E) R. For both H(E) and S(E) the elements of (E) need to be considered in the same order. Using outer joins results in a natural one-to-one correspondence between the tuples from the evaluation of the extended envelope expression and the tuples from the original envelope. Fact 4. For a given database instance I and -free expression E, the map t t[1, |E|] is a one-to-one map of QA H(E) (I) onto QA F (E) (I). Extending knowledge gathering to null tuples KG(R, ( , . . . , )) = allows us to state that using the extended envelope expression we can determine correctly all relevant facts without querying the database. 422 Theorem 5 (Soundness, completeness of ext. KG). For any database instance I and a -free expression E the following holds: t QA H(E) (I).R(t ) TC(E, t[1, |E|]). R(t ) KG(S(E), t) I |= R(t ). We note that in the case of {, }-expressions this approach doesn't unnecessarily extend the expression. 4.3 Other possibilities of optimizations 4.3.1 Negative knowledge gathering Knowledge gathering KG (as defined in Section 4.1) is complete only for queries that translate to a conjunction of positive literals. However, it is possible to come up with a construction that will be complete for queries that translate to a conjunction of positive as well as negative literals. The following example presents this idea. Example 10. Suppose we have tables R(1, 2) and P (1, 2) and a set of constraints C. For the query E = R\P , we have F (E) = R \ P C . Take any tuple t QA F (E) (I) for some instance I. We can easily conclude that t I(R). Also, we can say that t QA P C (I). Having this and hypergraph G C,I = ( V I , E C,I ) we can easily find if t I(P ). Namely, if there exists an edge e E C,I that P (t) e, then t I(P ). And if the vertex P (t) is not involved in any conflict in E then t I(P ). Reasoning of that sort cannot be applied to a query E = R R \ P P . Given a tuple t = (t 1 , t 2 ) from the envelope we know that t 1 , t 2 R, but the fact (t 1 , t 2 ) QA P C P C (I) doesn't imply that t 1 QA P C (I) or t 2 QA P C (I). And therefore we are not able to find if t 1 I(P ) or t 2 I(P ). This mechanism hasn't been included in the tested implementation yet. Implementing only positive knowledge gathering allows us to better observe the benefits of extending the envelope expression. We notice here that the query rewriting approach to computing consistent query answers described in [2] works also only for queries that are conjunctions of literals. However, as shown below, our approach leads to faster computation of consistent answers than query rewriting. 4.3.2 Intersection Another possible venue of optimization comes from directly implementing derived operators of relational algebra. For example, for intersection the appropriate extensions of the operators F and G are very simple: F (E 1 E 2 ) = F (E 1 ) F (E 2 ), G(E 1 E 2 ) = G(E 1 ) G(E 2 ). Now R P is equivalent to R \(R \P ) b ut F (R P ) = R P is not equivalent to F (R \ (R \ P )) = R \ ( R C \P ). Thus the envelope constructed by the operator F becomes sensitive to the way the original query is formulated. EXPERIMENTAL RESULTS Among available methods for computing consistent query answers, only the query rewriting technique [2] seems to be feasible for large databases. This is why in this work we compare the following engines: SQL An engine that executes the given query on the underlying RDBMS, and returns the query result. This method doesn't return consistent query answers, but provides a baseline to observe the overhead of computing consistent query answers using the proposed methods. QR Using the SQL engine, we execute the rewritten query constructed as decribed in [2]. More details on this approach can be found in Section 6. KG This method constructs the basic envelope expression and uses knowledge gathering, as described in Section 4.1. ExtKG This engine constructs the extended envelope expression (Section 4.2) and uses extended knowledge gathering. 5.1.1 Generating test data Every test was performed with the database containing two tables P and Q, both having three attributes X, Y, Z. For the constraints, we took a functional dependency X Z in each table. The test databases had the following parameters : n : the number of base tuples in each table, m : the number of additional conflicting tuples, and had both tables constructed in the following way: 1. Insert n different base tuples with X and Z being equal and taking subsequent values 0, . . . , n - 1, and Y being randomly drawn from the set {0, 1}. 2. Insert m different conflicting tuples with X taking subsequent values {0, n/m , 2 n/m , . . . , (m - 1) n/m }, Z = X + 1, and Y being randomly drawn from the set {0, 1}. In addition, we define auxiliary tables (P core and Q core ) containing only non-conflicting tuples from the base tables (resp. P and Q). Those table were used as materialized views of the core expressions ( P C and Q C ). Example 11. We show how a table P with n = 4 and m = 2 can be generated: 1. First we insert the base tuples (0, 1, 0), (1, 0, 1), (2, 0, 2), (3, 1, 3) into P 2. Then we insert the following conflicting tuples (0, 1, 1), (2, 0, 3) into P 3. P core will hold the following tuples (1, 0, 1), (3, 1, 3). In every table constructed in such a way the number of tuples is n + m, and the number of conflicts is m. 5.1.2 The environment The implementation is done in Java2, using PostgreSQL (version 7.3.3) as the relational backend. All test have been performed on a PC with a 1.4GHz AMD Athlon processor under SuSE Linux 8.2 (kernel ver. 2.4.20) using Sun JVM 1.4.1. 423 5.2 Test results Testing a query with a given engine consisted of computing the consistent 2 answers to the query and then iterating over the results. Iteration over the result is necessary, as the subsequent elements of the consistent query answer set are computed by Hippo in a lazy manner (this allows us to process results bigger than the available main memory). Every test has been repeated three times and the median taken. Finally, we note that the cost of computing the conflict hypergraph, which is incurred only once per session, is ignored while estimating the time of the query evaluation. We take a closer look at the time required for hypergraph construction in Section 5.2.3. 5.2.1 Simple queries We first compared performance of different engines on simple queries: join, union, and difference. Because we performed the tests for large databases, we added a range selection to the given query to obtain small query results, factoring out the time necessary to write the outputs. As parameters in the experiments, we considered the database size, the conflict percentage, and the estimated result size. Figure 3 shows the execution time for join as a function of the size of the database. In the case of {, }-expressions (thus also joins), the execution times of KG, ExtKG and SQL are essentially identical. Since no membership queries have to be performed, it means that for simple queries the work done by HProver for all tuples is practically negligible. X&lt;200 (P X Q) Time (sec.) Database size (increases in 1k) 0 100 200 300 0 3 6 9 12 SQL QR KG ExtKG Conflicts: 2% Figure 3: Execution time for join. Figure 4 contains the results for union. It shows that basic knowledge gathering KG is not sufficient to efficiently handle union. The cost of performing membership queries for all tuples is very large. Note that query rewriting is not applicable to union queries. Figure 5 contains the results for set difference (the execution time for KG was relatively much larger than values of other solutions and in order to increase readability it has not been included on this figure). Here the execution time is a function of the percentage of conflicts. We note that ExtKG performs as well as QR and both are approximately twice slower than SQL. 2 Except when using SQL engine. X&lt;200 (P Q) Time (sec.) Database size (increases in 1k) 0 20 40 60 80 100 0 3 6 9 12 15 SQL KG ExtKG Conflicts: 2% Figure 4: Execution time for union. X&lt;200 (P \ Q); Time (sec.) Conflicts (%) 0 1 2 3 4 5 0 1 2 3 SQL QR ExtKG DB size: 100k Figure 5: Execution time for difference. 5.2.2 Complex queries In order to estimate the cost of extending the envelope we considered a complex union query X&lt;d (P X Q X P X Q Q X P X Q X P ), with d being a parameter that will allow us to control the number of tuples processed by each engine. To assure no membership queries will be performed, we have to add 8 outer joins. The main goal was to compare two versions of knowledge gathering: KG and ExtKG. We have also included the results for SQL. (It should be noted here that this query has common subexpressions and RDBMS might use this to optimize the query evaluation plan. PostgreSQL, however, does not perform this optimization.) As we can see in Figure 6, KG outperforms ExtKG only in the case when the number of processed tuples is small. As the result size increases, the execution time of ExtKG grows significantly slower than that of KG. We notice also that ExtKG needs 23 times more time than SQL but the execution times of both grow in a similar fashion. 5.2.3 Hypergraph computation The time of constructing the hypergraph is presented on Figure 7). It depends on the total number of conflicts and the size of the database. It should be noticed here that the time of hypergraph construction consists mainly of the execution time of conflict detection queries. Therefore the time of hypergraph computation depends also on the number of integrity constraints and their arity. 424 X&lt;d(P X Q X P X Q Q X P X Q X P ) Time (sec.) Result size estimation (d tuples) 0 30 60 90 120 150 0 10 20 30 SQL KG ExtKG DB size: 100k Conflicts: 2% Figure 6: Impact of the result size. Time (sec.) Database size (increments in 1k) 0 50 100 150 200 250 0 3 6 9 12 Conflicts: 0% 2% 4% Figure 7: Hypergraph computation time RELATED WORK The discussion of related work here is very brief and focuses mainly on the most recent research. For a comprehensive discussion, please see [7]. Bry [10] was the first to note that the standard notion of query answer needs to be modified in the context of inconsistent databases and to propose the notion of a consistent query answer. Bry's definition of consistent query answer is based on provability in minimal logic and expresses the intuition that the part of the database instance involved in an integrity violation should not be involved in the derivation of consistent query answers. This is not quite satisfactory, as one would like to have a semantic, model-theoretic notion of consistent query answer that parallels that of the standard notion of query answer in relational databases. Moreover, the data involved in an integrity violation is not entirely useless and reliable indefinite information can often be extracted from it, as seen in Example 1. Query rewriting [2, 12] rewrites the original query Q to another query Q with the property that the set of all the answers to Q in the original database is equal to the set of consistent answers to Q in that database. When applicable , this approach provides an easy way to compute consistent query answers, as the rewritten query Q can typically be evaluated using the same query engine as the query Q. Because the query Q is rewritten independently of the database, the existence of a rewriting shows that requesting consistent query answers instead of the regular ones does not increase data complexity. However, query rewriting has been found to apply only to restricted classes of queries: the {, , \}-subset [2] or the {, }-subset [13] of the relational algebra. No method is presently known to rewrite queries with projection considered together with the binary operators, or union. Also, the class of constraints is limited to binary universal constraints [2] or single functional dependencies [13]. The line of research from [2] is con-tinued in [17] where a class of tractable conjunctive queries, based on generalized perfect matching, is identified. It is proved that the consistent answers to queries in this class cannot be obtained by query rewriting. We note here that the nonexistence of query rewriting for conjunctive queries follows also from the fact that computing consistent query answers for such queries is a co-NP-complete problem [4, 13]. This is because the rewritten query is first-order and thus can be evaluated in AC 0 , while known NP-complete problems like SAT are not in AC 0 . Several different approaches have been developed to specify all repairs of a database as a logic program with disjunction and classical negation [3, 6, 15, 18, 19, 22]. Such a program can then be evaluated using an existing system like dlv [14]. These approaches have the advantage of generality, as typically arbitrary first-order queries and universal constraints (or even some referential integrity constraints [3]) can be handled. However, the generality comes at a price: The classes of logic programs used are p 2 -complete. Therefore , the approaches based on logic programming are unlikely to work for large databases. The paper [15] proposes several optimizations that are applicable to logic programming approaches. One is localization of conflict resolution, another - encoding tuple membership in individual repairs using bit-vectors, which makes possible efficient computation of consistent query answers using bitwise operators. However, it is known that even in the presence of one functional dependency there may be exponentially many repairs [4]. With only 80 tuples involved in conflicts, the number of repairs may exceed 10 12 ! It is clearly impractical to efficiently manipulate bit-vectors of that size. [11] describes several possible definitions of repair, including Definition 5, and analyzes the complexity of computing consistent query answers under those definitions. Key and inclusion dependencies are considered. The computational approaches proposed are based on combinations of repair enumeration and chase computation [1]. New tractability results are obtained for classes of databases that satisfy key constraints but may violate inclusion dependencies. Presently, our approach requires that the integrated database be materialized at a single site. It remains to be seen if it can be generalized to a scenario where data is pulled from different sites during the evaluation of queries rewritten using, for example, the LAV approach [20]. This problem has been considered in the context of a logic-program-based approach to the computation of consistent query answers [8, 9] but, as explained earlier, such an approach does not scale up to large databases. A new scenario for data integration, data exchange, has been recently proposed [16]. In this scenario, a target database is materialized on the basis of a source database using source-to-target dependencies. In the presence of target integrity constraints, a suitable consistent target database may not exist. This is a natural context for the application of the concepts of repair and consistent query answer. However , [16] does not consider the issue of the inconsistency of target databases. [11] addresses the problem of consistent query answering in a restricted data exchange setting. 425 CONCLUSIONS AND FUTURE WORK In this paper, we have presented a practical, scalable framework for computing consistent query answers for large databases. We have also described a number of novel optimization techniques applicable in this context and summa-rized experimental results that validate our approach. The approach, however, has a number of limitations. Only projection-free relational algebra queries and denial integrity constraints are currently supported. Adding projection to the query language is a difficult issue because the complexity of computing consistent query answers becomes in that case co-NP-complete [4, 13]. So, unless P=NP, we cannot hope for computing consistent query answers efficiently for arbitrary conjunctive queries and arbitrary database instances. However, the evaluation of queries with projection can make use of the conflict hypergraph representation of all repairs, and of the operators F and G introduced in Section 3. Moreover , we expect to be able to compute consistent answers to queries with projection in polynomial time if conflict hypergraphs are suitably restricted. We hope that such restrictions can be translated into corresponding restrictions on database instances and integrity constraints. In [4], we have studied scalar aggregation queries in the presence of functional dependencies, also making use of conflict graphs. It remains to be seen whether the techniques developed in [4] can be combined with those of the present paper. Going beyond denial constraints appears challenging, too. Essentially, integrity violations of denial constraints are due to the presence of some facts in the database, and thus can be compactly represented using the conflict hypergraph. If arbitrary universal constraints, for example tuple-generating dependencies [1, 21], are allowed, constraint violations may be due to the simultaneous presence and absence of certain tuples in the database. It is not clear how to construct in this case a compact representation of all repairs that can be used for the computation of consistent query answers. Also, repairs are no longer guaranteed to be subsets of the original database but can contain additional tuples. If referential integrity is to be captured, constraints have to contain existentially quantified variables, which leads to the unde-cidability of consistent query answers [11]. Only in very restricted cases this problem has been shown to be tractable [11, 13]. Another avenue of further research involves using preferences to reduce the number of repairs and consequently make the computation of consistent query answers more efficient. For example, in data integration, we may have a preference for certain sources or for more recent information. The issue of benchmarking systems that compute consistent query answers requires more work. It would be desirable to design mechanisms that generate inconsistent databases in a systematic way and to perform more extensive experimental comparisons between implemented systems REFERENCES [1] S. Abiteboul, R. Hull, and V. Vianu. Foundations of Databases. Addison-Wesley, 1995. [2] M. Arenas, L. Bertossi, and J. Chomicki. Consistent Query Answers in Inconsistent Databases. In ACM Symposium on Principles of Database Systems (PODS), pages 6879, 1999. [3] M. Arenas, L. Bertossi, and J. Chomicki. Answer Sets for Consistent Query Answering in Inconsistent Databases. Theory and Practice of Logic Programming, 3(45):393424, 2003. [4] M. Arenas, L. Bertossi, J. Chomicki, X. He, V. Raghavan, and J. Spinrad. Scalar Aggregation in Inconsistent Databases. Theoretical Computer Science, 296(3):405434, 2003. [5] M. Arenas, L. Bertossi, and M. Kifer. Applications of Annotated Predicate Calculus to Querying Inconsistent Databases. In International Conference on Computational Logic, pages 926941. Springer-Verlag, LNCS 1861, 2000. [6] P. Barcelo and L. Bertossi. Logic Programs for Querying Inconsistent Databases. In International Symposium on Practical Aspects of Declarative Languages (PADL), pages 208222. Springer-Verlag, LNCS 2562, 2003. [7] L. Bertossi and J. Chomicki. Query Answering in Inconsistent Databases. In J. Chomicki, R. van der Meyden, and G. Snake, editors, Logics for Emerging Applications of Databases, pages 4383. Springer-Verlag, 2003. [8] L. Bertossi, J. Chomicki, A. Cortes, and C. Gutierrez. Consistent Answers from Integrated Data Sources. In International Conference on Flexible Query Answering Systems (FQAS), pages 7185, Copenhagen, Denmark, October 2002. Springer-Verlag. [9] L. Bravo and L. Bertossi. Logic Programs for Consistently Querying Data Integration Systems. In International Joint Conference on Artificial Intelligence (IJCAI), pages 1015, 2003. [10] F. Bry. Query Answering in Information Systems with Integrity Constraints. In IFIP WG 11.5 Working Conference on Integrity and Control in Information Systems, pages 113130. Chapman &Hall, 1997. [11] A. Cali, D. Lembo, and R. Rosati. On the Decidability and Complexity of Query Answering over Inconsistent and Incomplete Databases. In ACM Symposium on Principles of Database Systems (PODS), pages 260271, 2003. [12] A. Celle and L. Bertossi. Querying Inconsistent Databases: Algorithms and Implementation. In International Conference on Computational Logic, pages 942956. Springer-Verlag, LNCS 1861, 2000. [13] J. Chomicki and J. Marcinkowski. Minimal-Change Integrity Maintenance Using Tuple Deletions. Information and Computation, 2004. To appear. Earlier version: Technical Report cs.DB/0212004, arXiv.org e-Print archive. [14] T. Eiter, W. Faber, N. Leone, and G. Pfeifer. Declarative Problem-Solving in DLV. In J. Minker, editor, Logic-Based Artificial Intelligence, pages 79103. Kluwer, 2000. [15] T. Eiter, M. Fink, G. Greco, and D. Lembo. Efficient Evaluation of Logic Programs for Querying Data Integration Systems. In International Conference on Logic Programming (ICLP), pages 163177, 2003. [16] R. Fagin, P. G. Kolaitis, R. J. Miller, and L. Popa. Data Exchange: Semantics and Query Answering. In International Conference on Database Theory (ICDT), pages 207224. Springer-Verlag, LNCS 2572, 2003. [17] A. Fuxman and R. Miller. Towards Inconsistency Management in Data Integration Systems. In IJCAI-03 Workshop on Information Integration on the Web (IIWeb-03), 2003. [18] G. Greco, S. Greco, and E. Zumpano. A Logic Programming Approach to the Integration, Repairing and Querying of Inconsistent Databases. In International Conference on Logic Programming (ICLP), pages 348364. Springer-Verlag, LNCS 2237, 2001. [19] G. Greco, S. Greco, and E. Zumpano. A Logical Framework for Querying and Repairing Inconsistent Databases. IEEE Transactions on Knowledge and Data Engineering, 15(6):13891408, 2003. [20] A. Y. Halevy. Answering Queries Using Views: A Survey. VLDB Journal, 10(4):270294, 2001. [21] P. C. Kanellakis. Elements of Relational Database Theory. In Jan van Leeuwen, editor, Handbook of Theoretical Computer Science, volume B, chapter 17, pages 10731158. Elsevier/MIT Press, 1990. [22] D. Van Nieuwenborgh and D. Vermeir. Preferred Answer Sets for Ordered Logic Programs. In European Conference on Logics for Artificial Intelligence (JELIA), pages 432443. Springer-Verlag, LNAI 2424, 2002. [23] J. Wijsen. Condensed Representation of Database Repairs for Consistent Query Answering. In International Conference on Database Theory (ICDT), pages 378393. Springer-Verlag, LNCS 2572, 2003. 426

Knowledge gathering;Conflict hypergraph;inconsistency;Denial constraints;Inconsistent database;Optimization;Relational algebra;Polynomial time;Consistent Query answer;Disjunctive query;integrity constraints;query processing;Repair

55

Consistent Query Answering under Key and Exclusion Dependencies: Algorithms and Experiments

Research in consistent query answering studies the definition and computation of "meaningful" answers to queries posed to inconsistent databases, i.e., databases whose data do not satisfy the integrity constraints (ICs) declared on their schema. Computing consistent answers to conjunctive queries is generally coNP-hard in data complexity, even in the presence of very restricted forms of ICs (single, unary keys). Recent studies on consistent query answering for database schemas containing only key dependencies have an-alyzed the possibility of identifying classes of queries whose consistent answers can be obtained by a first-order rewriting of the query, which in turn can be easily formulated in SQL and directly evaluated through any relational DBMS. In this paper we study consistent query answering in the presence of key dependencies and exclusion dependencies. We first prove that even in the presence of only exclusion dependencies the problem is coNP-hard in data complexity , and define a general method for consistent answering of conjunctive queries under key and exclusion dependencies, based on the rewriting of the query in Datalog with negation . Then, we identify a subclass of conjunctive queries that can be first-order rewritten in the presence of key and exclusion dependencies, and define an algorithm for computing the first-order rewriting of a query belonging to such a class of queries. Finally, we compare the relative efficiency of the two methods for processing queries in the subclass above mentioned. Experimental results, conducted on a real and large database of the computer science engineering degrees of the University of Rome "La Sapienza", clearly show the computational advantage of the first-order based technique.

INTRODUCTION Suppose to have a database whose data violate the integrity constraints (ICs) declared on its schema. What are the answers that have to be returned to queries posed to such a database? The standard approach to this problem is through data cleaning, i.e., by explicitly modifying the data in order to eliminate violation of ICs: only when data are "repaired", i.e., are consistent with the ICs, queries can be answered. However, in many situations it would be much more desirable to derive significant information from the database even in the presence of data inconsistent with the ICs. Indeed, in many application scenarios, the explicit repair of data is not convenient, or even not possible. This happens, for instance, in data integration applications, which provide a unified, virtual view of a set of autonomous information sources [5]. This alternative approach is the one followed by research in consistent query answering, which studies the definition (and computation) of "meaningful" answers to queries posed to databases whose data do not satisfy the ICs declared on the database schema [1, 14, 4]. All these approaches are based on the following principle: schema is stronger than data. In other words, the database schema (i.e., the set of integrity constraints) is considered as the actually reliable information (strong knowledge), while data are considered as information to be revised (weak knowledge). Therefore, the problem amounts to deciding how to "repair" (i.e., change) data in order to reconcile them with the information expressed in the schema. Therefore, the intuitive semantics of consistent query answering can be expressed as follows: a tuple t is a consistent answer to a query q in an inconsistent database D if t is an answer to q in all the repairs of D, i.e., in all the possible databases obtained by (minimally) modifying the data in D to eliminate violations of ICs. Example 1. Let D = {r(a, b)} be a database whose schema contains the declaration of a key dependency on the first attribute of r. Since the database instance does not violate the key dependency on r, the only repair of the database 792 is D itself. Hence, the following query q(X, Y ) : r(X, Y ) has the consistent answer t = a, b . Now, let D be the database instance obtained by adding the fact r(a, c) to D. D is inconsistent with the key dependency, and has two possible repairs: {r(a, b)} and {r(a, c)}. Since there is no tuple which is an answer to q in both repairs, it follows that there are no consistent answers to the query q in D . In contrast, observe that the query q (X) : r(X, Y ) has the answer a both in D and in D , which can be therefore considered consistent . Recent studies in this area have established declarative semantic characterizations of consistent query answering over relational databases, decidability and complexity results for consistent query answering, as well as techniques for query processing [1, 6, 14, 4, 3, 5]. In particular, it has been shown that computing consistent answers of conjunctive queries (CQs) is coNP-hard in data complexity, i.e., in the size of the database instance, even in the presence of very restricted forms of ICs (single, unary keys). From the algorithmic viewpoint, the approach mainly followed is query answering via query rewriting: (i) First, the query that must be processed (usually a conjunctive query) is reformulated in terms of another, more complex query. Such a reformulation is purely intensional, i.e., the rewritten query is independent of the database instance; (ii) Then, the reformulated query is evaluated over the database instance. Due to the semantic nature and the inherent complexity of consistent query answering, Answer Set Programming (ASP) is usually adopted in the above reformulation step [14, 3, 5], and stable model engines like DLV [15] can be used for query processing. An orthogonal approach to consistent query answering is the one followed by recent theoretical works [1, 6, 13], whose aim is to identify subclasses of CQs whose consistent answers can be obtained by rewriting the query in terms of a first-order (FOL) query. The advantage of such an approach is twofold: first, this technique allows for computing consistent answers in time polynomial in data complexity (i.e., for such subclasses of queries, consistent query answering is compu-tationally simpler than for the whole class of CQs); second, consistent query answering in these cases can be performed through standard database technology, since the FOL query synthesized can be easily translated into SQL and then evaluated by any relational DBMS. On the other hand, this approach is only limited to polynomial subclasses of the problem . In particular, Fuxman and Miller in [13] have studied databases with key dependencies, and have identified a broad subclass of CQs that can be treated according to the above strategy. In this paper we study consistent query answering in the presence of key dependencies and exclusion dependencies, a well-known class of ICs. Notice that exclusion dependencies are not only typical of relational database schemas, but are also relevant and very common in languages for conceptual modeling, e.g., ontology languages [2]: indeed such dependencies allow for modeling partitioning/disjointness of entities . This makes the study of exclusion dependencies particularly important for the broad applicability of consistent query answering. Our contribution can be summarized as follows: 1. We prove that consistent answering of conjunctive queries for databases with only exclusion dependencies is coNP-hard in data complexity, i.e., the problem presents the same complexity lower bound already known for databases with only key dependencies [6, 4]. 2. We define a method for consistent query answering under key dependencies and exclusion dependencies based on the rewriting of the query in Datalog [10], a well-known extension of Datalog that allows for using negation in the body of program rules. The rewriting extends the one defined in [5] to the presence of exclusion dependencies. The rewriting is used by INFOMIX , 1 a system for the integration of inconsistent data, based on the use of the DLV system. 3. We extend the work of [13] to the presence of exclusion dependencies in the database schema. In particular , we identify the class of KE-simple queries (a subclass of CQs) that can be first-order rewritten in the presence of both key dependencies and exclusion dependencies, and define an algorithm for computing the first-order rewriting of a query belonging to such a class of queries. We point out that our algorithm, though inspired by the one of [13], in the presence of only key dependencies applies to a broader class of queries than the class considered first-order rewritable in [13]. Therefore, the technique of the present paper is relevant also for consistent query answering under only key dependencies. 4. We compare the relative efficiency of these two methods for processing KE-simple queries. To this aim, we have realized a software module that implements the above two rewriting methods. Then, we have compared query answering based on the rewriting in Datalog and evaluation in the DLV system [15] with the method based on first-order rewriting and query evaluation on MySQL DBMS. We have conducted our experiments on a real and large database of the computer science engineering degrees of the University of Rome "La Sapienza". Our experimental results clearly show, for KE-simple queries, the computational advantage of the specialized first-order based technique over the more general one based on Datalog . In particular, the results indicate that the advantage of the first-order based technique grows with the number of database tuples that violate the ICs. Such results thus provide, in a general sense, an experimental validation of the first-order based approach: its computational advantage is not only theoretical, but also can be effectively measured when applied to practical, realistic scenarios. However, it turns out that the general method based on Datalog , although not specifically tailored for KE-simple queries, proves particularly efficient in the presence of few data inconsistencies . In the next section, we briefly introduce the formal framework of consistent query answering. In Section 3, we prove coNP-hardness of consistent query answering under only exclusion dependencies, and present our Datalog rewriting and our algorithm for first-order rewriting in the presence of key and exclusion dependencies. In Section 4, we present our experimental results, and in Section 5 we address related work and conclude the paper. 1 http://sv.mat.unical.it/infomix. 793 INCONSISTENT DATABASES AND CONSISTENT ANSWERS Syntax. A database schema S is a triple A, K, E , where: A is a relational signature. K is a set of key dependencies over A. A key dependency (KD) over A is an expression of the form key(r) = {i 1 , . . . , i k }, where r is a relation of A, and, if n is the arity of r, 1 i j n for each j such that 1 j k. We assume that at most one KD is specified over a relation r. E is a set of exclusion dependencies over A. An exclusion dependency (ED) over A is an expression of the form r 1 [i 1 , . . . , i k ] r 2 [j 1 , . . . , j k ] = , where r 1 , r 2 are relations of A, and, if n 1 and n 2 are the arities of r 1 and r 2 respectively, for each such that 1 k, 1 i n 1 and 1 j n 2 . A term is either a variable or a constant symbol. An atom is an expression of the form p(t 1 , . . . , t n ) where p is a relation symbol of arity n and t 1 , . . . , t n is a sequence of n terms (either variables or constants). An atom is called fact if all the terms occurring in it are constants. A database instance D for S is a set of facts over A. We denote as r D the set {t | r(t) D}. A conjunctive query of arity n is an expression of the form h(x 1 , . . . , x n ) : a 1 , . . . , a m , where the atom h(x 1 , . . . , x n ), is called the head of the query (denoted by head(q)), and a 1 , . . . , a m , called the body of the query (and denoted by body(q)), is a set of atoms, such that all the variables occurring in the query head also occur in the query body. In a conjunctive query q, we say that a variable is a head variable if it occurs in the query head, while we say that a variable is existential if it only occurs in the query body. Moreover, we call an existential variable shared if it occurs at least twice in the query body (otherwise we say that it is non-shared). A FOL query of arity n is an expression of the form {x 1 , . . . , x n | (x 1 , . . . , x n )}, where x 1 , . . . , x n are variable symbols and is a first-order formula with free variables x 1 , . . . , x n . Semantics. First, we briefly recall the standard evaluation of queries over a database instance. Let q be the CQ h(x 1 , . . . , x n ) : a 1 , . . . , a m and let t = c 1 , . . . , c n be a tuple of constants. A set of facts I is an image of t w.r.t. q if there exists a substitution of the variables occurring in q such that (head(q)) = h(t) and (body(q)) = I. Given a database instance D, we denote by q D the evaluation of q over D, i.e., q D is the set of tuples t such that there exists an image I of t w.r.t. q such that I D. Given a FOL query q and a database instance D, we denote by q D the evaluation of q over D, i.e., q D = {t 1 , . . . , t n | D |= (t 1 , . . . , t n )}, where each t i is a constant symbol and (t 1 , . . . , t n ) is the first-order sentence obtained from by replacing each free variable x i with the constant t i . Then, we define the semantics of queries over inconsistent databases. A database instance D violates the KD key(r) = {i 1 , . . . , i k } iff there exist two distinct facts r(c 1 , . . . , c n ), r(d 1 , . . . , d n ) in D such that c i j = d i j for each j such that 1 j k. Moreover, D violates the ED r 1 [i 1 , . . . , i k ] r 2 [j 1 , . . . , j k ] = iff there exist two facts r 1 (c 1 , . . . , c n ), r 2 (d 1 , . . . , d m ) in D such that c i = d j for each such that 1 k. Let S = A, K, E be a database schema. A database instance D is legal for S if D does not violate any KD in K and does not violate any ED in E. A set of ground atoms D is a repair of D under S iff: (i) D D; (ii) D is legal for S; (iii) for each D such that D D D, D is not legal for S. In words, a repair for D under S is a maximal subset of D that is legal for S. Let q be a CQ. A tuple t is a consistent answer to q in D under S iff, for each repair D of D under S, t q D . Example 2. Consider the database schema S = A, K, E , where A comprises the relations Journal (title, editor), ConfPr (title, editor) and Editor (name, country), K comprises the dependencies key(Journal) = {1}, key(ConfPr ) = {1}, key(Editor ) = {1}, E comprises the dependency Journal [1] ConfPr [1] = . Consider the database instance D described below {Journal(TODS, ACM), Journal(TODS, IEEE), Editor (ACM, USA), ConfPr (PODS05, ACM), ConfPr (PODS05, SV), Editor (IEEE, USA)}. It is easy to see that D is not consistent with the KDs on Journal and ConfPr of S. Then, the repairs of D under S are: {Journal(TODS, ACM), ConfPr (PODS05, ACM), Editor (ACM, USA), Editor (IEEE, USA)} {Journal(TODS, ACM), ConfPr (PODS05, SV), Editor (ACM, USA), Editor (IEEE, USA)} {Journal(TODS, IEEE), ConfPr (PODS05, ACM), Editor (ACM, USA), Editor (IEEE, USA)} {Journal(TODS, IEEE), ConfPr (PODS05, SV), Editor (ACM, USA), Editor (IEEE, USA)}. Let q(x, z) : Journal (x, y), Editor (y, z) be a user query. The consistent answers to q in D under S are { TODS, USA }. QUERY ANSWERING Computational Complexity. The problem of computing consistent answers to conjunctive queries over inconsistent databases in the presence of KDs (under the repair semantics introduced in Section 2) is coNP-hard in data complexity [4, 6]. In the following, we prove that such a problem is coNP-hard in data complexity also for schemas in which only EDs occur 2 . Theorem 3. Let S = A, , E be a database schema containing only EDs, D a database instance for S, q a CQ of arity n over S, and t an n-tuple of constants. The problem of establishing whether t is a consistent answer to q in D under S is coNP-hard with respect to data complexity. Proof (sketch). We prove coNP-hardness by reducing the 3-colorability problem to the complement of our problem. Consider a graph G = V, E with a set of vertices V and edges E. We define a relational schema S = A, , E where A consists of the relation edge of arity 2, and the relation col of arity 5, and E contains the dependencies col[3]col[4] = , col[3] col[5] = , col[4] col[5] = . The instance D is defined as follows: D = {col(n, 1, n, , ), col(n, 2, , n, ), col(n, 3, , , n)| n V } {edge(x, y)| x, y E}. 2 We consider the decision problem associated to query answering (see e.g., [6]) 794 Where each occurrence of the meta-symbol denotes a different constant not occurring elsewhere in the database. Intuitively , to represent the fact that vertex n V is assigned with color i {1, 2, 3}, D assigns to col a tuple in which i occurs as second component and n occurs as first and also as 2 + i-th component. The EDs of S impose that consistent instances assign no more than one color to each node. Finally, we define the query q edge(x, y), col(x, z, w 1 , w 2 , w 3 ), col(y, z, w 4 , w 5 , w 6 ). On the basis of the above construction it is possible to show that G is 3-colorable (i.e., for each pair of adjacent vertices, the vertices are associated with different colors) if and only if the empty tuple is not a consistent answer to q in D under S (i.e., the boolean query q has a negative answer). Datalog Rewriting. We now provide a sound and complete query rewriting technique for consistent query answering in the presence of key and exclusion dependencies. To this aim, we make use of Datalog , i.e., Datalog enriched with (unstratified) negation, under stable model semantics [10]. From a computational point of view, Datalog is coNP-complete with respect to data complexity, and therefore is well suited for dealing with the high computational complexity of our problem. The rewriting that we present in the following extends the one proposed in [4] for CQs specified over database schemas with KDs, in order to properly handle the presence of EDs. The rewriting is employed in the system INFOMIX. Anal-ogously to other proposals that solve consistent query answering via query rewriting (although for different classes of constraints and query languages, see, e.g., [14, 3]), the basic idea of the technique is to encode the constraints of the relational schema into a Datalog program, such that the stable models of the program yield the repairs of the database instance D. Definition 4. Given a CQ 3 q and a schema S, the Datalog program (q, S) is defined as the following set of rules 4 : 1. the rule corresponding to the definition of q; 2. for each relation r S, the rules r(~ x, ~ y) : r D (~ x, ~ y) , not r(~ x, ~ y) r(~ x, ~ y) : r D (~ x, ~ y) , r(~ x, ~ z) , y 1 = z 1 r(~ x, ~ y) : r D (~ x, ~ y) , r(~ x, ~ z) , y m = z m where: in r(~ x, ~ y) the variables in ~ x correspond to the attributes constituting the key of the relation r; ~ y = y 1 , . . . , y m and ~ z = z 1 , . . . , z m . 3. for each exclusion dependency (r[i 1 , . . . , i k ] s[j 1 , . . . , j k ]) = in E, with r = s, the rules: r(~ x, ~ y) : r D (~ x, ~ y) , s(~ x, ~ z) s(~ x, ~ y) : s D (~ x, ~ y) , r(~ x, ~ z) 3 The present rewriting is not actually restricted to CQs, since it can be immediately extended to general Datalog queries. 4 Without loss of generality, we assume that the attributes in the key precede all other attributes in r, that i 1 = j 1 = 1, . . . , i k = j k = k, 1 = 1, . . . , h = h, and m 1 = h + 1, . . . , m h = h + h. where ~ x = x 1 , . . . , x k , i.e., the variables in ~ x correspond to the sequence of attributes of r and s involved in the ED. 4. for each exclusion dependency r[ 1 , . . . , h ] r[m 1 , . . . , m h ] = in E, the rules: r(~ x, ~ y, ~ z) : r D (~ x, ~ y, ~ z) , r(~ y, ~ w 1 , ~ w 2 ) , r(~ x, ~ y, ~ z) : r D (~ x, ~ y, ~ z) , r( ~ w 1 , ~ x, ~ w 2 ) , r(~ x, ~ x, ~ z) : r D (~ x, ~ x, ~ z). Furthermore, we denote with (D) the database instance obtained from D by replacing each predicate symbol r with r D . Informally, for each relation r, (q, S) contains (i) a relation r D that represents r D ; (ii) a relation r that represents a subset of r D that is consistent with the KD for r and the EDs that involve r; (iii) an auxiliary relation r that represents the "complement" of r, i.e., the subset of r D that together with r results inconsistent with the EDs and KDs on the schema. Notice that the extension of r depends on the choice made for r (and vice-versa), and that such choices are made in a non-deterministic way (enforced by the use of the unstratified negation). The above rules force each stable model M of (q, S) (D) to be such that r M is a maximal subset of tuples from r D that are consistent with both the KD for r and the EDs in E that involve r. Example 2.(contd.) The Datalog rewriting (q, S) of the query q(x, z) : Journal(x, y), Editor (y, z) is the following program: q(x, z) : Journal(x, y), Editor (y, z) Journal(x, y) : Journal D (x, y) , not Journal(x, y) Editor (x, y) : Editor D (x, y) , not Editor (x, y) ConfPr (x, y) : ConfPr D (x, y) , not ConfPr (x, y) Journal(x, y) : Journal D (x, y) , Journal(x, z) , z = y Editor (x, y) : Editor D (x, y) , Editor (x, z) , z = y ConfPr (x, y) : ConfPr D (x, y) , ConfPr (x, z) , z = y Journal(x, y) : Journal D (x, y) , ConfPr (x, z) ConfPr (x, y) : ConfPr D (x, y) , Journal(x, z) The first rule of the rewriting encodes the query. The second , third and fourth rule establish the relationship between each relation and the corresponding complementary predicate . The fifth, sixth, and seventh rule encode the KDs of S, whereas the last two rules encode the ED. We now state correctness of our encoding with respect to the semantics of consistent query answering. Theorem 5. let S = A, K, E be a database schema, D be a database instance for S, and q be a CQ over S. A tuple t is a consistent answer to q in D under S iff t q M for each stable model M of (q, S) (D). From the above theorem and Theorem 3 it follows that the consistent query answering problem under KDs and EDs is coNP-complete in data complexity. FOL Rewriting. Let us now consider a different approach to consistent query answering, which aims at identifying subclasses of queries for which the problem is tractable. This is the line followed in [1, 6, 13]. In particular, in [13] the authors define a subclass of CQs, called C tree , for which 795 they prove tractability of consistent query answering in the presence of KDs, and provide a FOL rewriting technique. The class C tree is based on the notion of join graph: a join graph of a query q is the graph that contains (i) a node N i for every atom in the query body, (ii) an arc from N i to N j iff an existential shared variable occurs in a non-key position in N i and occurs also in N j , (iii) an arc from N i to N i iff an existential shared variable occurs at least twice in N i , and one occurrence is in a non-key position. According to [13], C tree is the class of conjunctive queries (a) without repeated relation symbols, (b) in which every join condition involves the entire key of at least one relation and (c) whose join graph is acyclic. As pointed out in [13], this class of queries is very common, since cycles are rarely present in queries used in practice. However, no repeated symbols may occur in the queries, and queries must have joins from non-key attributes of a relation to the entire key of another one. We now extend the work of [13] as follows: We refine the class C tree by allowing join conditions in which not necessarily the entire key of one relation has to be involved, but it is sufficient that, for each pair of attributes, at least one attribute must belong to a key (i.e., we allow for joins involving portions of key). In such a way, we obtain a new class, called C + tree , larger than C tree , for which consistent query answering is polynomial in the presence of KDs. In other words, C + tree is the class of conjunctive queries for which only condition (a) and (c) above hold. We refine the class C + tree in order to obtain a class of queries, called KE-simple, for which consistent query answering is polynomial in the presence of both KDs and also EDs. We provide a new algorithm for computing the FOL rewriting for KE-simple queries. In the algorithm, we exploit the notion of join graph of [13], but we enrich the structure of the graph by associating to each node an adornment which specifies the different nature of terms in the atoms (see below), in order to deal with KE-simple queries. Let us describe in detail our technique. Henceforth, given a CQ q, we denote by R q the set of relation symbols occurring in body(q). Given a database schema S = A, K, E and a CQ q, we denote by O E (q) the set of relation symbols O E (q) = {s | r[j 1 , . . . , j k ] s[ 1 , . . . , k ] = E and r R q }. In words, O E (q) contains each relation symbol s A such that there exists an exclusion dependency between s and r in E, where r is a relation symbol occurring in body(q). Definition 6. Let S = A, K, E be a database schema. A conjunctive query q is KE-simple if q C + tree , and there exists no pair of relation symbols r, s in O E (q) such that there exists an exclusion dependency between r and s in E, there exists no relation symbol r in O E (q) such that there exists r[i 1 , . . . , i k ] s[j 1 , . . . , j k ] = in E, and either key(r) {i 1 , . . . , i k } or key(s) {j 1 , . . . , j k }, where s is a relation symbol in R q . In words, a query q is KE-simple if it belongs to the class C + tree , and if both there are no EDs between relations that are in O E (q), and each ED between a relation r R q and a relation s O E (q) does not involve non-key attributes of r or s. Notice that this last condition does not limit the applicability of our approach in many practical cases. For example, in relational databases obtained from ER-schemas, EDs are typically specified between keys. For KE-simple CQs, we present in the following a query rewriting algorithm which, given a query q, produces a FOL rewriting, whose evaluation over any database instance D for the database schema S returns the consistent answers to q in D under S. The basic idea of the algorithm is to specify a set of conditions, expressible in FOL, that, if verified over a database instance D, for a given tuple t, guarantee that in any repair of D there is an image of t w.r.t q, i.e., t is a consistent answer to q in D. We point out that, for non-KE -simple CQs, such conditions cannot be specified in FOL. Observe that, in our approach, the FOL rewriting is then in turn translated into SQL, and query evaluation is performed by means of standard DBMS query answering techniques. This further encoding does not present particular difficulties, and due to space limit we omit such transformation. In order to construct our join graph we need the following definition. Definition 7. Let S = A, K, E be a database schema, q be a CQ, and a = r(x 1 , . . . , x n ) be an atom (of arity n) occurring in R q . Then, let key(r) = {i 1 , . . . , i k } belong to K, and let 1 i n. The type of the i-th argument of a in q, denoted by type(a, i, q) is defined as follows: 1. If i 1 i i k , then: if x i is a head variable of q, a constant, or an existential shared variable, then type(a, i, q) = KB; if x i is an existential non-shared variable of q, then type(a, i, q) = KU. 2. Otherwise (i / {i 1 , . . . , i k }): if x i is a head variable of q or a constant, then type(a, i, q) = B; if x i is an existential shared variable of q, then type(a, i, q) = S; if x i is an existential non-shared variable of q, then type(a, i, q) = U. Terms typed by KB or B are called bound terms, otherwise they are called unbound. We call the typing of a in q the expression of the form r(x 1 /t 1 , . . . , x n /t n ), where each t i is the type of the argument x i in q. The following algorithm KEFolRewrite computes the FOL rewriting to a KE-simple conjunctive query q. In the algorithm , JG(q) denotes the join graph of q, in which each node N i is labelled with the typing of the corresponding atom a i in q. Furthermore, roots(JG(q)) denotes the set of nodes that are roots in JG(q) (notice that for KE-simple queries the join graph is a forest, since it is acyclic). Algorithm KEFolRewrite(q, S) Input: KE-simple CQ q (whose head variables are x 1 , . . . , x n ); schema S = A, K, E Output: FOL query (representing the rewriting of q) begin 796 Algorithm FolTree(N ,E) Input: node N of JG(q); set of EDs E Output: FOL formula begin let a = r(x 1 /t 1 , . . . , x n /t n ) be the label of N ; for i := 1 to n do if t i {KB, B} then v i := x i else v i := y i , where y i is a new variable if each argument of a is of type B or KB then f 1 := r(x 1 , . . . , x n ) else begin let i 1 , . . . , i m be the positions of the arguments of a of type S, U, KU; f 1 := y i 1 , . . . , y i m . r(v 1 , . . . , v n ) end; for each ED r[j 1 , . . . , j k ] s[ 1 , . . . , k ] = E do begin let m be the arity of s; for i := 1 to m do if i { 1 , . . . , k } then if i = c then z i = v j c else z i = y i where y i is a new variable; let y i 1 , . . . , y i k be the new variables above introduced; f 1 = f 1 y i 1 , . . . , y i k . s(z 1 , . . . , z m ) end if there exists no argument in a of type B or S then return f 1 else begin let p 1 , . . . , p c be the positions of the arguments of a of type U, S or B; let 1 , . . . , h be the positions of the arguments of a of type B; for i := 1 to c do if t p i = S then z p i := x p i else z p i := y i , where y i is a new variable for i := 1 to n do if t i {KB, KU} then w i := v i else w i := z i ; f 2 := z p 1 , . . . , z p c . r(w 1 , . . . , w n ) N jgsucc(N ) FolTree(N ) i{ 1 ,..., h } w i = x i return f 1 f 2 end end Figure 1: The algorithm FolTree compute JG(q); return {x 1 , . . . , x n | N roots(JG(q)) FolTree(N, E)} end Basically, the algorithm builds the join graph of q and then builds the first-order query by invoking the algorithm FolTree on all the nodes that are roots of the join graph. The algorithm FolTree is defined in Figure 1. Roughly speaking, the algorithm FolTree(N, E) returns a first-order formula that constitutes the encoding of the whole subtree of the join graph of the query whose root is the node N . To do that, the algorithm computes two subformulas f 1 and f 2 . The formula f 1 contains an atom whose predicate is the predicate r labelling the node N , in which the unbound variables of r are renamed with new existentially quantified variables. Furthermore, f 1 contains an atom of the form y i 1 , . . . , y i k . s(z 1 , . . . , z m ) for each ED that involves r and a relation s. Intuitively, when evaluated over a database instance D, each such atom checks that there are no facts of the form s(t s ) D that violate the ED together with a fact of the form r(t r ) D, which is in an image I of a tuple t w.r.t. the input query q, i.e., the atom guarantees that I is not contradicted w.r.t. the ED. The formula f 2 is empty only when all non-key arguments of the atom r are existential non-shared variables (i.e., of type U ). Otherwise, the formula f 2 is a universally quantified implication. In such an implication, the antecedent is an atom whose predicate is r, and the consequent is a conjunction of equality conditions and other subformulas: more precisely, there is an equality condition for each non-key argument in r of type B, and a subformula for each successor N of N in the join graph of q, computed by recursively invoking FolTree on N . Intuitively, f 2 enforces the joins between r and each atom labelling the successors of r in the join graph of q. At the same time f 2 ensures that, when evaluated over a database instance D, if there exists a fact of the form r(t r ) D that violates the KD specified on r together with a fact of the form r(t r ) D, which is in the image of a tuple t w.r.t. q, r(t r ) belongs to another image of t w.r.t. q. In other words, the atom guarantees that in any repair there exists an image of t (w.r.t. the KD on r). Such a check is iterated for other KDs by recursively invoking FolTree. The following example illustrates the way the algorithm works. Example 2.(contd.) It is easy to verify that the query q(x, z) : Journal(x, y), Editor (y, z) is KE-simple. Journal(x/KB, y/S) (N 1) - (N 2) Editor (y/KB, z/B) Now, by applying the algorithm KEFolRewrite and FolTree we obtain: KEFolRewrite(q) = {x, z | FolTree(N 1)} FolTree(N 1) = y 2 . Journal(x, y 2 ) y 2 . ConfPr (x, y 2 ) y. Journal(x, y) (FolTree(N 2)) FolTree(N 2) = Editor (y, z) y 2 . Editor (y, y 2 ) y 2 = z. 797 relations integrity constraints faculty/3 exam plan/10 key(f aculty) = {1, 2} key(plan status) = {1} course assignment/3 degree/5 key(exam plan) = {1} key(positioning) = {1} positioning/2 course/4 key(university) = {1} key(prof data) = {1} plan status/2 key(exam type) = {1} key(degree) = {1} prof data/3 key(course) = {1} key(exam) = {2} university/3 key(master exam) = {1} bachelor exam/2 key(bachelor exam) = {1} master exam/2 course assignment[2] professor [1] = exam type/2 master exam[1] bachelor exam[1] = exam/4 course[3, 4] bachelor exam[1, 2] = Figure 2: A portion of the test database schema By evaluating the rewriting over D we get { TODS, USA }, i.e., the set of consistent answers to q in D under S. Next, we state soundness and completeness of the algorithm. Theorem 8. Let S = A, K, E be a database schema, q be a KE-simple conjunctive query over S, and q r be the FOL rewriting returned by KEFolRewrite(q). Then, for every database instance D for S, a tuple t is a consistent answer to q in D under S iff t q D r . As a corollary, consistent query answering for KE-simple conjunctive queries over database schemas with KDs and EDs is polynomial in data complexity. EXPERIMENTS We now present some experimental results comparing the FOL and the Datalog rewriting previously described. To perform the experiments, we implemented a rewriting module that translates CQs issued over the database schema into both FOL queries and Datalog queries. FOL queries are in turn translated by the module into SQL queries. Then, we ran the SQL queries on a MySQL 4.1.10 instance of the test database, while we executed Datalog queries on DLV [15]. The experiments were conducted on a double processor machine, with 3 GHz Pentium IV Xeon CPU and 2 GB of main memory, running the Linux operating system. The test database holds information about the computer science engineering degrees of the university of Rome "La Sapienza" and contains 27 tables with an overall size of over 200.000 tuples. In Figure 2, we present the portion of the test database schema that is relevant for the queries (in the figure, "r/n" indicates that relation r is of arity n). Due to space limits, we only report details about three of the queries we tested: Q 0 = q(C) : -f aculty(C, U, INGEGNERIA ). Q 2 = q(S, D, P ) : -positioning(P S, P ), plan status(ST, DE), exam plan(C, S, P S, DT, ST, 1 , U 1, U 2, U 3, U 4). Q 3 = q(N, D, N P, CP ) : -master exam(C, N, T, 5 ), exam type(T, D), The queries have been posed on various instances of the test database with an increasing number of pairs of tuples violating some ICs. Figure 3, shows experimental results. In the charts 3(a), 3(b) and 3(c), the execution time of the SQL encoding and of the Datalog program are compared for queries Q 0 , Q 2 , and Q 3 . As expected, from a certain inconsistency level on, the execution time of the Datalog encoding has an exponential blow-up; in contrast, the execution time for the SQL encoding is constant on the average, and for Q 3 (Figure 3(b)) it decreases: although this might be surprising, it turns out that some inconsistency allows the SQL engine to prune the search space for query answering. Moreover, the chart presented in Figure 3(d) compares, on a logarithmic scale, the execution time of all queries at the highest inconsistency level. It shows that the SQL encoding is always more efficient when the degree of data inconsistency grows; however, it turns out that the method based on Datalog and DLV proves particularly efficient in the presence of few data inconsistencies. CONCLUSIONS The present work provides a general experimental validation of the first-order rewriting approach to the optimization of consistent query answering. Of course, the applicability of our technique is limited to the class of KE-simple queries. For general CQs, the use of a more expressive, and compu-tationally harder, query language like Datalog is necessary. Very recently, the first prototype implementations of consistent query answering have appeared, and the first efforts towards optimization of query processing are emerging. Within INFOMIX, several optimizations are currently under development to improve consistent query answering for more expressive classes of queries [9, 8]. In this respect, binding propagation techniques based on magic sets might significantly reduce execution time for Datalog programs on DLV [11], even if the coNP structure of the Datalog encoding suggests that the efficiency of the SQL rewriting can be hardly reached (especially for a large number of inconsistencies ). The ConQuer system [12] implements an extension of the technique of [13] which allows to rewrite in SQL queries belonging to the class C tree enriched with aggregates. Experiments show that the overhead of evaluating rewritten queries is not onerous if compared with evaluation of the original query over the inconsistent database. Therefore, [12] focuses on comparing standard query answering and consistent query answering, while our experiments compare two different query answering techniques. In this respect, we point out that optimization of our SQL rewriting was outside the scope of the present paper. Finally, Hippo [7] is a system for consistent answering of union of conjunctive queries without existential variables in the presence of denial constraints. Hence, this approach is different from our in terms of both query language and integrity constraints allowed. Moreover, Hippo techniques are not based on rewritings. As future work, we aim at extending our approach to other forms of ICs (e.g., foreign keys) and at optimizing the SQL rewriting produced by KEFolRewrite. 798 (a) Q 0 execution time (b) Q 3 execution time (c) Q 2 execution time (d) SQL vs. Datalog Figure 3: Experimental Results ACKNOWLEDGMENTS This research has been partially supported by the Project INFOMIX (IST-2001-33570) funded by the EU. REFERENCES [1] Marcelo Arenas, Leopoldo E. Bertossi, and Jan Chomicki. Consistent query answers in inconsistent databases. In Proc. of PODS'99, pages 6879, 1999. [2] Franz Baader, Diego Calvanese, Deborah McGuinness, Daniele Nardi, and Peter F. Patel-Schneider, editors. The Description Logic Handbook: Theory, Implementation and Applications. Cambridge University Press, 2003. [3] Loreto Bravo and Leopoldo Bertossi. Logic programming for consistently querying data integration systems. In Proc. of IJCAI 2003, pages 1015, 2003. [4] 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 2003, pages 260271, 2003. [5] Andrea Cal`i, Domenico Lembo, and Riccardo Rosati. Query rewriting and answering under constraints in data integration systems. In Proc. of IJCAI 2003, pages 1621, 2003. [6] Jan Chomicki and Jerzy Marcinkowski. On the computational complexity of minimal-change integrity maintenance in relational databases. In Inconsistency Tolerance, pages 119150, 2005. [7] Jan Chomicki, Jerzy Marcinkowski, and Slawomir Staworko. Computing consistent query answers using conflict hypergraphs. In Proc. of CIKM 2004, pages 417426, 2004. [8] Chiara Cumbo, Wolfgang Faber, Gianluigi Greco, and Nicola Leone. Enhancing the magic-set method for disjunctive datalog programs. In Proc. ICLP 2004), pages 371385, 2004. [9] 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 163177, 2003. [10] Thomas Eiter, Georg Gottlob, and Heikki Mannilla. Disjunctive Datalog. ACM Trans. on Database Systems, 22(3):364418, 1997. [11] Wolfgang Faber, Gianluigi Greco, and Nicola Leone. Magic sets and their application to data integration. In Proc. of ICDT 2005, pages 306320, 2005. [12] Ariel Fuxman, Elham Fazli, and Renee J. Miller. Conquer: Efficient management of inconsistent databases. In Proc. of SIGMOD 2005, pages 155166, 2005. [13] Ariel Fuxman and Renee J. Miller. First-order query rewriting for inconsistent databases. In Proc. of ICDT 2005, pages 337351, 2005. [14] Gianluigi Greco, Sergio Greco, and Ester Zumpano. A logical framework for querying and repairing inconsistent databases. IEEE Trans. on Knowledge and Data Engineering, 15(6):13891408, 2003. [15] Nicola Leone, Gerald Pfeifer, Wolfgang Faber, Thomas Eiter, Georg Gottlob, Simona Perri, and Francesco Scarcello. The DLV system for knowledge representation and reasoning. ACM Trans. on Computational Logic, 2005. To appear. 799

relational database;Query Rewriting;integrity constraints;query rewriting;consistent query answering;Computational Complexity;conjunctive queries;inconsistent database;Inconsistency;database schemas

56

Context-Aware Web Information Systems

Apart from completeness usability, performance and maintainability are the key quality aspects for Web information systems. Considering usability as key implies taking usage processes into account right from the beginning of systems development. Context-awareness appears as a promising idea for increasing usability of Web Information Systems. In the present paper we propose an approach to context-awareness of Web Information Systems that systematically distinguishes among the various important kinds of context. We show how parts of this context can be operationalized for increasing customers' usage comfort. Our approach permits designing Web information systems such that they meet high quality expectations concerning usability, performance and maintainability. We demonstrate the validity of our approach by discussing the part of a banking Web Information System dedicated to online home-loan application.

Introduction 1.1 Generations of Web Services Understanding Web Information Systems (WIS) as monolithic and presentation-oriented query-answer systems would be too simplistic. Implementing the individual services of a WIS only on the basis of XML or (D)HTML suites suffices for the interface accessible by a particular customer. The quality of service provided by a WIS both expected and implemented, however, evolved over the last decade and has evolved beyond mere completeness. Extending the classification in (Berger 2003, p.146) we distinguish between different generations of WIS. First generation (1G): "build it, and they will come" First develop a WIS, then customers will come, because they believe that it is useful. Many of the 1G-WIS were informational, i.e., they weren't interactive . Second generation (2G): "advertise online sales, and they will come" Develop a WIS and market it. Customers will Copyright c 2004, Australian Computer Society, Inc. This paper appeared at First Asia-Pacific Conference on Conceptual Modelling (APCCM 2004), Dunedin, New Zealand. Conferences in Research and Practice in Information Technology, Vol. 31. Sven Hartmann, John Roddick, Ed. Reproduction for academic , not-for profit purposes permitted provided this text is included. come, because the advertisement convinced them about the WIS's usability. The WIS may be transactional, i.e., contain interactive interfaces to company products and services. A standard interface is provided but hard to learn. No particular customer usage aid is offered. Third generation (3G): "realize a pleasant use of high quality services, and they will come" Customers will find using the WIS helpful. They will do the marketing. 3G-WIS's typical characteristics are: high value and up-to-date content, high performance, brand value of the provider, and pleasant and easy use for casual as well as for frequent customers. Many WIS including several banking WIS are still 2G. However, impressive and well-developed WIS, e.g., the Amazon web-site, demonstrate the feasibil-ity of 3G-WIS. The success of such WIS is based on deep understanding of the application area, the customers needs, abilities and habits. Adaptation to customers -- if provided -- is based on allocating the most suited subspace of the WIS application space to the customer. WIS can be classified into e-business, e-learning, edutainment, community, information and personality WIS. In the e-business class the B2B systems have been more successful than B2C systems. This success results from well-understood usage scenarios built into the WIS. We observe that usage scenarios are better understood for B2B-WIS than for B2C-WIS . Storyboarding is a design approach focusing on usage scenarios. However, so far it is mainly used employing pinboard approaches, see e.g. (Siegel 1998, Van Duyne et al. 2003). Pinboard approaches map a number of scenarios observed in the application onto tree-structured web sites. Storyboarding in the movie business is used to design much more complex scenarios . To overcome this limitation the storyboard specification language SiteLang has been introduced in (Thalheim and D usterh oft 2001). Until now it has been applied in more that two dozen WIS projects of the Cottbus InfoTeam since 1999. Our development experience implies that implementing 3G-WIS requires sophisticated database support , see (Thalheim 2000a). Our approach to guarantee for this support is based on the theory of media types, which generalize database views (see e.g. (Schewe and Thalheim 2001)). Another finding from our practical experiences is that customer behavior has changed. They are no more patiently waiting until their needs are met. They require personal interfaces . Customization of system interfaces to users is 37 known for quite a while. However, WIS are targeting new and casual customers. These customers are not capable or willing to arrange for system adaptation. Internet service providers report customers frequently complaining about insufficient user-friendliness and unsophisticated WIS. 1.2 Problems of Complex Applications Modern applications, in particular WIS often appear to be relatively simple, if only their interface is considered . Their point-and-click operating mode is de-liberately set up in a way that causes the impression of simplicity. Internally, however, things may be quite different. A client-server multi-tier architecture with HTML-server, database server and application server might be used. This implies some non-trivial development tasks done such as database design and development of an application programmer interface or similar. In addition, several customer types may be known to the application system. A WIS may appear very different to customers of different type. The functionality they access, however, is still the basic functionality as implemented by the servers mentioned before . Consequently as many function schemas and data schemas need to be developed as there are anticipated customer types. These schemas need to be integrated to develop a consistent view of the key application functionalities. Views that are based on these schemas need to be generated allowing the individual customers to operate with the application in the way that is most natural for them. The development of WIS thus can be a quite complex process. Among others the complexity of this development process depends on the degree of use of an underlying database, from which dynamic web pages are created. Furthermore, the complexity of this process depends on the number of versions of usage processes of the WIS that need to be anticipated. Since different usage processes may lead to different data and functionality accessible to customers. Additional complexity comes in -- e.g. in the case of modern retail banking -- when a requirement is set in place that various access channels -- e.g. channels needed for cell-phone - or PDA-access -- should be made available to customers. Apart from the purely technical problem arising from discretionary access channels the problem of layout for these channels has to be solved. 1.3 An Application Example An example of a WIS in retail banking showing a relatively high diversity of the usage process is online loan application, if considered in full generality as we do it here. For an introduction to lending in general , of which the loan business is just a part, see e.g. (Valentine 1999). Not all banks offer online home-loan application facilities. Those that provide such facilities do not necessarily allow customers to deal with them completely online. Banks that offer effective online loan application are the Swiss UBS AG and the New Zealand and Australia based ASB Bank. The acceptance of a longer interruption of service at the ASB site indicated that at least for this bank online home-loan application is not yet considered a major part of their business. Complexity in home-loan applications results from the fact that the applicant not necessarily is exactly one natural person. For each of the applicants properties and debts need to be identified and valuated. Often banks would accept only home-loan applications of at most two people, in general the couple that is going to live in the home financed with the loan. Complexity is further increased by the loan not necessarily being a fresh one but being already granted to someone who due to his or her financial conditions has chosen to move the loan to a different bank. Furthermore, the properties offered for securing the loan may belong to a variety of types. Some of these types, e.g. real estate property may require physical inspection to determine the value they can cover. Other properties such as financial instruments, i.e. shares, options or accounts, may only need an inquiry to the respective depot or account. If cash is a security, then it might even be impossible to finish the process electronically, as the cash needs to be brought to the bank branch, counted and deposited. It is similar with debts. On real estate properties there might be liabilities that require an assessment of the actual value. Of course there is quite a number of so-called loan structures (see (Valentine 1999, p.226f.)) distinguishing between loans. For instance, they may differ from each other in their term, frequency of repayments, borrower's authorization to increase the debt (e.g. overdraft facility), the minimum security ratio or the repayment structure. The latter addresses the schema of how capital and interests are paid for by the customer. Independently of the loan structure a customer might chose among several loan options that specify how the interests develop over time, i.e. they may be fixed for a particular time period or they may float like general interest rates in banking. Apart from these principal choices in a home-loan there are a number of tools available for customer's use throughout online home loan applications such as a borrowing power calculator, a repayment schedule calculator, etc. Additionally, dictionaries of banking terms, act excerpts and comments as well as descriptions of the applying financial instruments need to be accessible to customers. All the possible options in the financial instruments and the respective variations of the WIS usage will only be considered by a small number of customers . In more technical terms we have to deal with a generic process type. Most of its instances realize only a part of the possible variations. At present online home-loan application systems are not typical retail banking applications. Automated clearing house (ACH), i.e. direct deposit of payments, withdrawing monthly mortgage payments, etc. are more typical. According to (Berger 2003, p.150f.) its use in the US is steeply increasing and has after starting problems even increased productivity. However, ACH is a back-office activity, whereas online home-loan application is a customer-home or front-office activity. According to (Berger 2003, p.149) internet-only banks performed more poorly than conventional banks did. If this finding implies that online home-loan applications are less productive than conventional home-loan application processing then we believe that this is only a temporary phenomenon. We believe that cultural obstacles concerning internet-banking will disappear when 3G-WIS have become more popular. According to (Berger 2003) there is empirical evidence for an increasing market share of electronic payment . According to studies reported in (Berger 2003, p.162) there is even empirical evidence for increased productivity due to investment in IT labor, while there is no empirical evidence for IT investments increasing efficiency in general. This is consistent with the basic insight that not the mere use of IT but the kind and quality of this use can increase productivity . Our paper shall help making 3G-WIS more popular and thus contributes to internet banking more completely covering the business at a higher level of quality. 38 1.4 Related Work A lot of related work has been done on the development of web information systems. The work in (Atzeni et al. 1998) emphasizes the design of content leading to databases, navigation leading to hypertext , and presentation leading to the pages layout. Other authors (see for example (Baresi et al. 2000), (Bonifati et al. 2000), (G adtke and Turowski 1999) and (Rossi et al. 1999)) follow the same lines of thought or concentrate on the "add-on" to database design, emphasizing mainly the hypertext design dealing with navigation structures (see (Garzotto et al. 1993) and (Schwabe and Rossi 1998)). The work in (Feyer et al. 1998) presents the forerunner of the theory of media types (see (Schewe and Thalheim 2001)). Media types provide a theoretically sound way to integrate databases, external views, navigation structures , operations, and even support adaptivity to different users, environments and channels. The adaptivity feature distinguishes them from the dialogue types that are used to integrate database systems with their user interfaces (see (Schewe and Schewe 2000)). The work in (Schewe and Thalheim 2001) already emphasizes that conceptual abstraction from content, functionality, and presentation of an intended site is not sufficient for the adequate conceptual modelling of web-based systems, even if complex media types are taken into consideration. Some of the approaches mentioned before (see (Atzeni et al. 1998), (Baresi et al. 2000), (Bonifati et al. 2000), (G adtke and Turowski 1999), (Rossi et al. 1999), (Garzotto et al. 1993) and (Schwabe and Rossi 1998)) miss out on the important aspect of story boarding, which is needed to capture the business content of the system. Story boarding in a process-oriented holistic manner focusses on user intentions. In more recent work some of the authors (Kaschek et al. 2003a) started to investigate this idea more thoroughly. Conceptual modelling traditionally considered more ontolog-ical aspects than epistemological ones. Since web information systems in two respects considerably differ from non-web information systems epistemological aspects, however, need to be taken more seriously: Web information systems are open in the sense that actual users virtually may be just anyone. In non-web system there was traditionally a much stricter access control preventing non-staff from using the system . The business idea, however, has changed and customers need to be attracted and pre-selected by a web information system. Furthermore, web information systems are open in the sense that it is very easy to use them for accessing other web systems. This introduces more competition among those who offer services on the web. Quality of web information systems in the sense of fitness for users' use thus tends to be more important than it was for non-web systems. Web information systems partly substitute staff-customer interaction by customer-computer interaction . Consequently, web information systems must focus on aiding customers in doing the business the system provider is engaged in. Clearly this only can be done on the basis of a customer model. User profiling together with story boarding is a holistic manner for this. In (Schewe and Thalheim 2001) it is suggested that story boarding be supported through directed graphs called scenarios, in which the nodes represent the scenes and the edges correspond either to navigation or to actions issued by the user. This extends the work in (Feyer et al. 1998), where simply partially ordered sets have been used. In addition, user profiling is approached by using user dimensions capturing various aspects of how to characterise users. This has been extended in (Srinivasa 2001) to a formal description of interactive systems. The work in (D usterh oft and Thalheim 2001) presents a formalised language SiteLang to support the specification of story boards. the work also indicates ideas how to exploit word fields for designing dialogue steps in story boards. In (Schewe et al. 1995) and (Schewe 1996) refinement primitives for dialogues have been discussed. Due to the connection between dialogues and scenarios, this approach to refinement is also useful for story boarding. The work in (Schewe et al. 2002) applies story boarding and user profiling to the area of on-line loan systems. 1.5 Outline In section 2 we discuss WIS specification, in particular story spaces and scenarios, we further discuss media objects, dialogue-step specification and context. In the following section 3 we discuss database design for WIS, utilization of context for WIS and a stepwise WIS generation approach called "onion generation". Finally, in section 4 we continue the discussion of our example and show how our approach can be applied to modelling of WIS. Due to space restrictions, however, we can only discuss the storyboarding part. WIS Specification 2.1 Story Spaces and Scenario Modelling usage processes right from the beginning of systems development requires using a sufficiently expressive high level semantic model as a respective conceptual framework. Storyboarding uses the metaphor "story" to conceptualize usage processes. We presuppose that a story (for the source of the interrogatives used here refer to (Zachman 1987, Sowa and Zachman 1992)) tells what happened, why and where, as well as who did it how and when. The story of customer-WIS interaction thus is the intrigue or plot of a narrative work or an account of events. Within a story one can distinguish threads of activity , so-called scenarios, i.e., paths of scenes that are connected by transitions. See figure 2.1 for an example scenario. We do not intend to model branching stories. These require managing a number of activities at the same time, i.e., in parallel. A capability that -as we believe- many casual customers won't have. With the term story space we mean the integration of all scenarios in a story. We define the story space W of a WIS W as the 7-tuple (S W , T W , E W , G W , A W , W , W ) where S W , T W , E W , G W and A W are the set of scenes created by W , the set of scene transitions and events that can occur, the set of guards and the set of actions that are relevant for W , respectively. Thus, T W is a subset of S W S W . Furthermore W : S W SceneSpec is a function associating a scene specification with each scene in S W , and W : T W E W G W A W , t (e, g, a) is a function associating with each scene transition t occurring in W the event e that triggers transition t, the guard g, i.e. a logical condition blocking the transition if it evaluates to false on occurrence of e, and the action a that is performed while the transition takes place. The language SiteLang, see (Thalheim and D usterh oft 2001), offers concepts and notation for specification of story spaces, scene and scenarios in them. Scenes and their specifications are discussed in subsection 2.2. 2.2 Scenes We consider scenes as the conceptual locations at which the customer-WIS interaction, i.e., dialogue 39 sc 1 - sc 2 - sc 3 - sc 4 - sc 5 - ... y 6 ? 9 side story Figure 2.1: Scenario with a loop representing a side story takes place. Dialogues can be specified using so-called dialogue-step expressions. Scenes can be distinguished from each other by means of their identifier : Scene-ID. With each scene there is associated a media object and the set of actors that are involved in it. Furthermore, with each scene a representation specification is associated as well as a context. Scenes therefore can be specified using the following frame: Scene = ( Scene-ID DialogueStepExpression MediaObject Actors ActorID Right Tasks Assigned Roles Representation (styles, defaults, emphasis, ...) Context (equipment, channel, particular) Dialogue-step expressions consist of dialogues and operators applied to them. Dialogue steps are discussed in subsection 2.4 below. The provided operators are based on the basic dialogue step algebra introduced in (Thalheim and D usterh oft 2001): Basic control commands are sequence ; (execution of steps in sequence), parallel split | | (execute steps in parallel), exclusive choice | | (choose one execution path from many alternatives), synchronization | sync | (synchronize two parallel threads of execution by an synchronization condition sync , and simple merge + (merge two alternative execution paths). The exclusive choice is considered to be the default parallel operation and is denoted by ||. Structural control commands are arbitrary cycles (execute steps w/out any structural restriction on loops), arbitrary cycles + (execute steps w/out any structural restriction on loops but at least once), optional execution [ ] (execute the step zero times or once), implicit termination (terminate if there is nothing to be done), entry step in the scene and termination step in the scene . Advanced branching and synchronization control commands are multiple choice | ( m,n) | (choose between m and n execution paths from several alternatives ), multiple merge (merge many execution paths without synchronizing), discriminator (merge many execution paths without synchronizing , execute the subsequent steps only once) n-out-of-m join (merge many execution paths, perform partial synchronization and execute subsequent step only once), and synchronizing join (merge many execution paths, synchronize if many paths are taken, simple merge if only one execution path is taken). We also may define control commands on multiple objects (CMO) such as CMO with a priori known design time knowledge (generate many instances of one step when a number of instances is known at the design time), CMO with a priori known runtime knowledge (generate many instances of one step when a number of instances can be determined at some point during the runtime (as in FOR loops)), CMO with no a priori runtime knowledge (generate many instances of one step when a number of instances cannot be determined (as in a while loop)), and CMO requiring synchronization (synchronization edges) (generate many instances of one activity and synchronize afterwards). State-based control commands are deferred choice (execute one of the two alternative threads, the choice which tread is to be executed should be implicit), interleaved parallel executing (execute two activities in random order, but not in parallel ), and milestone (enable an activity until a milestone has been reached). Finally, cancellation control commands are used, e.g. cancel step (cancel (disable) an enabled step) and cancel case (cancel (disable) the step). These control composition operators are generalizations of workflow patterns ( see, e.g. (Workflow Management Coalition 1999, Jablonski 1996)) and follow approaches developed for Petri net algebras. A graphical representation of a login scene is given in figure 2.2. We are interested in well-formed dialogues and do not allow specifications which lead to and-split or or-split common in workflow specifications . This scene is specified by the dialogue step expression Enter login ; ( Customer login ; [ Change profile ; ] ( Service kind selection ; Service selection ; Service customization) || Join cooperating group || Join bank club || Join bank programs || General customer information ) | | ( Anonymous Login ; [Extend adding identity ; ] ( Program selection ; Module selection ; Unit selection) ) Enter Login : U Change profile Y j General customer information Anonymous login j : Extend by adding identity K Service kind selection j Service seeking selection U Service customization Customer login K y j U : Join cooperating group Join bank program Join bank club Login Scene With Adaptation of System Facilities Figure 2.2: Scene for Login Into a Bank WIS 2.3 Media Objects A scene is supported by media objects following the codesign approach. Media objects are instances of media types. 40 Bank Service Customer Service Role Customer Login Customer Profile Profile Type Task Portfolio Type Customer Portfolio Web Address Account Login History Figure 2.3: Cutout of the profiling schema The core of a media type is defined by a view on some underlying database schema, i.e. it consists of a view schema and a defining query. However, this query must be able to create identifiers in order to create links between the various media objects. This core of a media type -- called raw media type in (Schewe and Thalheim 2001) -- is extended in three directions : As a first extension operations are added to the view in the same way as d-operations were added to dialogue objects in (Schewe and Schewe 2000). Basically, the use of operations just adds dynamics to the media objects. So, if a media object is associated with a scene, the operations of the media object define the available dynamic functionality . The second extension provides adaptivity and hierarchies. Adaptivity to the user deals with needs arising from different users. Adaptivity to the technical environment copes with technical restrictions of end-devices. Adaptivity to the communication channel deals with adaptation to needs arising from various communication channels. For all three forms of adaptivity media types provide mechanisms for a controlled form of information loss, which is coupled with algorithms for the splitting of information content. The hierarchies are adopted from dimension hierarchies in OLAP. The third extension simply covers ordering and other presentation options. Thus, roughly speaking media objects consist of abstract containers, supported DBMS processes and database manipulations requests. Basic media objects (Schewe and Thalheim 2000) are characterized by syntactic expressions, have a semantical meaning and are used within a certain pragmatical framework. Media objects can be parameterized. Typical parameters are the representation style, the actor frame, and the context frame. Therefore we distinguish between media objects and runtime media objects in which all parameters are instantiated. During runtime, the media object is extended by specific escort information (Thalheim 2000). This escort information is represented for user support. It allows the user to see the history of steps performed before being in the current state. Escort information is further generated from the story space. In this case a user is informed on alternative paths which could be used to reach the given scene and which might be used for backtracking from the current scene. For the generation of media objects and their composition on the basis of information units we extend the classical SQL frame to the frame generate Mapping : Vars Structure from Views where Selection condition represent using Style guide & Abstraction browsing definition Condition & Navigation The views and therefore the media object may have hidden parameters (for instance, EventID) which are not visible to the actor. They can be parameterized by variables (for instance, @Today). For media objects we reuse ideas developed for OLAP technology (Thalheim 2000): views on ER schemata (abstraction on schemata (aggregation, scoping, ...), versions), variations of generation functions, display with canonical functionality (drill-down, roll-up, rotate, pivoting, push, pull, dimension, aggregation), using generic evaluation functions and models, implicit incorporation of hierarchies and implicit incorporation of time, space, .... Furthermore, involved actors are specified in dependence on their profiles, tasks assigned to them, their access and manipulation rights, and their roles to be taken while visiting the scene. This specification is based on (Altus 2000) and similar to profiles of actors in information systems. It is our aim to specify generic scenes. Thus, we add the representation styles which can be applied to the media object of the scene. Representation depends on the equipment of the actor. In the city site projects, we have gained experience with different representation styles: internet display with high-speed channel, internet-display with medium speed display (default style), videotext and WAP display. For instance, for videotext any graphical information is cut out or replaced by textual information. Finally, the context of access is specified. Access determines the display facilities. Channels can be of high or low speed. The particular usage of a scene by an actor depends on the scenario history. The login scene in Figure 2.2 is based on the schema in Figure 2.3. The corresponding media object specification has the following structure: MediaObject( @Customer ID) = generate (ID, profile, portfolio, context) from Customer 1 Login Account History 1 Customer Profile 1 Customer Portfolio 1 ... where Customer.ID = @Customer ID ... represent using 41 XSL style.Ident = Profile Type.Preference.StyleIdent & createVarsFor(profile, portfolio, context) browsing definition Customer portfolio ... & Navigation none The representation styles determine the order and the tailoring of the elements of the media object. 2.4 Dialogue Steps We conceptualize the customer-WIS interaction as a dialogue between these two. Therefore the customer-WIS interaction unfolds in a sequences of dialogue steps, i.e., elementary communication acts. The basic WIS-state transformations triggered by actors can thus be understood as caused by dialogue steps. These may access the media object that is associated to the scene within which the dialogue step occurs. Comparable to (Goldin et al. 2000) we use the following frame for specifying the control of dialogue steps: on precond if event and guard do action result in postcond Consequently dialogue steps may be specified by the following frame: DialogueStep( Identification ) = ( sub-unit = view on media object of the scene enabled processes = subset of supplied processes, manipulation requests actor = subset of enabled actors in a given context control = ( precondition, enabling event, guard, postcondition) ) Dialogue step specifications can be represented graphically as shown in figure 2.4. The figure for the scene 'anonymous login' represents the specification of dialogue step 'login'. Anonymous login) BankSurveyView ServiceOfferView SelectModule, SelectCommunication NewSession AddToLog LogChannelData LogUserEngineData Anonymous User, Visitor 6 ? (ClickAnonymous,ServicesAvailable,ServiceSelected CustomerStyleSelected ClickOnOneOption) Figure 2.4: Dialogue Step for Anonymous Login Based on the properties of the actions we conclude, for instance, that after withdrawal a previous member of a cooperating group cannot participate in the discussions in the community. A task property frame is defined by a task name, reasons for task involvement, an aim, a postcondition (enabled next activities), the information from the database, the information for the database, the resources (actor, resources, partner ), and a starting situation (precondition, activity, priority, frequency, repetition rate). We use graphical representations of scene specifications as indicated by figure 2.5. Scenes are represented by frameboxes and dialogue steps by ellipses. The transitions among dialogue steps are represented by arrows between these. We use the graphical notation developed for state charts, e.g., the default start step of a scene is denoted by a solid circle, the end state by a solid circle surrounded by an empty circle, the history entry into a scene is denoted by an `H' surrounded by an empty circle. Furthermore, we can adopt refinement and clustering, concurrency, delays and time-outs, transient states, event priorities and parameterized states. For more detail on state charts see, e.g. (Harel and Naamad 1996) and for their application (Rumbaugh et al. 1991). dialogue step next dialogue step sub -unit enabled process manipulation sub-request enabled actor 6 ? control : U dialogue scene expression transition according to scene involved actors story scene sequence media object representation style context, task 6 6 6 6 6 6 Figure 2.5: Representation of scene specifications 2.5 Context Context has been usually defined within the object sets of the database (Bell 2001, Connolly 2001). There only very few trials to consider context of the scenarios or stories (Whitsey 2003). In (Thalheim 2000a) context has been defined for media types. For dealing more complete and justifiable with context we start with a dictionary definition of context of something as that what one needs to understand the something . This implies our understanding of context as a three place predicate C(S, H, A) which if true says that actor A needs helper H to act reasonably on S. If the actor is an individual then we stay with the focus on understanding. For non-human actors, however, we focus on acting according to predefined quality aspects and behavior rules. The something we consider here as relevant are WIS-parts. The helpers we here take into account are the various data that are relevant for the WIS-parts in question. The actors we consider here are the WIS and the individuals occupying the roles: customer, vendor and developer with respect to the WIS at hand. We thus distinguish the following contexts: Customer's scenario context , i.e., that what the customer needs to understand for efficiently and ef-fectively solve his/her business problem. Vendor's WIS-context , i.e., that what the vendor needs to understand how to run the WIS economically . Data that typically is part of this context are: the intention of the provider, the theme of the web site, the mission or corporate identity of the site, and the occasion and purpose of the visits of actors . Developer's WIS-context , i.e., that what the developer needs to understand for being capable of implementing the WIS. Data that typically is part of this context are: the potential environment, e.g. hard- and software, channels, the information system, especially the associated databases, 42 Context Dialogue Expression AcceptCond ID ID Do Condition Event Particular Default Obligat Usage Emphasis Profile Group (1,1) (1,n) (1,1) ? enabled 6 ? involved : Actor Dialogue Step uses 6 ? used Right : Right Category Media Object Task Role Category Task Assignment ? Representation Style 6 basedOn Scene Activity Sequence ? ? Story in Figure 3.1: The Structure of the Web Site Database the story space, scenes, dialogue steps, roles, and rights, the tasks to be performed within the story, and the roles in the scenario. WIS's scene context , i.e., that what the WIS needs to be capable of making solving certain business problems easy and pleasant for customers. Data that typically is part of this context are: History and current usage allow context adaptation to scenarios which are played at present by the current user. Adaptation to the current environment is defined as context adaptation to the current channel, to the client infrastructure and to the server load. Users are grouped to actors. Therefore, we can define the current user by instantiation of the actor. Goals and particular, policy (exceptions, social , organizational) define a specialization of the content, structuring and functionality of a web page. A WIS is supported by media objects that belong to media types. The collection of all media types is called suite. Our framework offers four hooks for dealing with the context we need to consider: 1. Specialization of media type suite for usage and user adaptation: The database types may have subtypes specializing the database types. Media types are defined on the basis of views. Therefore, we can follow the approach discussed in (Thalheim 2000a) for specialization of types. Specialization is defin-able through specialization of types, instantiation of parameters. extension of types, and restriction and constraint application. 2. Application of rules towards generation of extended suites: Suites can be extended by providing view rules defining views on top of the media types. This approach supports portfolio extension and container extension. 3. Instantiation of explicit context parameters can be used for adaptation of web sites to the current profile, to the current environment or to the current workload. 4. Storage of utilization profile similar to login track supports to use history of previous utilization of the web site by the user. We extend the web site database by explicit utilization logs for used media objects, preferences of usage, users workspace or work rooms, and variations of users media objects Developing the Database Used to Generate the Web Site 3.1 Database Modelling for WIS Web site management becomes a nightmare whenever a web site has been developed in a handicraft approach. For this reason, generation of web sites is currently based on web site content management . Content management systems currently support the representation of web pages on a take-and-place metaphor: Select or compile the content objects of a page, compile the navigation structure and place the content objects using page frames. Our web site team also used this approach. This approach is entirely satisfying the needs as long as the general structure of a web site is stable and no adaptation to the user is required. In order to dynamically generate the web site we decided to store the web site stories in a database. The structure of this database is displayed in figure 3.1. We specify the web site story space based on the our web SiteLang. This specification is inserted into the database by the SiteLang editor. We can now extract the page under consideration by an instantiated query from this database. Context may be infused directly depending on the query result. Similar to the context infusion, users of a web site have their own profile, their own portfolio and their history. This information is used for adapting the content of the web site to the current usage. 3.2 Context Infusion in Scenarios Typical business processes have a very large number of variants. Classically, workflow approaches have 43 been used for specification of such varieties. Since the complexity of variants might be much higher the workflow approach did not succeed in providing a sound basis for the specification of all variants. We observe, however, that in practice these varieties are internally structured. They may be composed, extended or filtered by smaller scenarios. e-banking challenges storyboarding by its orthogonality and variety . Instead of specifying all possible variants we prefer to model the generation mechanism of the very large variety of scenarios. This generation supports runtime adaptation to the current scenario, the context and other parameters. At the same time, banking sites are threatened to expose customers to the "lost in hyperspace syndrome". Therefore, customers should be supported in tracking back onto the right path. Our solution to this challenge is based on generic parameters that are instantiated depending on the customer, the history, the context etc. Each set of media objects is specified by a context-free expression with a set of parameters. These parameters are instantiated depending on the customer profile, the customer task portfolio, the customer computational environment, the presentation environment, and the available and accessible media objects. Instead of providing a full generation rule set we illustrate our approach on the basis of an example. A customer of a bank provides his/her identity e 1 , inserts some data e 2 ,1 and e 2 ,2 in any order or signs that the bank may request these data from somewhere else e 2 ,3 . Then the customer seeks a loan and fills the corresponding forms e 3 . The customer gives bail data in different variants (e 5 ,1 || (e 5 ,2 ; e 5 ,3 )). The scenario is supported by the eight media objects. Now we can inject the context into the media object expression of the scenario. For instance, we may have the following stepwise refinements: Media objects of a scenario: e 1 ; ((e 2 ,1 ||e 2 ,2 ) | | e 2 ,3 ) ; e 3 ; (e 5 ,1 || (e 5 ,2 ; e 5 ,3 )) Extending by objects syntactic verbal context and meta-information: e 16 ; [ e 21 ; ] e 1 ; ((e 2 ,1 ||e 2 ,2 ) | | e 2 ,3 ) ; e 9 ; e 3 ; (e 10 ||e 11 ) ; (e 5 ,1 || (e 5 ,2 ; e 5 ,3 )) Extending by story space associations, e.g., side paths, , filtering against availability and compiling against the customer profile e 16 ; [ e 21 ; ] e 1 ; ((e 2 ,1 ||( SB e 2 ,2 || CB e 2 ,2 )) | | e 2 ,3 ) ; [( e 17 ; e 18 ; )] e 9 ; Gr e 3 ,1 ; An e 3 ,2 ; Inf e 3 ,3 ; F orm e 3 ; (e 10 ||e 11 ) ; (e 5 ,1 || (e 5 ,2 ; e 5 ,3 )) Filtering with or extending by the web site context: e 16 ; [ e 21 ; ] e 1 ; ( SB e 2 ,2 | | e 2 ,3 ) ; [( e 17 ; e 18 ; )] e 9 ; Gr e 3 ,1 ; An e 3 ,2 ; Inf e 3 ,3 ; F orm e 3 ; (e 10 ||e 11 ) ; ( e 5 ,2 ; e 5 ,3 ) Coping customer's history - already finished dialogue steps and repeating dialogue steps: e Repe 1 ; [( e 17 ; e 18 ; )] e Repe 9 ; Gr e 3 ,1 ; An e 3 ,2 ; Inf e 3 ,3 ; F orm e 3 ; (e 10 ||e 11 ) ; (e 5 ,2 ; e 5 ,3 ) Coping with customers history negotiation steps and pragmatical elements: e Repe 1 ; e 25 ; [( e 17 ; e 18 ; )] e Repe 9 ; Gr e 3 ,1 ; An e 3 ,2 ; Inf e 3 ,3 ; F orm e 3 ; (e 10 ||e 11 ) ; (e 5 ,2 ; e P rak 5 ,2 ; e 5 ,3 ) 3.3 The Onion Generation XML documents provide a universal structuring mechanism. XSL rules allow to generate XML documents from XML suites. This opportunity supports a multi-layer generation of web information systems. Thus we use the multi-layer onion generation presented in Figure 3.2. presentation engine actor profile adaptation, equipment adaptation, channel adaptation, decomposer, style extension container engine services packages, wrapping functions, dialogue scene and scenario functions units engine survey, landmark, indexing, I/O, navigation, integration etc. functions view handler virtual materialized views update views DBS ... DBMS Figure 3.2: The Onion Approach to Stepwise WIS-Generation The onion generation approach is based on the layered structure of the WIS arising from the use of SiteLang and media objects. On the outermost shell the presentation facilities are introduced. This shell deals with style presentation functions. Containers used in the next inner shell are used to ship information from the web-server to the user. Thus, this shell deals with the adaptation to the user and his/her environment. The next inner shell handles the information units, i.e. the core media objects. Inside this shell we find further shells dealing with views on the underlying database, and innermost we find the database itself. The onion approach fits nicely into a translational approach, which generates consistent sets of XML documents. In our projects we used the XML extender of the database system DB2 to generate XML documents. Thus, the layering approach to the generation of XML displayed in Figure 3.2 allows to use another strategy to generate XML documents. This facility is displayed in Figure 3.3. This transformation approach has been success-fully used in two of our e-learning projects and our community services projects. These project require sophisticated context adaptation. The approach implements an XML suite on top of the relational DBMS DB2. The extended ER model (Thalheim 2000) provides a better approach to XML suite generation than relational models or the classical ER model for a number of reasons: Structures can be defined already in complex nested formats. Types of higher order are supported. The model uses cardinality constraints with participation semantics. 44 conceptual representation abstract XML representation XML implementation on top of DB2 dynamic scene object XML scene onion reflective adaptations container media object meta functions views functions = = = database schema (HERM) functors for XSLT functors for XSLT container onion media object onion XML suite DTD functors for XSLT functors for XSLT enriched XML suite enriched XML suite enriched XML suite XML documents DAC for DB2 access ? ? ? ? ? ? ? ? j j j ? ? ? ? Figure 3.3: The General Procedure for Translation from SiteLang to XML An Advanced e-Banking Application 4.1 Banking and Mortgages According to (Wierichs and Smets 2001) a bank is an ". . . institution that as part of an economy offers financial services. The economical function of banks is to create a liquidity equalization in the cash flow that is reverse to the product and service flow. The focal points of the bank operational activity are conducting payments, acceptance of money for investment, and granting credits." Furthermore the particular liquidity equalization that is chosen out of the set of possible such equalizations is a preferable one. The respective preference structure is worked out by banks on base of an assessment involving financing cost and interests, see (Matthews et al. 2003). A loan according to (Wierichs and Smets 2001) is the "relinquishment of money or other fungible (...) properties connected with the obligation of the debtor to give back the relinquished in equal kind, quality and quantity." An enhanced version of the model of the loan process is represented in figure 4.1 as a UML sequence diagram. It shows the roles involved in the loan process as the labels inside the rectangular boxes on the top of the diagram. It further indicates the concurrency that may be utilized in this process. It achieves this by means of showing the communication between the roles. This communication is represented by the arrows starting at the dashed lines that represent logical time, i.e., life lines of the roles. The labels attached to the arrows indicate the content of the message associated with the respective arrow. The bottom level rectangle containing the messages 'Payback()' and 'CheckPayback()' signifies that these messages are to be repeatedly sent until the stop condition signified by the asterisk and displayed below the rectangle 'debit position balanced' becomes true. 4.2 Mortgages and variants The figure 4.1 from a bank technical point of view schematizes the process. This process clearly is not fully suited as the only base of application development . For aiding development more information is needed about how customers are anticipated to interact with the system under construction. We use here the function W of the story space W of a WIS W to show how the customer interaction with the WIS changes the appearance of it for the customer Story boarding is a useful technique to obtain the required information. Our respective starting point is the investigation of the Web site of the Australian and New Zealand based ASB Bank. From earlier work, see (Kaschek et al. 2003) we knew that it offered an online loan application facility. We investigated this Web site more closely and found that this site at each of its pages essentially offered customers data that can be typed as follows: advertisement, i.e., information about ASB Bank including a welcome and a logo. disclaimer, i.e., a statement limiting the legal responsibility of ASB Bank with respect to the data displayed and the implications customers might draw from it. search, i.e., a facility taking an unlimited customer input and returning those ASB Bank pages that best met this search expression. highlights, i.e., the main contents that ASB Bank wants to be displayed at each particular of its Web pages. path, i.e., a redundancy eliminated sequence of ASB Bank Web pages visited so far by the customer interacting with the site and supposed to be used as a navigation aid. reference, i.e., a couple of links the target of which offer more information about the page actually visited by the customer. business branch selector, i.e., a navigation bar that breaks down the information space of the site into subspaces according to the business branches of ASB Bank. subspace selector, i.e., a navigation bar that for each subspace that corresponds to a business branch breaks down the subspace into 2nd. level subspaces. subspace navigator, i.e., for each 2nd. level subspace a navigation bar breaking down the subspace in a number of information space locations . 45 Marketer Customer ProductAd() Inquiry() RevisedProductAd() Analyst Application() Lender ApplicationApproval() Notification() Service Documentation() Documentation() SignedContract() SignedContract() Accounts AdvanceFunds() PositionsGenerated() Payback() UseLoan() Moniter Start() CheckPayback() *[debit position balanced] Figure 4.1: UML diagram representing the loan process We have then represented the navigation structure offered by the ASB Web site as a state chart the states of which represent scenes. The state transitions are presupposed to be triggered by customers clicking links, i.e., navigation events. The labels attached to the state transitions are a string representing the navigation event and an action carried out throughout the transition. This action is prefixed by a slash, i.e., by "/". The semantics of the action is specified in form of a programming language like assignment and assigns new values to variables holding the data listed above. In this way one can specify what data and functionality is accessible to a customer at a particular scene. Explanations and tables or the like are presupposed to be just text. All other transition labels used as values in assignments are presupposed to be links. If a variable is supposed to hold several links then these are connected by a plus sign, i.e., by "+". If more than one action has to take place at a transition then all these actions are connected by a &-sign. The variables used in figure 4.2 are D, S, H, A, R and P respectively representing values of type disclaimer , search facility, highlight, advertisement, references and path. Those of them being displayed at a particular page are represented as non delimited string, i.e., if all of them occur the string DSHARP is attached as label to the state representing the page. Furthermore the variables BS, SS and SN are used to respectively represent values of type business branch selector, subspace selector and subspace navigator. The initial state in the figure is reached after moving onto the home page of ASB Bank and clicking BS.Personal which signifies the business branch of retail banking. The other 1st. level subspaces of the application's information space are "All", "Business", "Institutional" and "Rural" in the obvious meaning . The subspace selector of "BS.Personal" allows to chose from 18 different 2nd. level subspaces. One of them is Home loans. Clicking it, i.e., "BS.Personal-SS .Home loans" leads to the initial state of figure 4.2. If required we could add further navigation detail including the impact of navigation on the variables occurring in the figure. Furthermore if it would be required we could add further variables to represent data of types here not dealt with. 4.3 Adaptation to customers, context and specific case Adaptation to customers is a must if optimal customer support is aimed at. ASB Bank realizes a limited customer adaptation in that it offers in the subspace selector of BS.Personal second level subspaces both for kids and for young folks. ASB Bank concerning the home loan subspace of its information space does not offer much adaptation to customers. It only offers a bank technical terms dictionary and specif-ically addresses first home buyers. Neither are all New Zealand official language versions of the Web site available nor can it be tuned to meet any kind of disabilities such as weak eyesight or color blindness. The approach to adaptation to customers taken by sites like the one under investigation consists in identifying the subspace of the information space they create that most likely will fit best the needs of a particular customer. The match between customer and subspace is then done such that the customer is asked to give some characteristics of him or her into the system and based on that the respective subspace is chosen. ASB Bank does so concerning kids and young folks. Other banks have additionally the customer type student or wealthy individual. This strategy is suggested by the fact that the site vendor in general does not know much about the individuals accessing its site. A technique to consider knowledge about customers to the design process is the creation and use of personas, i.e., archetypical customers and design the navigation structure as well as the page layout such that it fits optimally to the personas used. Concerning more detail about personas in particular their construction see, e.g. (Wodtke 2003, pp. 159) Adaptation of the business case at hand of course can only be achieved in response to the customer-site interaction. In the navigation structure diagram in figure 4.2 we have used variable of data types that were chosen with respect to the site at hand, i.e., ASB Bank's Web site. We expect that this adaptation can always be achieved the way we have proposed here. Once the analysis has shown what data and functionality shall be accessible to customers data and functionality can be typed and variables of the respective type can be used to describe how the site adapts to the actual use. A type level adaptation that can be carried out while customers are interacting with a site is semi automatic reconsideration of the type of customer : Customers in this respect are presupposed to be characterized by a value for each of a number of 46 Lending Calculators / H:=Affordability Calculator + Amount Requested Calculator + Home Loan Options Calculator + Arrange your loan Home loans DSHARP BS, SS,SN Arranging a loan DSHARP, BS, SS,SN Affordability Calculator Amount Required Calculator Home Loan Options Calculator H.Affordability Calculator / H:= Amount Required Calculator + Home Loan Options Calculator SN.Arranging a home loan H.Home Loan Options Calculator / H:= Amount Required Calculator + Affordability Calculator H.Amount Required Calculator / H:= Apply by phone + Apply online + We come 2 u + U come 2 us + Send inquiry & refine SN calculator input form calculator nput form calculator input form / H:= Affordability Calculator + Home Loan Options Calculator Online home loan application H.Apply online SN.Home Loan Calculator SS.Home loans.Introduction / H:= Home loan rates + Buying your first home? + Loan top up needed? + Fixed rate loan expiring & SN := Introducyion + Loans at a glance + Interest rate options + Interest rates + Home loan calculators + Home buyers guide + Review your home loan + Move your loan to us + Fixed rate expiry + Arranging a home loan + Mobile lending service application form SN.Loans at a glance / H:= Loan types + loan options & refine SN (types, options) Interest rate options SN.Interest rate options / H:= variable rates + fixed rates + explanation DSHARP BS, SS,SN Interest rates DSHARP BS, SS,SN SN.Interest rates / H:= latest rates table Home buyers guide SN.Home buyers guide / H:= Houses 4 sale + Home buyers inspection list + Home valuation + Property information + Priority checklist + Glossary of terms & refine SN DSHARP BS, SS,SN Review your loan DSHARP BS, SS,SN SN.Review your loan / H:= explanation & R:= Credit cards + Omni cards + Moneymaker account Move your loan to us SN.Move your oan to us / H:= explanation DSHARP BS, SS,SN Loans at a glance DSHARP BS, SS,SN SN.Fixed rate expiry Fixed rate expiry DSHARP BS, SS,SN / H:= explanation & refine SN DSHARP BS, SS,SN Mobile lending service DSHARP BS, SS,SN SN.Mobile lending service / H:= explanation + contact phone numbers Figure 4.2: Navigation structure of a part of ASB Bank's Web site dimensions. The customer type according to (Schewe and Thalheim 2001) can be defined as a convex region in the multi dimensional space create as cartesian product of the scales associated to the customer dimensions. Based on an automatic customer assessment that in response to his or her site-interaction updates the scores in each of the dimension throughout customer-site interaction one can then track how a customer's trace moves through this space and detect when a modified type would better fit the customer's behavior than the actual type does. Clearly such update should only be done with customer permission Conclusion Banking services such as online home-loan application require a very sophisticated and well-adapted internet interface. Customers want to focus on solving their business problem, i.e., the goal they want to achieve by means of interacting with the WIS. They consider WIS as tools that shall be easy to handle, completely cover the business and do not add technical complexities to it. Customers want to have a pleasant usage experience, in particular they do not want to be treated like everybody else. They want WIS remember and exploit their usage peculiarities in authenticated and where adequate in anonymous sessions. This paper shows how this can be achieved. As a guiding principle we introduce considering context and using it to simplify WIS-handling. We have shown how WIS's scene context can be injected into the WIS and how XML suites can be generated using the story board of the site and available customer data. Acknowledgements We would like to thank Hans-J urgen Engelbrecht from the Department of Applied & International Economics at Massey University for pointing to us work on the the economic effects of technical progress in the banking industry. References Altus, M. Decision support for conceptual database design based on the evidence theory - An intelligent dialogue interface for conceptual database design. PhD thesis, Faculty of Mathematics, Natural Sciences and Computer Science of BTU Cottbus, Cottbus, 2000. Atzeni, P., Gupta, A., and Sarawagi, S. Design and maintenance of data-intensive web-sites. In Proceeding EDBT'98, vol. 1377 of LNCS. Springer-Verlag, Berlin, 1998, pp. 436450. Baresi, L., Garzotto, F., and Paolini, P. From web sites to web applications: New issues for conceptual modeling. In ER Workshops 2000, vol. 1921 of LNCS. Springer-Verlag, Berlin, 2000, pp. 89100. Bell, J. Pragmatic reasoning: Inferring contexts . In Proc. Context'1999 (1999), LNAI 1688, Springer, pp. 4253. Berger, A. N. The Economic Effects of Technolog-ical Progress: Evidence from the Banking Industry . Journal of Money, Credit, and Banking 35, 2 (2003), 141 176. Bonifati, A., Ceri, S., Fraternali, P., and Maurino, A. Building multi-device, content-centric applications using WebML and the W3I3 tool suite. In ER Workshops 2000, vol. 1921 of LNCS. Springer-Verlag, Berlin, 2000, pp. 6475. Workflow Management Coalition , Ed. The Workflow Management Coalition specification: Workflow Management Coalition terminology & glossary. Workflow Management Coalition, 47 Winchester, United Kingdom, 1999. Document Status Issue 3.0. Connolly, J. H. Context in the study of human languages and computer programming languages: A comparison. In Proc. Context'2001 (2001), LNAI 2116, Springer, pp. 116128. D usterh oft, A., and Thalheim, B. SiteLang: Conceptual modeling of internet sites. In Conceptual Modeling ER 2001, H. S. K. et al., Ed., vol. 2224 of LNCS. Springer-Verlag, Berlin, 2001, pp. 179192. Feyer, T., Schewe, K.-D., and Thalheim, B. Conceptual modelling and development of information services. In Conceptual Modeling ER'98, T. Ling and S. Ram, Eds., vol. 1507 of LNCS. Springer-Verlag, Berlin, 1998, pp. 720. G adke, M., and Turowski, K. Generic web-based federation of business application systems for e-commerce applications. In EFIS 1999. 1999, pp. 2542. Garzotto, F., Paolini, P., and Schwabe, D. HDM - a model-based approach to hypertext application design. ACM ToIS 11, 1 (1993), 126. Goldin, D., Srinivasa, S., and Thalheim, B. Is = dbs + interaction - towards principles of information systems. In Proc. ER'2000 (2000), LNCS 1920, Springer, pp. 140153. Harel, D., and Naamad, A. The STATEMATE Semantics of Statecharts. ACM Transactions on Software Engineering and Methodology 5, 4 (Ok-tober 1996), 293333. Jablonski, S. Workflow-Management-Systeme: Modellierung und Architektur. Thomson's Ak-tuelle Tutorien. International Thomson Publishing , Bonn, Germay et al., 1996. Kaschek, R., Matthews, C., and Wallace, C. e-Mortgages: NZ State of the Art and Perspectives . In Proceedings of SCI 2003 (2003). Kaschek, R., Schewe, K.-D., Thalheim, B., Zhang, L. Modelling contexts in web information systems. Proc. WES 2003. Matthews, C. D., Kaschek, R. H., Wallace , C. M., and Schewe, K. D. IST in Lending: Unlimited potential but limited practice. Available from: http://cbs.dk/staff/lars.heide/ISTOS/program.htm, 2003. Paper presented at ISTOS Workshop in Barcelona, Spain, 28-30 March 2003. Rossi, G., Schwabe, D., and Lyardet, F. Web application models are more than conceptual models. In Advances in Conceptual Modeling, P. C. et al., Ed., vol. 1727 of LNCS. Springer-Verlag , Berlin, 1999, pp. 239252. Rumbaugh, J., Blaha, M., Premerlani, W., Eddy, F., and Lorensen, W. Object-Oriented Modeling and Design. Prentice-Hall, Inc., Engle-wood Cliffs, New Jersey, 1991. Schewe, B. Kooperative Softwareentwicklung. Deutscher Universit atsverlag, Wiesbaden, Germany , 1996. Schewe, B., Schewe, K.-D., and Thalheim, B. Objektorientierter Datenbankentwurf in der Entwicklung betrieblicher Informationssysteme. Informatik Forschung und Entwicklung 10 (1995), 115127. Schewe, K.-D., Kaschek, R., Matthews, C., and Wallace, C. Modelling web-based banking systems: Story boarding and user profiling. In Proceedings of the Workshop on Conceptual Modelling Approaches to E-Commerce, H. Mayr and W.-J. Van den Heuvel, Eds. Springer-Verlag, 2002. Schewe, K.-D., and Schewe, B. Integrating database and dialogue design. Knowledge and Information Systems 2, 1 (2000), 132. Schewe, K.-D., and Thalheim, B. Modeling interaction and media objects. In Proc. NLDB' 2000 (2000), LNCS 1959, Springer, pp. 313324. Schewe, K.-D., and Thalheim, B. Modeling interaction and media objects. In Advances in Conceptual Modeling, E. M etais, Ed., vol. 1959 of LNCS. Springer-Verlag, Berlin, 2001, pp. 313 324. Schwabe, D., and Rossi, G. An object oriented approach to web-based application design. TAPOS 4, 4 (1998), 207225. Siegel, D. The secrets of successful web sites. Markt und Technik, M unchen, 1998. Sowa, J. F., and Zachman, J. A. Extending and formalizing the framework for information systems architecture. IBM Systems Journal 31, 3 (1992), 590 616. Srinivasa, S. A Calculus of Fixed-Points for Char-acterising Interactive Behaviour of Information Systems. PhD thesis, BTU Cottbus, Fachbereich Informatik, Cottbus, 2001. Thalheim, B. Entity-relationship modeling Foundations of database technology. Springer, 2000. See also http://www.informatik.tu-cottbus .de/ thalheim/HERM.htm. Thalheim, B. Readings in fundamentals of interaction in information systems. Reprint BTU Cottbus , 2000. Collection of papers by C. Binder, W. Clau, A. D usterh oft, T. Feyer, T. Gutacker, B. Heinze, J. Lewerenz, M. Roll, B. Schewe, K.-D. Schewe, K. Seelig, S. Srinivasa, B. Thalheim. Accessible through http://www.informatik.tu-cottbus.de/ thalheim . Thalheim, B.,

scenarios;Web Information Systems;web site;Web services;usability;story boarding;context-awareness;context-aware information systems;web information system;media objects;scenes;media type;SiteLang

57

Contour-based Partial Object Recognition using Symmetry in Image Databases

This paper discusses the problem of partial object recognition in image databases. We propose the method to reconstruct and estimate partially occluded shapes and regions of objects in images from overlapping and cutting. We present the robust method for recognizing partially occluded objects based on symmetry properties, which is based on the contours of objects. Our method provides simple techniques to reconstruct occluded regions via a region copy using the symmetry axis within an object. Based on the estimated parameters for partially occluded objects, we perform object recognition on the classification tree. Since our method relies on reconstruction of the object based on the symmetry rather than statistical estimates, it has proven to be remarkably robust in recognizing partially occluded objects in the presence of scale changes, rotation, and viewpoint changes.

INTRODUCTION Most existing methods for object recognition are based on full objects. However, many images in electronic catalogs contain multiple objects with occluded shapes and regions. Due to the occlusion of objects, image retrieval can provide incomplete, uncertain, and inaccurate results. To resolve this problem, we propose new method to reconstruct objects using symmetry properties since most objects in a given image database are represented by symmetrical figures. Even though there have been several efforts in object recognition with occlusion, currents methods have been highly sensitive to object pose, rotation, scaling, and visible portion of occluded objects [12] [9] [17] [3] [15]. In addition, many appearance-based and model-based object recognition methods assumed that they have known occluded regions of objects or images through extensive training processes with statistical approach. However, our approach is not limited to recognizing occluded objects by pose and scale changes, and does not need extensive training processes. Unlike existing methods, our method finds shapes and regions to reconstruct occluded shapes and regions within objects. Our approach can handle object rotation and scaling for dealing with occlusion, and does not require extensive training processes. The main advantage of our approach is that it becomes simple to reconstruct objects from occlusions. We present the robust method, which is based on the contours of objects, for recognizing partially occluded objects based on symmetry properties. The contour-based approach finds a symmetry axis using the maximum diameter from the occluded object. In experiments, we demonstrate how our method reconstruct and recognize occluded shapes and regions using symmetry. Experiments use rotated and scaled objects for dealing with occlusion. We also evaluate the recognition rate of the reconstructed objects using symmetry and the visible portion of the occluded objects for recognition. The rest of this paper is organized as follows. In Section 2, we briefly review work related to this study. In Section 3, we describe a method to recognize partial objects from given classes. In Section 4, we describe experimental results for partial object recognition. Finally, we summarize this paper in Section 5. RELATED WORK There have been several research efforts in object recognition for dealing with occlusion. Krumm [13] proposed a new algorithm for detecting objects in images which uses models based on training images of the object, with each model representing one pose. Williams [23] proposed a method for the reconstruction of solid-shape from image contour using the Huffman labeling scheme. For object recognition, Chang and Krumm [3] used the color cooccurrence histogram based on pairs of pixels. Schiele et al. [20] proposed a method to perform partial object recognition using statistical methods, which are based on multidimensional receptive field histograms. In addition, Rajpal et al. [17] introduced a method for partial object recognition using neural network based indexing. In appearance-based object recognition, Edwards and Murase [6] addressed the occlusion problem inherent in appearance-based methods using a mask to block out part of the basic eigenimages and the input image. Leonardis and Bischof [14] handled occlusion, scaling, and translation by randomly selecting image points from the scene and their corresponding points in the basis eigenvectors. Rao [18] applied the adaptive learning of eigenspace basis vectors in appearance-based methods. Ohba and Ikeuchi [16] were able to handle translation and occlusion of an object using eigenwindows. 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. SAC'05, March 13-17, 2005, Santa Fe, New Mexico, USA. Copyright 2005 ACM 1-58113-964-0/05/0003...$5.00. 1190 2005 ACM Symposium on Applied Computing Current methods for dealing with occlusion have been based on template matching, statistical approaches using localized invariants, and recognition of occluded regions based on local features. In addition, there are many efforts in ellipse construction and detection [7][9][22]. In this paper, we propose unique methodologies in object recognition for dealing with occlusion based on symmetry properties through the ellipse reconstruction. Even though there have been several efforts in object recognition with occlusion, current methods have been highly sensitive to object pose and scaling. In addition, many appearance-based and model-based object recognition methods assumed that they have known occluded regions of objects or images through extensive training processes. However, our method is not limited to recognizing occluded objects by pose and scale changes, and do not require extensive training processes. THE PROPOSED METHOD We discuss the object reconstruction and the parameter estimation method to find the best matching class of input objects using the classification tree method [4]. We extracted shape parameters from reconstructed objects using RLC lines, such as roundness, aspect ratio, form factor, surface regularity [5]. The approach tries to find occluded shapes within partially occluded objects. The basic assumption is that most objects are represented by symmetrical figures. When a symmetric object is partially occluded, we use the symmetry measure to evaluate the symmetric shape. We estimate the most similar parameters of occluded shape and region of objects, and we retrieve objects that have the estimated parameters of occluded objects. A basic idea of reconstruction and estimation of occluded objects is to use symmetry properties within objects and use to the contour of objects. Fortunately, most products in electronic catalogs have symmetry in their shapes and they are represented by symmetrical figures. Symmetrical descriptions of shape or detection of symmetrical features of objects can be useful for shape matching, model-based object matching, and object recognition [2] [1]. In the given database, we have elliptical and roughly-rounded objects such as plates, cups, pans, and pots, depending on their poses and shapes. First, we consider elliptical objects in which the occlusion changes values of measurements and parameters related to diameters. We assume that we can get diameters from elliptical objects, which are partially occluded. Figure 3.1 Three-Spoke from the Triangle. However, the elliptical objects are limited to the shape of objects. Therefore, it may not be applied to other types of shape such as irregular shapes. In this case, since we cannot easily detect the symmetry axes, we introduce the three-spoke type symmetry method as shown in Figure 3.1. We apply this approach to roughly-rounded objects such as cups. For roughly-rounded objects, we use the three-spoke type method, which is derived from the triangle. The triangle is a basic model to represent figures such as circle, rectangle, and polygon. We use extended lines of the triangle to make axes as shown in Figure 3.1. The three-spoke type symmetry axes, which are equally assigned by 120 degrees, provide the possibility to detect proper symmetry axes on roughly-rounded objects. Therefore, this method can detect symmetry axes in roughly-rounded objects. In order to perform the following procedures, we assume that objects are represented by symmetrical figures. 1. We have an occluded elliptical object in Figure 3.2 and roughly-rounded object in Figure 3.6, we can get cutting points of the occlusion (x,y)' and (x,y)'', that are given by overlapping or cutting. Figure 3.2 The Occlusion Area Estimation using Symmetry: Get cutting points (x,y)' and (x,y)'' and get a distance l'. Figure 3.3 The Occlusion Area Estimation using Symmetry: Get the maximum diameter and the symmetry axis. Figure 3.4 The Occlusion Area Estimation using Symmetry: Get the estimated region a' using a line l' and the symmetry axis. Figure 3.5 The Occlusion Area Estimation using Symmetry: Add region a' to occluded shape and region and re-captured the estimated shape of an object. 2. Compute a distance between two cutting points from (x,y)' and (x,y)'', which is called a line l' as in Figure 3.2 and 3.6. 3. Based on a line l', make a connection between two points, fill the concave region and re-captured the shape. It is important to compute a centroid in an object. 4. Get the maximum diameter from re-captured shape using extremal points as shown in Figure 3.4 and 3.7. Two extremal points (r, l) and (r, l)' from re-captured shape as in Figure 3.7. The distance between two extreme boundary points are represented by the maximum diameter. 5. In elliptical objects, one of the maximum and minimum diameters can be a symmetry axis. In roughly-rounded objects, we use the three-spoke type symmetry, one spoke can be a 1191 symmetry axis to find occluded region within an object. 6. Centroid Detection: In case of elliptical objects, we find a centroid based on the maximum diameter and a line perpendicular to the maximum diameter, which is located in the center of the length of the maximum diameter. We select symmetry axes based on one of these lines as in Figure 3.3. In roughly-rounded objects, we get a centroid, based on whole region of an object. Equation 2 is adapted from Russ [19]. If the centroid is calculated by equation 1 using the boundary pixels only, the results may not be correct. The calculated points will be biased toward whichever part of the boundary is most complex and contains the most pixels. The correct centroid location uses the pairs of coordinates i x , i y for each point in the shape boundary. The centroid of an irregular shape is calculated correctly using all of the pixels in an object. Area y C Area x C k i i y k i i x = = = = 0 0 , (1) Area x x y y C Area y y x x C k i i i i i y k i i i i i x + = + = 0 2 1 2 1 0 2 1 2 1 ) ( ) ( , ) ( ) ( (2) 7. In roughly-rounded objects, a centroid is put at the same position at the center of the three-spoke type symmetry axes. Figure 3.6 The occlusion of a cup: Get a centroid after re-captured a shape. Figure 3.7 Get extremal points (r,l), (r,l)' and (r,l)'',(r,l)''' and the maximum diameter of an object. Figure 3.8 Use the three spoke type symmetry: Match a center of the spoke to a centroid and parallel one of axes to the maximum diameter. Figure 3.9 Extend axes and make symmetry axes. Figure 3.10 Select a symmetry axis based on two regions, which are A and B. Figure 3.11 Find a region a' of occluded shape using a symmetry axis and add to a occluded shape. 8. Axis Detection: The midpoint of the major axis is called the center of the ellipse. The minor axis is the line segment perpendicular to the major axis which also goes through the center and touches the ellipse at two points. In elliptical objects, we detect a symmetry axis based on the maximum diameter or the minimum diameter. To find a symmetry axis in roughly-rounded objects, one of axes of the three-spoke type symmetry axes is in parallel with the maximum diameter of an object as shown in Figure 3.8. Based on occluded shape and region, we select a symmetry axis to estimate this region within an object. Figures 3.9 and 3.10 show how to select a symmetry axis. When we select an axis in roughly-rounded objects, we consider conditions as follows: Select axes, which don't intersect the occluded region. 3.9 and 3.10 show how to select a symmetry axis. Select axes, which have a region with the maximum diameter l'. Area and perimeter are invariants as in equation 3, compare the proportion of region A and B. B A Area Perimeter Area Perimeter (3) 9. Using mirror symmetry, we can get points across an axis. We find points on the contour across an axis which have the same length l' and the same angle corresponding to the axis that is perpendicular to a symmetry axis, but the distance between axis and points may or may not be the same. 10. Capture a region a', move the captured region to the occluded shape using the mirror symmetry, and add to these regions as shown in Figure 3.4, 3.5, and 3.11. 11. Re-compute shape measurements such as area, diameters, and perimeter using RLC lines from re-captured shape of an object. Then, re-compute shape parameters based on measurements. 12. Apply to a classifier. From the above discussions, we described how to reconstruct and estimate the partially occluded shape and region of an object and how to find the best matching class of partially occluded objects after the estimation. 1192 EXPERIMENTAL RESULTS In the sections, we evaluate and describe the results of partial object recognition by our proposed a method. We have selected 190 partially occluded objects of images from electronic catalogs on the Internet as well as manipulated images. We assume that occluded objects have more than 50% visibility of objects, and images of catalogs contain partially occluded objects. The objects are categorized by semantic meanings such as cup and plate. In addition, our approaches and experiments are limited to cups and plates since we use roughly-rounded or elliptical objects. More precisely, the database contains 32 objects from different viewpoints and images of 97 objects comprising image plane rotations and scale changes. In sample images, we have extracted image features of partially occluded objects such as shape and texture. We experimented with shape reconstruction based on the contour of objects using symmetry properties. We assumed that inputs are not correctly classified and have occlusion. We experimented with samples such as plates and cups to reconstruct the occluded shape of objects as shown in Figure 4.1 and 4.2. In Figure 4.2, it is correctly classified after the reconstruction with an occlusion about 30%. On the other hand, Figure 4.1 is not correctly classified after the reconstruction since the width of plate is too narrow. This experiment shows that our method heavily relies on shape of objects. Figure 4.1 Example of the occlusion with a Plate. Figure 4.2 Example of the manipulated occlusion with a Cup. Finally, we performed an experiment for the relationships between visible portion of objects and recognition rates. In order to evaluate the visibility of objects, we used manipulated images of cups and plates. Figure 4.3 shows the pattern of object recognition in the presence of partial occlusion of objects and the results obtained by the symmetric recognition. A visible portion of approximately 67% is sufficient for the recognition of objects based on the contour. Figure 4.3 Object recognition in the presence of the occlusion of objects based on the contour. There are many efforts in object recognition for dealing with occlusion. The visible portion of objects required to recognize occluded objects are shown in Table 4.1. Table 4.1 shows a simple comparison between our method and other existing methods. The probabilistic method based on local measurements requires small portions of objects to recognize the whole objects, but it required extensive training processes to recognize occluded objects [21] [20]. Our method shows good visibility of partial object recognition and do not need extensive training processes. Table 4.1 The visibility of object recognition in the presence of partial occlusion. Methods Visibility Training processes Appearance matching techniques using adaptive masks 90% not required Probabilistic technique using Chi-square 72% required Probabilistic technique using local measurements 34% required Contour-based approach using symmetry 67% not required In order to measure the influence of occlusion and compare its impact on the recognition performance of the different methods, we performed an experiment as follows. Figure 4.4 summarizes the recognition results for different visible object portions. For each test object, we varied the visible object portion from 20% to 100% and recorded the recognition results using Chi-square divergence and our method. Figure 4.4 Experimental results with occlusion. 1193 The results show that our method clearly obtains better results than Chi-square divergence. Using only 60% of the object area, almost 80% of the objects are still recognized. This confirms that our method is capable of reliable recognition in the presence of occlusion. Table 4.2 Summary of Object Recognition Methods for dealing with Occlusion. Methods Occlusion Scale changes Object Pose Rotation Bischof et al. [1] Yes Yes No No Edwards et al. [6] Yes Yes No Yes(limited) Ohba et al. [16] Yes No Yes No Rao [18] Yes No Yes No Jacob et al. [11] Yes No Yes No Krumm [13] Yes No No NO Contour-based using symmetry Yes Yes Yes(limited) Yes Table 4.2 summarizes the various object recognition methods. The table indicates whether the methods can handle occlusion, rotation, pose, and changes in the size of objects in the database. Unlike the other methods, our method can handle scale change, object pose, and rotated objects with occlusion, even though our method has minor limitations of object poses. CONCLUSION In this paper, we have discussed how to estimate parameters and to reconstruct the occluded shape of partial objects in image databases. In order to reconstruct occluded shapes, we used symmetry, which provides powerful method for the partial object recognition. Unlike the existing methods, our method tried to reconstruct occluded shapes and regions within objects, since most objects in our domain have symmetrical figures. However, we have limitations in the shape of objects and the occluded region of objects. For example, if a pan has an occlusion in handle, it cannot correctly reconstruct and be recognized. Another minor limitation of our method is that a method is sensitive to the pose of an object. For example, if we cannot see an ellipse due to the object's pose, we cannot recognize the object. After estimation, we have applied inputs, which include estimated parameters, to the existing classification trees, to get to the best matching class. All experiments are performed based on the classifier in earlier work. In experiments, the results show that the recognition of the occluded object is properly reconstructed, estimated, and classified, even though we have limited to the size of samples. In addition, we have experienced the power of the symmetry through experiments. REFERENCES [1] H. Bischof and A. Leonardis. Robust recognition of scaled eigenimages through a hierachical approach. In IEEE Conference on Computer Vision and Pattern Recognition, 1998. [2] H. Blum and R.N. Nagel. Shape description using weighted symmetric axis features. Pattern Recognition, 1978. [3] P. Chang and J. Krumm. Object Recognition with Color Cooccurrence Histograms. In IEEE Conference on Computer Vision and Pattern Recognition, 1999. [4] J. Cho and N. Adam. Efficient Splitting Rules based on the Probabilities of Pre-Assigned Intervals. In IEEE Conference on Data Mining, 2001. [5] J. Cho, A. Gangopadhyay and N. Adam. Feature Extraction for Content-based Image search in Electronic Commerce. In MIS/OA International Conference, 2000. [6] J. Edwards and H. Murase. Appearance matching of occluded objects using coarse-to-fine adaptive masks. In IEEE Conference on Computer Vision and Pattern Recognition, 1997. [7] A. W. Fitzgibbon, M. Pilu, and R. B. Fisher. Direct least squares fitting of ellipses. In International Conference on Pattern Recognition, 1996. [8] M. Fleck. Local Rotational Symmetries. In IEEE Conference on Computer Vision and Pattern Recognition, 1986. [9] C. Ho and L. Chan. A fast ellipse/circle detector using geometric symmetry. Pattern Recognition, 1995. [10] Joachim Hornegger, Heinrich Niemann, and Robert Risack. Appearance-based object recognition using optimal feature transforms. Pattern Recognition, 2000. [11] David W. Jacobs and Ronen Basri. 3D to 2D recognition with regions. In IEEE Conference on Computer Vision and Pattern Recognition, 1997. [12] Grinnell Jones and Bir Bhanu. Recognition of articulated and occluded objects. IEEE Transaction on Pattern Analysis and Machine Intelligence, 1999. [13] John Krumm. Object detection with vector quantized binary features. In IEEE Conference on Computer Vision and Pattern Recognition, 1997. [14] Ales Leonardis and Horst Bishof. Dealing with Occlusions in the Eigenspace Approach. In IEEE Conference on Computer Vision and Pattern Recognition, 1996. [15] David G. Lowe. Object Recognition from Local Scale-Invariant Features. In International Conference on Computer Vision, 1999. [16] K. Ohba and K. Ikeuchi. Detectability, uniqueness, and reliability of eigen windows for stable verification of partially occluded objects. IEEE Trans. Pattern Anal. Mach, 1997. [17] N. Rajpal, S. Chaudhury, and S. Banerjee. Recognition of partially occluded objects using neural network based indexing. Pattern Recognition, 1999. [18] R. Rao. Dynamic appearance-based recognition. In IEEE Conference on Computer Vision and Pattern Recognition, 1997. [19] John C. Russ. The Image Processing Handbook. CRC Press, 3rd edition, 1998. [20] Bernt Schiele and Alex Pentland. Probabilistic Object Recognition and Localization. In International Conference on Computer Vision, 1999. [21] H. Schneiderman and T. Kanade. Probabilistic modeling of local appearance and spatial relationships for object recognition. In IEEE Conference on Computer Vision and Pattern Recognition, 1998 [22] W. Wu and M. J. Wang. Elliptical object detection by using its geometrical properties. Pattern Recognition 1993. [23] Lance R. Williams. Topological reconstruction of a smooth manifold-solid from its occluding contour. Journal of Computer Vision, 1997. 1194

object recognition;reconstruction;Object;contour;Recognition;Symmetry;Image;Contour;occlusion;estimation;symmetry

58

COOLCAT: An entropy-based algorithm for categorical clustering

In this paper we explore the connection between clustering categorical data and entropy: clusters of similar poi lower entropy than those of dissimilar ones. We use this connection to design an incremental heuristic algorithm, COOLCAT , which is capable of efficiently clustering large data sets of records with categorical attributes, and data streams. In contrast with other categorical clustering algorithms published in the past, COOLCAT's clustering results are very stable for different sample sizes and parameter settings. Also, the criteria for clustering is a very intuitive one, since it is deeply rooted on the well-known notion of entropy. Most importantly, COOLCAT is well equipped to deal with clustering of data streams (continuously arriving streams of data point) since it is an incremental algorithm capable of clustering new points without having to look at every point that has been clustered so far. We demonstrate the efficiency and scalability of COOLCAT by a series of experiments on real and synthetic data sets.

INTRODUCTION Clustering is a widely used technique in which data points are partitioned into groups, in such a way that points in the same group, or cluster, are more similar among themselves than to those in other clusters. Clustering of categorical attributes (i.e., attributes whose domain is not numeric) is a difficult, yet important task: many fields, from statistics to psychology deal with categorical data. In spite of its importance, the task of categorical clustering has received scant attention in the KDD community as of late, with only a handful of publications addressing the problem ([18, 14, 12]). Much of the published algorithms to cluster categorical data rely on the usage of a distance metric that captures the separation between two vectors of categorical attributes, such as the Jaccard coefficient. In this paper, we present COOLCAT (the name comes from the fact that we reduce the entropy of the clusters, thereby "cooling" them), a novel method which uses the notion of entropy to group records. We argue that a classical notion such as entropy is a more natural and intuitive way of relating records, and more importantly does not rely in arbitrary distance metrics. COOLCAT is an incremental algorithm that aims to minimize the expected entropy of the clusters. Given a set of clusters , COOLCAT will place the next point in the cluster where it minimizes the overall expected entropy. COOLCAT acts incrementally, and it is capable to cluster every new point without having to re-process the entire set. Therefore , COOLCAT is suited to cluster data streams (contin-uosly incoming data points) [2]. This makes COOLCAT applicable in a large variety of emerging applications such as intrusion detection, and e-commerce data. This paper is set up as follows. Section 2 offers the background and relationship between entropy and clustering, and formulates the problem. Section 3 reviews the related work. Section 4 describes COOLCAT, our algorithm. Section 5 presents the experimental evidence that demonstrates the advantages of COOLCAT. Finally, Section 6 presents conclusions and future work. BACKGROUND AND PROBLEM FOR-MULATION In this section, we present the background of entropy and clustering and formulate the problem. 2.1 Entropy and Clustering Entropy is the measure of information and uncertainty of a random variable [28]. Formally, if X is a random variable, S(X) the set of values that X can take, and p(x) the prob-582 ability function of X, the entropy E(X) is defined as shown in Equation 1. E(X) = x S(X) p(x)log(p(x)) (1) The entropy of a multivariate vector ^ x = {X 1 , , X n } can be computed as shown in Equation 2, where p(^ x) = p(x 1 , , x n ) is the multivariate probability distribution. E(^ x) x 1 S(X 1 ) ... x n S(X n ) p(^ x)logp(^ x) (2) Entropy is sometimes referred to as a measure of the amount of "disorder" in a system. A room with socks strewn all over the floor has more entropy than a room in which socks are paired up, neatly folded, and placed in one side of your sock and underwear drawer. 2.2 Problem formulation The problem we are trying to solve can be formulated as follows. Given a data set D of N points ^ p 1 , , ^ p N , where each point is a multidimensional vector of d categorical attributes , i.e., ^ p j = (p 1 j , , p d j ), and given an integer k, we would like to separate the points into k groups C 1 , , C k , or clusters, in such a way that we minimize the entropy of the whole arrangement. Unfortunately, this problem is NP-Complete , and moreover, difficult to approximate [13]. In fact, the problem is NP-Complete for any distance function d(x, y), defined over pairs of points x, y, such that the function maps pairs of points to real numbers (and hence, our entropy function qualifies), therefore we need to resort to heuristics to solve it. We first have to resolve the issue of what we mean by the "whole entropy of the system." In other words, we have to make our objective function clear. We aim to minimize the expected entropy, whose expression is shown in Equation 3, where E(C 1 ), , E(C k ), represent the entropies of each cluster, C i denotes the points assigned to cluster i, C i D, with the property that C i C j = , for all i, j = 1, .., k i = j. The symbol C = {C 1 , , C k } represents the clustering. E( C) = k ( |C k | |D| (E(C k ))) (3) This function, as we will see later, allows us to implement an incremental algorithm that can effectively deal with large datasets, since we do not need to look at the entire set of points to decide about the entropy of an arrangement. Rather, we will be able to decide for each point, how it would affect the entropy of each of the existing clusters if placed in each one of them. The solution we propose in this paper (and present in Section 4) is a heuristic based in finding a set of initial clusters (using the entropic criteria), and then incrementally (greed-ily ) add points to the clusters according to a criteria that minimizes Equation 3. Furthermore, we make a simplification in the computation of entropy of a set of records. We assume independence of the attributes of the record, transforming Equation 2 into Equation 5. In other words, the joint probability of the combined attribute values becomes the product of the prob-members E Exp. Entropy Cluster0 {"red", "heavy"} 1.0 0.66 {"red", "medium"} Cluster1 {"blue", "light"} 0 Cluster0 {"red", "heavy"} 2.0 1.33 {"blue", "light"} Cluster1 {"red", "medium"} 0 Cluster0 {"red", heavy"} 0 1.33 Cluster1 {"red", "medium"} 2.0 {"blue", "light"} Figure 1: Three different clusterings for the set v 1 , v 2 , v 3 . Clustering 1 minimizes the expected entropy of the two clusters. abilities of each attribute, and hence the entropy can be calculated as the sum of entropies of the attributes. E(^ x) = x 1 S(X 1 ) x n S(X n ) (4) i (p(x i ))log( i p(x i )) = E(X 1 ) + E(X 2 ) + + E(X n ) (5) Assume that we have a set of three records, v 1 = {"red" , "heavy" }, v 2 = {"blue", "light"}, and v 3 = {"red", "medium" }, and we want to form two clusters with them. Figure 1 shows all the possible arrangements, with the entropy of each cluster, and the expected entropy in each arrangement . As we can see, the minimum expected entropy is that of arrangement 1, which obviously is the correct way of clustering the records (using two clusters). Even though the assumption of attribute independence is not true in every data set, it proves to work very well in practice (as shall be shown in the experimental section of this paper). Moreover, in the cases we can demonstrate that there is a correlation between two or more attributes of the data set, we can always change the data points by creating attributes that reflect these correlations and then apply Equation 5 to compute the join entropy. For instance, if the data set is composed of records of attributes A, B, C, D, E, F and we know that (A, B), (A, C) and (E, F ) are correlated. we can convert the data set into one having records with attributes AB, AC, D, EF and compute the entropy assuming that these new attributes are independent. Notice that for the grouped attributes, we are in effect computing their joint probabilities. The correlations between attributes can be easily found by techniques such as the Chi-Square and likelihood ratio tests. In our experimental experience, the gains obtained by doing this are small enough to justify the usage of the independence assumption. 2.3 Expected entropy and the Minimum Description Length principle The Minimum Description Length principle (MDL) [26, 27] recommends choosing the model that minimizes the sum of the model's algorithmic complexity and the description of the data with respect to that model. This principle is widely 583 used to compare classifiers (see [23]) but it has not been used much to deal with clustering. Formally, the complexity of the model can be stated as shown in Equation 6, where K() indicates the complexity, h is the model, and D denotes the data set. The term K(h) denotes the complexity of the model, or model encoding, while K(D using h) is the complexity of the data encoding with respect to the chosen model. K(h, D) = K(h) + K(D using h) (6) Consider first the term K(h). To encode the model, we need to encode for each cluster the probability distribution for the attribute values. This can be done by encoding the number of times each attribute value appears in the cluster, and the number of points in each cluster. Assuming that there are d attributes in the data, and that attribute A j can assume v j different values. As usual, k represents the number of clusters in the model. K(h) can be written as shown in Equation 7. In each cluster i, we need to encode c i = d-1 j=0 v j values. So, the total number of values we need to encode is k-1 i=0 c i = k, where is a constant. We also need to encode the number of points in each cluster, or k values. The number of bits needed to encode the number of times each attribute value occurs in the cluster, or the number of points in a cluster is equal to log( |D|), since the maximum number for these values is the size of the entire data set. Therefore K(h) is a linear function of k, with a constant that represents all the contributions described above. K(h) = klog( |D|) (7) On the other hand, the encoding of the data given the model can be stated as shown in Equation 8. Once the probabilities of occurrence of each attribute value in each cluster are known, an optimal code (Huffman) can be chosen to represent each attribute value in the cluster. Each point is simply represented by the encoding of its attributes' values . The optimal code is achieved by giving to each value a number of bits proportional to log(Pijl), where P (ijl) is the probability that the l value of attribute j occurs in cluster i. The second term in the equation simply indicates the membership of all the points, needing log(k) for the encoding of the individual memberships. K(D using h) = k-1 i=0 |C i | |D| d-1 j=0 v-1 l=0 Pijllog(Pijl) + Dlog(k) (8) Noticing that the first term of Equation 8 is simply the expected entropy of the clustering, we can write K(h, D) as shown in Equation 9. Notice that for a fixed k, the MDL principle indicates that the best model can be found by minimizing the expected entropy of the clustering, which is pre-cisely our goal. K(h, D) = log( |D|) + Dlog(k) + E( C) (9) 2.4 Evaluating clustering results A frequent problem one encounters when applying clustering algorithms in practice is the difficulty in evaluating the solutions. Different clustering algorithms (and sometimes multiple applications of the same algorithm using slight variations of initial conditions or parameters) result in very different solutions, all of them looking plausible. This stems from the fact that there is no unifying criteria to define clusters , and more often than not, the final clusters found by the algorithm are in fact the ones that correspond to the criteria used to drive the algorithm. Methods to evaluate whether or not the structure found is a property of the data set and not one imposed by the algorithm are needed. Authors have pondered about good ways to validate clusters found by algorithms (e.g., see [21, 1]). Two widely used methods are the following: Significance Test on External Variables This technique calls for the usage of significance tests that compare the clusters on variables not used to generate them. One way of doing this is to compute the entropy of the solution using a variable that did not participate in the clustering. (A class attribute.) The entropy of an attribute C in a cluster C k is computed as shown in Equation 10, where V j denotes one of the possible values that C can take. The evaluation is performed by computing the expected entropy (taken into consideration the cluster sizes). The smaller the value of E(C k ), the better the clustering fares. E(C k ) = j P (C = V j )logP (C = V j ) (10) The category utility function The category utility (CU) function [15] attempts to maximize both the probability that two objects in the same cluster have attribute values in common and the probability that objects from different clusters have different attributes. The expression to calculate the expected value of the CU function is shown in Equation 11, where P (A i = V ij |C k ) is the conditional probability that the attribute i has the value V ij given the cluster C k , and P (A i = V ij ) is the overall probability of the attribute i having the value V ij (in the entire set). The function aims to measure if the clustering improves the likelihood of similar values falling in the same cluster. Obviously, the higher the value of CU, the better the clustering fares. CU = k C k |D| i j [P (A i = V ij |C k ) 2 - P (A i = V ij ) 2 ] (11) We have used both techniques in validating our results, as shall be seen in the experimental section. 2.5 Number of clusters The issue of choosing the number of clusters is one common to all clustering methods, and our technique is no exception . Many methods have been proposed for determining the right number of clusters (e.g.,[4, 9]). Unfortunately many of these methods (e.g., [4]) assume that it is possible to compute a centroid for each cluster, which in categorical data is not easy. We consider this issue out of the scope of 584 this paper since we plan to examine good ways of selecting the optimal number of clusters in the context of our metric. RELATED WORK Clustering is an extensively researched area not only by data mining and database researchers [31, 11, 17, 18, 3], but also by people in other disciplines [10]. Among the numerical clustering algorithms, ENCLUS [6] uses entropy as a criteria to drive the algorithm. However, ENCLUS follows a completely different algorithm to our approach, dividing the hyperspace recursively. For each subspace, ENCLUS estimates its density and entropy and determines if it satisfies the goodness criteria: its entropy has to be lower than a threshold. However, it is not possible to translate either the algorithm or the relationships to the area of categorical clustering , since the notion of density has no intuitive meaning when the attributes are categorical. In a recent paper [16], the authors use Renyi's definition of entropy [25] to define a clustering evaluation function that measures the distance between clusters as the information potential [24] between them. Using this function, they describe an algorithm that, starting with a random placing of points in clusters, perturbs the placement until the improvement on the information potential is not appreciable. This algorithm, however, cannot scale to large data sets since it requires all points to perform the calculation of the distance. In the area of clustering categorical records, a few recent publications are worth mentioning. In [19], the authors address the problem of clustering transactions in a market basket database by representing frequent item sets as hyper-edges in a weighted hypergraph. The weight of the graph is computed as the average of the confidences for all possible association rules that can be generated from the item set. Then, a hypergraph partitioning algorithm is employed to partition the items, minimizing the weight of the cut hyper-edges . The algorithm does not produce a clustering of the transactions and it is not obvious how to obtain one from the item clusters. A related paper by Gibson et al [14] also treats categorical clustering as hypergraph partitioning, but uses a less combinatorial approach to solving it, based on non-linear dynamical systems. CACTUS [12], is an agglomerative algorithm that uses the author's definitions of support, strong connection and similarity to cluster categorical data. Support for an attribute value pair (a i , a j ), where a i is in the domain of attribute A i and a j in the domain of attribute A j is defined as the number of tuples that have these two values. The two attributes a i , a j are strongly connected if their support exceeds the value expected under the attribute-independence. This concept is then extended to sets of attributes. A cluster is defined as a region of attributes that are pairwise strongly connected, no sub-region has the property, and its support exceeds the expected support under the attribute-independence assumption. ROCK [18] computes distances between records using the Jaccard coefficient. Using a threshold, it determines, for each record, who are its neighbors. For a given point p, a point q is a neighbor of p if the Jaccard coefficient J(p, q) exceeds the threshold. Then, it computes the values of a matrix LIN K, in which the entries link(p, q) are the number of common neighbors between p and q. The algorithm then proceeds to cluster the records in an agglomerative way, trying to maximize for the k clusters (k is a predefined integer ) the function k i=1 n i p,qC i link(p,q) n 1 + 2f () i , where is the threshold, and f () is a function selected by the user. The choice of f () is critical in defining the fitness of the clusters formed the the ROCK algorithm, and, as the authors point out, the function is dependent on the data set as well as on the kind of cluster the user is interested in. We feel that choosing the function is a delicate and difficult task for users that may be a roadblock to using ROCK efficiently. Snob [29, 30] is an unsupervised learning algorithm based on the notion of Minimum Message Length (MML). MML is an information theoretic criterion for parameter estimation and model selection. Although MML is similar to the MDL criterion of Rissanen, MML is a Bayesian criterion and therefore uses an a-priori distribution of parameter values . Snob is in the category of mixture model algorithms [22]. Snob is iterative in nature and therefore does not scale with large data sets. Moreover, contrary to COOLCAT, it is difficult to envision how Snob can be used to cluster data streams. AUTOCLASS [5] also uses mixture models and Bayesian criteria to cluster data sets. Again, AUTOCLASS does not scale well with large data sets. OUR ALGORITHM Our entropy-based algorithm, COOLCAT, consists of two steps: initialization and incremental step. 4.1 Initialization The initialization step "bootstraps" the algorithm, finding a suitable set of clusters out of a sample S, taken from the data set ( |S| &lt;&lt; N), where N is the size of the entire data set. We first find the k most "dissimilar" records from the sample set by maximizing the minimum pairwise entropy of the chosen points. We start by finding the two points ps 1 , ps 2 that maximize E(ps 1 , ps 2 ) and placing them in two separate clusters (C 1 , C 2 ), marking the records (this takes O( |S| 2 )). From there, we proceed incrementally, i.e., to find the record we will put in the j-th cluster, we choose an unmarked point ps j that maximizes min i=1,..,j-1 (E(ps i , ps j )). The rest of the sample unmarked points ( |S| - k), as well as the remaining points (outside the sample), are placed in the clusters using the incremental step. We are interested in determining the size of the sample that guarantees with high probability the existence in the sample of at least one member of each cluster, given the number of clusters. In [17], the authors address the same problem and use Chernoff bounds[7] to bound the size of the sample given an estimate of the size of the smallest cluster with respect to the average size ( |D| k ), and the confidence level for the probability of finding at least a member of each cluster. The estimate of the size of the smallest cluster with respect to the average size is given in the form of a parameter = |D| k m , where m is the size of the smallest cluster. The parameter is then a number greater than 1. The bound on the size of the sample is then given by Equation 12. s = k + klog( 1 ) + k (log( 1 )) 2 + 2log( 1 ) (12) It is important to remark that Equation 12 does not depend on the size of the data set, which makes the bound 585 1.Given an initial set of clusters C = C 1 , , C k 2.Bring points to memory from disk and for each point p do 3. For i = 1, .., k 4. Tentatively place p in C i and compute E( C i ) where C i denotes the clustering obtained by placing p in cluster C i 5. Let j = argmin i ( E( C i )) 6. Place p in C j 7. Until all points have been placed in some cluster Figure 2: Incremental step. very favorable for larger sets (and unfavorable for small ones, but this is not a problem since for small sets we can simply use the entire set as a sample). 4.2 Incremental Step After the initialization, we process the remaining records of the data set (the rest of the sample and points outside the sample) incrementally, finding a suitable cluster for each record. This is done by computing the expected entropy that results of placing the point in each of the clusters and selecting the cluster for which that expected entropy is the minimum. We proceed in the incremental step by bringing a buffer of points to main memory and clustering them one by one. The order of processing points has a definite impact on the quality of the clusters obtained. It is possible that a point that seems a good fit for a cluster at a given point in time, becomes a poor fit as more points are clustered. In order to reduce this effect, we enhanced the heuristic by re-processing a fraction of the points in the batch. After a batch of points is clustered, we select a fraction m of points in the batch that can be considered the worst fit for the clusters they were put in. We proceed to remove these points from their clusters and re-cluster them. The way we figure out how good a fit a point is for the cluster where it landed originally, is by keeping track of the number of occurrences of each of its attributes' values in that cluster. That is, at the end of the batch, we know the values of q ij , for each record i in the batch and each attribute j, where q ij represent the number of times that the value V ij appears in the cluster where i was placed. We convert these numbers into probabilities by dividing q ij by the cluster size (i.e., C l , where C l is the cluster where i was placed). Let us call these numbers p ij . For each record, we can compute a fitting probability p i = j (p ij ). Notice that the lower the p i is, the worst fit the record is in that cluster (we can say that the global combination of attributes is not very common in the cluster). We then sort records according to p i and select the m records in the batch with lowest p i as the records to be reprocessed. Each re-processed record is placed in the cluster that minimizes the expected entropy (as done originally in the incremental step). EXPERIMENTAL RESULTS Our experiments were run in a DELL server equipped with a Pentium III running at 800 MHz, and 1 Gigabyte of main memory, running Red Hat Linux 2.2.14. We used two kinds of data sets: real data sets (for evaluating the quality of our algorithm) and synthetic data sets (for the evaluation of scalability). The experiments were conducted using the following datasets (plus a synthetically generated data set to test the scalability of the algorithm). Archaeological data set Our first data set is a hypothetical collection of human tombs and artifacts from an archaeological site. Although the data set is not "real," it is realistic enough and so we include it in this section. It has also the property of being small, so brute force can be used to find the optimal clustering. The data set is taken from [1] The first attribute (not used for clustering but for verification) indicates the sex (M for male, F for female) of the individuals buried. The other eight attributes are binary (1 present, 0 non-present), and represent artifacts types (e.g., ceramics, bracelets, arrow points) that were found (or not found) in the tomb. Congressional votes This data set was obtained from the UCI KDD Archive ([20]) and contains the United States Congressional Voting Records for the year 1984. Each record contains a Congressman's votes on 16 issues . All the attributes are boolean ("yes" or "no"), with a few of the votes containing missing values. We decided to treat missing values as another domain value for the attribute. A classification field with the labels "Democrat," or "Republican" is provided for each record, which are not used for clustering, but can be loosely used for quality measuring. (Some congress-men "crossed" parties to vote.) There are 435 records in the set (267 Democrats and 168 Republicans). KDD Cup 1999 data This data set can be obtained from the UCI Archive [20], and was used for the the Third International Knowledge Discovery and Data Mining Tools Competition. This database contains a standard set of network audit data, which includes a wide variety of simulated intrusions. Each record, corresponding to a connection, contains 42 features, some of them categorical, and the rest continuous variables. We transformed the continuous variables in categorical by a simple process of discretization: we computed the median of each attribute, and assigned any value below and including the median a label "0," while the rest of the values were assigned a label "1." There are many intrusion data sets in the repository, some of them to be used as training sets and some as test sets. We utilized the set that corresponds to 10% of the training data. In this set, records have an extra attribute (class), labeled with a "1" if the connection is part of an attack, or a "0" if it is not. We use this attribute for the evaluation of external entropy (not in the clustering process). 5.1 Archaeological Data Figure 3 show the results of using COOLCAT in the archaeological data set. We performed experiments with 2 clusters, since the attribute with which we evaluate the external entropy (not used in the clustering) is Sex (and the 586 Alg. m CU Ext E. Expected (sex) entropy COOLCAT 0% 0.7626 0 4.8599 10% 0.7626 0 4.8599 20% 0.7626 0 4.8599 Brute Force 0 .7626 0 4.8599 ROCK 0 .3312 0.9622 n/a Figure 3: Results for COOLCAT, ROCK and brute force in the Archaeological data set. data set is small, so we believed that the clustering could effectively separate the two sexes). We conducted experiments with the original data set (which we label "independent"), and a modified data set in which we grouped attributes in the following way: (1), (24), (26), (34), (35), (46), (78), to reflect the correlations found among the attributes of the set (found by using a Likelihood ratio test). However, we only report the results for independent data, since the correlated set results are essentially the same. (The same phenomena was observed in the other experiments.) We also conducted "brute force" experiments, in which we found the optimum clustering, i.e., that for which the expected entropy was the minimum. We did this to compare how well our heuristic (COOLCAT) performed. We also report in the table the best results found by ROCK (which have to be found by varying the parameter over a range of values). The results shown in Figure 3 show that the expected entropy function does an excellent job in clustering this data. The results obtained by COOLCAT (in terms of CU , and external entropy with respect to the variable sex, which is not used in the clustering), and expected entropy are the same obtained by the brute force (optimal) approach. In all cases, both the CU function and the external entropy of the COOLCAT solutions are better than those found for the best ROCK solution. Particularly encouraging is the fact that the external entropy for the variable SEX (which the authors of the data set indicated as the one being more correlated with the clusters), is 0 in all the COOLCAT solutions , so a perfect separation is achieved. (ROCK's solution does not achieve this, resulting in a high external entropy.) In this data set, the re-processing step does not have any effect, as seen by the fact that the results are the same for all the values of m. This is attributed to the size of the data set (only 20 records). Both COOLCAT and ROCK took 0.01 seconds to find a solution for this data set. 5.2 Congressional Voting results Figure 4 summarizes the results obtained by COOLCAT in the Congressional Voting records (no grouping of attributes was performed), for three values of m. The results obtained for various sample sizes are extremely stable. The CU values for the clusterings obtained with COOLCAT are, in all the cases superior to the one obtained by ROCK. The values show no fluctuations on our results as m changes, while the value for CU is 11% better than ROCK's value. The external entropy for the COOLCAT solutions is slightly better than the value in ROCK's solution. The buffer size (batch) in this experiment was 100 records, making the num-Alg . m CU Ext.Ent. Expected Running (pol. affl.) entropy time (sec.) COOL 0% 2.9350 0.4975 13.8222 0.16 CAT 10% 2.9350 0.4975 13.8222 0.26 20% 2.9350 0.4975 13.8222 0.28 ROCK 2 .6282 0.4993 N/A 0.51 Figure 4: Results for COOLCAT and ROCK in the Congressional Voting data set ber of re-processed points 0,10, and 20 (m = 0%, 10%, 20%). (Again, these numbers correspond to the means of 500 runs.) The running time of COOLCAT is significantly better than the one for ROCK (a decrease of 45% in the slowest case, m = 20%, of COOLCAT). 5.3 KDD Cup 1999 data set Since we did not have explicit knowledge of how many clusters we could find in this data set, we decided to find clusterings for many k values, and report, in each case, the expected entropy, external entropy (with respect to the attribute that denotes whether the record is an attack or not), and CU . The results are shown in the form of a graph in Figure 5. In the figure, the left hand side scale is used for expected entropy and CU , while the right hand side is used for external entropy (the values of external entropy are between 0 and 1, while the other parameters have larger ranges). The figure shows that all the parameters tend to an asymptotic limit as k grows. The saturation starts to occur in the value k = 10, which exhibits an external entropy of 0.09, which indicates that most of the clusters are "inhabited" by either attack records or attack-free records. In other words, the clustering achieves a good separation of the points. The experiments were conducted using a sample size of 1,000 points, which guarantees a level of confidence of 95% (for = 10). 5.4 Synthetic data set We used a synthetic data generator ([8]) to generate data sets with different number of records and attributes. We used these data sets to test the scalability of COOLCAT. The results are shown in the graph of Figure 7, where the y-axis shows the execution time of COOLCAT in seconds, and the x-axis the number of records (in multiples of 10 3 ), for four different number of attributes (A = 5, 10, 20, 40). In all the cases, COOLCAT behaves linearly with respect to the number of records, due to the incremental nature of the algorithm (it processes each record in the data set at most twice: those that are selected for re-processing are clustered twice, the rest only once; moreover, points are brought from disk to memory only once). We used for these experiments an m equal to 20%, and a buffer size of 300 records. Notice that in this experiment, we do not report running times for ROCK. The reason for this is that ROCK is designed to be a main memory algorithm. In [18], the authors make it explicit that ROCK deals with large data sets by using random sampling (not by looking at the entire set). Therefore, it would have been unfair to compare COOLCAT's running times with those of ROCK (over samples of the sets). We performed another experiment with synthetic data 587 Figure 5: Expected entropy, external entropy and CU vs. Number of Clusters (k) in the KDD Cup 1999 data set. The left scale (y-axis) is used for xpected entropy and CU , while the right one is used for external entropy. sets generated by [8]. In this experiment, each synthetic set contained 8,124 records of 23 attributes each. Twenty two of the attributes are used for clustering and one (Indx) for the evaluation of external entropy. Each data set was generated using 21 different types of rules. A rule involves 12 attributes. An example of the rules used is: A = c&C = a&D = b&K = c&N = a&O = c&Q = b&R = c&S = c&T = b&U = b Indx = r 1 . This rule says that when the 11 attributes on the left hand side take the values shown, the attribute Indx takes the value r 1 . Every record obeys one of the rules. Two sets were generated, using different probability distributions for the rules. In the first one (uniform ), every rule is used in the same number of records in the data set. (In other words the number of records that obey a particular rule is equal to the size of the data set divided by 21.) In the normal distribution, the populations are distributed following a Gaussian distribution (some rules receive more records than others). The 23rd attribute takes the value of the rule number (rule index). The external entropy is calculated using this attribute (which does not participate in the clustering). Figure 6 shows the evaluation of clusters obtained by COOLCAT over different synthetic data sets. The table shows also the results obtained by using ROCK. As we can see, COOLCAT results are significantly better than those obtained by ROCK for both data sets. Particularly significant is the fact that the external entropy for the COOLCAT solutions in the Uniform case with m = 10%, 20% are 0, indicating a perfect separation of rules. The values for other cases are extremely close to 0 as well. As expected, re-processing (increasing m) helps in finding a better clustering. However, the impact is more marked when going from no re-processing (m = 0) to re-processing 10% of the points, leveling out from then on. The running times of COOLCAT are more than one order of magnitude smaller than those of ROCK. CONCLUSIONS In this paper we have introduced a new categorical clustering algorithm, COOLCAT, based in the notion of entropy . The algorithm groups points in the data set trying to minimize the expected entropy of the clusters. The ex-Dist . m CU Ext.Ent. Expected Running (rule index ) entropy time (sec.) COOLCAT Uniform 0% 6.9187 0.00816 17.4302 6.73 10% 6.9268 0.00000 17.2958 11.85 20% 6.9268 0.00000 17.3958 12.95 Normal 0% 6.8893 0.02933 17.4969 6.88 10% 6.8996 0.00813 17.4458 11.99 20% 6.9008 0.00742 17.4328 13.07 ROCK Uniform 6 .6899 0.09861 n/a 207.37 Normal 6 .2749 0.34871 n/a 223.49 Figure 6: Results for COOLCAT and ROCK in the synthetic data sets 0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000 22000 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 execution time Number of records x 1000 A = 5 A = 10 A = 20 A = 40 Figure 7: COOLCAT's performance for the synthetic data sets: response time (in seconds) vs. the number of records in the data set (in multiples of 10 3 ), for different number of attributes (A = 5, 10, 20, 40). perimental evaluation supports our claim that COOLCAT is an efficient algorithm, whose solutions are stable for different samples (and sample sizes) and it is scalable for large data sets (since it incrementally adds points to the initial clusters). We have evaluated our results using category utility function, and the external entropy which determines if the clusters have significance with respect to external variables (i.e., variables not used in the clustering process). In our comparisons with ROCK, COOLCAT always shows a small advantage in terms of the quality measures (CU and external entropy). However, the real advantage of COOLCAT resides in the fact that ROCK is extremely difficult to tune (finding the right ), while COOLCAT's behavior to its only parameter (m) is extremely stable: small values of m are sufficient to obtain a good result. In the largest data set for which we compared both techniques (Mushrooms), COOLCAT had a significantly better running time. The incremental nature of COOLCAT makes it possible to apply the algorithm to data streams, and as the results in scalability show, the algorithm can cope with large volumes of data. We are currently doing research in tracking evolving clusters using COOLCAT. 588 ACKNOWLEDGMENTS We like to thank Vipin Kumar and Eui-Hong (Sam) Han for lending us their implementation of ROCK. REFERENCES [1] M.S. Aldenderfer and R.K. Blashfield. Cluster Analysis. Sage Publications, (Sage University Paper series on Quantitative Applications in the Social Sciences, No. 44), 1984. [2] D. Barbar a. Requirements for clustering data streams. SIGKDD Explorations (Special Issue on Online, Interactive, and Anytime Data Mining), 3(2), 2002. [3] D. Barbar a and P. Chen. Using the fractal dimension to cluster datasets. In Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Boston, MA, August 2000. [4] R.B. Calinski and J. Harabasz. A dendrite method for cluster analysis. Communications in Statistics, pages 127, 1974. [5] P. Cheeseman and J. Stutz. Bayesian classification (AUTOCLASS): Theory and Results. In U.M. Fayyad, G. Piatetsky-Shapiro, P. Smyth, and R. Uthurusamy, editors, Advances in Knowledge Discovery and Data Mining. AAAI Press, Menlo Park, 1995. [6] C. CHen, A.W. Fu, and Y. Zhang. Entropy-based Subspace Clustering for Mining Numerical Data. In Proceedings of ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Diego, CA, August 1999. [7] H. Chernoff. A Measure of Asymptotic Efficiency for Tests of a Hypothesis Based on the Sum of Observations. Annals of Mathematical Statistics, pages 493509, 1952. [8] DataGen. Data Generator: Perfect data for an imperfect world. http://www.datasetgenerator.com/. [9] R.C. Dubes and A.K. Jain. Validity studies in clustering methodologies. Pattern Recognition, pages 235254, 1979. [10] R.O. Duda and P.E. Hart. Pattern Classification and Scene Analysis. Wiley-Interscience, New York, 1973. [11] M. Ester, H.P. Kriegel, and X. Wu. A density-based algorithm for discovering clusters in large spatial database with noise. In Proceedings of the International Conference on Knowledge Discovery and Data Mining, Portland, Oregon, August 1996. [12] V. Ganti, J. Gehrke, and R. Ramakrishnan. CACTUS-Clustering Categorical Data Using Summaries. In Proceedings of the ACM-SIGKDD International Conference on Knowledge Discovery and Data Mining, San Diego, CA, 1999. [13] M. Garey and D. Johnson. Computers and Intractability: A Guide to the Theory of NP-Completeness. W.H. Freeman, 1979. [14] D. Gibson, J. Kleinberg, and P. Raghavan. Clustering Categorical Data: An Approach Based on Dynamical Systems. In Proceedings of the International Conference on Very Large Databases (VLDB), New York, NY, September 1998. [15] A. Gluck and J. Corter. Information, uncertainty, and the utility of categories. In Proceedings of the Seventh Annual Conference of the Cognitive Science Society, 1985. [16] E. Gokcay and J.C. Principe. Information Theoretic Clustering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(2), February 2002. [17] S. Guha, R. Rastogi, and K. Shim. CURE: A clustering algorithm for large databases. In Proceedings of the ACM SIGMOD Conference on Management of Data, Seattle, WA, May 1998. [18] S. Guha, R. Rastogi, and K. Shim. ROCK: A Robust Clustering Algorithm for Categorical Attributes. In Proceedings of the 15th International Conference on Data Engineering, Sydney, Australia, April 1999. [19] E.H. Han, G. Karypis, V. Kumar, and B. Mobasher. Clustering based on association rule hypergraphs. In Proceedings of the SIGMOD Workshop on Research Issues on Data Mining and Knowledge Discovery, June 1997. [20] S. Hettich(librarian). UCI KDD Archive. http://kdd.ics.uci.edu/. [21] A.K. Jain and R.C. Dubes. Algorithms for clustering data. Prentice Hall, 1988. [22] G.J McLachlan and K.E. Basford. Mixture Models. Marcel Dekker, New York, 1988. [23] T.M. Mitchell. Machine Learning. McGraw-Hill, 1997. [24] J.C. Pincipe, D. Xu, and J. Fisher. Information theoretic learning. In S. Haykin, editor, Unsupervised Adaptive Filtering. John Wiley & Sons, 2000. [25] A. Renyi. On Measures of Entropy and Information. In Proc. of the Fourth Berkeley Symp. Math., Statistics, and Probability, 1960. [26] J. Rissanen. A universal prior for integers and estimation by minimum description length. The Annals of Statistics, 1983. [27] J. Rissanen. Stochastic complexity in statistical inquiry. World Scientific Pub., 1989. [28] C.E. Shannon. A mathematical theory of communication. Bell System Techical Journal, pages 379423, 1948. [29] C.S. Wallace and D.M. Boulton. An information measure for classification. The Computer Journal, 11(2), 1968. [30] C.S. Wallace and D.L. Dowe. Intrinsic classification by MML, the Snob program. In Proceedings of the 7th Australian Joint Conference on Artificial Intelligence, 1994. [31] R. Zhang, R. Ramakrishnan, and M.Livny. Birch: An efficient data clustering method for very large databases. In Proceedings of the ACM SIGMOD Conference on Data Management, Montreal, Canada, June 1996. 589

data streams;incremental algorithm;COOLCAT;categorical clustering;data stream;entropy;clustering

59

Coupling and Cohesion Measures for Evaluation of Component Reusability

This paper provides an account of new measures of coupling and cohesion developed to assess the reusability of Java components retrieved from the internet by a search engine. These measures differ from the majority of established metrics in two respects: they reflect the degree to which entities are coupled or resemble each other, and they take account of indirect couplings or similarities. An empirical comparison of the new measures with eight established metrics shows the new measures are consistently superior at ranking components according to their reusability.

INTRODUCTION The work reported in this paper arose as part of a project that retrieves Java components from the internet [1]. However, components retrieved from the internet are notoriously variable in quality. It seems highly desirable that the search engine should also provide an indication of both how reliable the component is and how readily it may be adapted in a larger software system. A well designed component, in which the functionality has been appropriately distributed to its various subcomponents, is more likely to be fault free and easier to adapt. Appropriate distribution of function underlies two key concepts: coupling and cohesion. Coupling is the extent to which the various subcomponents interact. If they are highly interdependent then changes to one are likely to have significant effects on others. Hence loose coupling is desirable. Cohesion is the extent to which the functions performed by a subsystem are related. If a subcomponent is responsible for a number of unrelated functions then the functionality has been poorly distributed to subcomponents. Hence high cohesion is a characteristic of a well designed subcomponent. We decided that the component search engine should provide the quality rankings of retrieved components based on measures of their coupling and cohesion. There is a substantial literature on coupling and cohesion metrics which is surveyed in the next section. We then describe in detail the metrics we have developed which attempt to address some of the limitations of existing metrics. In particular, we consider both the strength and transitivity of dependencies. The following section describes an empirical comparison of our proposed metrics and several popular alternatives as predictors of reusability. Section 5 presents an analysis of the results which demonstrate that our proposed metrics consistently outperform the others. The paper concludes with a discussion of the implications of the research. COUPLING AND COHESION METRICS Cohesion is a measure of the extent to which the various functions performed by an entity are related to one another. Most metrics assess this by considering whether the methods of a class access similar sets of instance variables. Coupling is the degree of interaction between classes. Many researches have been done on software metrics [8], the most important ones are selected used in our comparative study. Table 1 and Table 2 summarize the characteristics of these cohesion and coupling metrics. Table 1. Coupling metrics Name Definition CBO [4][5][11] Classes are coupled if methods or instance variables in one class are used by the other. CBO for a class is number of other classes coupled with it. RFC [4][5] Count of all methods in the class plus all methods called in other classes. CF [3][6] Classes are coupled if methods or instance variables in one class are used by the other. CF for a software system is number of coupled class pairs divided by total number of class pairs. DAC[9] The number of attributes having other classes as their types. 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. MSR'06, May 22-23, 2006, Shanghai, China. Copyright 2006 ACM 1-59593-085-X/06/0005...$5.00. 18 Table 2. Cohesion metrics Name Definition LCOM [5] Number of non-similar method pairs in a class of pairs. LCOM3[7][ 9] Number of connected components in graph whose vertices are methods and whose edges link similar methods. RLCOM [10] Ratio of number of non-similar method pairs to total number of method pairs in the class. TCC [2] Ratio of number of similar method pairs to total number of method pairs in the class. All of these measures have two important features in common. First, they treat relationship between a pair of classes or methods as a binary quantity; second, they treat coupling and cohesion as an intransitive relation; that is no account is taken of the indirect coupling and cohesion, although two of cohesion (LCOM3 [7][9] and TCC [2]) have suggested extensions to incorporate indirect relationships between methods. In cohesion metrics, it should be noted that three of them (LCOM, LCOM3 and RLCOM) are in fact measures of lack of cohesion. TCC [2], in contrast to the other three metrics, measures cohesion rather than its absence. In other respects it is similar to RLCOM, being the number of similar method pairs divided by the total number of method pairs. PROPOSED NEW METRICS The study suggested that none of these measures was very effective in ranking the reusability of Java components. We therefore decided to develop alternative coupling and cohesion metrics in the hope of achieving superior performance. One obvious step was to develop measures that reflected the extent to which a pair of classes was coupled or a pair of methods resembled each other. Because none of the measures treated coupling or similarity as transitive relations, we decided that such indirect dependencies should be incorporated into our metrics. 3.1 Cohesion We develop a cohesion metric that takes account of both the degree of cohesion and transitive (i.e indirect) cohesion between methods. Methods are said to be similar if the sets of instance variables that they access overlap. We adopt a graph theoretical approach. The methods of the class are the vertices. Suppose a class has a set of method members M { M 1 , M 2 ,...M m } and let. V j {V j,1 , V j,2 , .... V j,n } be the instance variables accessed by method M j . Then the edge from M j to M i exists if and only if V j V i is not null. Thus an edge of the graph reflects the similarity of the methods in that they have at least one instance variable in common. The similarity graph is undirected because intersection is a symmetric relation. The next step is to associate a number with each edge that reflects the extent to which the two methods have instance variables in common. We therefore define SimD(i,j), our measure of direct similarity of two methods, M i and M j , as ( ) j i j i V V V V j i SimD = , where i j (SimD(j,j) is defined to be zero). Note that 1 SimD(i,j) 0. The extension of the measure to include indirect similarity proceeds along the same lines as we employed for indirect coupling. The strength of similarity provided by a path between two methods is the product of the SimD values of the edges that make up the path. Thus we define SimT(i,j, ), the transitive similarity between methods M i and M j due to a specific path , as ( ) = = t s t s e t s t s e V V V V t s SimD j i SimT , , , ) , , ( where e s , t denotes the edge between vertices s and t. As in the case of coupling, the path with the highest SimT value is selected to define the similarity of the two methods, Sim(i,j). ) , , ( ) , ( max j i SimT j i Sim = where and is the set of all paths from M i to M j . This measure is used to provide a measure of the cohesion of the class, ClassCoh, by summing the similarities of all method pairs and dividing by the total number of such pairs: ) , , ( max arg ) , ( max j i SimT j i = m m j i Sim ClassCoh m j i = = 2 1 , ) , ( where m is the number of methods in the class. Finally, the weighted transitive cohesion of the complete software system, WTCoh, is defined as the mean cohesion of all the classes of which it is comprised: n ClassCoh WTCoh n j j = = 1 where n is the number of classes in the system. 3.2 Coupling As with cohesion measure, we regard software system as a directed graph, in which the vertices are the classes comprising the system. Suppose such a system comprises a set of classes C {C 1 , C 2 ,...C m }. Let M j {M j,1 , M j,2 , .... M j,n } be the methods of the class C j , and R j,i the set of methods and instance variables in class C i invoked by class C j for j i (R j,j is defined to be null). Then the edge from C j to C i exists if and only if R j,j is not null. Thus an edge of the graph reflects the direct coupling of one class to another. The graph is directed since R j,i is not necessarily equal to Ri,j. The next step is to associate a number with each edge that reflects the extent of direct coupling from one class to another. We define CoupD(i,j), as the ratio of the number of methods in class j invoked by class I to the total number of methods in class I, which indicates the impact of class j to class i. ( ) i i j i M R R j i CoupD + = , , Then the indirect coupling between classes is included. Suppose that CoupD(i,j) and CoupD(j,k) have finite values but that CoupD(i,k) is zero. Thus although there is no direct coupling between classes C i and C k , there is a dependency because C i invokes methods in C j which in turn invokes methods in C k . The strength of this dependency depends on the two direct couplings of which it is composed, a reasonable measure is defined as: 19 CoupD(i,j) CoupD(j,k). This notion is readily generalised. A coupling between two classes exists if there is a path from one to the other made up edges whose CoupD values are all non-zero. Thus we define CoupT(i,j, ), the transitive coupling between classes C i and C j due to a specific path , as ( ) + = = t s t s e s s t s e M R R t s CoupD j i CoupT , , , , ) , , ( e s , t denotes the edge between vertices s and t. Note first that CoupT includes the direct coupling, which corresponds to path of length one, and second that, because the CoupD values are necessarily less than one, transitive couplings due to longer paths will typically have lower values. In general there may be more than one path having a non-zero CoupT value between any two classes. We simply select the path with largest CoupT value and hence define Coup(i,j), the strength of coupling between the two classes, C i and C j to be: ) , , ( ) , ( max j i CoupT j i Coup = where ) , , ( max arg ) , ( max j i CPT j i = and is the set of all paths from C i to C j . The final step is to use measure between each pair of classes as a basis for a measure of the total coupling of a software system. The weighted transitive coupling (WTCoup) of a system is thus defined m m j i Coup WTCoup m j i = = 2 1 , ) , ( where m is the number of classes in the system. AN EXPERIMENTAL COMPARISON In our study, the metrics are used for a specific purpose: predicting how much effort would be required to reuse a component within a larger system. We therefore chose to measure reusability as simply the number of lines of code that were added, modified or deleted (NLOC) in order to extend its functionality in a prescribed way. The more lines required, the lower the reusability. This appears to us to be a crude but reasonable measure of the effort that would be required to adapt a component for use within a larger system. Three case studies were carried out: Case 1 HTML Parser: The original components analysed HTML documents, eliminated tags and comments and output the text. The required extension was to count and output the number of tags found during parsing. Case 2 Lexical Tokenizer: The original components tokenized a text document using user supplied token rules and output the tokens on a web interface. The required extension was to count and output the number of tokens retrieved. Case 3 Barcode: The original components accepted a sequence of alphanumeric characters and generated the corresponding barcode. The required extension was to count the number of letters. For each case, 20 Java components were retrieved from a repository of about 10,000 Java components retrieved form the internet. The requisite extensions were then implemented by a very experienced Java programmer and NLOC counted. Despite the relative simplicity of the extensions, there was considerable variation in the quantity of extra code required. We then proceeded to investigate how successful the various measures of coupling and cohesion are in predicting this quantity. Our proposed metrics are compared with all the metrics reviewed in section 2. In order to present the results on the same graph, those measures that do not produce values in the range (0,1) (i.e. CBO, RFC, DAC, LCOM and LCOM3) were divided by 100. RESULTS Two approaches were used to evaluate the performance of the various measures in predicting reusability: linear regression and rank correlation. 5.1 Linear Regression The regression lines obtained for the five cohesion measures when applied to the HTML parser components are shown in Figure 1. The results for the other two sets of components were similar. It is clear that some measures provide much more consistent predictors than others. There are no obvious systematic departures from linearity so the use of simple regression appears reasonable. The regression lines obtained for coupling measures demonstrate the same situation. The coefficient of determination, R 2 , provides a measure of how much of the variation in NLOC is accounted for by the measures. Table 3 and Table 4 display the values of R 2 obtained for each of the coupling and cohesion measures on all three sets of components. In each case, our proposed new measure, WTCoup and WTCoh gave the largest value of R 2 , indicating that it was the best linear predictor of reusability. The remaining measures produced at least one R 2 value so low as to indicate that that the correlation was not significantly above chance at the 5% level. Figure 1. Regression of cohesion measures against reusability Table 3. R 2 values for coupling measure regression lines. Cases WTCoup CF CBO RFC DAC HTML Parser .846 .621 .259 .793 .254 Lexical Token. .836 .098 .004 .729 .738 Barcode Gen. .958 .693 .121 534 .507 20 Table 4. R 2 values for cohesion measure regression lines. Cases WTCoh RLCOM LCOM3 LCOM TCC H. Parser .847 .319 .259 .564 .178 L. Token. .838 .783 .002 .709 .646 B. Gen. .892 .702 .177 .101 .785 5.2 Spearman Rank Correlation Although these results provide a strong indication that the proposed new measures are better predictors of reusability than the alternatives, our primary purpose is simply to rank a set of components retrieved from the repository. We therefore also computed the Spearman rank correlation coefficients between the rankings determined by NLOC and those produced by the various coupling and cohesion measures (Tables 5 and 6). Table 5. Rank correlations values for coupling measures. Cases WTCoup CF CBO RFC DAC HTML Parser .975 .882 .465 .896 .507 Lexical Token. .952 .291 .117 .822 .817 Barcode Gen. .974 .758 .485 .656 .800 Table 6. Rank correlations values for cohesion measures. Cases WTCoh RLCOM LCOM3 LCOM TCC H. Parser -.993 .522 .218 .564 -.343 L. Token. .838 .783 .002 .709 .646 Bar. Gen. .892 .702 .177 .101 .785 The relative performance of the various measures is consistent with the regression studies. In all cases, the two proposed measures, WTCoup and WTCoh, produced the highest rank correlations. They are in fact extremely high; no value was lower than 0.95. DISCUSSION These results clearly demonstrate that our proposed metrics for coupling and cohesion are very good predictors of the number of lines of code required to make simple modifications to Java components retrieved from the internet and are superior to other measures. The majority of coupling and cohesion metrics treat coupling and similarity as simple binary quantities and ignore the transitive relationship. Both our proposed measures concern these issues: First, they are weighted; that is, they use a numeric measure of the degree of coupling or similarity between entities rather than a binary quantity. Second they are transitive; that is, they include indirect coupling or similarity mediated by intervening entities. It is reasonable to enquire whether both these characteristics are necessary to achieve good prediction performance. In fact our investigations suggest that both contribute to the performance. Although both WTCoup and WTCoh are good predictors, it is worth considering whether a linear combination might not produce even better results. Multiple regression for the Lexical Tokenizer components produced an R 2 of 0.981; the ranking produced using the regression coefficients to weight the terms had a Spearman correlation of 0.986. These are superior to the results produced by each metric alone but not by a great margin simply because there original results leave only modest scope for improvement. Developing such a composite quality measure would entail assuming the relative weighting of the two metrics should be the same for all types of component. This work arose from, and is intended primarily as a contribution to, search engine technology. Nevertheless, we believe it may be of interest to a wider body of researchers: in particular, those involved in developing and evaluating software metrics. ACKNOWLEDGMENTS We are grateful to the four UK higher education funding bodies (for England, Scotland, Wales and Northern Ireland) for an Overseas Research Studentship (ORS/2002015010) awarded to G. Gui. REFERENCES [1] Gui, G. and Scott, P. D. Vector Space Based on Hierarchical Weighting: A Component Ranking Approach to Component Retrieval. In Proceedings of the 6th International Workshop on Advanced Parallel Processing Technologies (APPT'05) [2] Bieman, J. M. and Kang, B-Y. Cohesion and Reuse in an Object-Oriented System. In Proc. ACM Symposium on Software Reusability (SSR'95). (April 1995) 259-262. [3] Briand, L., Devanbu, P. and Melo, W. An investigation into coupling measures for C++. Proceedings of ICSE 1997. [4] Brito e Abreu, F. and Melo, W. Evaluating the impact of OO Design on Software Quality. Proc. Third International Software Metrics Symposium. (Berin 1996). [5] Chidamber, S. R. and Kemerer, C. K. A Metrics Suite for Object Oriented Design. IEEE Transactions on Software Engineering, Vol. 20 (June 1994), 476-493. [6] Harrison, R., S.J.Counsell, & R.V.Nith. An Evaluation of the MOOD Set of Object-Oriented Software Metrics. IEEE Transactions on Software Engineering, Vol. 24 (June 1998), 491-496. [7] Hitz, M. and Montazeri, B. Measuring coupling and cohesion in object-oriented systems. Proceedings of International Symposium on Applied Corporate Computing. (Monterrey, Mexico, 1995). [8] Kanmani, S., Uthariraj, R., Sankaranarayanan, V. and Thambidurai, P. Investigation into the Exploitation of Object-Oriented Features. ACM Sigsoft, Software Engineering Notes, Vol. 29 (March 2004). [9] Li, W. & Henry, S. Object-Oriented metrics that predict maintainability. Journal of Systems and Software. 23(2) 1993 111-122. [10] Li, X., Liu, Z. Pan, B. & Xing, B. A Measurement Tool for Object Oriented Software and Measurement Experiments with It. In Proc. IWSM 2000, 44-54. [11] Subramanyam, R. & Krishnan, M. S. Empirical Analysis of CK Metrics for Object-Oriented Design Complexity: Implications for Software Defects. IEEE Transactions on Software Engineering, Vol. 29 (April 2003), 297-310. 21

Binary Quantity;Experimentary Comparsion;Component search engine;Search Engine Technology;Spearman Rank Correlation;Intransitive Relation;Reusability;Coupling;Cohesion Metric;Linear Regression;Cohesion;Java components

6

A Distributed 3D Graphics Library

We present Repo-3D, a general-purpose, object-oriented library for developing distributed, interactive 3D graphics applications across a range of heterogeneous workstations. Repo-3D is designed to make it easy for programmers to rapidly build prototypes using a familiar multi-threaded, object-oriented programming paradigm. All data sharing of both graphical and non-graphical data is done via general-purpose remote and replicated objects, presenting the illusion of a single distributed shared memory. Graphical objects are directly distributed, circumventing the "duplicate database" problem and allowing programmers to focus on the application details. Repo-3D is embedded in Repo, an interpreted, lexically-scoped, distributed programming language, allowing entire applications to be rapidly prototyped. We discuss Repo-3D's design, and introduce the notion of local variations to the graphical objects, which allow local changes to be applied to shared graphical structures. Local variations are needed to support transient local changes, such as highlighting, and responsive local editing operations. Finally, we discuss how our approach could be applied using other programming languages, such as Java.

INTRODUCTION Traditionally, distributed graphics has referred to the architecture of a single graphical application whose components are distributed over multiple machines [14, 15, 19, 27<A href="6.html#1">] (Figure 1 a ). By taking advantage of the combined power of multiple machines, and the particular features of individual machines, otherwise impractical applications became feasible. However, as machines have grown more powerful and application domains such as Computer 1. {bm,feiner}@cs.columbia.edu, http://www.cs.columbia.edu/graphics Supported Cooperative Work (CSCW) and Distributed Virtual Environments (DVEs) have been making the transition from research labs to commercial products, the term distributed graphics is increasingly used to refer to systems for distributing the shared graphical state of multi-display/multi-person, distributed, interactive applications<A href="6.html#1"> (Figure 1b). This is the definition that we use here. While many excellent, high-level programming libraries are available for building stand-alone 3D applications (e.g. Inventor [35], Performer [29], Java 3D [33]), there are no similarly powerful and general libraries for building distributed 3D graphics applications . All CSCW and DVE systems with which we are familiar (e.g., [1, 7, 11, 12, 16, 28, 30, 31, 32, 34, 37, 41]) use the following approach: A mechanism is provided for distributing application state (either a custom solution or one based on a general-purpose distributed programming environment, such as ISIS [4] or Obliq [8]), and the state of the graphical display is maintained separately in the local graphics library. Keeping these "dual databases" synchronized is a complex, tedious, and error-prone endeavor. In contrast , some non-distributed libraries, such as Inventor [35], allow programmers to avoid this problem by using the graphical scene description to encode application state. Extending this "single database" model to a distributed 3D graphics library is the goal of our work on Repo-3D. Repo-3D is an object-oriented, high-level graphics package, derived from Obliq-3D [25]. Its 3D graphics facilities are similar to those of other modern high-level graphics libraries. However, the objects used to create the graphical scenes are directly distribut-able --from the programmer's viewpoint, the objects reside in one large distributed shared memory (DSM) instead of in a single process. The underlying system replicates any of the fine-grained objects across as many processes as needed, with no additional effort on the part of the programmer. Updates to objects are automatically reflected in all replicas, with any required objects automatically distributed as needed. By integrating the replicated objects into the programming languages we use, distributed applications may be built using Repo-3D with little more difficulty than building applications in a single process. Figure 1: Two meanings of distributed graphics: (a) a single logical graphics system with distributed components, and (b) multiple distributed logical graphics systems. We use the second definition here. No matter how simple the construction of a distributed application may be, a number of differences between distributed and monolithic applications must be addressed. These include: Distributed control. In a monolithic application, a single component can oversee the application and coordinate activities among the separate components by notifying them of changes to the application state. This is not possible in a non-trivial distributed application. Therefore, we must provide mechanisms for different components to be notified of changes to the distributed state. Interactivity. Updates to distributed state will be slower than updates to local state, and the amount of data that can be distributed is limited by network bandwidth. If we do not want to sacrifice interactive speed, we must be able to perform some operations locally. For example, an object could be dragged locally with the mouse, with only a subset of the changes applied to the replicated state. Local variations. There are times when a shared graphical scene may need to be modified locally. For example, a programmer may want to highlight the object under one user's mouse pointer without affecting the scene graph viewed by other users. Repo-3D addresses these problems in two ways. First, a programmer can associate a notification object with any replicated object. The notification object's methods will be invoked when the replicated object is updated. This allows reactive programs to be built in a straightforward manner. To deal with the second and third problems, we introduce the notion of local variations to graphical objects. That is, we allow the properties of a graphical object to be modified locally, and parts of the scene graph to be locally added, removed, or replaced. In<A href="6.html#2"> Section 2 we describe how we arrived at the solution presented here.<A href="6.html#2"> Section 3 discusses related work, and<A href="6.html#3"> Section 4 offers a detailed description of the underlying infrastructure that was used. The design of Repo-3D is presented in <A href="6.html#5">Section 5, followed by some examples and concluding remarks in Sectio<A href="6.html#8">ns 6 <A href="6.html#8">and 7. BACKGROUND Repo-3D was created as part of a project to support rapid prototyping of distributed, interactive 3D graphical applications, with a particular focus on DVEs. Our fundamental belief is that by providing uniform high-level support for distributed programming in the languages and toolkits we use, prototyping and experimenting with distributed interactive applications can be (almost) as simple as multi-threaded programming in a single process. While care must be taken to deal with network delays and bandwidth limitations at some stage of the program design (the languages and toolkits ought to facilitate this), it should be possible to ignore such issues until they become a problem. Our view can be summarized by a quote attributed to Alan Kay, "Simple things should be simple; complex things should be possible." This is especially true during the exploration and prototyping phase of application programming. If programmers are forced to expend significant effort building the data-distribution components of the application at an early stage, not only will less time be spent exploring different prototypes, but radical changes in direction will become difficult, and thus unlikely. For example, the implementation effort could cause programs to get locked into using a communication scheme that may eventually prove less than ideal, or even detrimental, to the program's final design. Since we are using object-oriented languages, we also believe that data distribution should be tightly integrated with the language's general-purpose objects. This lets the language's type system and programming constructs reduce or eliminate errors in the use of the data-distribution system. Language-level integration also allows the system to exhibit a high degree of network data transparency, or the ability for the programmer to use remote and local data in a uniform manner. Without pervasive, structured, high-level data-distribution support integrated into our programming languages and libraries, there are applications that will never be built or explored, either because there is too much programming overhead to justify trying simple things ("simple things are not simple"), or because the added complexity of using relatively primitive tools causes the application to become intractable ("com-plex things are not possible"). Of the tools available for integrating distributed objects into programming languages, client-server data sharing is by far the most common approach, as exemplified by CORBA [26], Modula-3 Network Objects [5], and Java RMI [39]. Unfortunately, interactive graphical applications, such as virtual reality, require that the data used to refresh the display be local to the process doing the rendering or acceptable frame refresh rates will not be achieved. Therefore, pure client-server approaches are inappropriate because at least some of the shared data must be replicated. Furthermore, since the time delay of synchronous remote method calls is unsuitable for rapidly changing graphical applications, shared data should be updated asynchronously. Finally, when data is replicated, local access must still be fast. The most widely used protocols for replicated data consistency, and thus many of the toolkits (e.g., ISIS [4] and Visual-Obliq [3]), allow data updates to proceed unimpeded, but block threads reading local data until necessary updates arrive. The same reason we need replicated data in the first place--fast local read access to the data--makes these protocols unsuitable for direct replication of the graphical data. Of course, these protocols are fine for replicating application state that will then be synchronized with a parallel graphical scene description, but that is what we are explicitly trying to avoid. Fortunately, there are replicated data systems (e.g., Orca [2] or COTERIE [24]) that provide replicated objects that are well suited to interactive applications, and it is upon the second of these systems that Repo-3D is built. RELATED WORK There has been a significant amount of work that falls under the first, older definition of distributed graphics. A large number of systems, ranging from established commercial products (e.g., IBM Visualization Data Explorer [21]) to research systems (e.g., PARADISE [19] and ATLAS [14]), have been created to distribute interactive graphical applications over a set of machines. However, the goal of these systems is to facilitate sharing of application data between processes, with one process doing the rendering. While some of these systems can be used to display graphics on more than one display, they were not designed to support high-level sharing of graphical scenes. Most high-level graphics libraries, such as UGA [40], Inventor [35] and Java 3D [33], do not provide any support for distribution. Others, such as Performer [29], provide support for distributing components of the 3D graphics rendering system across multiple processors, but do not support distribution across multiple machines. One notable exception is TBAG [13], a high-level constraint-based, declarative 3D graphics framework. Scenes in TBAG are defined using constrained relationships between time-varying functions. TBAG allows a set of processes to share a single, replicated constraint graph. When any process asserts or retracts a constraint, it is asserted or retracted in all processes. However, this means that all processes share the same scene, and that the system's scalability is limited because all processes have a copy of (and must evaluate) all constraints, whether or not they are interested in them. There is also no support for local variations of the scene in different processes. Machiraju [22] investigated an approach similar in flavor to ours, but it was not aimed at the same fine-grained level of interactivity and was ultimately limited by the constraints of the implementation platform (CORBA and C++). For example, CORBA objects are heavyweight and do not support replication, so much of their effort was spent developing techniques to support object migration and "fine-grained" object sharing. However, their fine-grained objects are coarser than ours, and, more importantly, they do not support the kind of lightweight, transparent replication we desire. A programmer must explicitly choose whether to replicate, move, or copy an object between processes when the action is to occur (as opposed to at object creation time). Replicated objects are independent new copies that can be modified and used to replace the original --simultaneous editing of objects, or real-time distribution of changes as they are made is not supported. Of greater significance is the growing interest for this sort of system in the Java and VRML communities. Java, like Modula-3, is much more suitable as an implementation language than C or C++ because of its cross-platform compatibility and support for threads and garbage collection: Without the latter two language features, implementing complex, large-scale distributed applications is extremely difficult. Most of the current effort has been focused on using Java as a mechanism to facilitate multi-user VRML worlds (e.g., Open Communities [38]). Unfortunately, these efforts concentrate on the particulars of implementing shared virtual environments and fall short of providing a general-purpose shared graphics library. For example, the Open Communities work is being done on top of SPLINE [1], which supports only a single top-level world in the local scene database. Most DVEs [11, 12, 16, 31, 32] provide support for creating shared virtual environments, not general purpose interactive 3D graphics applications. They implement a higher level of abstraction , providing support for rooms, objects, avatars, collision detection , and other things needed in single, shared, immersive virtual environments. These systems provide neither general-purpose programming facilities nor the ability to work with 3D scenes at a level provided by libraries such as Obliq-3D or Inventor. Some use communication schemes that prevent them from scaling beyond a relatively small number of distributed processes, but for most the focus is explicitly on efficient communication. SIMNET [7], and the later NPSNet [41], are perhaps the best known large-scale distributed virtual-environment systems. They use a fixed, well-defined communication protocol designed to support a single, large-scale, shared, military virtual environment. The techniques for object sharing implemented in recent CSCW toolkits [28, 30, 34, 37] provide some of the features we need, particularly automatic replication of data to ease construction of distributed applications. However, none of these toolkits has integrated the distribution of data into its programming language's object model as tightly as we desire. As a result, they do not provide a high enough level of network data transparency or suffi-ciently strong consistency guarantees. In groupware applications, inconsistencies tend to arise when multiple users attempt to perform conflicting actions: the results are usually obvious to the users and can be corrected using social protocols. This is not an acceptable solution for a general-purpose, distributed 3D graphics toolkit. Furthermore, none of these CSCW systems provides any support for asynchronous update notification, or is designed to support the kind of large-scale distribution we have in mind. Finally, while distributed games, such as Quake, have become very popular, they only distribute the minimum amount of application state necessary. They do not use (or provide) an abstract, high-level distributed 3D graphics system. UNDERLYING INFRASTRUCTURE Our work was done in the Modula-3 programming language [18]. We decided to use Modula-3 because of the language itself and the availability of a set of packages that provide a solid foundation for our infrastructure. Modula-3 is a descendant of Pascal that corrects many of its deficiencies, and heavily influenced the design of Java. In particular, Modula-3 retains strong type safety, while adding facilities for exception handling, concurrency, object-oriented programming, and automatic garbage collection 2 . One of its most important features for our work is that it gives us uniform access to these facilities across all architectures. Repo-3D relies on a number of Modula-3 libraries, as illustrated in<A href="6.html#3"> Figure 2. Distributed data sharing is provided by two packages, the Network Object client-server object package [5], and the Replicated Object shared object package [2<A href="6.html#3">4] (see Section 4.1). DistAnim-3D is derived from Anim-3D [25], a powerful, non-distributed , general-purpose 3D library originally designed for 3D algorithm animation (see<A href="6.html#5"> Section 4.2). Finally, Repo itself is a direct descendant of Obliq [8], and uses the Replicated Object package to add replicated data to Obliq (se<A href="6.html#5">e Section 4.3). 4.1 Distributed Shared Memory Repo-3D's data sharing mechanism is based on the Shared Data-Object Model of Distributed Shared Memory (DSM) [20]. DSM allows a network of computers to be programmed much like a mul-tiprocessor , since the programmer is presented with the familiar paradigm of a common shared memory. The Shared Data-Object Model of DSM is particularly well suited to our needs since it is a high-level approach that can be implemented efficiently at the application level. In this model, shared data is encapsulated in user-defined objects and can only be accessed through those objects' method calls. The DSM address space is partitioned implicitly by the application programmer, with an object being the smallest unit of sharing. All shared data is fully network transpar-2 . The Modula-3 compiler we used is available from Critical Mass, Inc. as part of the Reactor programming environment. The compiler, and thus our system, runs on all the operating systems we have available (plus others): Solaris, IRIX, HP-UX, Linux, and Windows NT and 95. Figure 2: The architecture of Repo-3D. Aside from native graphics libraries (X, Win32, OpenGL, Renderware) the Modula-3 runtime shields most of the application from the OS. The Replicated Object package uses an Event communication package and the Network Object package. DistAnim-3D is implemented on top of a variety of native graphics libraries and Replicated Objects. Repo exposes most of the useful Modula-3 packages, as well as using Network Objects and Replicated Objects to present a distributed shared memory model to the programmer. Operating System Services Network Objects Replicated Objects Modula-3 Runtime Events Native Graphics DistAnim-3D Repo Repo-3D Network ent because it is encapsulated within the programming language objects. Distribution of new objects between the processes is as simple as passing them back and forth as parameters to, or return values from, method calls--the underlying systems take care of the rest. 3 Objects are only distributed to new processes as necessary, and (in our system) are removed by the garbage collector when they are no longer referenced. Furthermore, distributed garbage collection is supported, so objects that are no longer referenced in any process are removed completely. There are three kinds of distributed object semantics in our DSM: Simple objects correspond to normal data objects, and have no special distributed semantics. When a simple object is copied between processes, a new copy is created in the destination process that has no implied relationship to the object in the source process. Remote objects have client-server distribution semantics. When a remote object is copied between processes, all processes except the one in which the object was created end up with a proxy object that forwards method invocations across the network to the original object. Replicated objects have replicated distribution semantics. When a replicated object is passed between processes, a new replica is created in the destination process. If any replica is changed, the change is reflected in all replicas. The Network Object package provides support for remote objects. It implements distributed garbage collection, exception propagation back to the calling site, and automatic marshalling and unmarshalling of method arguments and return values of virtually any data type between heterogeneous machine architectures. The package is similar to other remote method invocation (RMI) packages developed later, such as the Java RMI library [39]. All method invocations are forwarded to the original object, where they are executed in the order they are received. The Replicated Object package supports replicated objects. Each process can call any method of an object it shares, just as it can with a simple or remote object. We will describe the Replicated Object package in more detail, as Repo-3D relies heavily on its design, and the design of a replicated object system is less straightforward than a remote one. The model supported by the Replicated Object package follows two principles: All operations on an instance of an object are atomic and serializable. All operations are performed in the same order on all copies of the object. If two methods are invoked simultaneously , the order of invocation is nondeterministic, just as if two threads attempted to access the same memory location simultaneously in a single process. The above principle applies to operations on single objects. Making sequences of operations atomic is up to the programmer . The implementation of the Replicated Object package is based on the approach used in the Orca distributed programming language [2]. A full replication scheme is used, where a single object is either fully replicated in a process or not present at all. Avoiding partial replication significantly simplifies the implementation and the object model, and satisfies the primary rationale for replication: fast read-access to shared data. To maintain replication consistency an update scheme is used, where updates to the object are applied to all copies. The method of deciding what is and is not an update is what makes the Orca approach particularly interesting and easy to implement. All methods are marked as either read or update methods by the programmer who creates the object type. Read methods are assumed to not change the state of the object and are therefore applied immediately to the local object without violating consistency . Update methods are assumed to change the state. To distribute updates, arguments to the update method are marshalled into a message and sent to all replicas. To ensure all updates are applied in the same order, the current implementation of the Replicated Object package designates a sequencer process for each object. There may be more than one sequencer in the system to avoid overloading one process with all the objects (in this case, each object has its updates managed by exactly one of the sequencers.) The sequencer is responsible for assigning a sequence number to each message before it is sent to all object replicas. The replicas then execute the incoming update messages in sequence. The process that initiated the update does not execute the update until it receives a message back from the sequencer and all updates with earlier sequence numbers have been executed. There are three very important reasons for choosing this approach. First, it is easy to implement on top of virtually any object-oriented language, using automatically generated object subtypes and method wrappers that communicate with a simple runtime system. We do this in our Modula-3 implementation, and it would be equally applicable to an implementation in C++ or Java. For example, the JSDT [36] data-sharing package in Java uses a similar approach. Second, the Replicated Object package does not pay attention to (or even care) when the internal data fields of an object change. This allows the programmer great flexibility in deciding exactly what constitutes an update or not, and what constitutes the shared state 4 . For example, objects could have a combination of global and local state, and the methods that change the local state could be classified as read methods since they do not modify the global state. Alternatively, read methods could do some work locally and then call an update method to propagate the results, allowing time-consuming computation to be done once and the result distributed in a clean way. We took advantage of both of these techniques in implementing Repo-3D. Finally, the immediate distribution of update methods ensures that changes are distributed in a timely fashion, and suggests a straightforward solution to the asynchronous notification problem. The Replicated Object package generates a Notification Object type for each Replicated Object type. These new objects have methods corresponding to the update methods of their associated Replicated Object. The arguments to these methods are the same as the corresponding Replicated Object methods, plus an extra argument to hold the Replicated Object instance. These notifiers can be used by a programmer to receive notification of changes to a Replicated Object in a structured fashion. To react to updates to a Replicated Object instance, a programmer simply overrides the methods of the corresponding Notification Object with methods that react appropriately to those updates, and associates an instance 3. An important detail is how the communication is bootstrapped. In the case of the Network and Replicated Object packages, to pass a first object between processes, one of them exports the object to a special network object demon under some known name on some known machine. The second process then retrieves the object. 4. Of course, it falls squarely on the shoulders of the programmer to ensure that the methods provided always leave the object in a consistent state. This is not significantly different than what needs to be done when building a complex object that is simultaneously accessed by multiple threads in a non-distributed system. For example, if a programmer reads an array of numbers from inside the object and then uses an update method to write a computed average back into the object, the internal array may have changed before the average is written, resulting in a classic inconsistency problem. In general, methods that perform computations based on internal state (rather than on the method arguments) are potentially problematic and need to be considered carefully. of it with the Replicated Object instance. Each time an update method of the Replicated Object is invoked, the corresponding method of the Notifier Object is also invoked. Notification Objects eliminate the need for object polling and enable a "data-driven" flow of control. 4.2 Obliq-3D Obliq-3D is composed of Anim-3D, a 3D animation package written in Modula-3, and a set of wrappers that expose Anim-3D to the Obliq programming language (see<A href="6.html#5"> Section 4.3). Anim-3D is based on three simple and powerful concepts: graphical objects for building graphical scenes, properties for specifying the behavior of the graphical objects, and input event callbacks to support interactive behavior. Anim-3D uses the damage-repair model: whenever a graphical object or property changes (is damaged), the image is repaired without programmer intervention. Graphical objects (GOs) represent all the logical entities in the graphical scene: geometry (e.g., lines, polygons, spheres, polygon sets, and text), lights and cameras of various sorts, and groups of other GOs. One special type of group, the RootGO , represents a window into which graphics are rendered. GOs can be grouped together in any valid directed acyclic graph (DAG). The GO class hierarchy is shown i<A href="6.html#5">n Figure 3 . A property is a defined by a name and a value. The name determines which attribute is affected by the property, such as "Texture Mode" or "Box Corner1". The value specifies how it is affected and is determined by its behavior, a time-variant function that takes the current animation time and returns a value. Properties, property values, and behaviors are all objects, and their relationships are shown in<A href="6.html#5"> Figure 4. When a property is created, its name and value are fixed. However, values are mutable and their behavior may be changed at any time. There are four kinds of behaviors for each type of properties: constant (do not vary over time), synchronous (follow a programmed set of requests, such as "move from A to B starting at time t=1 and taking 2 seconds"), asynchronous (execute an arbitrary time-dependent function to compute the value) and dependent (asynchronous properties that depend on other properties). Synchronous properties are linked to animation handles and do not start satisfying their requests until the animation handle is signalled. By linking multiple properties to the same handle, a set of property value changes can be synchronized. Associated with each GO g is a partial mapping of property names to values determined by the properties that have been associated with g. A property associated with g affects not only g but all the descendants of g that do not override the property. A single property may be associated with any number of GOs. It is perfectly legal to associate a property with a GO that is not affected by it; for example, attaching a "Surface Color" property to a GroupGO does not affect the group node itself, but could potentially affect the surface color of any GO contained in that group. A RootGO sets an initial default value for each named property. There are three types of input event callbacks in Anim-3D, corresponding to the three kinds of interactive events they handle: mouse callbacks (triggered by mouse button events), motion callbacks (triggered by mouse motion events) and keyboard callbacks (triggered by key press events). Each object has three callback stacks, and the interactive behavior of an object can be redefined by pushing a new callback onto the appropriate stack. Any event that occurs within a root window associated with a RootGO r will be delivered to the top handler on r's callback stack. The handler could delegate the event to one of r's children, or it may handle it itself, perhaps changing the graphical scene in some way. DistAnim-3D is a direct descendant of Anim-3D. In addition to the objects being distributed, it has many additional facilities that are needed for general-purpose 3D graphical applications, such as texture mapping, indexed line and polygon sets, choice groups, projection and transformation callbacks, and picking. Since DistAnim-3D is embedded in Repo instead of Obliq (see <A href="6.html#5">Section 4.3), the resulting library is called Repo-3D. 4.3 Obliq and Repo Obliq [8] is a lexically-scoped, untyped, interpreted language for distributed object-oriented computation. It is implemented in, and tightly integrated with, Modula-3. An Obliq computation may involve multiple threads of control within an address space, multiple address spaces on a machine, heterogeneous machines over a local network, and multiple networks over the Internet. Obliq uses, and supports, the Modula-3 thread, exception, and garbage-collection facilities. Its distributed-computation mechanism is based on Network Objects, allowing transparent support for multiple processes on heterogeneous machines. Objects are local to a site, while computations can roam over the network. Repo [23] is a descendant of Obliq that extends the Obliq object model to include replicated objects. Therefore, Repo objects have state that may be local to a site (as in Obliq) or replicated across multiple sites. DESIGN OF REPO-3D Repo-3D's design has two logical parts: the basic design and local variations. The basic design encompasses the changes to Obliq-3D to carry it into a distributed context, and additional enhancements that are not particular to distributed graphics (and are therefore not discussed here). Local variations are introduced to handle two issues mentioned in<A href="6.html#1"> Section 1: transient local changes and responsive local editing. Figure 3: The Repo-3D GO class hierarchy. Most of the classes are also in Obliq-3D; the italicized ones were added to Repo-3D. GroupGO GO CameraGO LightGO NonSurfaceGO SurfaceGO RootGO ChoiceGroupGO OrthoCameraGO PerspCameraGO AmbientLightGO VectorLightGO PointLightGO SpotLightGO LineGO MarkerGO TextGO PolygonGO BoxGO SphereGO CylinderGO DiskGO TorusGO QuadMeshGO IndexedPolygonSetGO Text2DGO IndexedLineSetGO Figure 4: The relationship between properties, names, values, and behaviors. Each oval represents an object and arrows show containment . Value Behavior Property Name Request Request . . . 5.1 Basic Repo-3D Design The Anim-3D scene-graph model is well suited for adaptation to a distributed environment. First, in Anim-3D, properties are attached to nodes, not inserted into the graph, and the property and child lists are unordered (i.e., the order in which properties are assigned to a node, or children are added to a group, does not affect the final result). In libraries that insert properties and nodes in the graph and execute the graph in a well-defined order (such as Inventor), the siblings of a node (or subtree) can affect the attributes of that node (or subtree). In Anim-3D, and similar libraries (such as Java 3D), properties are only inherited down the graph, so a node's properties are a function of the node itself and its ancestors--its siblings do not affect it. Therefore, subtrees can be added to different scene graphs, perhaps in different processes, with predictable results. Second, the interface (both compiled Anim-3D and interpreted Obliq-3D) is programmatical and declarative. There is no "graphi-cal scene" file format per se: graphical scenes are created as the side effect of executing programs that explicitly create objects and manipulate them via the object methods. Thus, all graphical objects are stored as the Repo-3D programs that are executed to create them. This is significant, because by using the Replicated Object library described in<A href="6.html#3"> Section 4.1 to make the graphical objects distributed, the "file format" (i.e., a Repo-3D program) is updated for free. Converting Anim-3D objects to Replicated Objects involved three choices: what objects to replicate, what methods update the object state, and what the global, replicated state of each object is. Since replicated objects have more overhead (e.g., method execution time, memory usage, and latency when passed between processes), not every category of object in Repo-3D is replicated. We will consider each of the object categories described in <A href="6.html#5">Figure 4.2 in turn: graphical objects (GOs), properties (values, names, behaviors, animation handles) and callbacks. For each of these objects, the obvious methods are designated as update methods , and, as discussed in<A href="6.html#3"> Section 4.1, the global state of each object is implicitly determined by those update methods. Therefore, we will not go into excessive detail about either the methods or the state. Finally, Repo-3D's support for change notification will be discussed. 5.1.1 Graphical Objects GOs are the most straightforward. There are currently twenty-one different types of GOs, and all but the RootGOs are replicated. Since RootGOs are associated with an onscreen window, they are not replicated--window creation remains an active decision of the local process. Furthermore, if replicated windows are needed, the general-purpose programming facilities of Repo can be used to support this in a relatively straightforward manner, outside the scope of Repo-3D. A GO's state is comprised of the properties attached to the object, its name, and some other non-inherited property attributes. 5 The methods that modify the property list are update methods. Group GOs also contain a set of child nodes, and have update methods that modify that set. 5.1.2 Properties Properties are more complex. There are far more properties in a graphical scene than there are graphical objects, they change much more rapidly, and each property is constructed from a set of Modula-3 objects. There are currently 101 different properties of seventeen different types in Repo-3D, and any of them can be attached to any GO. A typical GO would have anywhere from two or three (e.g., a BoxGO would have at least two properties to define its corners) to a dozen or more. And, each of these properties could be complex: in the example in <A href="6.html#8">Section 6, a single synchronous property for a long animation could have hundreds of requests enqueued within it. Consider again the object structure illustrated<A href="6.html#5"> in Figure 4. A property is defined by a name and a value, with the value being a container for a behavior. Only one of the Modula-3 objects is replicated, the property value. Property values serve as the replicated containers for property behaviors. To change a property, a new behavior is assigned to its value. The state of the value is the current behavior. Animation handles are also replicated. They tie groups of related synchronous properties together, and are the basis for the interaction in the example in<A href="6.html#8"> Section 6. In Anim-3D, handles have one animate method, which starts an animation and blocks until it finishes. Since update methods are executed everywhere, and block access to the object while they are being executed, they should not take an extended period of time. In creating Repo-3D, the animate method was changed to call two new methods: an update method that starts the animation, and a non-update method that waits for the animation to finish. We also added methods to pause and resume an animation, to retrieve and change the current relative time of an animation handle, and to stop an animation early. The state of an Animation handle is a boolean value that says if it is active or not, plus the start, end, and current time (if the handle is paused). Most of the Modula-3 objects that comprise a property are not replicated, for a variety of reasons: Properties represent a permanent binding between a property value and a name. Since they are immutable, they have no synchronization requirements and can simply be copied between processes. Names represent simple constant identifiers, and are therefore not replicated either. Behaviors and requests are not replicated. While they can be modified after being created, they are treated as immutable data types for two reasons. First, the vast majority of behaviors, even complex synchronous ones, are not changed once they have been created and initialized. Thus, there is some justification for classifying the method calls that modify them as part of their initialization process. The second reason is practical and much more significant. Once a scene has been created and is being "used" by the application, the bulk of the time-critical changes to it tend to be assignments of new behaviors to the existing property values. For example, an object is moved by assigning a new (often constant) behavior to its GO_Transform property value. Therefore, the overall performance of the system depends heavily on the performance of property value behavior changes. By treating behaviors as immutable objects, they can simply be copied between processes without incurring the overhead of the replicated object system. 5.1.3 Input Callbacks In Repo-3D, input event callbacks are not replicated. As discussed in<A href="6.html#5"> Section 4.2, input events are delivered to the callback stacks of a RootGO. Callbacks attached to any other object receive input events only if they are delivered to that object by the programmer, perhaps recursively from another input event callback (such as the one attached to the RootGO). Therefore, the interactive behavior of a root window is defined not only by the callbacks attached to its RootGO, but also by the set of callbacks associated with the graph rooted at that RootGO. Since the RootGOs are not replicated, the 5. Some attributes of a GO, such as the arrays of Point3D properties that define the vertices of a polygon set, are not attached to the object, but are manipulated through method calls. callbacks that they delegate event handling to are not replicated either. If a programmer wants to associate callbacks with objects as they travel between processes, Repo's general-purpose programming facilities can be used to accomplish this in a straightforward manner. 5.1.4 Change Notification The final component of the basic design is support for notification of changes to distributed objects. For example, when an object's position changes or a new child is added to a group, some of the processes containing replicas may wish to react in some way. Fortunately , as discussed in<A href="6.html#3"> Section 4.1, the Replicated Object package automatically generates Notification Object types for all replicated object types, which provide exactly the required behavior. The Notification Objects for property values allow a programmer to be notified of changes to the behavior of a property, and the Notification Objects for the various GOs likewise allow notification of updates to them. 5.2 Local Variations Repo-3D's local variations solve a set of problems particular to the distributed context in which Repo-3D lives: maintaining interactivity and supporting local modifications to the shared scene graph. If the graphical objects and their properties were always strictly replicated, programmers would have to create local variations by copying the objects to be modified, creating a set of Notification Objects on the original objects, the copies of those objects, and all their properties (to be notified when either change), and reflecting the appropriate changes between the instances. Unfortunately, while this process could be automated somewhat, it would still be extremely tedious and error prone. More seriously, the overhead of creating this vast array of objects and links between them would (a) (b) (c) (d) Figure 5: Simultaneous images from a session with the distributed CATHI animation viewer, running on four machines, showing an animation of an engine. (a) Plain animation viewer, running on Windows NT. (b) Overview window, running on Windows 95. (c) Animation viewer with local animation meter, running on IRIX. (d) Animation viewer with local transparency to expose hidden parts, running on Solaris. make this approach impractical for short transient changes, such as highlighting an object under the mouse. To overcome this problem, Repo-3D allows the two major elements of the shared state of the graphical object scene--the properties attached to a GO and the children of a group--to have local variations applied to them. (Local variations on property values or animation handles are not supported, although we are considering adding support for the latter.) Conceptually, local state is the state added to each object (the additions, deletions, and replacements to the properties or children) that is only accessible to the local copies and is not passed to remote processes when the object is copied to create a new replica. The existence of local state is possible because, as discussed in<A href="6.html#3"> Section 4.1, the shared state of a replicated object is implicitly defined by the methods that update it 6 . Therefore, the new methods that manipulate the local variations are added to the GOs as non-update methods. Repo-3D combines both the global and local state when creating the graphical scene using the underlying graphics package. As mentioned above, local variations come in two flavors: Property variations. There are three methods to set, unset, and get the global property list attached to a GO. We added the following methods to manipulate local variations: add or remove local properties (overriding the value normally used for the object), hide or reveal properties (causing the property value of the parent node to be inherited), and flush the set of local variations (removing them in one step) or atomically apply them to the global state of the object. Child variations. There are five methods to add, remove, replace, retrieve, and flush the set of children contained in a group node. We added the following ones: add a local node, remove a global node locally, replace a global node with some other node locally, remove each of these local variations, flush the local variations (remove them all in one step), and atomically apply the local variations to the global state. This set of local operations supports the problems local variations were designed to solve, although some possible enhancements are discussed in<A href="6.html#8"> Section 7. EXAMPLE AN ANIMATION EXAMINER As an example of the ease of prototyping distributed applications with Repo-3D, we created a distributed animation examiner for the CATHI [6] animation generation system. CATHI generates short informational animation clips to explain the operation of technical devices. It generates full-featured animation scripts, including camera and object motion, color and opacity effects, and lighting setup. It was reasonably straightforward to modify CATHI to generate Repo-3D program files, in addition to the GeomView and Render-Man script files it already generated. The resulting output is a Repo-3D program that creates two scene DAGs: a camera graph and a scene graph. The objects in these DAGs have synchronous behaviors specified for their surface and transformation properties. An entire animation is enqueued in the requests of these behaviors, lasting anywhere from a few seconds to a few minutes. We built a distributed, multi-user examiner over the course of a weekend. The examiner allows multiple users to view the same animation while discussing it (e.g., via electronic chat or on the phone)<A href="6.html#7">. Figure 5 shows images of the examiner running on four machines, each with a different view of the scene. The first step was to build a simple "loader" that reads the animation file, creates a window, adds the animation scene and camera to it, and exports the animation on the network, requiring less than a dozen lines of Repo-3D code. A "network" version, that imports the animation from the network instead of reading it from disk, replaced the lines of code to read and export the animation with a single line to import it. <A href="6.html#7">Figure 5(a) shows an animation being viewed by one of these clients. The examiner program is loaded by both these simple clients, and is about 450 lines long. The examiner supports: Pausing and continuing the animation, and changing the current animation time using the mouse. Since this is done by operating on the shared animation handle, changes performed by any viewer are seen by all. Because of the consistency guarantees , all users can freely attempt to change the time, and the system will maintain all views consistently. A second "overview" window<A href="6.html#7"> (Figure 5(b)), where a new camera watches the animation scene and camera from a distant view. A local graphical child (representing a portion of the animation camera's frustum) was added to the shared animation camera group to let the attributes of the animation camera be seen in the overview window. A local animation meter (bottom of<A href="6.html#7"> Figure 5(c)), that can be added to any window by pressing a key, and which shows the current time offset into the animation both graphically and numerically. It was added in front of the camera in the animation viewer window, as a local child of a GO in the camera graph, so that it would be fixed to the screen in the animation viewer. Local editin<A href="6.html#7">g (Figure 5(d)), so that users can select objects and make them transparent (to better see what was happening in the animation) or hide them completely (useful on slow machines, to speed up rendering). Assorted local feedback (highlighting the object under the mouse and flashing the selected object) was done with local property changes to the shared GOs in the scene graph. Given the attention paid to the design of Repo-3D, it was not necessary to be overly concerned with the distributed behavior of the application (we spent no more than an hour or so). Most of that time was spent deciding if a given operation should be global or a local variation. The bulk of programming and debugging time was spent implementing application code. For example, in the overview window, the representation of the camera moves dynamically, based on the bounding values of the animation's scene and camera graphs. In editing mode, the property that flashes the selected node bases its local color on the current global color (allowing a user who is editing while an animation is in progress to see any color changes to the selected node.) CONCLUSIONS AND FUTURE WORK We have presented the rationale for, and design of, Repo-3D, a general-purpose, object-oriented library for developing distributed, interactive 3D graphics applications across a range of heterogeneous workstations. By presenting the programmer with the illusion of a large shared memory, using the Shared Data-Object model of DSM, Repo-3D makes it easy for programmers to rapidly prototype distributed 3D graphics applications using a familiar object-oriented programming paradigm. Both graphical and general-purpose, non-graphical data can be shared, since Repo-3D is embedded in Repo, a general-purpose, lexically-scoped, distributed programming language. Repo-3D is designed to directly support the distribution of graphical objects, circumventing the "duplicate database" problem and allowing programmers to concentrate on the application function-6 . The local state is not copied when a replicated object is first passed to a new process because the Repo-3D objects have custom serialization routines (or Picklers, in Modula-3 parlance). These routines only pass the global state, and initialize the local state on the receiving side to reasonable default values corresponding to the empty local state. ality of a system, rather than its communication or synchronization components. We have introduced a number of issues that must be considered when building a distributed 3D graphics library, especially concerning efficient and clean support for data distribution and local variations of shared graphical scenes, and discussed how Repo-3D addresses them. There are a number of ways in which Repo-3D could be improved. The most important is the way the library deals with time. By default, the library assumes all machines are running a time-synchronization protocol, such as NTP, and uses an internal animation time offset 7 (instead of the system-specific time offset) because different OSs (e.g., NT vs. UNIX) start counting time at different dates. Hooks have been provided to allow a programmer to specify their own function to compute the "current" animation time offset within a process. Using this facility, it is possible to build inter-process time synchronization protocols (which we do), but this approach is not entirely satisfactory given our stated goal of relieving the programmer of such tedious chores. Future systems should integrate more advanced solutions, such as adjusting time values as they travel between machines, so that users of computers with unsynchronized clocks can collaborate 8 . This will become more important as mobile computers increase in popularity , as it may not be practical to keep their clocks synchronized. The specification of local variations in Repo-3D could benefit from adopting the notion of paths (as used in Java 3D and Inventor, for example). A path is an array of objects leading from the root of the graph to an object; when an object occurs in multiple places in one or more scene graphs, paths allow these instances to be differ-entiated . By specifying local variations using paths, nodes in the shared scene graphs could have variations within a process as well as between processes. One other limitation of Repo-3D, arising from our use of the Replicated Object package, is that there is no way to be notified when local variations are applied to an object. Recall that the methods of an automatically generated Notification Object correspond to the update methods of the corresponding Replicated Object. Since the methods that manipulate the local variations are non-update methods (i.e., they do not modify the replicated state), there are no corresponding methods for them in the Notification Objects. Of course, it would be relatively straightforward to modify the Replicated Object package to support this, but we have not yet found a need for these notifiers. A more advanced replicated object system would also improve the library. Most importantly, support for different consistency semantics would be extremely useful. If we could specify semantics such as "all updates completely define the state of an object, and only the last update is of interest," the efficiency of the distribution of property values would improve significantly; in this case, updates could be applied (or discarded) when they arrive, without waiting for all previous updates to be applied, and could be applied locally without waiting for the round trip to the sequencer. There are also times when it would be useful to have support for consistency across multiple objects, either using causal ordering (as provided by systems such as ISIS and Visual-Obliq), or some kind of transaction protocol to allow large groups of changes to be applied either as a unit, or not at all. It is not clear how one would provide these features with a replicated object system such as the one used here. While a library such as Repo-3D could be built using a variety of underlying platforms, the most likely one for future work is Java. Java shares many of the advantages of Modula-3 (e.g., threads and garbage collection are common across all architectures) and the packages needed to create a Repo-3D-like toolkit are beginning to appear. While Java does not yet have a replicated object system as powerful as the Replicated Object package, a package such as JSDT [36] (which focuses more on data communication than high-level object semantics) may be a good starting point. Work is also being done on interpreted, distributed programming languages on top of Java (e.g., Ambit [9]). Finally, Java 3D is very similar to Anim-3D, even though its design leans toward efficiency instead of generality when there are trade-offs to be made. For example, the designers chose to forgo Anim-3D's general property inheritance mechanism because it imposes computational overhead. By combining packages such as Java 3D, JSDT, and Ambit, it should be possible to build a distributed graphics library such as Repo-3D in Java. Acknowledgments We would like to thank the reviewers for their helpful comments, as well as the many other people who have contributed to this project. Andreas Butz ported CATHI to use Repo-3D and helped with the examples and the video. Clifford Beshers participated in many lively discussions about the gamut of issues dealing with language-level support for 3D graphics. Tobias Hllerer and Steven Dossick took part in many other lively discussions. Xinshi Sha implemented many of the extensions to Obliq-3D that went into Repo-3D. Luca Cardelli and Marc Najork of DEC SRC created Obliq and Obliq-3D, and provided ongoing help and encouragement over the years that Repo and Repo-3D have been evolving. This research was funded in part by the Office of Naval Research under Contract N00014-97-1-0838 and the National Tele-Immersion Initiative, and by gifts of software from Critical Mass and Microsoft. References [1] D. B. Anderson, J. W. Barrus, J. H. Howard, C. Rich, C. Shen, and R. C. Waters. Building Multi-User Interactive Multimedia Environments at MERL. Technical Report Research Report TR95-17, Mit-subishi Electric Research Laboratory, November 1995. [2] H. Bal, M. Kaashoek, and A. Tanenbaum. Orca: A Language for Parallel Programming of Distributed Systems. IEEE Transactions on Software Engineering , 18(3):190205, March 1992. [3] K. Bharat and L. Cardelli. Migratory Applications. In ACM UIST '95, pages 133-142, November 1995. [4] K. P. Birman. The Process Group Approach to Reliable Distributed Computing. CACM, 36(12):3653, Dec 1993. [5] A. Birrell, G. Nelson, S. Owicki, and E. Wobber. Network Objects. In Proc. 14th ACM Symp. on Operating Systems Principles, 1993. [6] A Butz, Animation with CATHI, In Proceedings of AAAI/IAAI '97, pages 957962, 1997. [7] J. Calvin, A. Dickens, B. Gaines, P. Metzger, D. Miller, and D. Owen. The SIMNET Virtual World Architecture. In Proc. IEEE VRAIS '93 , pages 450455, Sept 1993. [8] L. Cardelli. A Language with Distributed Scope. Computing Systems , 8(1):2759, Jan 1995. [9] L. Cardelli and A. Gordon. Mobile Ambients. In Foundations of Software Science and Computational Structures , Maurice Nivat (Ed.), LNCE 1378, Springer, 140155. 1998. [10] R. Carey and G. Bell. The Annotated VRML 2.0 Reference Manual. Addison-Wesley, Reading, MA, 1997. [11] C. Carlsson and O. Hagsand. DIVE--A Multi-User Virtual Reality System. In Proc. IEEE VRAIS '93, pages 394400, Sept 1993. [12] C. F. Codella, R. Jalili, L. Koved, and J. B. Lewis. A Toolkit for Developing Multi-User, Distributed Virtual Environments. In Proc. IEEE VRAIS '93 , pages 401407, Sept 1993. 7. Computed as an offset from January 1, 1997. 8. Implementation details of the combination of Network and Replicated Objects made it difficult for us to adopt a more advanced solution. [13] C. Elliott, G. Schechter, R. Yeung and S. Abi-Ezzi. TBAG: A High Level Framework for Interactive, Animated 3D Graphics Applications, In Proc. ACM SIGGRAPH 94, pages 421434, August, 1994. [14] M. Fairen and A. Vinacua, ATLAS, A Platform for Distributed Graphics Applications, In Proc. VI Eurographics Workshop on Programming Paradigms in Graphics, pages 91102, September, 1997. [15] S. Feiner, B. MacIntyre, M. Haupt, and E. Solomon. Windows on the World: 2D Windows for 3D Augmented Reality. In Proc. ACM UIST '93, pages 145155, 1993. [16] T. A. Funkhouser. RING: A Client-Server System for Multi-User Virtual Environments. In Proc. 1995 ACM Symp. on Interactive 3D Graphics, pages 8592, March 1995. [17] G. Grimsdale. dVS--Distributed Virtual Environment System. In Proc. Computer Graphics '91 Conference, 1991. [18] S. P. Harbison. Modula-3. Prentice-Hall, 1992. [19] H.W. Holbrook, S.K. Singhal and D.R. Cheriton, Log-Based Receiver-Reliable Multicast for Distributed Interactive Simulation, Proc. ACM SIGCOMM '95, pages 328341, 1995. [20] W. Levelt, M. Kaashoek, H. Bal, and A. Tanenbaum. A Comparison of Two Paradigms for Distributed Shared Memory. Software Practice and Experience, 22(11):9851010, Nov 1992. [21] B. Lucas. A Scientific Visualization Renderer. In Proc. IEEE Visualization '92, pp. 227-233, October 1992. [22] V. Machiraju, A Framework for Migrating Objects in Distributed Graphics Applications, Masters Thesis, University of Utah, Department of Computer Science, Salt Lake City, UT, June, 1997. [23] B. MacIntyre. Repo: Obliq with Replicated Objects. Programmers Guide and Reference Manual. Columbia University Computer Science Department Research Report CUCS-023-97, 1997.} [24] B. MacIntyre, and S. Feiner. Language-level Support for Exploratory Programming of Distributed Virtual Environments. In Proc. ACM UIST '96, pages 8394, Seattle, WA, November 68, 1996. [25] M. A. Najork and M. H. Brown. Obliq-3D: A High-level, Fast-turnaround 3D Animation System. IEEE Transactions on Visualization and Computer Graphics, 1(2):175145, June 1995. [26] R. Ben-Natan. CORBA: A Guide to the Common Object Request Broker Architecture, McGraw Hill, 1995. [27] D. Phillips, M. Pique, C. Moler, J. Torborg, D. Greenberg. Distributed Graphics: Where to Draw the Lines? Panel Transcript, SIGGRAPH 89, available at: http://www.siggraph.org:443/publications/panels/siggraphi89/ [28] A. Prakash and H. S. Shim. DistView: Support for Building Efficient Collaborative Applications Using Replicated Objects. In Proc. ACM CSCW '94, pages 153162, October 1994. [29] J. Rohlf and J. Helman, IRIS Performer: A High Performance Multiprocessing Toolkit for Real-Time {3D} Graphics, In Proc. ACM SIGGRAPH 94, pages 381394, 1994. [30] M. Roseman and S. Greenberg. Building Real-Time Groupware with GroupKit, a Groupware Toolkit. ACM Transactions on Computer-Human Interaction, 3(1):66106, March 1996. [31] C. Shaw and M. Green. The MR Toolkit Peers Package and Experiment. In Proc. IEEE VRAIS '93, pages 1822, Sept 1993. [32] G. Singh, L. Serra, W. Png, A. Wong, and H. Ng. BrickNet: Sharing Object Behaviors on the Net. In Proc. IEEE VRAIS '95, pages 1925, 1995. [33] H. Sowizral, K. Rushforth, and M. Deering. The Java 3D API Specification, Addison-Wesley, Reading, MA, 1998. [34] M. Stefik, G. Foster, D. G. Bobrow, K. Kahn, S. Lanning, and L. Suchman. Beyond The Chalkboard: Computer Support for Collaboration and Problem Solving in Meetings. CACM, 30(1):32 47, January 1987. [35] P. S. Strauss and R. Carey, An Object-Oriented 3D Graphics Toolkit, In Computer Graphics (Proc. ACM SIGGRAPH 92), pages 341349, Aug, 1992. [36] Sun Microsystems, Inc. The Java Shared Data Toolkit, 1998. Unsupported software, available at: http://developer.javasoft.com/developer/earlyAccess/jsdt/ [37] I. Tou, S. Berson, G. Estrin, Y. Eterovic, and E. Wu. Prototyping Synchronous Group Applications. IEEE Computer, 27(5):4856, May 1994. [38] R. Waters and D. Anderson. The Java Open Community Version 0.9 Application Program Interface. Feb, 1997. Available online at: http://www.merl.com/opencom/opencom-java-api.html [39] A. Wollrath, R. Riggs, and J. Waldo. A Distributed Object Model for the Java System, In Proc. USENIX COOTS '96, pages 219231, July 1996. [40] R. Zeleznik, D. Conner, M. Wloka, D. Aliaga, N. Huang, P. Hubbard, B. Knep, H. Kaufman, J. Hughes, and A. van Dam. An Object-oriented Framework for the Integration of Interactive Animation Techniques. In Computer Graphics (SIGGRAPH '91 Proceedings), pages 105112, July, 1991. [41] M. J. Zyda, D. R. Pratt, J. G. Monahan, and K. P. Wilson. NPSNET: Constructing a 3D Virtual World. In Proc. 1992 ACM Symp. on Interactive 3D Graphics, pages 147156, Mar. 1992.

Data sharing;programming language;Distributed graphics;Data structures;Interactive graphical application;3D graphics library;Change notification;Library;Replicated object;Object representation;distributed virtual environments;Shared memory;Syncronisation;data distribution;object-oriented graphics;Java;Programming;duplicate database;local variations;multi-threaded programming;Heterogeneous workstation;Multi-user interaction;3D graphics application;Repo-3D;3D graphics;Callbacks;Prototyping;Distributed systems;Client-server approach;object-oriented library;Programming language;shared-data object model;prototype;Graphical objects;Client-Server;Object-oriented;Local variation;3D Graphics;Object oriented;Distributed applications;distributed shared memory;Distributed processes;Properties

60

Coupling Feature Selection and Machine Learning Methods for Navigational Query Identification

It is important yet hard to identify navigational queries in Web search due to a lack of sufficient information in Web queries, which are typically very short. In this paper we study several machine learning methods, including naive Bayes model, maximum entropy model, support vector machine (SVM), and stochastic gradient boosting tree (SGBT), for navigational query identification in Web search. To boost the performance of these machine techniques, we exploit several feature selection methods and propose coupling feature selection with classification approaches to achieve the best performance. Different from most prior work that uses a small number of features, in this paper, we study the problem of identifying navigational queries with thousands of available features, extracted from major commercial search engine results, Web search user click data, query log, and the whole Web's relational content. A multi-level feature extraction system is constructed. Our results on real search data show that 1) Among all the features we tested, user click distribution features are the most important set of features for identifying navigational queries. 2) In order to achieve good performance, machine learning approaches have to be coupled with good feature selection methods. We find that gradient boosting tree, coupled with linear SVM feature selection is most effective. 3) With carefully coupled feature selection and classification approaches, navigational queries can be accurately identified with 88.1% F1 score, which is 33% error rate reduction compared to the best uncoupled system, and 40% error rate reduction compared to a well tuned system without feature selection.

INTRODUCTION Nowadays, Web search has become the main method for information seeking. Users may have a variety of intents while performing a search. For example, some users may already have in mind the site they want to visit when they type a query; they may not know the URL of the site or may not want to type in the full URL and may rely on the search engine to bring up the right site. Yet others may have no idea of what sites to visit before seeing the results. The information they are seeking normally exists on more than one page. Knowing the different intents associated with a query may dramatically improve search quality. For example, if a query is known to be navigational, we can improve search results by developing a special ranking function for navigational queries. The presentation of the search results or the user-perceived relevance can also be improved by only showing the top results and reserving the rest of space for other purposes since users only care about the top result of a navigational query. According to our statistics, about 18% of queries in Web search are navigational (see Section 6). Thus, correctly identifying navigational queries has a great potential to improve search performance. Navigational query identification is not trivial due to a lack of sufficient information in Web queries, which are normally short. Recently, navigational query identification, or more broadly query classification, is drawing significant attention . Many machine learning approaches that have been used in general classification framework, including naive Bayes classifier, maximum entropy models, support vector machines , and gradient boosting tree, can be directly applied here. However, each of these approaches has its own advantages that suit certain problems. Due to the characteristics of navigational query identification (more to be addressed in Section 2 ), it is not clear which one is the best for the task of navigational query identification. Our first contribution in this paper is to evaluate the effectiveness of these machine learning approaches in the context of navigational query identification. To our knowledge, this paper is the very first attempt in this regard. 682 Machine learning models often suffer from the curse of feature dimensionality. Feature selection plays a key role in many tasks, such as text categorization [18]. In this paper , our second contribution is to evaluate several feature selection methods and propose coupling feature selection with classification approaches to achieve the best performance : ranking features by using one algorithm before another method is used to train the classifier. This approach is especially useful when redundant low quality heterogeneous features are encountered. Most previous studies in query identification are based on a small number of features that are obtained from limited resources [12]. In this paper, our third contribution is to explore thousands of available features, extracted from major commercial search engine results, user Web search click data, query log, and the whole Web's relational content. To obtain most useful features, we present a three level system that integrates feature generation, feature integration, and feature selection in a pipe line. The system, after coupling features selected by SVM with a linear kernel and stochastic gradient boosting tree as classification training method, is able to achieve an average performance of 88.1% F1 score in a five fold cross-validation. The rest of this paper is organized as follows. In the next section, we will define the problem in more detail and describe the architecture of our system. We then present a multi-level feature extraction system in Section 3. We describe four classification approaches in Section 4 and three feature selection methods in Section 5. We then conduct extensive experiments on real search data in Section 6. We present detailed discussions in Section 7. We discuss some related work in Section 8. Finally, we conclude the paper in Section 9. PROBLEM DEFINITION We divide queries into two categories: navigational and informational. According to the canonical definition [3, 14], a query is navigational if a user already has a Web-site in mind and the goal is simply to reach that particular site. For example, if a user issues query "amazon", he/she mainly wants to visit "amazon.com". This definition, however, is rather subjective and not easy to formalize. In this paper, we extend the definition of navigational query to a more general case: a query is navigational if it has one and only one perfect site in the result set corresponding to this query. A site is considered as perfect if it contains complete information about the query and lacks nothing essential. In our definition, navigational query must have a corresponding result page that conveys perfectness, uniqueness, and authority. Unlike Broder's definition, our definition does not require the user to have a site in mind. This makes data labeling more objective and practical. For example, when a user issues a query "Fulton, NY", it is not clear if the user knows the Web-site "www.fultoncountyny.org". However, this Web-site has an unique authority and perfect content for this query and therefore the query "Fulton, NY" is labeled as a navigational query. All non-navigational queries are considered informational. For an informational query, typically there exist multiple excellent Web-sites corresponding to the query that users are willing to explore. To give another example, in our dataset, query "national earth science teachers association" has only one perfect corresponding URL "http://www.nestanet.org/" and therefore is labeled as navigational query. Query "Canadian gold maple leaf" has several excellent corresponding URL's, including "http://www. goldfingercoin.com/ catalog gold/ canadian maple leaf.htm", "http://coins.about.com/ library/weekly/ aa091802a.htm" and "http://www.onlygold.com/Coins/ Cana-dianMapleLeafsFullScreen .asp". Therefore, query "Canadian gold maple leaf" is labeled as non-navigational query. Figure 1 illustrates the architecture of our navigational query identification system. A search engine takes in a query and returns a set of URLs. The query and returned URLs are sent into a multi-level feature extraction system that generates and selects useful features; details are presented in the next section. Selected features are then input into a machine learning tool to learn a classification model. MULTI-LEVEL FEATURE EXTRACTION The multiple level feature system is one of the unique features of our system. Unlike prior work with a limited number of features or in a simulated environment [11, 12], our work is based on real search data, a major search en-gine's user click information and a query log. In order to handle large amount of heteorgeneous features in an efficient way, we propose a multi-level feature system. The first level is the feature generation level that calculates statistics and induces features from three resources: a click engine, a Web-map and a query log. The second level is responsible for integrating query-URL pair-wise features into query features by applying various functions. The third level is a feature selection module, which ranks features by using different methods. Below we present the details of the first two levels. The third level will be presented separately in Section 5 since those feature selection methods are standard. 3.1 Feature Generation Queries are usually too short and lack sufficient context to be classified. Therefore, we have to generate more features from other resources. We use three resources to generate features: a click engine, a Web-map, and query logs. The click engine is a device to record and analyze user click behavior. It is able to generate hundreds of features automatically based on user click through distributions [16]. A Web-map can be considered as a relational database that stores hundreds of induced features on page content, anchor text, hyperlink structure of webpages, including the inbound, outbound URLs, and etc. Query logs are able to provide bag-of-words features and various language model based features based on all the queries issued by users over a period of time. Input to feature generation module is a query-URL pair. For each query, the top 100 ULRs are recorded and 100 query-URLs are generated. Thus for each query-URL pair, we record a total of 197 features generated from the following four categories: Click features: Click features record the click information about a URL. We generate a total number of 29 click features for each query-URL pair. An example of a click feature is the click ratio (CR). Let n i k denote the number of clicks on URL k for query i and total number of clicks n i = X k n i k . 683 Webmap Click engine Query log Entropy Max Min ... SGBT Naive Bayes MaxEnt SVM Search engine query Classifier Classification module Feature generation Feature selection module Feature integration Information gain SVM feature ranking Boosting feature selection Integrated feature query-url feature Selected feature query-URL Figure 1: Diagram of Result Set Based Navigational Query Identification System The click ratio is the ratio of number of clicks on a particular URL K for query i to the total number of clicks for this query, which has the form CR(i, K) = n i K n i . URL features: URL features measure the characteristics of the URL itself. There are 24 URL based features in total. One such feature is a URL match feature, named urlmr, which is defined as urlmr = l(p) l(u) where l(p) is the length of the longest substring p of the query that presents in the URL and l(u) is the length of the URL u. This feature is based on the observation that Web-sites tend to use their names in the URL's. The distributions confers uniqueness and authority. Anchor text features: Anchor text is the visible text in a hyperlink, which also provides useful information for navigational query identification. For example, one anchor text feature is the entropy of anchor link distribution [12]. This distribution is basically the histogram of inbound anchor text of the destination URL. If an URL is pointed to by the same anchor texts, the URL is likely to contain perfect content. There are many other anchor text features that are calculated by considering many factors, such as edit distance between query and anchor texts, diversity of the hosts, etc. In total, there are 63 features derived from anchor text. Since we record the top 100 results for each query and each query URL pair has 197 features, in total there are 19,700 features available for each query. Feature reduction becomes necessary due to curse of dimensionality [5]. Before applying feature selection, we conduct a feature integration procedure that merges redundant features. 3.2 Feature Integration We design a feature integration operator, named normalized ratio r k of rank k, as follows: r k (f j ) = max(f j ) - f jk max(f j ) - min(f j ) , k = 2, 5, 10, 20. (1) The design of this operator is motivated by the observation that the values of query-URL features for navigational query and informational query decrease at different rates. Taking the urlmr feature for example and considering a navigational query "Walmart" and an informational query "Canadian gold maple leaf", we plot the feature values of top 100 URLs for both queries, as shown in Figure 2. As we can see, the feature value for the navigational query drops quickly to a stable point, while an information query is not stable. As we will see in the experiment section, this operator is most effective in feature reduction. Besides this operator, we use other statistics for feature integration, including mean, median, maximum, minimum, entropy, standard deviation and value in top five positions of the result set query-URL pair features. In total, we now have 15 measurements instead of 100 for the top 100 URLs for each query. Therefore, for each query, the dimension of a feature vector is m = 15 197 = 2955, which is much smaller than 197, 000. CLASSIFICATION METHODS We apply the most popular generative (such as naive Bayes method), descriptive (such as Maximum Entropy method), and discriminative (such as support vector machine and stochastic gradient boosting tree) learning methods [19] to attack the problem. 4.1 Naive Bayes Classifier A simple yet effective learning algorithm for classification 684 0 20 40 60 80 100 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 Rank Result Set Feature: URLmr Query: &quot;Walmart&quot; 0 20 40 60 80 100 0 0.1 0.2 0.3 0.4 0.5 Rank Result Set Feature: URLmr Query: &quot;Canadian gold maple leaf&quot; Figure 2: urlmr query-URL feature for navigational query (upper) and a informational query (lower) is based on a simple application of Bayes' rule P (y |q) = P (y) P (q|y) P (q) (2) In query classification, a query q is represented by a vector of K attributes q = (v 1 , v 2 , ....v K ). Computing p(q |y) in this case is not trivial, since the space of possible documents q = (v 1 , v 2 , ....v K ) is vast. To simplify this computation, the naive Bayes model introduces an additional assumption that all of the attribute values, v j , are independent given the category label, c. That is, for i = j, v i and v j are conditionally independent given q. This assumption greatly simplifies the computation by reducing Eq. (2) to P (y |q) = P (y) Q K j=1 P (v j |y) P (q) (3) Based on Eq. (3), a maximum a posteriori (MAP) classifier can be constructed by seeking the optimal category which maximizes the posterior P (c |d): y = arg max yY ( P (y) K Y j=1 P (v j |y) ) (4) = arg max yY ( K Y j=1 P (v j |y) ) (5) Eq. (5) is called the maximum likelihood naive Bayes classifier , obtained by assuming a uniform prior over categories. To cope with features that remain unobserved during training , the estimate of P (v j |y) is usually adjusted by Laplace smoothing P (v j |y) = N y j + a j N y + a (6) where N y j is the frequency of attribute j in D y , N y = P j N y j , and a = P j a j . A special case of Laplace smoothing is add one smoothing, obtained by setting a j = 1. We use add one smoothing in our experiments below. 4.2 Maximum Entropy Classifier Maximum entropy is a general technique for estimating probability distributions from data and has been successfully applied in many natural language processing tasks. The over-riding principle in maximum entropy is that when nothing is known, the distribution should be as uniform as possible, that is, have maximal entropy [9]. Labeled training data are used to derive a set of constraints for the model that characterize the class-specific expectations for the distribution . Constraints are represented as expected values of features. The improved iterative scaling algorithm finds the maximum entropy distribution that is consistent with the given constraints. In query classification scenario, maximum entropy estimates the conditional distribution of the class label given a query. A query is represented by a set of features. The labeled training data are used to estimate the expected value of these features on a class-by-class basis. Improved iterative scaling finds a classifier of an exponential form that is consistent with the constraints from the labeled data. It can be shown that the maximum entropy distribution is always of the exponential form [4]: P (y |q) = 1 Z(q) exp( X i i f i (q; y)) where each f i (q; y) is a feature, i is a parameter to be estimated and Z(q) is simply the normalizing factor to ensure a proper probability: Z(q) = P y exp(P i i f i(q; y)). Learning of the parameters can be done using generalized iterative scaling (GIS), improved iterative scaling (IIS), or quasi-Newton gradient-climber [13]. 4.3 Support Vector Machine Support Vector Machine (SVM) is one of the most successful discriminative learning methods. It seeks a hyperplane to separate a set of positively and negatively labeled training data. The hyperplane is defined by w T x + b = 0, where the parameter w R m is a vector orthogonal to the hyperplane and b R is the bias. The decision function is the hyperplane classifier H(x) = sign(w T x + b). The hyperplane is designed such that y i (w T x i + b) 1 i , i = 1, ..., N, where x i R m is a training data point and y i {+1, -1} denotes the class of the vector x i . The margin is defined by the distance between the two parallel hyperplanes w T x + b = 1 and w T x + b = -1, i.e. 2/||w|| 2 . The margin is related to the generalization of the classifier [17]. The SVM training problem is defined as follows: minimize (1/2)w T w + 1 T subject to y i (w T x i + b) 1 i , i = 1, ..., N 0 (7) 685 where the scalar is called the regularization parameter, and is usually empirically selected to reduce the testing error rate. The basic SVM formulation can be extended to the nonlinear case by using nonlinear kernels. Interestingly, the complexity of an SVM classifier representation does not depend on the number of features, but rather on the number of support vectors (the training examples closest to the hyperplane ). This property makes SVMs suitable for high dimensional classification problems [10]. In our experimentation, we use a linear SVM and a SVM with radial basis kernel. 4.4 Gradient Boosting Tree Like SVM, gradient boosting tree model also seeks a pa-rameterized classifier. It iteratively fits an additive model [8] f t (x) = T t (x; 0 ) + T X t=1 t T t (x; t ), such that certain loss function L(y i , f T (x + i) is minimized, where T t (x; t ) is a tree at iteration t, weighted by parameter t , with a finite number of parameters, t and is the learning rate. At iteration t, tree T t (x; ) is induced to fit the negative gradient by least squares. That is ^ := arg min N X i ( -G it t T t (x i ); ) 2 , where G it is the gradient over current prediction function G it = L(y i , f (x i ) f (x i ) f=f t-1 . The optimal weights of trees t are determined t = arg min N X i L(y i , f t-1 (x i ) + T (x i , )). If the L-2 loss function [y i -f(x i )] 2 /2 is used, we have the gradient G(x i ) = -y i + f (x i ). In this paper, the Bernoulli loss function -2 X i (y i f (x i ) - log(1 + exp(f(x i )))) is used and the gradient has the form G(x i ) = y i 1 1 + exp( -f(x i )) . During each iteration of gradient boosting, the feature space is further partitioned. This kind of rectangular partition does not require any data preprocessing and the resulting classifier can be very robust. However, it may suffer from the dead zoom phenomenon, where prediction is not able to change with features, due to its discrete feature space partition . Friedman (2002) found that it helps performance by sampling uniformly without replacement from the dataset before estimating the next gradient step [6]. This method was called stochastic gradient boosting. FEATURE SELECTION Many methods have been used in feature selection for text classification, including information gain, mutual information , document frequency thresholding, and Chi-square statistics. Yang and Pedersen [18] gives a good comparison of these methods. Information gain is one of the most effective methods in the context of text categorization. In addition to information gain, we also use feature selection methods based on SVM's feature coefficients and stochastic gradient boosting tree's variable importance. 5.1 Information Gain Information gain is frequently used as a measure of feature goodness in text classification [18]. It measures the number of bits of information obtained for category prediction by knowing the presence or absence of a feature. Let y i : i = 1..m be the set of categories, information gain of a feature f is defined as IG(f ) = m X i=1 P (y i )logP (y i ) + P (f ) m X i=1 P (y i |f)logP (y i |f) + P (f ) m X i=1 P (y i |f)logP (y i |f) where f indicates f is not present. We compute the information gain for each unique feature and select top ranked features. 5.2 Linear SVM Feature Ranking Linear SVM (7) produces a hyperplane as well as a normal vector w. The normal vector w serves as the slope of the hyperplane classifier and measures the relative importance that each feature contribute to the classifier. An extreme case is that when there is only one feature correlated to sample labels, the optimal classifier hyperplane must be perpendicular to this feature axle. The L-2 norm of w, in the objective, denotes the inverse margin. Also, it can be viewed as a Gaussian prior of random variable w. Sparse results may be achieved by assuming a laplace prior and using the L-1 norm [2]. Unlike the previous information gain method, the linear SVM normal vector w is not determined by the whole body of training samples. Instead, it is determined by an optimally determined subset, support vectors, that are critical to be classified. Another difference is obvious: normal vector w is solved jointly by all features instead of one by one independently. Our results show that linear SVM is able to provide rea-sonably good results in feature ranking for our navigational query identification problem even when the corresponding classifier is weak. 5.3 Stochastic Gradient Boosting Tree Boosting methods construct weak classifiers using subsets of features and combines them by considering their predica-tion errors. It is a natural feature ranking procedure: each feature is ranked by its related classification errors. Tree based boosting methods approximate relative influence of a feature x j as J 2 j = X splits on x j I 2 k 686 where I 2 k is the empirical improvement by k-th splitting on x j at that point. Unlike the information gain model that considers one feature at a time or the SVM method that considers all the feature at one time, the boosting tree model considers a set of features at a time and combines them according to their empirical errors. Let R( X ) be a feature ranking function based on data set X . Information gain feature ranking depends on the whole training set RInfo(X ) = RInfo(Xtr). Linear SVM ranks features is based on a set of optimally determined dataset. That is, RSVM(X ) = RSVM(XSV), where XSV is the set of support vectors. The stochastic gradient boosting tree (GSBT) uses multiple randomly sampled data to induce trees and ranks feature by their linear combination. Its ranking function can be written as RSGBT(X ) = P T t=1 t R t SGBT(X t ), where X t is the training set randomly sampled at iteration t. EXPERIMENTS A total number of 2102 queries were uniformly sampled from a query log over a four month period. The queries were sent to four major search engines, including Yahoo, Google, MSN, and Ask. The top 5 URL's returned by each search engine were recorded and sent to trained editors for labeling (the number 5 is just an arbitrary number we found good enough to measure the quality of retrieval). If there exists one and only one perfect URL among all returned URLs for a query, this query is labeled as navigational query. Otherwise, it is labeled as non-navigational query. Out of 2102 queries, 384 queries are labeled as navigational . Since they are uniformly sampled from a query log, we estimate there are about 18% queries are navigational. The data set were divided into five folders for the purpose of cross-validation. All results presented in this section are average testing results in five fold cross validations. 6.2 Evaluation Classification performance is evaluated using three metrics : precision, recall and F1 score. In each test, Let n ++ denotes the number of positive samples that correctly classified (true positive); n -+ denotes the number of negative samples that are classified as positive (false positive); n +-denotes the number of false positive samples that are classified as negative (false negative); and n -denotes the number of negative samples that are correctly classified (true negative). Recall is the ratio of the number of true positives to the total number of positives samples in the testing set, namely recall = n ++ n ++ + n + . Precision is the ratio of the number of true positive samples to the number samples that are classified as positive, namely precision = n ++ n ++ + n -+ . F1 is a single score that combines precision and recall, defined as follows: F 1 = 2 precsion recall precsion + recall . 6.3 Results 6.3.1 Feature Selection Results Table 1 shows the distributions of the top 50 features selected by different methods. All methods agree that click features are the most important. In particular, linear SVM and boosting tree select more click features than information gain. On the other hand, information gain select many features from anchor text and other metrics such as spam scores. Table 1: Distributions of the Selected Top 50 Features According to Feature Categories Feature Set Info. Gain Linear SVM Boosting Click 52% 84% 74% URL 4% 2% 6% Anchor Text 18% 2% 12% Other metrics 26% 12% 8% Table 2 shows the distribution of the selected features according to feature integration operators. It shows which operators applied to result set query-URL pair wise features are most useful. We group the 15 operators into 5 types: vector, normalized ratios (r k , k = 2, 5, 10, 20), min/max, en-tropy/stand deviation, and median/mean. Vector group includes all query-URL pair features in top 5 positions; normalized ratios are defined in (1). As we can see from the table, all feature integration operators are useful. Table 2: Distributions of the Selected Top 50 Features According to Integration Operators Operators Info. Gain Linear SVM Boosting vector 40% 22% 28% normalized ratios 8% 38% 22% min/max 6% 20% 16% entropy/std 20% 16% 18% mean/median 26% 4% 16% The number of selected features directly influence the classification performance. Figure 3 shows relationship between the boosting tree classification performance and the number of selected features. As we can see, performance increases with cleaner selected features. However, if the number of selected feature is too small, performance will decrease. A number of 50 works the best in our work. 6.3.2 Classification Results We first apply four different classification methods: naive Bayes, maximum entropy methods, support vector machine and stochastic gradient boosting tree model over all available features. The results are reported in Table 3. As we can see, stochastic gradient boosting tree has the best performance with an F1 score of 0.78. We then apply those methods to machine selected features . We test 4 different feature sets with 50 number of features , selected by information gain, linear SVM and boosting tree. The combined set consists of 30 top features selected by linear SVM and 29 top features selected by boosting tree. Please note that the total number of features are still 50 since linear SVM and boosting tree selected 9 same features in their top 30 feature set. 687 0 500 1000 1500 2000 2500 3000 0.78 0.79 0.8 0.81 0.82 0.83 0.84 0.85 0.86 Number of Features Selected By Boosting Tree F1 Score of Boosting Tree Classifier Classification Performance VS Number of Features Figure 3: Classification performance F1 against number of features: 25, 50, 100, 200, 400, 800, and 2955 (all features) Table 3: Results of Various Classification Methods over All Features Recall Precision F1 Naive Bayes 0.242 0.706 0.360 SVM (Linear Kernel) 0.189 1.000 0.318 Maximum Entropy 0.743 0.682 0.711 SVM (RBF Kernel) 0.589 0.485 0.528 Boosting Trees 0.724 0.845 0.780 Table 4 presents the results of the coupled feature selection and classification methods. It is obvious that the performance of each method is improved by applying them to machine selected clean features, except naive Bayes classifier. Surprisingly, the features selected by linear SVM are the best set of features. The results show that even if the underlying problem is not linear separable, the linear coefficients of the large margin linear classifier still convey important feature information. When the stochastic gradient boosting tree is applied over this set of features, we get the best performance with 0.881 F1 score among all cross-methods evaluations. Without feature ablation, SGBT is only able to achieve 0.738 F1 score. That is, feature selection has an effect of error reduction rate 40%. Without introducing linear SVM in feature ablation, if SGBT works on the feature set selected by its own variable importance ranking, it achieves 0.848 F1 score. That is to say, a cross methods coupling of feature selection and classification causes a 33% error reduction. DISCUSSION An interesting result from Table 1 is the features selected for navigational query identification. Those features are mostly induced from user click information. This is intu-itively understandable because if a query is navigational, the navigational URL is the most clicked one. On the other hand, it might be risky to completely rely on click information . The reasons might be 1) user click features may be easier to be spammed, and 2) clicks are often biased by various presentation situation such as quality of auto abstraction , etc. From Table 4, we observe that linear SVM and boosting tree have better feature selection power than information gain. The reason that information gain performs inferior to linear SVM and boosting tree is probably due to the fact that information gain considers each feature independently while linear SVM considers all features jointly and boosting tree composites feature rank by sum over all used features. The results show that URL, anchor text and other metrics are helpful only when they are considered jointly with click features. The most important result is that the stochastic gradient boosting tree coupled with linear SVM feature selection method achieves much better results than any other combination . In this application, the data has very high dimension considering the small sample size. The boosting tree method needs to partition an ultra-high dimensional feature space for feature selection. However, the stochastic step does not have enough data to sample from [6]. Therefore, the boosted result might be biased by earlier sampling and trapped in a local optimum. Support vector machine, however, is able to find an optimally determined subset of training samples, namely support vectors, and ranks features based on those vectors. Therefore, the SVM feature selection step makes up the disadvantage of the stochastic boosting tree in its initial sampling and learning stages that may lead to a local optimum. As expected, naive Bayes classifier hardly works for the navigational query identification problem. It is also the only classifier that performs worse with feature selection. Naive Bayes classifiers work well when the selected features are mostly orthogonal. However, in this problem, all features are highly correlated. On the other hand, classification methods such as boosting tree, maximum entropy model and SVM do not require orthogonal features. RELATED WORK Our work is closely related to query classification, a task of assigning a query to one or more categories. However, general query classification and navigational query identification are different in the problems themselves. Query classification focuses on content classification, thus the classes are mainly topic based, such as shopping and products. While in navigational query identification, the two classes are intent based. In the classification approaches regard, our work is related to Gravano, et al. [7] where authors applied various classification methods, including linear and nonlinear SVM, decision tree and log-linear regression to classify query locality based on result set features in 2003. Their work, however, lacked carefully designed feature engineering and therefore only achieved a F1 score of 0.52 with a linear SVM. Beitzel, et al.[1] realized the limitation of a single classification method in their query classification problem and pro-posed a semi-supervised learning method. Their idea is to compose the final classifier by combining classification results of multiple classification methods. Shen, et al. [15] also trained a linear combination of two classifiers. Differ-ently , instead of combining two classifiers for prediction, we couple feature selection and classification. In the feature extraction aspect, our work is related to Kang and Kim 2003 [11] where authors extracted heterogenous features to classify user queries into three categories: topic relevance task, the homepage finding task and service finding task. They combined those features, for example URL feature and content feature, by several linear empiri-688 Table 4: F1 Scores of Systems with Coupled Feature Selection and Classification Methods Methods Info. Gain Linear SVM Boosting Combined Set SVM (Linear Kernel) 0.124 0.733 0.712 0.738 Naive Bayes 0.226 0.182 0.088 0.154 Maximum Entropy 0.427 0.777 0.828 0.784 SVM (RBF Kernel) 0.467 0.753 0.728 0.736 Boosting Tree 0.627 0.881 0.848 0.834 cal linear functions. Each function was applied to a different binary classification problem. Their idea was to emphasize features for different classification purposes. However, the important features were not selected automatically and therefore their work is not applicable in applications with thousands of features. CONCLUSION We have made three contributions in the paper. First, we evaluate the effectiveness of four machine learning approaches in the context of navigational query identification. We find that boosting trees are the most effective one. Second , we evaluate three feature selection methods and propose coupling feature selection with classification approaches. Third, we propose a multi-level feature extraction system to exploit more information for navigational query identification . The underlying classification problem has been satisfacto-rily solved with 88.1% F1 score. In addition to the successful classification, we successfully identified key features for recognizing navigational queries: the user click features. Other features, such as URL, anchor text, etc. are also important if coupled with user click features. In future research, it is of interest to conduct cross methods co-training for the query classification problem to utilize unlabeled data, as there is enough evidence that different training methods may benefit each other. REFERENCES [1] S. Beitzel, E. Jensen, D. Lewis, A. Chowdhury, A. Kolcz, and O. Frieder. Improving Automatic Query Classification via Semi-supervised Learning. In The Fifth IEEE International Conference on Data Mining, pages 2730, New Orleans, Louisiana, November 2005. [2] C. Bhattacharyya, L. R. Grate, M. I. Jordan, L. El Ghaoui, and I. S. Mian. Robust Sparse Hyperplane Classifiers: Application to Uncertain Molecular Profiling Data. Journal of Computational Biology, 11(6):10731089, 2004. [3] A. Broder. A Taxonomy of Web Search. In ACM SIGIR Forum, pages 310, 2002. [4] S. della Pietra, V. della Pietra, and J. Lafferty. Inducing Features of Random Fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(4), 1995. [5] R. O. Duda, P. E. Hart, and D. G. Stork. Pattern Classification. John Wiley, New York, NY, 2nd edition, 2000. [6] J. H. Friedman. Stochastic Gradient Boosting. Computational Statistics and Data Analysis, 38(4):367378, 2002. [7] L. Gravano, V. Hatzivassiloglou, and R. Lichtenstein. Categorizing Web Queries According to Geographical Locality. In ACM 12th Conference on Information and Knowledge Management (CIKM), pages 2730, New Orleans, Louisiana, November 2003. [8] T. Hastie, R. Tibshirani, and J. Friedman. The Elements of Statistical Learning: Data Mining, Inference, and Predication. Springer Verlag, New York, 2001. [9] E. T. Jaynes. Papers on Probability, Statistics, and Statistical Physics. D. Reidel, Dordrecht, Holland and Boston and Hingham, MA, 1983. [10] T. Joachims. Text Categorization with Support Vector Machines: Learning with Many Relevant Features. In Proceedings of the 10th European Conference on Machine Learning (ECML), pages 137142, Chemnitz, Germany, 1998. [11] I.-H. Kang and G. Kim. Query Type Classification for Web Document Retrieval. In Proceedings of the 26th annual international ACM SIGIR conference on Research and development in informaion retrieval, pages 64 71, Toronto Canada, July 2003. [12] U. Lee, Z. Liu, and J. Cho. Automatic Identification of User Goals in Web Search. In Proceedings of the 14th International World Wide Web Conference (WWW), Chiba, Japan, 2005. [13] R. Malouf. A Comparison of Algorithms for Maximum Entropy Parameter Estimation. In Proceedings of the Sixth Conference on Natural Language Learning (CoNLL), Taipei, China, 2002. [14] D. E. Rose and D. Levinson. Understanding User Goals in Web Search. In Proceedings of The 13th International World Wide Web Conference (WWW), 2004. [15] D. Shen, R. Pan, J.-T. Sun, J. J. Pan, K. Wu, J. Yin, and Q. Yang. Q2C at UST: Our Winning Solution to Query Classification in KDDCUP 2005. SIGKDD Explorations, 7(2):100110, 2005. [16] L. Sherman and J. Deighton. Banner advertising: Measuring effectiveness and optimizing placement. Journal of Interactive Marketing, 15(2):6064, 2001. [17] V. Vapnik. The Nature of Statistical Learning Theory. Springer Verlag, New York, 1995. [18] Y. Yang and J. Pedersen. An Comparison Study on Feature Selection in Text Categorization. In Proceedings of the 20th annual international ACM SIGIR conference on Research and development in informaion retrieval, Philadelphia, PA, USA, 1997. [19] S.C. Zhu. Statistical modeling and conceptualization of visual patterns. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(6):619712, 2003. 689

Stochastic Gradient Boosting Tree;Linear SVM Feature Ranking;Gradient Boosting Tree;Information Gain;Naive Bayes Classifier;Support Vector Machine;Experiments Results;Machine Learning;Maximum Entropy Classifier;Navigational Query Classification;Navigational and Informational query;Multiple Level feature system

61

Coverage Directed Test Generation for Functional Verification using Bayesian Networks

Functional verification is widely acknowledged as the bottleneck in the hardware design cycle. This paper addresses one of the main challenges of simulation based verification (or dynamic verification ), by providing a new approach for Coverage Directed Test Generation (CDG). This approach is based on Bayesian networks and computer learning techniques. It provides an efficient way for closing a feedback loop from the coverage domain back to a generator that produces new stimuli to the tested design. In this paper, we show how to apply Bayesian networks to the CDG problem. Applying Bayesian networks to the CDG framework has been tested in several experiments, exhibiting encouraging results and indicating that the suggested approach can be used to achieve CDG goals.

INTRODUCTION Functional verification is widely acknowledged as the bottleneck in the hardware design cycle [1]. To date, up to 70% of the design development time and resources are spent on functional verification . The increasing complexity of hardware designs raises the need for the development of new techniques and methodologies that can provide the verification team with the means to achieve its goals quickly and with limited resources. The current practice for functional verification of complex designs starts with a definition of a test plan, comprised of a large set of events that the verification team would like to observe during the verification process. The test plan is usually implemented using random test generators that produce a large number of test-cases , and coverage tools that detect the occurrence of events in the test plan, and provide information related to the progress of the test plan. Analysis of the coverage reports allows the verification team to modify the directives for the test generators and to better hit areas or specific tasks in the design that are not covered well [5]. The analysis of coverage reports, and their translation to a set of test generator directives to guide and enhance the implementation of the test plan, result in major manual bottlenecks in the otherwise highly automated verification process. Considerable effort is invested in finding ways to close the loop of coverage analysis and test generation. Coverage directed test generation (CDG) is a technique to automate the feedback from coverage analysis to test generation. The main goals of CDG are to improve the coverage progress rate, to help reaching uncovered tasks, and to provide many different ways to reach a given coverage task. Achieving these goals should increase the efficiency and quality of the verification process and reduce the time and effort needed to implement a test plan. In this paper, we propose a new approach for coverage directed test generation. Our approach is to cast CDG in a statistical inference framework, and apply computer learning techniques to achieve the CDG goals. Specifically, our approach is based on modeling the relationship between the coverage information and the directives to the test generator using Bayesian networks [9]. A Bayesian network is a directed graph whose nodes are random variables and whose edges represent direct dependency between their sink and source nodes. Each node in the Bayesian network is associated with a set of parameters specifying its conditional probability given the state of its parents. Simply stated, the CDG process is performed in two main steps. In the first step, a training set is used to learn the parameters of a Bayesian network that models the relationship between the coverage information and the test directives. In the second step, the Bayesian network is used to provide the most probable directives that would lead to a given coverage task (or set of tasks). Bayesian networks are well suited to the kind of modeling required for CDG, because they offer a natural and compact representation of the rather complex relationship between the CDG ingredients, together with the ability to encode essential domain knowledge. Moreover, adaptive tuning of the Bayesian network parameters provides a mean to focus on the rare coverage cases. We describe two experiments in which we tested the the ability of Bayesian networks to handle aspects of the CDG problem in various settings. The goals of the experiments were to increase the hitting rates in hard-to-reach coverage cases; design directives aimed at reaching uncovered tasks; and provide many different directives for a given coverage task. We used two settings for our experiments. In the first setting, we used a Bayesian network to generate instruction streams to an abstract model of the pipeline of 286 18.2 Random Test Generator Test Plan Fail Pass Information Directives Test Coverage Coverage Analysis Tool Reports Coverage Simulator DUT Figure 1: Verification process with automatic test generation an advanced super-scalar PowerPC processor. In the second setting , we used a Bayesian network to generate directives to an existing test generator of a storage control unit of a mainframe with a goal to cover all possible transactions from the CPUs connected to this unit. In both experiments we reached our goals. The encouraging results suggests that Bayesian networks may well be used to achieve the primary goals of CDG. The remainder of this paper is as follows. In Section 2, we briefly present the CDG framework and review related work. In Section 3, we describe Bayesian networks and their application to CDG. Sections 4 and 5 provide detailed descriptions of the experiments. We conclude with a few remarks and suggestions for future study. COVERAGE DIRECTED TEST GENERATION (CDG) In current industry practice, verification by simulation, or dynamic verification, is the leading technique for functional verification . Coverage is used to ensure that the verification of the design is thorough, and the definition of coverage events or testing requirements is a major part in the definition of the verification plan of the design. Often, a family of coverage events that share common properties are grouped together to form a coverage model [7]. Members of the coverage model are called coverage tasks and are considered part of the test plan. Cross-product coverage models [7] are of special interest. These models are defined by a basic event and a set of parameters or attributes, where the list of coverage tasks comprises all possible combinations of values for the attributes. Figure 1 illustrates the verification process with an automatic random test generation. A test plan is translated by the verification team to a set of directives for the random test generator. Based on these directives and embedded domain knowledge, the test generator produces many test-cases. The design under test (DUT) is then simulated using the generated test-cases, and its behavior is monitored to make sure that it meets its specification. In addition, coverage tools are used to detect the occurrence of coverage tasks during simulation. Analysis of the reports provided by the coverage tools allows the verification team to modify the directives to the test generator to overcome weaknesses in the implementation of the test plan. This process is repeated until the exit criteria in the test plan are met. The use of automatic test generators can dramatically reduce the amount of manual labor required to implement the test plan. Even so, the manual work needed for analyzing the coverage reports and translating them to directives for the test generator, can constitute a bottleneck in the verification process. Therefore, considerable effort is spent on finding ways to automate this procedure, and close the loop of coverage analysis and test generation. This automated feedback from coverage analysis to test generation, known as Coverage Directed test Generation (CDG), can reduce the manual work in the verification process and increase its efficiency. In general, the goal of CDG is to automatically provide directives that are based on coverage analysis to the test generator. This can be further divided into two sub-goals: First, to provide directives to the test generator that help in reaching hard cases, namely uncovered or rarely covered tasks. Achieving this sub-goal can shorten the time needed to fulfill the test plan and reduce the number of manually written directives. Second, to provide directives that allow easier reach for any coverage task, using a different set of directives when possible. Achieving this sub-goal makes the verification process more robust, because it increases the number of times a task has been covered during verification. Moreover, if a coverage task is reached via different directions, the chances to discover hidden bugs related to this task are increased [8]. In the past, two general approaches for CDG have been pro-posed : feedback-based CDG and CDG by construction. Feedback-based CDG relies on feedback from the coverage analysis to automatically modify the directives to the test generator. For example, in [2], a genetic algorithm is used to select and modify test-cases to increase coverage. In [13], coverage analysis data is used to modify the parameters of a Markov Chain that represents the DUT. The Markov Chain is then used to generate test-cases for the design. In [11], the coverage analysis results trigger a set of generation rules that modify the testing directives. In contrast, in CDG by construction, an external model of the DUT is used to generate test directives designed to accurately hit the coverage tasks. For example , in [14] an FSM model of pipelines is used to generate tests that cover instruction interdependencies in the pipes. COVERAGE DIRECTED TEST GENERATION USING BAYESIAN NETWORKS The random nature of automatic test-case generators imposes a considerable amount of uncertainty in the relationship between test directives and coverage tasks, e.g., the same set of directives can be used to generate many different test-cases, each leading to different coverage tasks. This inherent uncertainty suggests to cast the CDG setup in a statistical inference framework. To this end, Bayesian networks offer an efficient modeling scheme by providing a compact representation of the complex (possibly stochastic) relationships among the CDG ingredients, together with the possibility to encode essential domain knowledge. It should be noted that we do not suggest modeling the behavior of the design, typi-cally a large and complicated (deterministic) finite state machine. Rather, we model the CDG process itself, namely the trial-and-error procedure governed by the verification team, which controls the test generation at one end and traces the progress of covering the test plan at the other. 3.1 A Brief Introduction to Bayesian Networks A Bayesian network is a graphical representation of the joint probability distribution for a set of variables. This representation was originally designed to encode the uncertain knowledge of an expert and can be dated back to the geneticist Sewall Wright [15]. Their initial development in the late 1970s was motivated by the need to model the top-down (semantic) and bottom-up (perceptual) combinations of evidence (observations/findings). Their capability for bidirectional inferences, combined with a rigorous probabilistic foundation, led to the rapid emergence of Bayesian networks as the method of choice for uncertain reasoning in AI and expert systems, replacing ad hoc rule-based schemes. Bayesian networks also play a crucial role in diagnosis and decision support systems [10]. Obviously, there's a computational problem in dealing with many sources of uncertainty, i.e. the ability to perform probabilistic ma-nipulations in high dimensions (the "curse of dimensionality"). The main breakthrough emerged in the late 1980s and can be attributed to Judea Pearl [12], who introduced 'modularity', thus enabling 287 large and complex models and theirs associated calculations, to be split up into small manageable pieces. The best way to do this is via the imposition of meaningfully simplified conditional independence assumptions. These, in turn, can be expressed by means of a powerful and appealing graphical representation. A Bayesian network consists of two components. The first is a directed acyclic graph in which each vertex corresponds to a random variable. This graph represents a set of conditional independence properties of the represented distribution: each variable is probabilistically independent of its non-descendants in the graph given the state of its parents. The graph captures the qualitative structure of the probability distribution, and is exploited for efficient inference and decision making. The second component is a collection of local interaction models that describe the conditional probability p (X i |Pa i ) of each variable X i given its parents Pa i . Together , these two components represent a unique joint probability distribution over the complete set of variables X [12]. The joint probability distribution is given by the following equation: p (X) = n i =1 p (X i |Pa i ) (1) It can be shown that this equation actually implies the conditional independence semantics of the graphical structure given earlier. Eq. 1 shows that the joint distribution specified by a Bayesian network has a factored representation as the product of individual local interaction models. Thus, while Bayesian networks can represent arbitrary probability distributions, they provide a computational advantage for those distributions that can be represented with a simple structure. The characterization given by Eq. 1 is a purely formal characterization in terms of probabilities and conditional independence. An informal connection can be made between this characterization and the intuitive notion of direct causal influence. It has been noted that if the edges in the network structure correspond to causal relationships , where a variable's parents represent the direct causal influences on that variable, then resulting networks are often very concise and accurate descriptions of the domain. Thus it appears that in many practical situations, a Bayesian network provides a natural way to encode causal information. Nonetheless, it is often difficult and time consuming to construct Bayesian networks from expert knowledge alone, particularly because of the need to provide numerical parameters. This observation, together with the fact that data is becoming increasingly available and cheaper to acquire, has led to a growing interest in using data to learn both the structure and probabilities of a Bayesian network (cf. [3, 9, 12]). Typical types of queries that can be efficiently answered by the Bayesian network model are derived from applying the Bayes rule to yield posterior probabilities for the values of a node (or set of nodes), X , given some evidence, E, i.e. assignment of specific values to other nodes: p (X|E) = p(E|X) p(X) p (E) Thus, a statistical inference can be made in the form of either selecting the Maximal A Posteriori (MAP) probability, max p (X|E), or obtaining the Most Probable Explanation (MPE), arg max p (X|E). The sophisticated yet efficient methods that have been developed for using Bayesian networks provide the means for predictive and diagnostic inference 1 . A diagnostic query is such that the evidence 1 This is in contrast to standard regression and classification methods (e.g., feed forward neural networks and decision trees) that encode only the probability distribution of a target variable given several input variables. State Int Covearge Variables Directives Test Generator Core Enbable Cmd Type cp_cmd_type = {// val weight {read, 20}, {write, 20}, {RMW, 5}, ... }; cp_core_enable = {// val weight {Core 1, 10}, }; Op Mode Core Cmd Resp {Core 0, 10}, {Both, 100} Figure 2: Bayesian Network of CDG nodes E represent a cause, while the queried nodes, X , represent an effect. The reversed direction, i.e. evidence on the effect nodes which serves to determine the possible cause, is called abductive. These methods also allow Bayesian networks to reason efficiently with missing values, by computing the marginal probability of the query given the observed values. There are two important extensions of Bayesian networks: Dynamic Bayesian networks and influence diagrams. The first extension (see [6]) enables the incorporation of time, thus modeling temporal dependencies in a stochastic process. The second extension (see [3]) enriches the Bayesian network paradigm with decision making and utility considerations which create a powerful mechanism for dealing with decisions under uncertainty constraints. 3.2 A Bayesian Network for CDG The CDG process begins with the construction of a Bayesian network model that describes the relations between the test directives and the coverage space. Figure 2 illustrates a simple, yet typical, Bayesian network, which models a small excerpt of the CDG setup. The network describes the relationship between the directives that influence the type of command that is generated (cp cmd type) and the active cores inside a CPU (cp core enable), and the coverage attributes of a generated command (cmd), its response (resp), and the core that generated it (core). The network is comprised of input nodes (the white circles on the left) that relate to test directives that appear to their left and coverage nodes (the white squares on the right) that define the coverage space. In addition to these nodes, for which we have physical observations, the network may also contain hidden nodes, namely variables for which we don't have any physical evidence (observations) for their interactions. These variables are represented as shaded ovals in the figure. Hidden nodes are added to the Bayesian network structure primarily to reflect expert domain knowledge regarding hidden causes and functionalities which impose some structure on the interaction between the interface (observed) nodes 2 . The Bayesian network at Fig. 2 describes the causal relationships from the test generation directives (causes) to the coverage model space (effects). For example, it encodes the expert knowledge that indicates that there is an internal mode of operation for which we do not have any direct physical observation, yet it is determined by the combined values of the test generation attributes. On the other hand, the (hidden) mode of operation directly influences the choice of the resulting command and core, which are attributes of 2 Introducing hidden nodes to the network structure has the secondary impact of reducing the computational complexity by dimensionality reduction, and as a means for capturing non-trivial (higher order) correlations between observed events. 288 the coverage model. Note the absence of a direct link between the requested core (via the directive cp core enable) and the observed one (at Core), which captures our understanding that there is no direct influence between the directives and the coverage attribute . Another assumption encoded in the CDG Bayesian network structure at Fig. 2, is that the only information that governs the response for the command is the generated command itself, and this is encoded via the direct link from Cmd to Resp. In a nutshell, the design of the Bayesian network starts with identifying the ingredients (attributes) that will constitute the directives to the test generator on one hand, and to the coverage model on the other. These attributes are dictated by the interface to the simulation environment, to the coverage analysis tool, and by the specification of the coverage model in the test plan. These ingredients are used as the first guess about the nodes in the graph structure. Connecting these nodes with edges is our technique for expert knowledge encoding, as demonstrated in Fig. 2. Obviously, using a fully connected graph, i.e. with an edge between every pair of nodes, represents absolutely no knowledge about the possible dependencies and functionalities within the model. Hence, as the graph structure becomes sparser, it represents deeper domain knowledge. We discovered that a good practice in specifying a dependency graph is to remove edges for which we have strong belief that the detached nodes are not directly influencing one another. At this point, hidden nodes can be added to the structure, either to represent hidden causes, which contribute to a better description of the functionalities of the model, or to take on a role from the complexity stand point, by breaking the barges cliques in the graph (see [4]). After the Bayesian network structure is specified, it is trained using a sample of directives and the respective coverage tasks. To this end, we activate the simulation environment and construct a training set out of the directives used and the resulting coverage tasks. We then use one of the many known learning algorithms (cf. [3]) to estimate the Bayesian network's parameters (i.e. the set of conditional probability distributions). This completes the design and training of the Bayesian network model. In the evaluation phase, the trained Bayesian network can be used to determine directives for a desired coverage task, via posterior probabilities, MAP and MPE queries, which use the coverage task attributes as evidence. For example, in a model for which the directives are weights of possible outcomes for internal draws in the test generator (e.g. the directive cp cmd type in Fig. 2 specifies a preference to read commands, write commands , etc.), we can specify a desired coverage task assignment (evidence) for the coverage nodes (e.g. Resp = ACK) and calculate the posterior probability distribution for directive nodes (e.g. p (Cmd Type|Resp = ACK)), which directly translates to the set of weights to be written in the test generator's parameter file. Note, as the example demonstrates, we can specify partial evidence and/or determine a partial set of directives. INSTRUCTION STREAM GENERATION USING A DYNAMIC NETWORK To evaluate the feasibility of the suggested modeling approach to the CDG problem, we designed a controlled study that acts in a simple domain (small state space), where we have a deep understanding of the DUT's logic, direct control on the input, and a `ground truth' reference to evaluate performance. We conducted the experiment on a model of the pipeline of NorthStar , an advanced PowerPC processor. The pipeline of NorthStar contains four execution units and a dispatch unit that dispatches instructions to the execution units. Figure 3 illustrates the general Dispatch 0000 0000 0000 0000 0000 0000 1111 1111 1111 1111 1111 1111 Branch Pipe (B) Write Back Execute Data Fetch S3 S2 S1 Simple Arith Pipe (S) C2 C1 C3 Complex Arith Pipe (C) 0000 0000 0000 0000 0000 0000 1111 1111 1111 1111 1111 1111 Load/Store Pipe (L) Figure 3: The structure of the NorthStar pipeline structure of the NorthStar pipeline. For reasons of simplicity, our model contains only the simple arithmetic unit that executes simple arithmetic instructions such as add, and the complex arithmetic unit that can execute both simple and complex arithmetic instructions. Each execution unit consists of three pipeline stages: (1) Data fetch stage, in which the data of the instruction is fetched; (2) Execute stage, in which the instruction is executed; (3) Write back stage, where the result is written back to the target register. The flow of instructions in the pipeline is governed by a simple set of rules. For example, in-order dispatching of instructions to the execution units, and rules for stalling because of data dependency. Note, the complete set of rules is omitted to simplify the description. We developed a simple abstract model of the dispatch unit and two pipelines and used it to simulate the behavior of the pipeline. The input to our NorthStar model is a simplified subset of the PowerPC instruction set. Each instruction is modeled by four input variables. The first variable indicates the type of the instruction. There are five possible types: S - simple arithmetic; C1, C2, C3 - complex arithmetic; and NOP - instructions that are executed in other execution units. The second and third input variables constitute the source and target register of the instructions. For simplicity and in order to increase the possibility of register interdependency, we used only eight registers instead of the 32 registers available in PowerPC. The last input variable indicates whether the instruction uses the condition register. Due to restrictions on the legal combinations of the input variables (e.g., NOP instruction is not using registers), there are 449 possible instructions. We used a coverage model that examines the state of the two pipelines, and properties of the instructions in them. The coverage model consists of five attributes, the type of instruction at stage 1 of the simple and complex arithmetic pipelines (S1Type and C1Type, resp.), flags indicating whether stage 2 of the pipelines are occupied (S2Valid and C2Valid, resp.), and a flag indicating whether the instruction at stage 2 of the simple arithmetic pipeline uses the condition register (S2CR). The total number of legal coverage tasks in the model is 54 (out of 80 possible cases). The goal of the experiment was to generate instruction streams that cover the coverage model described above. Specifically, we concentrated on the ability to reach the desired coverage cases with many, yet relatively short, instruction sequences. We modeled the temporal dependencies between the instructions and coverage tasks and among the instructions using a two-slice Dynamic Bayesian Network (DBN) [6]. Rather than an accurate mapping of the specific state machine structure, the DBN encoded the general knowledge of an expert on the modus operandi of this type of DUT. Using an expert's domain knowledge proved to be vital in this setup because it provided essential information needed for the generation of instruction streams. Moreover, it enabled the use of hidden nodes, which effectively reduced the complexity through dimensionality reduction. The resulting DBN has 19 289 Time slice (cycle) t Time slice (cycle) t+1 Input Node Coverage Node Hidden Node type1 sr1 tg1 cr1 type2 sr2 tg2 cr2 type1 sr1 tg1 cr1 type2 sr2 tg2 cr2 im0 ir0 mv1 rv1 rcr1 im0 ir0 mv1 rv1 rcr1 Figure 4: two-slice DBN for the NorthStar experiment Rare Uncovered Instructions Cycles Instructions Cycles Training Set 6 7 DBN 4 5 4 5 Text Book 3 4 3 4 Table 1: NorthStar experiment results nodes per slice, 13 of which are observed, 15 intra (within a slice) edges, and 37 inter (between slices) edges (see Fig 4). The training set is composed of 1000 sequences of random instructions . The length of each sequence is 10 cycles. Note, the model the we used for the Bayesian network made it easier to measure length in terms of cycles instead of instructions. The training set contained 385 different instructions. During its simulation, 49 (out of 54) coverage cases were observed. The average number of instructions per sequence in the training set was 9.7 out of the 20 possible dispatches in 10 cycles (i.e., more than half of the dispatch slots in the sequence are empty). After training the Bayesian network, we tried to generate instruction sequences for all 54 coverage tasks in the coverage model. Each sequence was generated using the DBN, by solving the Most Probable Explanation (MPE) problem for the requested coverage task. All 49 coverage cases of the training set plus three additional uncovered cases were reached using instruction sequences designed by the DBN. In addition, we generated many different instruction sequences for each coverage task that was covered by the Bayesian network. The average number of cycles in a generated sequence dropped to 2.9, while the average number of instructions in a sequence reduced to 3.7. This reflects the fact that the generated instruction sequences cause less stall states en-route to reaching the desired coverage cases. Table 1 illustrates the details of reaching two difficult coverage cases--the rarest coverage task, which was seen only once in the training set, and an uncovered task. The table shows the number of cycles and instructions required to reach these tasks in the training set, the instruction sequences generated by the trained DBN, and the `text book' solution--the best possible sequence. The table indicates that the instruction sequences generated by the DBN are shorter, both in instructions and cycles, than the sequences in the training set. Overall, the results indicate that the trained DBN is able to generate many compact instruction sequences that are not far from the best possible solution. Resp Cmd Resp Cmd Resp Cmd Pipe 0 Pipe 1 Core 0 Core 1 Core 0 Core 1 Core 0 Core 1 Storage Control Element (SCE) Memory Subsystem CP0 CP1 CP7 Figure 5: The structure of SCE simulation environment STORAGE CONTROL EXPERIMENT USING A STATIC NETWORK The second experiment was conducted in a real-life setting. The design under test in the experiment is the Storage Control Element (SCE) of an IBM z-series system. Figure 5 shows the structure of the SCE and its simulation environment. The SCE handles commands from eight CPUs (CP0 CP7). Each CPU consists of two cores that generate commands to the SCE independently. The SCE handles incoming commands using two internal pipelines. When the SCE finishes handling a command, it sends a response to the commanding CPU. The simulation environment for the SCE contains, in addition to the SCE itself, behavioral models for the eight CPUs that it services , and a behavioral model for the memory subsystem. The behavioral models of the CPUs generate commands to the SCE based on their internal state and a directive file provided by the user. The directive file contains a set of parameters that affect the behavior of the system. Some of these parameters control the entire system while others are specific to certain components of the system, such as a specific CPU. Figure 2 shows an example of some parameters that are used in the simulation environment of the SCE. Each parameter contains a set of possible values that the parameter can receive. Each value has a weight associated with it. When the value of a parameter is needed, it is randomly chosen from the set of possible values according the weights of these values. For example, when a CPU generates a new command, it first uses the cp cmd type parameter to determine the type of command to generate, and then a specific parameter for that command type to determine the exact command to be used. In the experiment, we tried to cover all the possible transactions between the CPUs and the SCE. The coverage model contained five attributes: The CPU (8 possible values) and the core (2 values) in it that initiated the command, the command itself (31 values), its response (14 values), and the pipeline in the SCE that handled it (2 values). Overall, the cross product contains 13,888 cases and the coverage model contains 1968 legal coverage tasks. This experiment added many new challenges over the controlled experiment described in the previous section. First, our knowledge about the DUT in this experiment was very limited compared to the full understanding of the design in the first experiment. In addition , we were less able to observe and control the input and output nodes of the Bayesian network. For the test parameters, we could only specify the distribution of each parameter and we could not observe the values that were actually used, only their distribution. Moreover, in some cases the behavioral models ignored the parameters and generated commands based on their internal state. Thus, the actual distribution used was not exactly the provided distribu-290 0 50 100 150 200 250 300 350 400 450 0 50 100 150 200 250 Test-cases C overe d T as k s CDG Baseline 223 Figure 6: Coverage progress of the CDG process tion of the parameters. This type of observation (distribution instead of specific value) is known as a soft evidence. The coverage data that we got out of the simulation environment was a summary of all the coverage tasks that occurred during the simulation of a test-case. Therefore, it was hard to correlate between the observed coverage tasks and the parameters' values that caused them and between the different observed coverage tasks. Because we had limited knowledge about the DUT and the correlation between the parameters in the test directives and the coverage tasks, the first Bayesian network we constructed contained arcs between each of the coverage variables and each of the test parameters. We trained this network with 160 test-cases (each taking more than 30 minutes to execute). After the initial training, we analyzed the Bayesian network and found out that most of the test parameters were strongly correlated either to the command and response coverage variables or the pipe and core variables, but only a single variable was strongly correlated to all coverage variables. Therefore, we partitioned the Bayesian network into two networks, one for command and response and the other for core and pipe. The result of the inference on the common parameter from the first network was used as input for the second one. We trained the second network with the same training set of 160 test-cases. During the training, 1745 out of the 1968 tasks in the model were covered, while 223 remained uncovered. We checked the performance of the trained network and its ability to increase the coverage rate for the uncovered tasks in the training set. The baseline for comparison was the progress achieved by the best test directive file created by an expert user. We tried to maximize the coverage progress rate using a large number of test directive files aimed at specific sets of uncovered tasks. This approach is not realistic for a human user due the effort needed to create each set of directives. However, it is useful for the automatic creation of directives, because the inference time from the trained network is negligible. Our method to maximize the coverage progress rate was to randomly partition the uncovered tasks, use the trained network to create a test directive file for each partition, and simulate a single test-case for each directive file. This process was repeated until all the tasks were covered. The CDG process was able to cover all uncovered tasks after 250 test-cases, while the baseline case of the user defined test directives file covered only two thirds of them after over 400 test-cases (see Figure 6). CONCLUSIONS AND FUTURE WORK In this paper we demonstrated how Bayesian networks can be used to close the loop between coverage data and directives to test generators. The experiments described in the paper show that this modeling technique can be efficiently used to achieve the CDG goals of easier reach for hard coverage cases, diverse reach for average cases, and improved coverage progress rate. It should be noted that the suggested CDG method is not limited to the types of simulation environments handled in this paper (i.e., parameters-based test generation and direct stimuli generation). It can be used in other types of environments, such as test generators in which the control on the stimuli is embedded in the generator itself. Our future work has two distinct aspects: enhancing the learning capabilities and effectively applying the suggested framework to the verification process. From the learning perspective, we plan to explore other techniques that may increase our capabilities. For example, incremental structure learning as a means for encoding richer domain knowledge, and the efficient construction of good queries to boost targeting rare cases using selective sampling. To effectively deploy the CDG framework, we need to gain a better understanding of the type of knowledge that should be encoded in the model, and to identify in which areas the suggested approach may prove most beneficial to the verification process. REFERENCES [1] J. Bergeron. Writing Testbenches: Functional Verification of HDL Models. Kluwer Academic Publishers, January 2000. [2] M. Bose, J. Shin, E. M. Rudnick, T. Dukes, and M. Abadir. A genetic approach to automatic bias generation for biased random instruction generation. In Proceedings of the 2001 Congress on Evolutionary Computation CEC2001, pages 442448, May 2001. [3] R. G. Cowell, A. P. Dawid, S. L. Lauritzen, and D. J. Spiegelhalter. Probabilistic Networks and Expert Systems. Springer-Verlag, 1999. [4] G. Elidan, N. Lotner, N. Friedman, and D. Koller. Discovering hidden variables: A structure-based approach. In Proceedings of the 13th Annual Conference on Neural Information Processing Systems, pages 479485, 2000. [5] L. Fournier, Y. Arbetman, and M. Levinger. Functional verification methodology for microprocessors using the Genesys test-program generator. In Proceedings of the 1999 Design, Automation and Test in Europe Conference (DATE), pages 434441, March 1999. [6] Z. Ghahramani. Learning dynamic Bayesian networks. In Adaptive Processing of Sequences and Data Structures, Lecture Notes in Artificial Intelligence, pages 168197. Springer-Verlag, 1998. [7] R. Grinwald, E. Harel, M. Orgad, S. Ur, and A. Ziv. User defined coverage - a tool supported methodology for design verification. In Proceedings of the 35th Design Automation Conference, pages 158165, June 1998. [8] A. Hartman, S. Ur, and A. Ziv. Short vs long size does make a difference. In Proceedings of the High-Level Design Validation and Test Workshop, pages 2328, November 1999. [9] D. Heckerman. A tutorial on learning with Bayesian networks. Technical report, Microsoft Research, 1996. [10] D. Heckerman, A. Mamdani, and M. Wellman. Real-world applications of Bayesian networks. Communications of the ACM, 38(3):2430, 1995. [11] G. Nativ, S. Mittermaier, S. Ur, and A. Ziv. Cost evaluation of coverage directed test generation for the IBM mainframe. In Proceedings of the 2001 International Test Conference, pages 793802, October 2001. [12] J. Pearl. Probabilistic Reasoning in Intelligent Systems: Network of Plausible Inference. Morgan Kaufmann, 1988. [13] S. Tasiran, F. Fallah, D. G. Chinnery, S. J. Weber, and K. Keutzer. A functional validation technique: biased-random simulation guided by observability-based coverage. In Proceedings of the International Conference on Computer Design, pages 8288, September 2001. [14] S. Ur and Y. Yadin. Micro-architecture coverage directed generation of test programs. In Proceedings of the 36th Design Automation Conference, pages 175180, June 1999. [15] S. Wright. Correlation and causation. Journal of Agricultural Research, 1921. 291

Coverage directed test generation;conditional probability;Functional Verification;Bayesian Networks;bidirectional inferences;Maximal A Posteriori;Dynamic Bayesian Network;design under test;coverage model;Coverage Analysis;Most Probable Explanation;Markov Chain

62

Creating a Massive Master Index for HTML and Print

An index connects readers with information. Creating an index for a single book is a time-honored craft. Creating an index for a massive library of HTML topics is a modern craft that has largely been discarded in favor of robust search engines. The authors show how they optimized a single-sourced index for collections of HTML topics, printed books, and PDF books. With examples from a recent index of 24,000 entries for 7,000 distinct HTML topics also published as 40 different PDF books, the authors discuss the connections between modern technology and traditional information retrieval methods that made the index possible, usable, and efficient to create and maintain.

THE PROBLEM A project with a large library of existing documentation of about 40 books (several with multiple volumes) required a high-quality master index. The material was written using a proprietary SGML authoring tool and converted to both HTML for browser-based information and PDF. The library was being converted from a set of books into a task-oriented, topic-based set of documentation, so very little indexing time was available for the approximately 20 authors in two sites fully engaged in rewriting the documentation to conform to new guidelines for online topics, to incorporate new content, and to change content for the next release of their product. The four project editors responsible for the complete library were likewise busy assisting the structural conversion of the library. Customers were asking for more direct ways to find information. At a user conference, 52 percent of users said the former book indexes were their primary entry point into product information. Given these imposing (but sadly typical) resource constraints, the information architect for the project and an editor with extensive indexing experience worked together to develop an approach that would use technology to maximize the available efforts of both authors and editors. This paper describes the approach from the perspectives of the authors, editors, and, most importantly, the users of this set of documentation. The approach is described in enough detail to enable other projects to adapt the approach to the constraints of their situation. The paper also indicates the future directions that the project will explore. The challenges to producing a high-quality master index were not just posed by available resources or by available technology, but also by the writing culture of the project. The project had historically been heavily oriented towards writing and producing books--the HTML documentation set had been simply a collection of books converted to HTML. As a result, navigation through the HTML documentation was difficult; there were almost no links between books, and each book was organized under its own table of contents. Full-text search of the HTML books had been the only method of finding information across the complete library, if a user didn't know which book to consult directly. The product market share was expanding, and new users had to learn the product quickly. However, cultural attitudes towards writing reinforced the problem of books as separate silos of information: authors were responsible for, and took justifiable pride in, producing their individual books, and while consistency in style and ease of access across the library was encouraged, it was much less important to the writers' satisfaction and explicit goals than completing their self-contained sets of information well. Earlier, the product had offered a PostScript master index as a printed book and a file, in response to customer feedback about finding information across the growing library. There was never time to improve it, so it was eventually dropped, but at the same time, the need for better retrievability of information was increasing and search did not adequately meet that need. Users 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 to republish, to post on servers or to redistribute to lists, requires specific permission and/or a fee. SIGDOC '02, October 20-23, Toronto, Ontario, Canada. Copyright 2002 ACM 1-58113-543-2/02/0010...$5.00 186 demanded both an interactive information finder and an easily scanned printable index. The previous version of the library produced a master index in both PDF and PostScript formats to assist users on platforms on which search did not work. However, the PDF index was generated from the individual indexes of each PDF book after the content of the books had been frozen for translation, so there was no attempt or opportunity to enforce consistency in indexing style across books, even when editors were able to help writers improve individual book indexes. Any effort in directly editing the master index would have been lost on the library as a whole because the index entries were part of the source books. Even so, the process of creating the master index took days and required the cooperation of dozens of authors. The PDF master index was only a limited replacement for search, since neither the PDF nor the PostScript version could provide direct links to the indexed information. The PDF master index forced users to find and open the corresponding PDF or printed book, then find the referenced page themselves. As part of the effort to address the problems of navigating through the HTML documentation for the next release of the product, the information architect and the editors decided to produce a high-quality master index in HTML, even though the full-text search capability would be supported across all platforms for the next release. WHY SEARCH IS NOT ENOUGH A frequent question early in the project was "Why bother with an online index if you have a good search engine?" The problem with search is that typical search solutions are limited to the terms or keywords that are found within the content matter, making search results a happy coincidence between the user's terminology and the content's terminology. If synonyms or preferred terms are addressed at all in typical search solutions, they are implemented as meta keywords. This turns an explicit advantage of the index (the capability of training the user in a product's vocabulary using see or see also references) into an implicit method to improve search results. The terms used in an index reflect the indexers' knowledge of the subject matter and the users of that set of documentation. While search solutions typically provide the user with the title, a ranking, and occasionally keywords in context, an index's primary, secondary, and tertiary entries give users more data and context with which to select the right piece of information to meet their needs. PROBLEMS WITH THE PREVIOUS PDF MASTER INDEX A previous version of the product shipped a PDF master index that was generated simply by collecting and sorting the author-created, book-specific indexes from each of the 40 books. An analysis of the resulting master index revealed that the indexing approach varied greatly across writers. There were many cases of inconsistent terminology. Singular versus plural nouns, preferences for verbs or nouns, and capitalization differences were the easiest problems to spot. Some authors indexed every occurrence of a given term, other authors indexed only the most significant usage of a given term, and others indexed almost nothing at all. Some authors indexed every possible synonym for a given term, and others indexed only the term itself. Some authors relied heavily on secondary and tertiary entries to lend structure to their index, while others relied almost exclusively on primary index entries. Many writers clearly didn't understand the significance of such decisions on the customers who would ultimately use the indexes. The master index was over 350 pages, and seemingly small differences in primary entries sometimes resulted in users having to look through several pages to find a specific entry. A complicating cultural problem was that, for many authors, indexing was an activity to be performed only when the content had been written and checked for technical accuracy. In many cases, this meant that the index for an individual book received very little attention, and, in most cases, the master index received no attention at all. BARRIERS It was clear that there were indexing problems throughout the library. However, without painstakingly analyzing each index in conjunction with the master index, it was impossible to determine what significant material might not have been indexed or which books had inconsistent capitalization and use of plurals. In the wake of the first master index in PDF, the editors initiated an attempt to address some of the most glaring inconsistencies by setting indexing guidelines, educating authors in the art of indexing, and encouraging authors to collaborate with other authors in the project to define standard indexing terminology for new or existing problem areas (for example, using specific labels for program entities so they made sense in the context of other entities from across the library, such as "SQL data type" versus "data type"). The effort met with some success, but the process of standardizing terminology across the library was unwieldy: index entries were maintained within the source files for each book, and the average book had between 50 and 200 source files. So, making a single terminology change that affected 10 percent of the source files in half of the books in the library would require opening, editing, and saving approximately 200 source files that are kept within a version control system. At five minutes per source file, that's almost 17 hours of work! In addition, the source files were often accessible only to the author, so there was a potential bottleneck and a conflict between demands on the author's time to write new content or revise existing index entries. INDEX ENTRIES AS METADATA Recognizing these problems along with others related to the shift to a topic-based architecture, we proposed a solution that required both technical and cultural change. The new approach to indexing was a shift from viewing index entries as content owned by each author and created solely during the writing process to treating index entries as metadata to be created and edited during a separate process. The new topic-writing process involves adding all topics in the library to a relational database and storing index entries in a table related to the table that stores the topics. The database is affectionately known as Dobalina. The very nature of an index is a database, because they consist of records that are compiled into a readable, searchable format (Wright, 2001), or several for single-sourced information. The database maximized our ability to generate flexible outputs for different formats. The initial indexing pass took advantage of the index entries in the legacy documentation: a Perl script stripped the index 187 entries from the source files, replacing them with an SGML text entity unique to each file, and inserting the index entries into the database. Once the index entries were in the database, both authors and editors could then use a Web interface to the database to maintain the index entries. To build a PDF version of their books, the authors download a set of auxiliary SGML files from the database that define the SGML text entities as their corresponding index entries. The following example demonstrates how the index entries are dynamically generated from the database and incorporated into the SGML source. &lt;!-- Start of topic source file --&gt; &index1; &lt;!-- Index text entity --&gt; &lt;p&gt;Some topic content.&lt;/p&gt; &lt;!-- End of topic source file --&gt; &lt;!-- Index entity definitions generated by database --&gt; &lt;!ENTITY index1 "&lt;index&gt; &lt;primary&gt;SQL statements&lt;/primary&gt; &lt;secondary&gt;data definition&lt;/secondary&gt; &lt;primary&gt;data definition&lt;/primary&gt; &lt;secondary&gt;SQL statements&lt;/secondary&gt; &lt;/index&gt;"&gt; 5.1 Better living through automation Storing the index entries in a database gives the team the ability to quickly generate the HTML master index, freeing authors from the requirement to painstakingly transform each of their books to create the individual indexes that composed the PDF master index. The process of creating the master index now takes approximately 15 minutes instead of days. Dynamic access to the entire set of index entries enables the team to easily isolate certain consistency problems, such as plural nouns or incorrect capitalization; to immediately identify topics without index entries; to help ensure library-wide consistency for new index terms by providing drop-down selection of terms; and to effect social change by eliminating the need to access source files to add, edit, or delete index entries. Spell-checking the 24,000-odd master index entries takes about two hours. The changes to the index terms are made in the database, so the corrections are automatically reflected in the book indexes. The previous approach would have required communicating each correction to the writer, and the master index would still have been subject to the errors if a writer did not have time to integrate the changes throughout the affected source files. 5.2 Cultural and process changes The role of editors changed significantly; rather than providing guidance to authors, the editors collectively edit the master index by directly manipulating index entries for the entire library through a Web interface. More extensive or complicated changes to index entries are accomplished through Perl scripts that manipulate the database records. Maintaining index entries in a database enables new methods of collaboration; for example, a writing team might designate their most skilled indexer to index the new content for a release. THE AUTHOR EXPERIENCE Indexing training provided detailed guidelines for the authors with a hands-on experience in group indexing. The detailed guidelines are included in the Appendix. They consist of two tables--one listing general indexing guidelines and a content-oriented table that lists specific consistency rules for the library. They also include some guidelines to help writers improve the quality of their PDF indexes. After the index entries are in the database, authors can dynamically generate Web-based reports about the index entries for their book. One report simply lists the number of index entries per topic; as each topic should have at least one index entry, this is an easy way for writers to ensure that all topics are indexed and to see how extensively. This also helps writers meet the guideline that each PDF book should have no more than two locators for any index entry (primary-secondary-tertiary combination), to ensure that entries are at the best level of detail for users. Another report, shown in Figure 1, flags specific index entries that might contravene the indexing guidelines. For example, the report checks for index entries that differ only by case and primary index entries that contain a comma, as well as other conditions. A primary index entry that contains a comma typically suggests that the entry should be split into primary and secondary entries so it will nest with other secondary entries. Figure 1. Report on inconsistent index entries To ease the authors' job of verifying or addressing deficiencies noted in the reports, the reports provide links to both the content of the source file (View column) and the authors' indexing interface (Title column). The authors' screens are implemented as a Web-based form with two modes of work: updating existing index entries, and adding index entries. 6.1 Authors' update index screen The update screen, shown in Figure 2, is a set of text fields that display the current primary, secondary, and tertiary entries (shown as i1, i2, and i3, respectively). 188 Figure 2. Authors' updating screen When authors update an entry in a text field, the status of that entry automatically switches from No Change to Edit. When authors complete their changes, they click Submit to commit their changes for any flagged entries to the topic database. In Figure 2, for example, the writer can easily see that the third entry isn't done the same way as the first two, and that the third entry under names has an unnecessary for. Authors also have the option of deleting index entries by selecting the Delete radio button for any entries. After they submit their changes, they see a refreshed table with any deletions at the top followed by a list of changes. 6.2 Authors' add entry screen The Add Index Entries screen, shown in Figure 3, enables authors to add index entries easily in a way that maintains consistency with the entries already in the database. To add a new index entry for a topic, the author begins by clicking the initial character for the new index entry from a drop-down box. Figure 3. Selecting the initial character of a new index entry The author then selects the primary entry (the i1 entry) from the drop-down box that has been dynamically added to the form. This box lists all the primary entries that start with the initial character selected. When an author picks a primary entry, as in Figure 4, it is automatically copied into the i1 entry text box. The author then drills down through the available secondary and tertiary entries, picking the relevant entries. At any time in the process of adding a new index entry, authors can manually enter or change the primary, secondary, or tertiary entries in the text boxes if the existing index entries in the database do not suit their needs. This approach provides a flexible way to limit, but not rigidly control, authors' indexing vocabulary. The next builds of the individual book or deliverable, the HTML master index, and the PDF master index reflect any new, deleted, or updated index entries. Figure 4. Selecting the primary entry from the existing list THE EDITOR EXPERIENCE While the author interface focuses on a book-by-book (or deliverable-by-deliverable) view of the index, the role of the editors in the indexing effort is to improve the quality of the index across the entire library. Their primary focus is on the master index in HTML. During the editing phase, a new master index in HTML was generated every two hours to enable the editors to quickly view the results of their work. The editors' indexing interface reflects their needs and approach to the task of editing the master index. Editors begin by finding a problem index entry in the master index using a separate browser window. Then, they use a string or substring of the primary index entry, as shown in Figure 5, to bring up a listing of all primary index entries beginning with that string. The search is not case-sensitive, so any variations in capitalization are included in the search results. 189 Figure 5. Editors' screen for finding a primary index entry The primary entry search yields matching results from across the entire library, so the editor can quickly address inconsistencies between two primary entries and rationalize secondary and tertiary entries. The list of search results appears in a screen similar to the authors' Show/Update/Add Index Entries screen, as shown in Figure 6. Figure 6. Editors' indexing screen for updating entries When an editor submits a change, the editors' indexing interface refreshes with the results, making it easy to confirm that the expected change was made. The next builds of the affected individual books or deliverables, the HTML master index, and the PDF master index reflect any new, deleted, or changed index entries. USER EXPERIENCE HTML Several alternative presentation formats were considered for the HTML version of the master index. These included the format used by the individual book indexes in the previous version of the documentation, third-party formats, and an internally developed format. 8.1 Previous presentation format Previous versions of the HTML indexes generated by the project's standard tools were created for individual books. These indexes were presented as nested, unordered lists of index terms. Links to the HTML pages were displayed as arbitrary four-digit integers, as shown in Figure 7. If an index entry pointed to more than one location in the HTML book, the links were displayed as arbitrary integers separated by commas. Figure 7. Previous index presentation format One of the disadvantages of this presentation format is that users have no criteria by which to choose among the links that might be presented for a single index entry. The arbitrary integers might even give users the impression of being some sort of relevance ranking. Another disadvantage is that, for a master index composed of approximately 40 books, the large number of primary, secondary, and tertiary index entries makes the index very difficult to scan in an online medium. 8.2 Third-party formats Some existing systems that support indexes, such as Sun JavaHelpTM, limit index entries to a one-to-one mapping between each unique index entry and a corresponding locator. This was not an acceptable limitation for the project's large documentation set; each unique index entry needed to be able to map to multiple locators. Other systems, such as Microsoft HTML Help, do support multiple locators for a single index entry through the use of a pop-up window but do not support the operating systems supported by the project. Oracle Help for Java and Oracle Help for the Web (http://otn.oracle.com/tech/java/help/content.html) can display multiple topics per index entry using the topic titles, but they are currently limited to two levels of index entries. Oracle Help for the Web enables the user to type the first few characters of a primary entry to jump ahead to that section of the index, but requires the user to click on the index entry to display any links to topics. We decided to display the index in the content pane of a help system partly to enable portability. If we later decide to deliver our information in a different help system such as JavaHelp, the Eclipse help system, or Windows Help we can avoid any limitations in the help system's built-in support for indexes. Trying to drop 24,000 index entries into a navigation pane simply will not deliver acceptable performance. Assuming that we continue to deliver HTML-based documentation in future versions of the 190 product, this approach will provide users of future versions of the product with a consistent means of accessing the master index. 8.3 Internally developed format The solution for this project was to provide an expanding, collapsing master index, as shown in Figure 8. To assist in scanning, the HTML master index presents only the primary entries at first, but enables users to drill down to secondary and tertiary entries to find exactly the information they need. An index entry that maps to multiple topics displays the locators as a further nested list of links. To provide users with criteria by which they can judge which of several locators meets their needs, each locator is shown as the title of the corresponding topic, as illustrated below. Figure 8. Current index presentation format Nielsen's 1997 article "The Need for Speed" suggests that 34K is the maximum size of an optimal Web page for a ten-second response time for a dial-up connection to the Web. However, internal surveys of our users indicated that most users of our product access the documentation either from a local workstation or from a web server within their company's intranet. We wanted to give users the additional context provided by full topic titles, so we divided the content into multiple files. Given our typical user scenarios, we decided that we could allow somewhat larger files than recommended by Nielsen and divided the index by initial letter into 27 separate files, including one file for all index entries beginning with non-alphabetic characters. The resulting average size of the set of index files is 100K, the median size of the set of index files is 60K, and the largest index file is 400K. The average index file loads and displays in less than one second from an intranet or local workstation, while the largest index file takes just over three seconds to load and display. These times are within Nielsen's maximum threshold for web usability. While the longest delays fall outside the threshold of optimal usability response times of less than one second, we feel that the slightly increased initial load time is balanced by the ease of scanning an index file that contains all of the entries for an initial character. When we collected the complete set of 24,000 index entries in a single HTML file, the file size was over two megabytes and some browsers were incapable of displaying the file. The current index presentation format uses only two images, each smaller than one kilobyte, which are downloaded by the browser only once. Approximately five percent of the total size of the HTML index represents the JavaScript and CSS code that makes the index accessible by keyboard navigation. The majority of the content is due to the topic titles, which average 29 characters per title. Index entries average 13 characters per entry. Initial usability sessions on the current index presentation format indicate that users understand how to work with the expanding and collapsing lists, prefer this format to our previous index format, and find the performance of the index from an intranet or from a locally workstation acceptable. However, as these sessions drew on the experiences of only four users, we recognize the need to do further research to confirm the validity of this presentation format. USER EXPERIENCE PDF In PDF, the individual book indexes look like regular book indexes; that is, they display the index entries and one or more page numbers for those index entries. The decision to associate index entries with entire topics has the detrimental effect of forcing all index entries for a given topic to point to the beginning of that topic. For topics that are longer than a page, users might have to browse through subsequent pages to find the information represented by the index entry in which they are interested. However, length should not be an issue for most topics, as authors are encouraged to keep topics short to ensure good online display. The exception is some reference topics: a topic covering the syntax of a single command might be over 20 pages, and a user might legitimately be interested in just one of the optional arguments for that command. In future iterations of the topic database we hope to enable a more granular indexing strategy to address this shortcoming. We decided to optimize our indexing effort for the master index (both in HTML and PDF), rather than for individual book indexes, so we had to relax some traditional indexing guidelines. One decision we made was to allow individual book indexes to contain primary entries with only a single secondary entry. Rather than concatenating such entries with a comma, we decided, for the sake of the master index, to keep these entries as primary and secondary entries. The following example demonstrates how an individual book index is normally edited to concatenate the primary and secondary entries: SQL statements, overview 453 However, to optimize such entries for the master index, the individual book index must contain distinct primary entries with a lone secondary entry, as in the following example: SQL statements overview 453 We felt that this decline in usability of the individual book indexes would be recouped by reducing the number of comma-spliced entries in the master index, as shown in the following example from the PDF version (the identifiers in front of the page numbers are short forms of the book names): 191 SQL statements, issuing ADGv1: 238 SQL statements, overview ADMINv1: 453 The master index instead presents the optimal arrangement of nested primary and secondary entries: SQL statements issuing ADGv1: 238 overview ADMINv1: 453 One indexing decision driven by the requirements of PDF that posed a potential compromise to the quality of the HTML master index instead improved the master index. It was necessary to artificially subdivide extremely long index entries into separate primary and secondary entries in PDF because the books display the index in a three-column format. A limitation of the PDF transform technology used by the authoring tool is that extremely long index entries that contain no spaces, such as the API keyword SQL_DESC_DATETIME_INTERVAL_ PRECISION , bleed over column boundaries rather than wrap cleanly. In those cases where many similar entries also start with the same prefix, the prefix was turned into a primary entry and the remainder was indexed as a secondary entry. This indexing decision reduces the number of primary entries in the master index, making it much easier for the user to scan the primary entries in the collapsed HTML master index. So the preceding example becomes one of many index entries with the primary entry "SQL_DESC": SQL_DESC_ BIND_OFFSET_PRT DATETIME_INTERVAL_PRECISION DISPLAY_SIZE TECHNOLOGY The HTML master index is implemented in HTML Version 4.0 using Cascading Style Sheets level 2 (CSS2 at http://www.w3.org/Style/CSS/), Document Object Model (DOM at http://www.w3.org/DOM/), and ECMAScript 262 (http://www.ecma.ch/ecma1/stand/ecma-262.htm) to enable the expanding and collapsing behavior. Microsoft Internet Explorer 5.0 and higher and Netscape 6 and higher support the expanding and collapsing behavior. Other browsers degrade gracefully to display nested unordered lists of the index terms and associated topics. The topic database, known as "Dobalina," is implemented with a blend of open-source and proprietary products. The relational database is IBM DB2 Version 7.2 (http://www.software.ibm.com/data/db2/udb), running on a reasonably powerful server, but it could fairly easily be replaced by an open-source database such as MySQL or PostgreSQL. The source format for documentation is SGML (produced with a tool built on ArborText using a proprietary IBMIDDOC DTD), which enables us to dynamically generate index entries as text entities. XML or some sort of manipulated HTML would also fit easily into this model. The project uses the Apache Web server (http://httpd.apache.org) to display the Web scripting front end implemented in PHP (PHP: Hypertext Preprocessor, see http://php.apache.org). PHP also creates the auxiliary source files used to generate the PDF books. The HTML master index is generated by a Perl script (http://www.perl.org), which connects to the database through the Perl database interface module (http://dbi.perl.org). Most of the initial processing of the documentation source files was also performed with Perl scripts. FUTURE DIRECTIONS The index improvement process for such a large information set is planned over several phases, as shown in Table 1. In this project, we are now planning Phase 3. Table 1. Phases of the indexing effort Phase 1: Recognize the problem and build internal support Reassess master index quality problems Analyze information retrieval needs Experiment with scripts Build indexing requirements Develop indexing guidelines Collect user feedback Phase 2: Create first version for writers and customers Create the topic database Develop indexing interfaces Design the presentation of the master index and do initial user testing Establish flexible controlled vocabulary process and start adding see references Train authors in indexing and guidelines Edit master index (focus on primary entries) Conduct initial user testing of index presentation format Phase 3: Continue refining the master index Provide multiple entry points to the index in the product Implement See also references Review consistency of primary entries Edit master index (focus on subentries) Add more syntax and reference entries Implement and apply further reporting features Promote internal use and gather feedback on index content to refine user orientation of entries and to "fill holes" Do user testing Reuse user-centered design scenarios Phase 4: Ensure completeness of contributing content areas Continue to develop reporting features to improve overall consistency Ensure completeness of concept and task entries Edit PDF indexes to help ensure that content is adequately indexed Work with each small writing teams to improve index coverage Analyze problems with PDF and master indexes in other languages Phase 5: Maintain and continue improving index Establish iterative maintenance process Continue to improve presentation, technology, and content Build consensus to improve quality of localized indexes This table shows our actual process and not necessarily the best possible task flow. For example, ideally, See and See also references would be fully implemented in phase 2. 192 One area for future development is the creation of additional reports and scripts. For example, a regular report could identify all the new entries that were not part of the last published index for special editing attention. We are currently augmenting our lists of see references for synonyms, competitive terms, and obsolete terms. We plan to take full advantage of the capabilities of our relational database to create mappings between deprecated or less acceptable terms and acceptable terms, so that any PDF book with an index entry for a deprecated term automatically includes cross-references that lead to the acceptable term. A new indexing screen for authors, with some of the function of the editors' indexing screen, will facilitate improvements of the PDF indexes. We've put a lot of effort into making what's already in the index more consistent, usable, and predictable. But one of the biggest problems is to fill the remaining content holes, that is, what's not yet indexed. Customer, developer, and service feedback indicates the need to improve the granularity of indexing reference topics that document syntax. We will also ensure that the index provides easy access to information needed to implement business and customer scenarios developed by the user-centered design and development teams, and continue to develop the usability of the index interfaces in the product. In Phase 4, we plan to work with each small writing team on specific ways to improve the retrievability of their information. REFERENCES [1] Nielsen, Jakob, "The Need for Speed," 1997; available at http://www.useit.com/alertbox/9703a.html [2] Wright, Jan C., "Single-Source Indexing," Proceedings of the 19 th Annual International Conference on Systems Documentation, October 2001; available at http://www.portal.acm.org 193

book indexes;Information retrieval methods;Massive Master;drop-down selection of terms;Indexing;SQL data type;Search;primary index entry;Internally developed format;automation;indexing problems;Human factors;HTML master index;Online information;Navigation

63

Data Mining in Metric Space: An Empirical Analysis of Supervised Learning Performance Criteria

Many criteria can be used to evaluate the performance of supervised learning. Different criteria are appropriate in different settings, and it is not always clear which criteria to use. A further complication is that learning methods that perform well on one criterion may not perform well on other criteria. For example, SVMs and boosting are designed to optimize accuracy, whereas neural nets typically optimize squared error or cross entropy. We conducted an empirical study using a variety of learning methods (SVMs, neural nets, k-nearest neighbor, bagged and boosted trees, and boosted stumps) to compare nine boolean classification performance metrics: Accuracy, Lift, F-Score, Area under the ROC Curve, Average Precision, Precision/Recall Break-Even Point, Squared Error, Cross Entropy, and Probability Calibration. Multidimensional scaling (MDS) shows that these metrics span a low dimensional manifold. The three metrics that are appropriate when predictions are interpreted as probabilities: squared error, cross entropy, and calibration, lay in one part of metric space far away from metrics that depend on the relative order of the predicted values: ROC area, average precision, break-even point, and lift. In between them fall two metrics that depend on comparing predictions to a threshold: accuracy and F-score. As expected, maximum margin methods such as SVMs and boosted trees have excellent performance on metrics like accuracy , but perform poorly on probability metrics such as squared error. What was not expected was that the margin methods have excellent performance on ordering metrics such as ROC area and average precision. We introduce a new metric, SAR, that combines squared error, accuracy, and ROC area into one metric. MDS and correlation analysis shows that SAR is centrally located and correlates well with other metrics, suggesting that it is a good general purpose metric to use when more specific criteria are not known.

INTRODUCTION In supervised learning, finding a model that could predict the true underlying probability for each test case would be optimal. We refer to such an ideal model as the One True Model. Any reasonable performance metric should be optimized (in expectation, at least) by the one true model, and no other model should yield performance better than it. Unfortunately, we usually do not know how to train models to predict the true underlying probabilities. The one true model is not easy to learn. Either the correct parametric model type for the domain is not known, or the training sample is too small for the model parameters to be esti-mated accurately, or there is noise in the data. Typically, all of these problems occur together to varying degrees. Even if magically the one true model were given to us, we would have difficulty selecting it from other less true models. We do not have performance metrics that will reliably assign best performance to the probabilistically true model given finite validation data. In practice, we train models to minimize loss measured via a specific performance metric. Since we don't have metrics that could reliably select the one true model, we must accept the fact that the model(s) we select will necessarily be suboptimal. There may be only one true model, but there are many suboptimal models. There are different ways that suboptimal models can differ from the one true model tradeoffs can be made between different kinds of deviation from the one true model. Different performance metrics reflect these different tradeoffs. For example, ordering metrics such as area under the ROC curve and average precision do not care if the predicted values are near the true probabilities, but depend only on the relative size of the values. Dividing all predictions by ten does not change the ROC curve, and metrics based on the ROC curve are insensitive to this kind of deviation from truth. Metrics such as squared error and cross entropy, however, are greatly affected by scaling the predicted values, but are less affected by small changes in predicted values that might alter the relative ordering but not significantly change the deviation from the target values. Squared error and cross entropy reflect very different tradeoffs than metrics based on the ROC curve. Similarly, metrics such as accuracy depend on how the predicted values fall relative to a threshold. If predicted values are rescaled, accuracy will be unaffected if the threshold also is rescaled. But if small changes to 69 Research Track Paper 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 W2 W1 Max ACC 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 W2 W1 Max AUC 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 W2 W1 Min RMS 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 W2 W1 Min MXE 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 W2 W1 Min CAL 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 W2 W1 Max SAR Figure 1: Level curves for six error metrics: ACC, AUC, RMS, MXE, CAL, SAR for a simple problem. predicted values are made for cases near the threshold, this can have large impact on accuracy. Accuracy reflects yet another tradeoff in how deviation from truth is measured. The one true model, if available, would have (in expectation ) the best accuracy, the best ROC curve, and the best cross entropy, and the different tradeoffs made by these metrics would not be important. But once we accept that we will not be able to find the one true model, and must therefore accept suboptimal models, the different tradeoffs made by different performance metrics become interesting and important . Unfortunately, little is known about how different performance metrics compare to each other. In this paper we present results from an empirical analysis of nine widely used performance metrics. We perform this empirical comparison using models trained with seven learning algorithms: SVMs, neural nets, k-nearest neighbor , bagged and boosted trees, and boosted stumps. We use multidimensional scaling (MDS) and correlation analysis to interpret the results. We also examine which learning methods perform best on the different metrics. Finally, we introduce a new metric, SAR, that combines squared error, accuracy, and ROC area into a single, robust metric. THE PERFORMANCE METRICS We experiment with nine performance metrics for boolean classification: Accuracy (ACC), Lift (LFT), F-Score (FSC), Area under the ROC Curve (AUC), Average Precision (APR), the Precision/Recall Break-Even Point (BEP), Root Mean Squared Error (RMS), Mean Cross Entropy (MXE), and Probability Calibration (CAL). Definitions for each of the metrics can be found in Appendix A. Figure 1 shows level curves for six of the ten performance metrics for a model with only two parameters (W 1 and W 2) trained on a simple synthetic binary problem. Peak performance in the first two plots occurs along a ridge in weight space. In the other four plots peak performance is indicated by solid dots. Peak performance for some metrics nearly coincide: RMS and MXE peak at nearly the same model weights. But other metrics peak in different places: CAL has a local optimum near the optima for RMS and MXE, but its global optimum is in a different place. Also, the ridges for optimal ACC and optimal AUC do not align, and the ridges do not cross the optima for the other four metrics. Optimizing to each of these metrics yields different models, each representing different tradeoffs in the kinds of errors the models make. Which of these tradeoffs is best depends on the problem, the learning algorithm, and how the model predictions ultimately will be used. We originally divided the nine metrics into three groups: threshold metrics, ordering/rank metrics, and probability metrics. The three threshold metrics are accuracy (ACC), F-score (FSC) and lift (LFT). F-score is the harmonic mean of precision and recall at some threshold. Lift measures the true positive rate in the fraction of cases that fall above threshold. (See Appendix A for a definition of lift, and [3] for a description of Lift Curves. Lift is the same as precision at some threshold, but scaled so that it can be larger than 1.) Usually ACC and FSC use a fixed threshold. In this paper we use 0.5. With lift, often the threshold is adjusted so that a fixed percent, p, of cases are predicted as positive, the rest falling below threshold. Usually p depends on the problem. For example, in marketing one might want to send fliers to 10% of customers. Here we somewhat arbitrarily set p = 25% for all problems. Note that for all threshold metrics it is not important how close a prediction is too a threshold, only if the predicted value is above or below threshold. The ordering/rank metrics look at predictions differently from the threshold metrics. If cases are ordered by predicted value, the ordering/rank metrics measure how well the ordering ranks positive cases above negative cases. The rank metrics can be viewed as a summary of the performance of a model across all possible thresholds. The rank metrics we use are area under the ROC curve (AUC), average precision (APR), and precision/recall break even point (BEP). See [10] for a discussion of ROC curves from a machine learning perspective. Rank metrics depend only on the ordering of the predictions, not the actual predicted values. If the ordering is preserved it makes no difference if the predicted values range between 0 and 1 or between 0.29 and 0.31. Although we group Lift with the threshold metrics, and BEP with the ordering metrics, BEP and Lift are similar to each other in some respects. Lift is directly proportional to BEP if Lift is calculated at p equal to the proportion of positives in the data set. This threshold also is the break-even point where precision equals recall. BEP and Lift are similar to the ordering metrics because the threshold depends implicitly on the ordering, but also are similar to the threshold metrics because neither is sensitive to the orderings on either side of the threshold once that threshold has been defined. Results presented later suggest that both Lift and BEP are more similar to the ordering metrics than to the threshold metrics. The three probability metrics depend on the predicted values , not on how the values fall relative to a threshold or relative to each other. The probability metrics are uniquely min-imized (in expectation) when the predicted value for each case coincides with the true probability of that case being positive. The probability metrics we consider are squared error (RMS), cross entropy (MXE) and calibration (CAL). CAL measures the calibration of a model: if a model predicts 0.85 for a large number of cases, about 85% of those cases should prove to be positive if the model is well calibrated. See Appendix A for details of how CAL is calculated. 70 Research Track Paper We also experiment with a new performance metric, SAR, that combines squared error, accuracy, and ROC area into one measure: SAR = (ACC + AU C + (1 - RM S))/3. SAR behaves somewhat differently from ACC, AUC, and RMS alone, and is a robust metric to use when the correct metric is unknown. SAR is discussed further in Section 8. NORMALIZING THE SCORES Performance metrics such as accuracy or squared error have range [0, 1], while others (lift, cross entropy) range from 0 to q where q depends on the data set. For some metrics lower values indicate better performance. For others higher values are better. Metrics such as ROC area have baseline rates that are independent of the data, while others such as accuracy have baseline rates that depend on the data. If baseline accuracy is 0.98, an accuracy of 0.981 probably is not good performance, yet on another problem, if the Bayes optimal rate is 0.60, achieving an accuracy of 0.59 might be excellent performance. In order to compare performance metrics in a meaningful way, all the metrics need to be placed on a similar scale. One way to do this is to scale the performances for each problem and metric from 0 to 1, where 0 is poor performance, and 1 is good performance. For example, we might place baseline performance at 0, and the Bayes optimal performance at 1. Unfortunately, we cannot estimate the Bayes optimal rate on real problems. Instead, we can use the performance of the best observed model as a proxy for the Bayes optimal performance. We calculate baseline rate as follows: predict p for every case, where p is the percent of positives in the test set. We normalize performances to the range [0, 1], where 0 is baseline and 1 represents best performance. If a model performs worse than baseline, its normalized score will be negative. See Table 1 for an example of normalized scores. The disadvantage of normalized scores is that recovering the raw performances requires knowing the performances that define the top and bottom of the scale, and as new best models are found the top of the scale changes. CAL, the metric we use to measure probability calibration , is unusual in that the baseline model that predicts p for all cases, where p is the percent of positives in the test set, has excellent calibration. (Because of this, measures like CAL typically are not used alone, but are used in conjunction with other measures such as AUC to insure that only models with good discrimination and good calibration are selected. See Figure 1 for a picture of how unusual CAL's error surface is compared with other metrics.) This creates a problem when normalizing CAL scores because the baseline model and Bayes optimal model have similar CAL scores. This does not mean CAL is a poor metric it is effective at distinguishing poorly calibrated models from well calibrated models. We address this problem later in the paper. EXPERIMENTAL DESIGN The goal of this work is to analyze how the ten metrics compare to each other. To do this we train many different kinds of models on seven test problems, and calculate for each test problem the performance of every model on the ten metrics. We train models using seven learning algorithms: Neural Nets (ANN), SVMs, Bagged Decision Trees (BAG-DT), Boosted Decision Trees (BST-DT), Boosted Decision Stumps Table 1: Accuracy on ADULT problem model acc norm score bst-stmp 0.8556 1.0000 bag-dt 0.8534 0.9795 dt 0.8503 0.9494 svm 0.8480 0.9267 bst-dt 0.8464 0.9113 ann 0.8449 0.8974 knn 0.8320 0.7731 baseline 0.7518 0.0000 (BST-STMP), single Decision Trees (DT) and Memory Based Learning (KNN). For each algorithm we train many variants and many parameter settings. For example, we train ten styles of decision trees, neural nets of different sizes, SVMs using many different kernels, etc. A total of 2000 models are trained and tested on each problem. See Appendix B for a description of the parameter settings we use for each learning method. While this strategy won't create every possible model, and won't create a uniform sample of the space of possible models, we feel that this is an adequate sample of the models that often will be trained in practice. For each problem, the 2000 models are trained on the same train set of 4000 points. The performance of each model is measured on the same large test set for each of the ten performance metrics. In order put the performances on the same scale across different metrics and different problems, we transform the raw performance to normalized scores as explained in Section 3. In total, across the seven problems, we have 2000 7 = 14, 000 models and for each model we have it's score on each of the 10 performances metrics. DATA SETS We compare the algorithms on seven binary classification problems. ADULT, COVER TYPE and LETTER are from UCI Repository [1]. ADULT is the only problem that has nominal attributes. For ANNs, SVMs and KNNs we transform nominal attributes to boolean. Each DT, BAG-DT, BST-DT and BST-STMP model is trained twice, once with the transformed attributes and once with the original attributes . COVER TYPE has been converted to a binary problem by treating the largest class as the positive and the rest as negative. We converted LETTER to boolean in two ways. LETTER.p1 treats the letter "O" as positive and the remaining 25 letters as negative, yielding a very unbalanced binary problem. LETTER.p2 uses letters A-M as positives and the rest as negatives, yielding a well balanced problem. HYPER SPECT is the IndianPine92 data set [4] where the difficult class Soybean-mintill is the positive class. SLAC is a problem from collaborators at the Stanford Linear Accelerator and MEDIS is a medical data set. The characteristics of these data sets are summarized in Table 2. Table 2: Description of problems problem #attr train size test size % pos. adult 14/104 4000 35222 25% cover type 54 4000 25000 36% letter.p1 16 4000 14000 3% letter.p2 16 4000 14000 53% medis 63 4000 8199 11% slac 59 4000 25000 50% hyper spect 200 4000 4366 24% 71 Research Track Paper MDS IN METRIC SPACE Training 2000 models on each problem using seven learning algorithms gives us 14,000 models, each of which is eval-uated on ten performance metrics. This gives us 14,000 sample points to compare for each performance metric. We build a 10x14,000 table where lines represent the performance metrics, columns represent the models, and each entry in the table is the score of the model on that metric. For MDS, we treat each row in the table as the coordinate of a point in a 14,000 dimension space. The distance between two metrics is calculated as the Euclidean distance between the two corresponding points in this space. Because the coordinates are strongly correlated, there is no curse-of-dimensionality problem with Euclidean distance in this 14,000 dimensional space. We are more interested in how the metrics compare to each other when models have good performance than when models have poor performance. Because of this, we delete columns representing poorer performing models in order to focus on the "interesting" part of the space where models that have good performance lie. For the analyses reported in this paper we delete models that perform below baseline on any metric (except CAL). Ten metrics permits 10 9/2 = 45 pairwise comparisons. We calculate Euclidean distance between each pair of metrics in the sample space, and then perform multidimensional scaling on these pairwise distances between metrics. MDS is sensitive to how the performance metrics are scaled. The normalized scores described in Section 3 yield well-scaled performances suitable for MDS analysis for most metrics . Unfortunately, as discussed in Section 3, normalized scores do not work well with CAL. Because of this, we perform MDS two ways. In the first, we use normalized scores, but exclude the CAL metric. In the second, we include CAL, but scale performances to mean 0.0 and standard deviation 1.0 instead of using normalized scores. Scaling by standard deviation resolves the problem with CAL for MDS, but is somewhat less intuitive because scores scaled by standard deviation depend on the full distribution of models instead of just the performances that fall at the top and bottom of each scale. Figure 2 shows the MDS stress as a function of the number of dimensions in the MDS (when CAL is included). The ten metrics appear to span an MDS space of about 3 to 5 dimensions. In this section we examine the 2-D MDS plots in some detail. Figure 3 shows two MDS plots for the metrics that result when dimensionality is reduced to two dimensions. The plot on the left is MDS using normalized scores when CAL is excluded. The plot on the right is MDS using standard deviation scaled scores when CAL is included. Both MDS plots show a similar pattern. The metrics appear to form 4-5 somewhat distinct groups. In the upper right hand corner is a group that includes AUC, APR, BEP, LFT, and SAR. The other groups are RMS and MXE, ACC (by itself, or possibly with FSC), FSC (by itself, or possibly with ACC), and CAL (by itself). It is not surprising that squared error and cross entropy form a cluster. Also, presumably because squared error tends to be better behaved than cross entropy, RMS is closer to the other measures than MXE. We are somewhat surprised that RMS is so centrally located in the MDS plots. Perhaps this partially explains why squared error has proved so useful in many applications. 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 1 2 3 4 5 6 7 8 MDS Stress Number of MDS Dimensions Figure 2: MDS stress vs. number of dimensions It is somewhat surprising that accuracy does not appear to correlate strongly with any of the other metrics, except possibly with FSC. ACC does not fall very close to other metrics that use thresholds such as Lift and F-Score, even though F-Score uses the same 0.5 threshold as accuracy in our experiments. (The threshold for Lift is adjusted dynam-ically so that 25% of the cases are predicted as positive.) Accuracy is surprisingly close to RMS, and closer to RMS than to MXE, again suggesting that part of the reason why RMS has been so useful is because of its close relationship to a metric such as ACC that has been so widely used. The most surprising pattern in the MDS plot that includes CAL is that CAL is distant from most other metrics . There appears to be an axis running from CAL at one end to the ordering metrics such as AUC and APR at the other end that forms the largest dimension in the space. This is surprising because one way to achieve excellent ordering is to accurately predict true probabilities, which is measured by the calibration metric. However, one can achieve excellent AUC and APR using predicted values that have extremely poor calibration, yet accurately predict the relative ordering of the cases. The MDS plot suggests that many models which achieve excellent ordering do so without achieving good probabilistic calibration. Closer examination shows that some models such as boosted decision trees yield remarkably good ordering, yet have extremely poor calibration. We believe maximum margin methods such as boosting tradeoff reduced calibration for better margin . See Section 9 for further discussion of this issue. One also can achieve good calibration, yet have poor AUC and APR. For example, decision trees with few leaves may be well calibrated, but the coarse set of values they predict do not provide a basis for good ordering. Figure 4 shows 2-D MDS plots for six of the seven test problems. The seventh plot is similar and is omitted to save space. (The omitted plot is one of the two LETTER problems.) Although there are variations between the plots, the 2-D MDS plots for the seven problems are remarkably consistent given that these are different test problems. The consistency between the seven MDS plots suggests that we have an adequate sample size of models to reliably detect relationships between the metrics. Metrics such as ACC, FSC, and LFT seem to move around with respect to each other in these plots. This may be because they have different sensi-72 Research Track Paper acc fsc lft auc apr bep rms mxe sar Dim 2 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 Dim 1 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 acc fsc lft auc apr bep rms mxe cal sar Dim 2 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 Dim 1 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 Figure 3: 2D MDS plot using normalized scores (left) and standard deviation scaling (right). tivities to the ratio of positives to negatives in the data sets. For example, BEP is proprtional to LFT (and thus behaves similarly) when the percentage of positives in the dataset equals the fraction predicted above threshold (25% in this paper). Other than this, we have not been able to correlate differences we see in the individual plots with characteristics of the problems that might explain those differences, and currently believe that the MDS plots that combine all seven problems in Figure 3 represents an accurate summary of the relationships between metrics. Note that this does not mean that the performance of the different learning algorithms exhibits the same pattern on these test problems (in fact they are very different), only that the relationships between the ten metrics appear to be similar across the test problems when all the learning algorithms are considered at one time. CORRELATION ANALYSIS As with the MDS analysis in the previous section, we used each of the ten performance metrics to measure the performance of the 2000 models trained with the different learning methods on each of the seven test problems. In this section we use correlation analysis on these models to compare metrics instead of MDS. Again, to make the correlation analysis easier to interpret , we first scale performances to the range [0, 1] so that the best performance we observed with that metric on each problem with any of the learning methods is performance 1, and baseline performance with that metric and data set is performance 0. This eliminates the inverse correlation between measures such as accuracy and squared error, and normalizes the scale of each metric. Ten metrics permits 10 9/2 = 45 pairwise correlations. We do these comparisons using both linear correlation (excluding CAL) and rank correlation. The results from the linear and rank correlation analyses are qualitatively similar . We present the results for non-parametric rank correlation because rank correlation makes fewer assumptions about the relationships between the metrics, and because rank correlation is insensitive to how CAL is scaled. Table 3 shows the rank correlation between all pairs of metrics. Each entry in the table is the average rank correlation across the seven test problems. The table is sym-metric and contains only 45 unique pairwise comparisons. We present the full matrix because this makes it easier to scan some comparisons. The final column is the mean of the rank correlations for each metric. This gives a rough idea how correlated each metric is on average to all other metrics. Metrics with pairwise rank correlations near one behave more similarly than those with smaller rank correlations. Ignoring the SAR metric which is discussed in the next section, seven metric pairs have rank correlations above 0.90: 0.96: Lift to ROC Area 0.95: ROC Area to Average Precision 0.93: Accuracy to Break-even Point 0.92: RMS to Cross-Entropy 0.92: Break-Even Point to ROC Area 0.92: Break-Even Point to Average Precision 0.91: Average Precision to Lift We expected AUC and average precision to behave very similarly and thus have high rank correlation. But we are surprised to see that Lift has such high correlation to AUC. Note that because Lift has high correlation to AUC, and AUC has high correlation to average precision, it is not surprising that Lift also has high correlation to average precision . As expected, break-even point is highly correlated with the other two ordering metrics, AUC and average precision. But the high correlation between accuracy and break-even 73 Research Track Paper acc fsc lft auc apr bep rms mxe cal sar -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 COVER_TYPE -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 acc fsc lft auc apr bep rms mxe cal sar -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 ADULT -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 acc fsc lft auc apr bep rms mxe cal sar -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 LETTER.P1 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 acc fsc lft auc apr bep rms mxe cal sar -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 HYPER_SPECT -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 acc fsc lft auc apr bep rms mxe cal sar -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 MEDIS -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 acc fsc lft auc apr bep rms mxe cal sar -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 SLAC -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 Figure 4: 2-D MDS plots for six of the seven test problems. The seventh problem yields a similar plot and is omitted only to save space. The missing plot is for one of the LETTER problems. 74 Research Track Paper Table 3: Average rank correlations between metrics acc fsc lft auc apr bep rms mxe cal sar mean acc 1.00 0.87 0.85 0.88 0.89 0.93 0.87 0.75 0.56 0.92 0.852 fsc 0.87 1.00 0.77 0.81 0.82 0.87 0.79 0.69 0.50 0.84 0.796 lft 0.85 0.77 1.00 0.96 0.91 0.89 0.82 0.73 0.47 0.92 0.832 auc 0.88 0.81 0.96 1.00 0.95 0.92 0.85 0.77 0.51 0.96 0.861 apr 0.89 0.82 0.91 0.95 1.00 0.92 0.86 0.75 0.50 0.93 0.853 bep 0.93 0.87 0.89 0.92 0.92 1.00 0.87 0.75 0.52 0.93 0.860 rms 0.87 0.79 0.82 0.85 0.86 0.87 1.00 0.92 0.79 0.95 0.872 mxe 0.75 0.69 0.73 0.77 0.75 0.75 0.92 1.00 0.81 0.86 0.803 cal 0.56 0.50 0.47 0.51 0.50 0.52 0.79 0.81 1.00 0.65 0.631 sar 0.92 0.84 0.92 0.96 0.93 0.93 0.95 0.86 0.65 1.00 0.896 point is somewhat surprising and we currently do not know how to explain this. The weakest correlations are all between the calibration metric (CAL) and the other metrics. On average, CAL correlates with the other metrics only about 0.63. We are surprised how low the correlation is between probability calibration and other metrics, and are currently looking at other measures of calibration to see if this is true for all of them. acc fsc lft auc apr bep rms mxe cal sar Dim 2 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 Dim 1 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 Figure 5: MDS using rank correlation Figure 5 shows an MDS plot for the metrics when distance between metrics is calculated as 1 - rank correlation, making MDS insensitive to how the metrics are scaled. (Distances based on 1 - rank correlation do not respect the triangle inequality so this is not a proper metric space.) The overall pattern is similar to that observed in the MDS plots in Figure 3. CAL is at one end of the space far from the other metrics. Cross-entropy is closest to RMS, though not as close as in the other plots. Cross-entropy and RMS have high rank correlation, but because cross-entropy has lower rank-correlation to other most metrics than RMS, it is pushed far from RMS which is close to other metrics in the MDS plot. APR and AUC are at the other end of the space farthest from CAL. FSC is in the upper left side of the space. ACC and RMS are near the center of the space. SAR A GENERAL PURPOSE METRIC When applying supervised learning to data, a decision must be made about what metric to train to and what metric to use for model selection. Often the learning algorithm dictates what metrics can be used for training, e.g. it is difficult to train a neural net for metrics other than RMS or MXE. But there usually is much more freedom when selecting the metric to use for model selection, i.e. the metric used to pick the best learning algorithm and the best parameters for that algorithm. If the correct metric for the problem is known, model selection probably should be done using that metric even if the learning algorithm cannot be trained to it. What should be done when the correct metric is not known? The MDS plots and correlation analysis suggest that RMS is remarkably well correlated with the other measures, and thus might serve as a good general purpose metric to use when a more specific optimization criterion is not known. We wondered if we could devise a new metric more centrally located than RMS and with better correlation to the other metrics. Rather than devise a completely new metric , we tried averaging several of the well behaved metrics into a new metric that might be more robust than each one individually. SAR combines Squared error, Accuracy, and ROC area into one measure: SAR = (ACC + AU C + (1 RM S))/3. We chose these metrics for SAR for three reasons : 1. we wanted to select one metric from each metric group: the threshold metrics, the ordering metrics, and the probability metrics 2. ACC, AUC, and RMS seemed to be the most popular metric from each of these groups, respectively 3. these three metrics are well correlated to the other metrics in their groups, and in the MDS plots lie closest to the other metrics in their groups As can be seen from the MDS plots and in the tables, SAR behaves differently from ACC, AUC, and RMS alone. In Table 3 SAR has higher mean rank correlation to other metrics than any other metric. In the MDS plots, SAR tends to be more consistently centrally located than other metrics. And in Table 4 it is the metric that best reflects the ordering by mean performance of the seven learning methods. These results suggest that of the ten metrics we exam-ined , SAR is the metric that on average is most correlated with the other metrics, both separately, and in groups. SAR is even more representative than RMS (though RMS also is 75 Research Track Paper Table 4: Normalized scores for each learning algorithm by metric (average over seven problems) model acc fsc lft auc apr bep rms mxe cal mean sar ann 0.9399 0.9486 0.9623 0.9722 0.9538 0.9632 0.9043 0.9009 0.9963 0.9491 0.9516 svm 0.9010 0.9515 0.9642 0.9688 0.9523 0.9635 0.9024 0.9041 0.9881 0.9440 0.9524 bag-dt 0.8796 0.8986 0.9450 0.9765 0.9577 0.9464 0.8763 0.9087 0.9800 0.9299 0.9470 bst-dt 0.9506 0.9443 0.9843 0.9866 0.9779 0.9858 0.6400 0.6427 0.9399 0.8947 0.9171 knn 0.8127 0.9042 0.9248 0.9481 0.9052 0.9252 0.7954 0.7754 0.9871 0.8865 0.9012 dt 0.6737 0.8621 0.8393 0.8897 0.8169 0.8403 0.6292 0.6748 0.9731 0.7999 0.8160 bst-stmp 0.7929 0.8265 0.8721 0.9291 0.8799 0.8724 0.3181 0.3013 0.9477 0.7489 0.6966 very good). In an experiment where SAR was used for model selection, SAR outperformed eight of the nine metrics in selecting the models with the best overall, and tied with RMS. We believe our results suggest that SAR is a robust combination of three popular metrics that may bey appropriate when the correct metric to use is not known, though the benefit of SAR over RMS is modest at best. Attempts to make SAR better by optimizing the weights given to ACC, AUC, and RMS in the SAR average did not significantly improve SAR compared to equal weights for the three metrics. We are very impressed at how well behaved RMS alone is and are currently working to devise a better SAR-like metric that yields more improvement over RMS alone. PERFORMANCES BY METRIC Table 4 shows the normalized performance of each learning algorithm on the nine metrics. (CAL is scaled so that the minimum observed CAL score is 0.0 and the maximum observed CAL score is 1.0) For each test problem we find the best parameter settings for each learning algorithm and compute it's normalized score. Each entry in the table averages these scores across the seven problems. The last two columns are the mean normalized scores over the nine metrics , and the SAR performance. Higher scores indicate better performance. The models in the table are ordered by mean overall performance. We have written a separate paper to compare the performance of the learning methods to each other on these metrics, but there are a few interesting relationships between learning algorithms and metrics that are worth discussing in the context of this paper. Overall, the best performing models are neural nets, SVMs, and bagged trees. Surprisingly, neural nets outperform all other model types if one averages over the nine metrics. ANNs appear to be excellent general purpose learning methods . This is not to say that ANNs are the best learning algorithm they only win on RMS and CAL, but because they rarely perform poorly on any problem or metric, they have excellent overall performance. The SVMs perform almost as well as ANNs. Note that SVM predictions on [-, +] are not suitable for measures like cross entropy, calibration, and squared error. SVMs do well on these metrics because we use Platt's method [8] to transform SVM predictions to calibrated probabilities. Like neural nets, SVMs appear to be a safe, general purpose, high performance learning method once their predictions have been calibrated by a method such as Platt scaling. Although single decision trees perform poorly, bagged trees perform nearly as well as neural nets and SVMs. Bagging improves decision tree performance on all metrics, and yields particularly large improvements on the probability metrics. Like neural nets and SVMs, bagged trees appear to be a safe, general purpose, high performance learning method. Boosted trees outperform all other learning methods on ACC, LFT, ROC, APR, and BEP. Boosting wins 2 of 3 threshold metrics and 3 of 3 rank metrics, but performs poorly on the probability metrics: squared error, cross entropy , and calibration. Maximum margin methods such as boosted trees yield poorly calibrated probabilities. (SVMs perform well on these because Platt scaling "undoes" the maximum margin.) Overall, boosting wins 5 of the 6 metrics for which it is well suited, and would easily be the top performing learning method if we consider only the 6 threshold and ordering metrics. The KNN methods were not competitive with the better algorithms, but might done better with larger train sets. Single decision trees also did not perform as well as most other methods, probably because recursive partitioning runs out of data quickly with 4k train sets, and because small trees are not good at predicting probabilities [9]. We tested many different kinds of decision trees, including smoothed unpruned trees, and then picked the best, so the poor performance of trees here is not due to any one tree type being inferior, but because all of the many tree types we tested did not perform as well as other methods. Interestingly, boosting stump models does not perform as well as boosting full decision trees. Boosted stumps do outperform single trees on 5 of the 6 threshold and rank metrics. Their last-place ranking below decision trees is due to their extremely poor performance on the three probability measures. RELATED WORK There is not a large literature comparing performance metrics. The closest work to ours is by Flach [7]. In this work Flach uses the ROC space to understand and compare different metrics. He analyzes accuracy, precision, weighted relative accuracy and several decision tree splitting criteria. The STATLOG project [6] performed a large scale empirical evaluation of a number of learning algorithms in 1995. STATLOG compared the performance of the different algorithms , and also did an analysis of how the predictions made by the algorithms compared to each other. STATLOG, however , did not compare performance using different metrics. DISCUSSION AND CONCLUSIONS Our analysis allows us to draw a variety of conclusions which we summarize here. If the goal is to maximize accuracy , but the model needs a continuous performance metric (e.g. using backpropagation to train a neural net), it probably is better to train the model using squared error instead of cross entropy because squared error sits closer to accuracy in metric space. This result is surprising since cross entropy is the theoretically preferred loss function for binary classification . We suspect cross entropy is not as robust as squared 76 Research Track Paper error on real data sets because real data sometimes contains class noise that cross entropy is very sensitive to. Squared error is a remarkably robust performance metric that has higher average correlation to the other metrics than any other metric except SAR. Squared error appears to be an excellent general purpose metric. Many models achieve excellent performance on the ordering metrics AUC, APR, and BEP without making predictions that yield good probabilities. For example, the k-nearest neighbor models with the best ROC performance use values of K that are so large that most of the predictions are close to p, the fraction of positives in the data. This yields predictions that are poor when viewed as probabilities , yet small differences between these predicted values are sufficient to provide for good ordering. As expected, maximum margin methods such as boosting and SVMs yield excellent performance on metrics such as accuracy for which they are designed. Surprisingly, however, the maximum margin methods also yield excellent performance on the ordering metrics. We had not expected that maximizing distances to decision boundaries would provide a good basis for ordering cases that fall far from those boundaries . Although boosted trees perform well on accuracy and ROC, they perform poorly on probability metrics such as squared error and cross entropy. This poor performance on probability metrics is a consequence of boosting being a maximum margin method. SVMs do not exhibit this problem because we scale SVM predictions with Platt's method; Lin-early scaling SVM predictions to [0, 1] does not work well. Neural nets trained with backpropagation have excellent overall performance because, unlike boosting, they perform well on all metrics including the probability metrics RMS, MXE, and CAL. We believe part of the reason why the neural nets perform so well is that they were trained with backpropagation on squared error, and as we have seen squared error is an excellent metric. The three ordering metrics, AUC, APR, and BEP, cluster close in metric space and exhibit strong pairwise correlations . These metrics clearly are similar to each other and somewhat interchangeable. We originally grouped LFT with the threshold metrics ACC and FSC, but the results suggest that LFT behaves more like BEP, an ordering metric. We now would group LFT with BEP in the ordering metrics along with AUC and APR. The metric space for the ten metrics has three or more significant dimensions. The ten metrics do not all measure the same thing. Different performance metrics yield different tradeoffs that are appropriate in different settings. No one metric does it all, and the metric optimized to or used for model selection does matter. The SAR metric that combines accuracy, ROC area, and squared error appears to be a good, general purpose metric, but RMS is so good that SAR may not provide much benefit over using RMS alone. We hope that additional research in this area will enable us to design better metrics, and will shed more light on which metrics are most appropriate to use in different settings. ACKNOWLEDGMENTS Thanks to Geoff Crew and Alex Ksikes for help running some of the experiments. Thanks to the creators of XGVIS and XGOBI for the interactive MDS software used to generate the MDS plots. Thanks to collaborators at Stanford Linear Accelerator for the SLAC data, and to Tony Gualtieri at NASA Goddard for help with the Indian Pines data. REFERENCES [1] C. Blake and C. Merz. UCI repository of machine learning databases, 1998. [2] M. DeGroot and S. Fienberg. The comparison and evaluation of forecasters. Statistician, 32(1):1222, 1982. [3] P. Giudici. Applied Data Mining. John Wiley and Sons, New York, 2003. [4] A. Gualtieri, S. R. Chettri, R. Cromp, and L. Johnson. Support vector machine classifiers as applied to aviris data. In Proc. Eighth JPL Airborne Geoscience Workshop, 1999. [5] T. Joachims. Making large-scale svm learning practical. In Advances in Kernel Methods, 1999. [6] R. King, C. Feng, and A. Shutherland. Statlog: comparison of classification algorithms on large real-world problems. Applied Artificial Intelligence, 9(3):259287, May/June 1995. [7] P.A.Flach. The geometry of roc space: understanding machine learning metrics through roc isometrics. In Proc. 20th International Conference on Machine Learning (ICML'03), pages 194201. AAAI Press, January 2003. [8] J. Platt. Probabilistic outputs for support vector machines and comparison to regularized likelihood methods. In A. Smola, P. Bartlett, B. Schoelkopf, and D. Schuurmans, editors, Advances in Large Margin Classifiers, pages 6174, 1999. [9] F. Provost and P. Domingos. Tree induction for probability-based rankings. Machine Learning, 52(3), 2003. [10] F. J. Provost and T. Fawcett. Analysis and visualization of classifier performance: Comparison under imprecise class and cost distributions. In Knowledge Discovery and Data Mining, pages 4348, 1997. APPENDIX A. PERFORMANCE METRICS accuracy: probably the most widely used performance metric in Machine Learning. It is defined as the proportion of correct predictions the classifier makes relative to the size of the dataset. If a classifier has continuous outputs (e.g. neural nets), a threshold is set and everything above this threshold is predicted to be a positive. root-mean-squared-error (RMSE): widely used in regression , it measures how much predictions deviate from the true targets. 1 RMSE is defined as: RM SE = 1 N (P red(C) - T rue(C)) 2 (1) mean cross entropy (MXE): is used in the probabilistic setting when interested in predicting the probability 1 Root-mean-squared error is applicable to binary classification settings where the classifier outputs predictions on [0, 1] that are compared with the true target labels on {0, 1}. 77 Research Track Paper that an example is positive (1). It can be proven that in this setting minimizing the cross entropy gives the maximum likelihood hypothesis. mean cross entropy is defined as: M XE = 1 N (T rue(C) ln(P red(C)) + (1 - T rue(C)) ln(1 - P red(C))) (2) (The assumptions are that P red(C) [0, 1] and T rue(C) {0, 1}) receiver operating characteristic (ROC): has it's roots in WWII in the early days of radar where it was difficult to distinguish between true positives and false positives . ROC is a plot of sensitivity vs. (1-specificity) for all possible thresholds. Sensitivity is the defined as P (P red = positive|T rue = positive) and is approxi-mated by the fraction of true positives that are predicted as positive (this is the same as recall). Specificity is P (P red = negative|T rue = negative). It is approx-imated by the fraction of true negatives predicted as negatives. AUC, the area under the ROC curve, is used as a summary statistic. ROC has a number of nice properties that make it more principled than similar measures such as average precision. AUC is widely used in fields such as medicine, and recently has become more popular in the Machine Learning community. lift: often used in marketing analysis, Lift measures how much better a classifier is at predicting positives than a baseline classifier that randomly predicts positives (at the same rate observed for positives in the data). The definition is: LIF T = %of true positives above the threshold %of dataset above the threshold (3) Usually the threshold is set so that a fixed percentage of the dataset is classified as positive. For example, suppose a marketing agent wants to send advertising to potential clients, but can only afford to send ads to 10% of the population. A classifier is trained to predict how likely a client is to respond to the advertisement, and the ads are sent to the 10% of the population predicted most likely to respond. A classifier with optimal lift will get as many clients as possible that will respond to the advertisement in this set. precision and recall : These measures are widely used in Information Retrieval. Precision is the fraction of examples predicted as positive that are actually positive. Recall is the fraction of the true positives that are predicted as positives. These measures are trivially maxi-mized by not predicting anything, or predicting everything , respectively, as positive. Because of this these measures often are used together. There are different ways to combine these measures as described by the next 4 metrics. precision-recall F-score: for a given threshold, the F-score is the harmonic mean of the precision and recall at that threshold. precision at a recall level: as the name suggests, set the threshold such that you have a given recall and the precision for this threshold is computed. precision-recall break-even point: is defined as the precision at the point (threshold value) where precision and recall are equal. average precision: usually is computed as the average of the precisions at eleven evenly spaced recall levels. CAL is based on reliability diagrams [2]. It is calculated as follows: order all cases by their predicted value, and put cases 1-100 in the same bin. Calculate the percentage of these cases that are true positives. This approximates the true probability that these cases are positive. Then calculate the mean prediction for these cases. The absolute value of the difference between the observed frequency and the mean prediction is the calibration error for this bin. Now take cases 2-101, 3-102, .... and compute the errors in the same way for each of these bins. CAL is the mean of these binned calibration errors. B. PARAMETER SETTINGS We use the following parameter settings and algorithm variations for the seven learning methods: KNN: we use 26 values of K ranging from K = 1 to K = |trainset|. We use KNN with Euclidean distance and Euclidean distance weighted by gain ratio. We also use distance weighted KNN, and locally weighted averaging. The kernel widths for locally weighted averaging vary from 2 0 to 2 10 times the minimum distance between any two points in the train set. ANN: we train nets with gradient descent backprop and vary the number of hidden units {1, 2, 4, 8, 32, 128} and the momentum {0, 0.2, 0.5, 0.9}. We don't use validation sets to do weight decay or early stopping. Instead, for each performance metric, we examine the nets at many different epochs. DT: we vary the splitting criterion, pruning options, and smoothing (Laplacian or Bayesian smoothing). We use all of the tree models in Buntine's IND package: Bayes, ID3, CART, CART0, C4, MML, and SMML. We also generate trees of type C44 (C4 with no pruning), C44BS (C44 with Bayesian smoothing), and MMLLS (MML with Laplacian smoothing). See [9] for a description of C44. BAG-DT: we bag at least 25 trees of each type. With BST-DT we boost each tree type. Boosting can overfit, so we consider boosted DTs after {2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048} steps of boosting. With BST-STMP we use stumps (single level decision trees) with 5 different splitting criteria, each boosted {2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192} steps. SVMs: we use most kernels in SVMLight [5] {linear, polynomial degree 2 & 3, radial with width {0.001, 0.005, 0.01, 0.05, 0.1, 0.5, 1, 2}} and vary the regularization parameter C by factors of ten from 10 -7 to 10 3 . The output range of SVMs is [-, +] instead of [0, 1]. To make the SVM predictions compatible with other models, we use Platt's method to convert SVM outputs to probabilities by fitting them to a sigmoid [8]. Without scaling, SVMs would have poor RMS and it would not be possible to calculate MXE and CAL. 78 Research Track Paper

Lift;Precision;performance metric;ROC;Supervised Learning;supervised learning;squared error;SVMs;pairwise;Recall;algorithmns;Cross Entropy;ordering metric;Euclidean distance;Performance Evaluation;standard deviation;backpropagation;Metrics

64

Database Security Curriculum in InfoSec Program

Database Security course is an important part of the InfoSec curriculum. In many institutions this is not taught as an independent course. Parts of the contents presented in this paper are usually incorporated in other courses such as Network Security. The importance of database security concepts stems from the fact that a compromise of data at rest could expose an organization to a greater security threat than otherwise. Database vulnerabilities exposed recently in several high profile incidents would be a good reason to dedicate a full course to this important topic. In this paper we present key topics such as technologies for database protection, access control, multilevel security, database vulnerabilities and defenses, privacy and legal issues, impact of policies and some well known secure database models.

INTRODUCTION Information Security curriculum is receiving greater attention from many institutions, thanks to the standardization efforts by the Committee on National Security Systems (CNSS). The CNSS members come from the National Security Agency, Department of Defense, and the Department of Homeland Security, among others. The CNSS standardization efforts are based on the Presidential Decision Directive [24] issued in 1998 for training professionals to protect the nation's critical infrastructure. To achieve this goal, CNSS has developed five major standards known as the National Security Telecommunications Information Systems Security Instruction (NSTISSI). The NSTISSI standards are numbered 4011, 4012, 4013, 4014 and 4015 [8]. Additional standards under this sequence are in the offing as well. The relevance of these standards is that they include a vast number of topics that cover the entire gamut of information assurance and database security topics are included in many of these standards. First, we will briefly outline the main content of each of these standards and then move onto the main content of this paper. The 4011 standard covers the information security foundation topics such as wired and wireless communications basics, operations security, transmission security, information security from a policy perspective, cryptography, key management, legal aspects of security, contingency planning and disaster recovery, risk management, trust, auditing, and monitoring. At present, coverage of topics mentioned in this standard is considered essential by CNSS in every InfoSec curriculum. The 4012 standard is primarily aimed at training Designated Approving Authority personnel. A quick look at the following topics would show the relationship of these standards vis--vis database security. The primary topics of this standard include: liabilities, legal issues, security policy, sensitive data access policy, threats, vulnerabilities, incident response, life cycle management, configuration management, and contingency management. The purpose of 4013 standard is to provide a minimum set of topics necessary for certifying Systems Administrators in Information Systems Security. Some of the topics in this category include: development and maintenance of security policies and procedures, education, training and awareness of such policies, development of countermeasures for known attacks as well as development of safeguards. Also, configuration management is an important part of 4013 standard. The standard for training Information Systems Security Officers is 4014. This standard covers topics such as facilities planning, business continuity, and password management, access control policies, laws and regulations related to information security, privacy, encryption standards, intrusion detection, audit tools, and security reviews. The last standard currently in place in this series is numbered 4015. This standard is for training System Certifiers. Among the main topics here are: defining roles and responsibilities for personnel, certification of systems, identifying process 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. Information Security Curriculum Development (InfoSecCD) Conference '05, September 23-24, 2005, Kennesaw, GA, USA. Copyright 2005 ACM 1-59593-261-5...$5.00. 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. Information Security Curriculum Development (InfoSecCD) Conference '05, September 23-24, 2005, Kennesaw, GA, USA. Copyright 2005 ACM 1-59593-261-5/05/0009...$5.00. 79 boundaries, integration, security engineering, and applications security. These five standards have been in place since 1994 and are constantly getting updated by CNSS. INFOSEC FOUNDATION COURSES Traditionally, the following courses are considered as a set of foundation courses: Network Security, Information Security, and Cryptography. Usually these courses are augmented by additional courses such as Operating System Security, Database Security, Secure E-commerce, and Security Management. In our curriculum at the University of Louisville we are offering the three foundation courses listed above and the Database Security course. The main purpose of this paper is to identify several topics that could be included in a Database Security course. In the last quarter of 2004 and the first quarter of 2005, several incidents of theft or loss of data from databases of large organizations have brought to light the vulnerabilities in managing database systems. Every organization depends heavily on databases, both large and small, in order to manage inventory, human resources, and business functions on a day to day basis. Therefore, in order to mitigate risk, every organization must take adequate steps to protect the data that they hold. Issues related to technology as well as policies are important for the protection of data. Such topics form the core of this Database Security course, which we will discuss in greater detail in the remaining sections. INFOSEC AT U OF LOUISVILLE At the University of Louisville (U of L), InfoSec courses are offered in two departments. The Computer Information Systems (CIS) department in the College of Business offers an undergraduate concentration in InfoSec [36]. The Computer Science department in the college of Engineering offers graduate courses in InfoSec at the masters and doctoral levels. Database security course is offered as the second course in database, the first course being the standard database design and management course. Students taking the database security course are either juniors or seniors and are expected to have experience with one of the mainframe commercial databases such as Oracle or SQL Server 2000. The major course objectives were for students to: Learn the fundamental concepts in database security Understand how access controls work in a database Learn to develop and manage secure database architectures Be familiar with the laws governing computer privacy Understand the alternatives to encrypting data at rest Understand the security implementations and vulnerabilities in commercial database systems such as Oracle and SQL Server 2000 Learn security audit methods Learn about multi-level database security The course content was covered using material from many sources, primarily research papers. The Database Security book by Castano, et al is an out of print book as it was originally developed in 1994. The Database Security and Auditing book by Afyouni was printed in April 2005 and so was not available when the semester started. In the course we used two SQL Server Security books which were available in print and one Oracle Security book that was available in electronic form through Safari books. These books contributed to reinforcing concepts discussed by testing several attack methods. Another special feature of teaching the Database Security course was the availability of a dedicated InfoSec Lab. We will discuss the contribution of the InfoSec Lab later in this paper. The initial emphasis in the course was on incorporating database security concepts during the design phase. The primary reason for this emphasis was on the need for integration of various components of an information system. Since database security is heavily dependent on network security, operating system security, physical security and other applications security such an integrated approach is essential for an appreciation of design decisions. The course content was arranged in such a way that both technology and policy aspects were equally emphasized. This emphasis was motivated by the fact that there are several legal requirements to be met and people's privacy must be protected. A compromised database endangers the privacy of individuals by the release of personal information such as social security number, date of birth, credit card numbers, and health history. An important part of database security is access control. There are several types of access controls that are available for the database administrator to work with. More importantly, choosing the proper type of access control enables the allocation and revocation of privileges to individuals for various components of the database. The three types of access controls discussed related to Mandatory Access Control (MAC), Discretionary Access Control (DAC) and Role-based Access Control (RAC). A simple example of MAC would be that of using a suitable password for database access. However, practical uses of databases always require overriding a default access privilege for a specific event. In such instances one uses Discretionary Access Control. Since database privileges sometimes have the inheritance property it becomes essential to understand how the particular commercial system would handle DAC. The most important of access controls is Role-based Access Control. Discussion of this topic showed the various nuances involved in assigning access privileges in a timely manner without hindering productivity and at the same time providing security. All necessary database accesses could be associated with a specific role that an individual performs in an organization. Since roles change often and consequently access needs change as well, it is much easier to manage access control by associating privileges with roles. It is worth noting that these three types of access controls are not mutually exclusive but work in combinations that suit the organizational needs. Another important aspect of database security is authentication. Since databases provide many views of the data, suitable privileges for the ability to drill down data requires appropriate authentication. The authentication aspect of database access supports the confidentiality in the CIA (Confidentiality-Integrity -Availability) triangle that is basic to information security. Authentication models discussed include single-factor, two-factor, and three-factor authentication and the attendant performance issues. Among the many topics covered in this course, one of the important ones relates to Multi-Level Secure (MLS) databases [12, 14]. Commercial databases such as Oracle or SQL Server do not handle the MLS aspects in their software. However, it is an important aspect to be aware of. For example, in an 80 organization not every one has the access rights to view confidential information. Database queries are designed to pull all data that satisfy the query condition. In the MLS environment, the query condition would not necessarily reflect the security level of the person making the query. Since the security level authorization of the individual running the query is known from the login time, the MLS database is supposed to show all data that is cleared for that level of user. Usually the security levels could be unclassified, confidential, secret, or top secret. Not all fields of data in a record would need to be carrying a classification. Those sensitive data that have an associated security classification should be able to serve the needs of users with the appropriate security clearance and hide the data from others. A major problem to overcome in this context is known as polyinstantiation [13]. This concept refers to the fact that if persons with a lower clearance level have reason to suspect the existence of a record with hidden values, then they would be able to infer. Polyinstantiation could be addressed to a large extent by allowing certain redundancies in a database. Another common problem with MLS databases is the presence of inference channel. Inference channel leaks information about data classified at a higher level to users with lower level clearances [19]. Security policies also play an important role in protecting against inference channel leaks. A related approach to this problem is to develop classification constraints on data. These data classifications are then used at query time and then the appropriate level of the constraint is applied to the resulting data before it is presented to the user. In this context we discussed the security architecture for databases. This was broadly classified as those systems that use a Trusted Computing Base (TCB) that is external to the DBMS and those systems that manage access to data through the DBMS [22]. In the TCB architecture, the access controls were usually handled by the operating system or the network. In the DBMS control architecture, security design involved multi-factor authentication as well as security clearance and role-based access. As part of the secure architecture topic, we studied the Bell-LaPadula Model and the Biba Model [5]. Then we took a detailed look at the Seaview Model [17]. This is the first paper that studied in detail the security needs for database systems that contained data with various levels of security clearances. The major contribution of this paper was the application-independent nature of data integrity with particular reference to entity integrity, referential integrity and polyinstantiation integrity. We studied additional secure architecture topics with particular reference to commercial database systems. These topics include input validation, credential handling and encryption. Encryption is a major topic in itself in the security context. Usually encryption is an important tool for data in transit. However, the recent spate of incidents involving lost or stolen data [37] shows the need for protecting data at rest from falling into the wrong hands. One useful tool in this regard is encryption. We studied the impact of encrypted data with respect to performance. Usually, encryption of sensitive data at rest is a desirable feature provided the access to such data is not frequent. On the other hand, for data that is frequently used the better alternative to encryption would be to partially secure storage [31, 33] whereby the data management is handled by an independent system that works outside the operating system control. This technique protects the data from hackers as the access control is under an independent system that is not manipulated by the operating system, where most of the vulnerabilities are exploited. In this context we studied the FARSITE model that discusses the reliable storage aspects in an incompletely trusted environment such as the Internet [2]. This research, performed at Microsoft, shows how "to harness the collective resources of loosely coupled, insecure, and unreliable machines to provide logically centralized, secure, and reliable file-storage service." The next major topic covered was security audit for a database. The sources used for this topic were Jajodia [15], Andrews [4], and material from the Congressional Hearing reference provided in the References section. Audit involves different components such as login, access and storage. Commercial database systems such as Oracle and SQL Server facilitate login auditing in a simple way. For example, in SQL Server the user could set the login audit level to any one of four levels. Level 0 does not log any information about the logins, level 1 logs information about successful logins only, level 2 logs information about unsuccessful logins only and level 3 logs information about all attempted logins. This aspect of setting the appropriate level is related to the security policy of the organization. An organization might feel that they need to know only those people who attempted a login and failed as the ones who successfully logged in are considered authorized users. This is not a good assumption when it comes to computer forensics where one is trying to reconstruct an event that happened in the past. Consequently, organizations must consider the impact of their policies when it comes to information security. Auditing is also mandated by certain accreditation bodies. In order to satisfy certain data security requirements, some organizations might have to secure C2 level security rating from the National Computer Security Center (NCSC). The NCSC certification is measured according to the Department of Defense Trusted Computer System Evaluation Criteria [4]. We concluded the course with an analysis of database protection, copyright and privacy aspects both from a policy and legal perspective. First, we discussed the Congressional hearing on "Database and Collections of Information Misappropriation Act of 2003." This hearing showed the limitations of Copyright laws and how U.S. courts have interpreted the laws that protect privacy. We then studied the future of the database protection in U.S. and the laws to help in this regard. U.S. court rulings, including that of the Supreme Court, have shown that "sweat of the brow" argument does not offer protection for databases, rather the demonstration of some form of "originality" in data collection and dissemination is essential for database ownership. A court ruling in 2001 in United Kingdom in the case of the British Horseracing Board (BHB) has once again brought into focus the sweat of the brow argument. The U.K. court upheld the BHB's claim of ownership of data pertaining to horses and jockeys [10]. It remains to be seen how the U.S. courts would consider challenges to the sweat of the brow argument when it comes to protecting large databases from competitors. EVALUATION TOOLS In this course we used several different types of evaluation tools. Students were required to write three individual research reports on topics provided in class. The topics were: 81 1. Buffer overflows 2. Security audit 3. Sarbanes Oxley Act and its impact on Database Security On the testing side, we used a closed book, closed notes, midterm and final examinations. All questions were essay type. The students had access to a dedicated InfoSec lab where they could perform several different types of hands-on testing for vulnerabilities [32]. The InfoSec Lab has 16 workstations on a LAN connected to a Windows 2000 server. First the SQL Server 2000 was installed on the server. Two stand-alone computers that were not connected to the network were also provided to the students for testing. The first assignment provided a chance for the students to install SQL Server 2000 and choose appropriate security settings for various components of the database. The students then created new SQL Server accounts on the stand-alone computers and granted suitable privileges first and then tested the DENY and REVOKE features as well. The students had to install the latest SQL Server patches on the stand-alone computers and test for vulnerabilities. The dedicated lab environment provided an excellent facility for us to allow students to understand how a hacker would gain routine information about the database system. First the SQL Server 2000 was left unpatched and the students used the SQL Ping2 utility to gather information about the database system. This showed the port 1433 in use. Then the SQL Server 2000 was patched with version 3a and the students tried the same SQL Ping2 utility, this time finding a different type of information about the SQL Server. Next, the SQL Server was put in hide mode and the students found out this piece of information by noticing that the listening port had changed to 2433. We were able to accomplish this testing by making changes to the SQL Server every two days giving a short time between changes for testing. This was done as assignment 2. The third assignment involved testing Bulk Copy / Bulk Insert features of SQL Server. The fourth assignment involved a buffer overflow attack. A sample code was given to the students to try the buffer overflow attack on the patched server. The patched server foiled the attack. The students were then asked to test the same buffer overflow attack on the stand-alone computers where patches were not applied. The last assignment involved SQL Injection attack. The students were given a series of codes for the SQL Injection attack testing. The first part involved logging into a SQL Server database system knowing the userid of the user but not the password. The second part involved not knowing the userid or the password. The third part involved creating a new user and then exploiting the system. The fourth part involved finding the password of the sa account. The fifth part involved dropping the SQL Server from the server and shutting down the SQL Server via SQL Injection attack. The students were given the challenge in the fourth part of the SQL Injection attack testing to find out the strong password used on the server, which had all the latest patches both for the SQL Server part and the operating system part. This required more work beyond the SQL knowledge. One of the students succeeded in finding out the server password, not just the sa password, which was much easier to get using SQL Injection. CONCLUSION Overall, the students enjoyed the content of the course that involved learning many database security concepts and the ability to test many aspects of SQL Server installation, suitable settings, detect vulnerabilities, develop simple countermeasures and have the ability to use the logs to detect intrusion. ACKNOWLEDGEMENTS This research was supported in part by the NSF grant DUE-0416900 and the Kentucky Council on Postsecondary Education grant GB040955. REFERENCES [1] Abrams, M. D., Jajodia, S., Podell, H. J. 1995. Information Security: An integrated collection of essays, IEEE Computer Society Press, CA. [2] Adya, A., Bolosky, W.J., Castro, M., Cermak, G., Chaiken, R., Douceur, J., Howell, J., Lorch, J.R., Theimer, M. and Wattenhofer, R.P., 2002. "FARSITE: Federated, Available, and Reliable Storage for an Incompletely Trusted Environment," Proceedings of the 5 th Symposium on Operating Systems Design and Implementation, Boston, MA, December, 1 14. [3] Afyouni, H. A. 2006. Database Security and Auditing, Course Technology, MA. [4] Andrews, C., Litchfield, D., Grindlay, B. 2003. SQL Server Security Fundamentals, McGraw-Hill/Osborne, NY. [5] Castano, S., Fugini, M., Martella, G., Samarati, P. 1994. Database Security, ACM Press Books (Diane Publishing Co.), NY. [6] Cerrudo, C. "Manipulating Microsoft SQL Server Using SQL Injection" http://database.ittoolbox.com/browse.asp?c=DBPeerPubl ishing&r=%2Fpub%2FSG090202%2Epdf, Accessed on 07/25/2005 [7] CERT http://www.cert.org, Accessed on 05/20/2005 [8] CNSS Stds. "National IA Education Standards," http://www.nsa.gov/ia/academia/cnsstesstandards.cfm [9] Congressional Hearing, 2003. "Database and Collections of Information Misappropriation Act of 2003," September. http://www.copyright.gov/docs/regstat092303.html, Accessed on 04/10/2005 [10] Duke University, 2001. "The Future of Database Protection in U.S. Copyright Law" http://www.law.duke.edu/journals/dltr/articles/2001dltr0 017.html, Accessed on 04/15/2005 [11] Hinke, T., 1995. "Multilevel Secure Database Management Prototypes," in Information Security: An Integrated Collection of Essays, 1 edition st , Edited by Abrams, M.D., Jajodia, S.G., Podell, H.J., Essay 23, IEEE Computer Society Press, CA, 542-569. 82 [12] Jajodia, S. and Sandhu, R., 1995. "Toward a Multilevel Secure Relational Model," in Information Security: An Integrated Collection of Essays, 1 edition st , Edited by Abrams, M.D., Jajodia, S.G., Podell, H.J., Essay 20. IEEE Computer Society Press, CA, 460-492. [13] Jajodia, S., Sandhu, R. and Blaustein, B.T., 1995. "Solutions to the Polyinstantiation Problem" in Information Security: An Integrated Collection of Essays, 1 edition st , Edited by Abrams, M.D., Jajodia, S.G., Podell, H.J., Essay 21. IEEE Computer Society Press, CA, 493-529. [14] Jajodia, S. and Meadows, C. 1995. "Inference problems in multilevel secure database management systems," in Information Security: An Integrated Collection of Essays, 1 edition st , Edited by Abrams, M.D., Jajodia, S.G., Podell, H.J., Essay 24. IEEE Computer Society Press, CA, 570-584. [15] Jajodia, S., Gadia, S.K., and Bhargava, G., 1995. "Logical Design of Audit Information in Relational Databases" in Information Security: An Integrated Collection of Essays, 1 edition st , Edited by Abrams, M.D., Jajodia, S.G., Podell, H.J., Essay 25. IEEE Computer Society Press, CA, 585-595. [16] Lewis, M. 2004. "SQL Server Security Distilled," 2 nd edition, Apress, CA. [17] Lunt, T., Denning, D. E., Schell, R. R., Heckman, M. and Shockley, W. R. 1990. "The Seaview Security Model," IEEE Transactions on Software Engineering, 16 (#6), June, 593 607. [18] Mao, W. 2004. "Modern Cryptography," Prentice-Hall, NJ. [19] Meadows, C. and Jajodia, S., 1995. "Integrity in Multilevel Secure Database Management Systems," in Information Security: An Integrated Collection of Essays, 1 edition st , Edited by Abrams, M.D., Jajodia, S.G., Podell, H.J., Essay 22. IEEE Computer Society Press, CA, 530-541. [20] Nevins, S.C., 2003. "Database security breaches on the rise" http://www.snwonline.com/evaluate/database_ security_03-31-03.asp?article_id=224, Accessed on 04/15/2005. [21] Nessus http://www.nessus.org. Accessed on 05/19/2005. [22] Notargiacomo, L. "Architectures for MLS Database Management Systems" in Information Security: An Integrated Collection of Essays, 1 edition st , Edited by Abrams, M.D., Jajodia, S.G., Podell, H.J., Essay 19. IEEE Computer Society Press, CA. [23] O'Reilly Publishers. Developing a Database Security Plan http://www.oreilly.com/catalog/orasec/chapter/ch07.html Accessed on 03/10/2005. [24] PDD63, 1998. http://www.fas.org/irp/offdocs/pdd/pdd63.htm, Accessed on 05/22/2005. [25] Pernul, Gunther, 1994. "Database Security" chapter in `Advances in Computers,' Edited by M.C.Yovits, vol. 38, Academic Press, NY. [26] Rob, P. and Coronel, C. 2004. "Design, Implementation and Management," 6 th Edn., Course Technology, MA. [27] Sandhu, R. and Samarati, P., 1994. "Access Control: Principles and Practice," IEEE Communications Magazine, vol. 32, September, 40-48. [28] Sandhu, R., Coyne, E.J., Feinstein, H. L. and Youman, C.E., 1996. "Role-based Access Control Models," IEEE Computer, vol. 29, February, 38-47. [29] SANS http://www.sans.org, Accessed on 05/19/2005. [30] Solworth, J. A. 2004. "Integrating Discretionary and Mandatory Access Controls" http://parsys.cs.uic.edu/~solworth/integratingMacDac.pd f. Accessed on 04/15/2005. [31] Son, S. H., Chaney, C., and Thomlinson, N. P., "Partial Security Policies to Support Timeliness in Secure Real time Databases," 1998. Proceedings of the IEEE Symposium on Security and Privacy, May 3-6, 136 147. [32] Srinivasan, S. 2005. "Design and Development of an Information Security Laboratory," Proceedings of the 9 th Annual Colloquium on Information System Security Education, Atlanta, GA, June 6-9. [33] Strunk, J.D., Goodson, G.R., Scheinholtz, M.L., Soules, C.A.N. and Ganger, G.R., 2003. "Self-Securing Storage: Protecting Data in Compromised Systems," Foundations of Intrusion Tolerant Systems, 195 209. [34] Theriault, M. and Heney, W. 1998. "Oracle Security," O'Reilly Publishers, IN. [35] Tomson, B., 2004. "SQL Server 2000 Security Best Practices" http://wp.bitpipe.com/resource/org_1078177630_947/SQ Lserver2000.pdf. Accessed on 03/20/2005. [36] UofL InfoSec, 2005. "InfoSec Program website," http://www.louisville.edu/infosec [37] Wall Street Journal, 2005. "ChoicePoint struggles to gauge how much information fell into wrong hands," May 3, Page 1. 83

inference channel;access control;buffer overflows;CIA;privacy;polyinstantiation;database;inference;Database;encryption;multilevel security;authentication;policy;security

65

dBBlue: Low Diameter and Self-routing Bluetooth Scatternet

This paper addresses the problem of scatternet formation for single-hop Bluetooth based ad hoc networks, with minimal communication overhead. We adopt the well-known structure de Bruijn graph to form the backbone of Bluetooth scatternet, hereafter called dBBlue, such that every master node has at most seven slaves, every slave node is in at most two piconets, and no node assumes both master and slave roles. Our structure dBBlue also enjoys a nice routing property: the diameter of the graph is O(log n) and we can find a path with at most O(log n) hops for every pair of nodes without any routing table . Moreover, the congestion of every node is at most O(log n/n), assuming that a unit of total traffic demand is equally distributed among all pair of nodes. We discuss in detail a vigorous method to locally update the structure dBBlue using at most O(log n) communications when a node joins or leaves the network. In most cases, the cost of updating the scatternet is actually O(1) since a node can join or leave without affecting the remaining scatternet. The number of nodes affected when a node joins or leaves the network is always bounded from above by a constant. To facilitate self-routing and easy updating, we design a scalable MAC assigning mechanism for piconet, which guarantees the packet delivery during scatternet updating. The dBBlue scatternet can be constructed incrementally when the nodes join the network one by one. Previously no method can guarantee all these properties although some methods can achieve some of the properties.

INTRODUCTION Bluetooth [8] is a promising new wireless technology, which enables portable devices to form short-range wireless ad hoc networks based on a frequency hopping physical layer. Bluetooth ad-hoc networking presents some technical challenges, such as scheduling , network forming and routing. User mobility poses additional challenges for connection rerouting and QoS services. It has been widely predicted that Bluetooth will be the major technology for short range wireless networks and wireless personal area networks. This paper deals with the problem of building ad hoc networks using Bluetooth technology. According to the Bluetooth standard, when two Bluetooth devices come into each other's communication range, one of them assumes the role of master of the communication and the other becomes the slave. This simple one hop network is called a piconet, and may include more slaves. The network topology resulted by the connection of piconets is called a scatternet. There is no limit on the maximum number of slaves connected to one master, although the number of active slaves at one time cannot exceed . If a master node has more than slaves, some slaves must be parked. To communicate with a parked slave, a master has to unpark it, thus possibly parking another active slave instead. The standard also allows multiple roles for the same device. A node can be master in one piconet and a slave in one or more other piconets. However, one node can be active only in one piconet. To operate as a member of another piconet, a node has to switch to the hopping frequency sequence of the other piconet. Since each switch causes delay (e.g., scheduling and synchronization time), an efficient scatternet formation protocol can be one that minimizes the roles assigned to the nodes, without losing network connectivity. While several solutions and commercial products have been in-troduced for one-hop Bluetooth communication, the Bluetooth specification does not indicate any method for scatternet formation. The problem of scatternet formation has not been dealt with until very recently. The solutions proposed in literature can be divided into single-hop and multi-hop solutions. Several criteria could be set as the objectives in forming scatternet. First of all, the protocol should create degree limited scatternets, to avoid parking any node. Secondly, the number of piconets should be minimized to provide faster routing. Thirdly, the formation and maintenance of scatternet should have small communication overhead. Fourthly, the diameter of the scatternet should be small, i.e., the maximum number of hops between any two devices must be small. In this paper, we focus on scatternet formation for single-hop ad hoc networks. In a single-hop ad hoc network, all wireless devices are in the radio vicinity of each other, e.g., electronic devices in a laboratory, or laptops in a conference room. A single-hop network can be modeled by a complete graph. 22 Previous literature on scatternet formation assumed that devices are not able to communicate unless they have previously discovered each other by synchronizing their frequency hopping patterns. Thus, even if all nodes are within direct communication range of each other, only those nodes, which are synchronized with the transmitter , can hear the transmission. Synchronizing the frequency hopping patterns is apparently a time consuming and pseudo-random process [13]. In this paper we assume that the problem of discovering all neighbors within transmission radius of a device is resolved by separate Bluetooth protocol. One such protocol for discovering all one hop networks is described in [13, 3], while a protocol that provides two-hop information to every node is described in [12]. These protocols are applicable as the pre-phase of our scheme. This paper addresses the problem of scatternet formation for single-hop Bluetooth based ad hoc networks, with minimal communication overhead. We adopt the well-known structure de Bruijn graph to form the backbone of Bluetooth scatternet, hereafter called dBBlue, such that every master node has at most seven slaves, every slave node is in at most two piconets, and no node assumes both master and slave roles. Our structure dBBlue also enjoys a nice routing property: the diameter of the graph is and we can find a path with at most hops between every pair of nodes without any routing table. Moreover, the congestion of every node is at most , assuming that a unit of total traffic demand is evenly distributed among all pair of nodes. We discuss in detail a vigorous method to locally update the structure dBBlue using at most communications when a node joins or leaves the network. In most cases, the cost of updating the scatternet is actually since a node can join or leave without affecting the remaining scatternet. The number of nodes affected when a node joins or leaves the network is always bounded from above by a constant. To facilitate self-routing and easy updating, we design a scalable MAC assigning mechanism for piconet, which can guarantee the packet delivery even during updating. Our method can construct the structure dBBlue incrementally when the nodes join the network one by one. In addition, the structure formed by our method can sustain the faults of nodes and the network is still guaranteed to be connected. If a node detects a fault of some neighboring master node or bridge slave node, it can dynamically re-route the packets and the path traveled by the packet is still at most . Previously no method can guarantee all these properties although some methods can achieve some of the properties. The rest of the paper is organized as follows. Section 2 presents our new Bluetooth formation algorithms for single-hop ad hoc networks . We describe how to build a static scatternet of nodes based on de Bruijn graph and assign roles and labels to them. Section 3 proposes a vigorous method to locally and dynamically update the scatternet topology when node joins or leaves the network. Section 4 describes the routing method for our de Bruijn based scatternet which efficiently finds the next node need to go without any routing table. The related works is discussed in section 5. We conclude our paper in Section 6 by pointing out some possible future research directions. DBBLUE SCATTERNET CONSTRUCTION Our dBBlue scatternet first builds a backbone based on the well-known de Bruijn graph [5]. The de Bruijn graph, denoted by , is a directed graph with nodes. Assume that each node is assigned a unique label of length on the alphabet . There is an edge in from a node with label to any node with label , where . Figure 1 illustrates . It is well-known that the de Bruijn graph enables self-routing intrinsically. The self-routing path from the source with label to the target with label is . Observe that, we could find a shorter route by looking for the longest sequence that is both a suffix of and a prefix of . Suppose that is such longest sequence. The shortest path between the source and the target is . Clearly, the route between any two nodes is at most hops, i.e., has diameter , where is the number of nodes of the graph. 111 001 011 010 100 110 000 101 Figure 1: The de Bruijn graph . The classical de Bruijn graph is balanced in the sense that the labels of all nodes have the same length. The de Bruijn graph can be generalized to any set of vertices whose labels form a universal prefix set. In [7], Fraigniaud and Gauron proposed a novel method to construct an efficient topology for P2P network based on the generalized de Bruijn graph defined on a universal prefix set. "A universal prefix set is a set of labels on an alphabet such that, for any infinite word , there is a unique word in , which is a prefix of . The empty set is also a universal prefix set."[7] For instance, is a universal prefix set on alphabet , but and are not. There is a directed edge from node to another node in the generalized de Bruijn graph if is the prefix of the label of node . A generalized de Bruijn graph is pseudo-balanced if the lengths of the labels are different by at most one. For simplicity, we still denote a pseudo-balanced de Bruijn graph on alphabet by if the node labels have length at least bits and at most bits. We also say that a node from is at level if its label has bits. In this paper, we only consider the balanced or pseudo-balanced binary de Bruin graph . Node labels in a pseudo-balanced de Bruijn graph correspond to all the leaf nodes in a full binary tree, in which the depth difference between any two leaf nodes is at most one and each internal node has two children, Figure 2 illustrates the correspondence between them. In the figure, the pseudo-balanced de Bruijn graph is defined on the leaf nodes and directed edges. In a pseudo-balanced de Bruijn graph , each node has at most out-neighbors and in-neighbors. To route a packet from a node with label to another node with label , where . Node will forward the packet to its neighbor node with label , or , or . Notice that since the labels of the nodes are a universal prefix set, we know that exactly one of these three labels does exist. The following nodes keep forwarding the packet similarly until it reaches node . Consequently, the diameter of pseudo-balanced de Bruijn graph is still . In this paper, we propose a scalable scatternet structure based on pseudo-balanced de Bruijn graph . 23 root 0000 0001 0010 0011 001 000 011 010 100 101 110 111 10 11 01 00 1 0 Figure 2: The correspondence between full binary tree and pseudo-balanced de Bruijn graph. In a pseudo-balanced de Bruijn graph , two nodes are called critical pair if they only differ in the least significant bit of their labels. Let be the sequence of nodes visited by a traversal of all leaf nodes in the corresponding binary tree of . A node is called the successor of another node and is called the predecessor of another node . Here takes value . For example, in Figure 2, nodes and is a critical pair; node is the successor of the node . 2.2 MAC Address Assignment for Piconet Our method will construct a balanced (or pseudo-balanced) de Bruijn graph as the backbone of the network. Here the choosing of the integer is discussed later. We will ignore the direction of the edges in the de Bruijn graph . Thus, every node will have at most (or for pseudo-balanced de Bruijn graph ) edges incident. Every node in the backbone of dBBlue scatternet will be assigned a master role. We will add a bridge slave node for every pair of master nodes that are connected in the backbone. Thus, every master node will have at most six bridge slave nodes so far. We then add some free slave nodes to each master node, and call them pure slave nodes. Before we discuss in detail our scatternet construction methods, we present our novel rule of assigning the MAC address in a piconet . In our dBBlue scatternet, when we route a packet to a destination node , we only know the piconet ID of node , say , which is same as the label of its master node, and the MAC address, say , of this node in that piconet. The detail routing mechanism will be discussed in Section 4. When some node joins or leaves the scatternet, we often have to reorganize some piconets and thus re-assign the MACs of some nodes. Our method of assigning MAC addresses in a piconet and reorganizing the piconets guarantees that the new piconet (even the new MAC address) can be found by a simple appending or deleting the least significant bit, which keeps the label prefix of updating nodes un-changed so that even the delivery of the packets on the way to those updating nodes will not be interrupted. In a piconet, MAC is always reserved by the master node. For simplicity, we omit the MAC address of a master node hereafter while representing its label, i.e., the master node with label actually has a label if consistent labels with slave nodes are needed. Remember that, in a pseudo-balanced de Bruijn graph, any node has in-neighbors (except and ) and at most out-neighbors, so MAC addresses and are always reserved for the two bridge slaves to in-neighbors , MAC , , and are reserved for bridge slaves to out-neighbors if they exist, and is reserved for the th slave (it must be a pure slave) if it exists. Figure 3 illustrates all four possibilities for the piconet MAC address assignment according to the number of out-neighbors in scatternet backbone. In the figure, for simplicity, we use to denote a node with label or , whichever exists in the network. Notice that a master node in the constructed scatternet based on a pseudo-balanced de Bruijn graph always has two incoming neighbors. For example, a master node in level can have incoming neighbor or , but not both since the de Bruijn graph is built upon a universal prefix set; similarly another incoming neighbor is . Analogously, a master node in level has incoming neighbors and . On the other hand, the number of out-neighbors of a node in the pseudo-balanced de Bruijn graph could be . Only the node at level could have or out-neighbors and only the node at level could have out-neighbor (except nodes and if they exist). ... x m-1 1 x 1 0 ... x m-1 x 1 m (x ) m (x ) 001 010 100 101 110 011 111 ... x m+1 ... x x m+1 2 x 1 x 2 (a) One out-neighbor m m (x ) (x ) 1 ... x 1 x 2 x m ... ... x 1 x m-1 ... x 1 x m-1 001 010 100 101 110 011 111 1 x 0 2 x m 0 ... x 2 x m (b) Two out-neighbors (x ) m m (x ) x m ... ... x 1 x m-1 ... x 1 x m-1 x 2 x m 0 0 ... x 2 x m 0 ... 1 001 010 100 101 110 011 111 1 x 0 2 x m 1 ... x 1 x 2 (c) Three out-neighbors m (x ) (x ) m x m-1 ... x 1 x m-1 x 2 x m 0 0 ... x 2 x m 0 ... 1 x 2 x m ... 1 0 x 2 x m ... 1 1 001 010 100 101 110 011 111 1 x 0 1 x 2 x m ... ... x 1 (d) Four out-neighbors Figure 3: MAC address assignment for a piconet. 24 Table 1 summarizes the rule of assigning the MAC address to the bridge slave nodes in a piconet. Their MAC addresses can be decided uniquely according to the label bit difference between current piconet and neighbor piconet IDs. For example, if the master is labeled and its out-neighbor is labeled , then the MAC addresses of their bridge slave is assigned by , and assigned by . Remember that every bridge slave has one MAC address in each of the two piconets it resides. Table 1: The rule to assign MAC address to bridge slave nodes. In-Neighbor Out-Neighbor Node Notice that, in bluetooth scatternet, the bridge slave nodes have two independent piconet IDs and MAC addresses in two piconets respectively. However, since the routing mechanism in de Bruijn is directional, only their piconet ID and MAC address assigned by their in-master is public and meaningful for routing, saying label in the remaining paper, and the other one is only used for inter-communication in a piconet. Figure 4 illustrates one piconet in the scatternet. Here nodes , , , and assume master role and form the backbone for scatternet. These master nodes are connected in the de Bruijn graph by bridge slaves , , and respectively. Assume that node has label . Nodes , denote the two incoming neighbors of node , which has label and respectively. Nodes , denote the two outgoing neighbors of node , which has label and respectively. Nodes , , and are the pure slave nodes of in the scatternet. The label of node ( ) is where is the MAC address of node in this piconet, and and has public label and , respectively, which is consistent with the prefix of and respectively . Notice that the MACs of and in the piconet mastered by node are and respectively, which are used only by nodes in this piconet and not broadcasted to the network. 2 v 5 v 1 v 6 v 4 v 3 v 7 u I 1 I 2 2 O O v 1 Figure 4: An example of a static piconet (with nodes inside the shaded region) formed by our method. Here a master node is denoted by a square, a pure slave is denoted by a circle, and a bridge slave is denoted by a triangle. As will see later, our labeling rule makes the updating of the scatternet topology and nodes' labels much easier when new nodes join the network or some existing nodes leave the network. For incremental updating of the scatternet, there are two scenarios when a new node joins the network. The first case is that there is a master node who has free slot for a pure slave. We then directly assign the newly joined node as the pure slave of that master node. The second case is that no master node has free slot for a pure slave. We then have to split some piconet and in turn create some free slots for pure slaves. The splitting of a piconet is performed such that the resulting backbone (formed by master nodes and bridge slaves) is still a pseudo-balanced de Bruijn graph. When a piconet is aplit-ted , the labels of some nodes have to be updated. While updating the topology, it is possible that some packets are already on their way to the destinations (via or toward this splitting piconet). Our labeling rule makes sure that the packets can still be routed without any interruption, only the local nodes are assigned new labels, and the re-labeling are also conducted locally. 2.3 Static Scatternet Construction Given nodes currently distributed in the network, the section gives an efficient algorithm to construct our de Bruijn based scatternet dBBlue, which has low diameter and bounded node degree property. In other words, we first study the construction of the scatternet for a static -nodes network, which will serve the base for our dynamic construction. Our method will construct a balanced de Bruijn graph as the initial backbone of the network. We will choose integer such that . The choosing of guarantees that there are enough bridge slave nodes, which implies that no master node serves as bridge slave. Our method does not consider the detail of the neighbor discovering process. We assume that every node already knows the existence of the other nodes. A LGORITHM 1. Static DeBruijn-Based Scatternet 1. Assume that there is a leader already among these nodes . The leader could be the node with smallest ID. We give the token to the leader and call it token node. Token node randomly selects nodes (including itself) into the master set which assumes the master role in final scatternet topology, where and is the number of nodes in . Let , which is the total number of nodes that can be assigned as pure slaves. 2. Token node assigns itself with label , and each node in with a unique bits label in the range from to . The set of nodes forms a de Bruijn graph as the scatternet backbone. 3. Token node, with label , selects nodes 1 from the remaining as its bridge slaves, and assigns them labels and respectively. Here , will also serve as the Medium Access Code (MAC) for these two slaves in the piconet mastered by this token node. Token node uses its bridge slave node to connect with its out-neighbor and the bridge slave node to connect the out-neighbor node . 4. Assume that the current token node has label with value . The token node selects nodes 2 from the remaining as its slaves and assigns them with labels , 1 There are two special nodes and , which only have 1 out-neighbor , we then just use one bridge slave node to connect with its out-neighbor. 2 Node and may choose nodes as its pure slaves since they only have one in-neighbor and one out-neighbor. 25 and in the order if they exist . Let . Then the token is passed to its successor. 5. Repeat the above steps (3) and (4) until all nodes in are processed. After all nodes have been processed, the current token node passes the token back to node again. Once the initial topology construction is finished, the token node will be responsible for the following node joining and leaving issues. Master nodes form the backbone of bluetooth scatternet, and a piconet works like a node in de Bruijn graph. 111 001 011 010 100 110 000 101 Figure 5: dBBlue Bluetooth Scatternet. Figure 5 illustrates a dBBlue scatternet containing nodes based on graph. T HEOREM 1. In dBBlue scatternet, each master has no more than slaves and each slave assumes as bridge for at most piconets . And the number of piconets is at most and at least . Moreover, the computation cost is for static construction. P ROOF . From the topology construction, each master carries at most same prefix slaves, and at most different prefix slaves from its in-neighbors since each node in graph has at most in-neighbors , so each master has no more than slaves. And, each slave exists as a free slave or as the bridge between its same prefix master and one of 's out-neighbors, so the degree of a slave node is at most 2. Let , where and is the number of masters. Then implies . Thus, and . Consequently, , which implies . It is obvious that the total computation cost of constructing static dBBlue scatternet is . In this paper we always assume a bluetooth piconet consists of at most slaves and master. If future bluetooth technology allows a master to bring more slaves, say , our scatternet construction method can adapt easily as follows. The scatternet backbone will be still based on de Bruijn graph. However, is chosen such that . In other words, every master node will carry pure slaves and bridge slaves to connect to its two out-neighbors and two in-neighbors in the de Bruijn graph . It is not difficult to show that using de Bruijn graph will create a scatternet with less piconets than using for since each master node will carry less pure slaves in the later case. On the other hand, the scatternet based on for does provide a better fault tolerance since the degree of each master node is increased to . DYNAMIC SCATTERNET UPDATING In this section we describe a vigorous method to locally update the scatternet topology dynamically when node joins or leaves the network. Considering each piconet as an abstract node in the de Bruijn graph, our goal is to maintain a scalable pseudo-balanced de Bruijn graph. 3.1 Token Based Updating First consider the case when a node wants to join the network. We have to assign a role for this newly joined node. There are several possible scenarios about the existing scatternet. (1) the existing scatternet has a master node that has free slave slots, then we can simply assign this newly joined node as the pure slave of this master node. (2) all master nodes in the existing scatternet already have slaves, we then have to expand the backbone of the scatternet to incorporate this newly joined node. In other words, we have to split some piconet to two such that the two new piconets will have some free pure slave slots to hold this newly joined node. Several methods can be used to implement the above scheme. To make the updating efficient, we should be able to quickly find the master node with empty slot for pure slave if there is any. One approach is to keep the current scatternet compact and assign a special node the token in a way such that all master nodes with label less than the token node do not have empty slot, and all master nodes with label larger than the token node do have empty slot. When a new node joins the network, we can simply assign it the empty pure slave slot and then update the token node if necessary. This approach is efficient for node joining but suffers more cost for node leaving. When a node leaves the network, we have to update the scatternet to keep the scatternet compact. Thus, we possibly have to move some nodes to fill the slot emptied by this left node. The other approach is not to compact the scatternet. When a node leaves, we do nothing if the backbone of the scatternet is untouched . However, this approach suffers a large cost when node joins the network since we have to find where to put the newly joined node. One method is to use the broadcast method to travel the whole scatternet to find the master node with free pure slave slot. This may perform better if only a few of the existing piconets have free slots. The other method is to randomly select a master node and check if it has free slot. If it does not, we then select another random master node until one such master node is found. This approach performs better if the majority of the piconets have free slots. We omit the detail of performance analysis here, which will be presented in the full version of the paper. In this paper, we will adopt the compact approach. Before we present the detail of our methods of updating the scatternet, we first study the possible status of the scatternet, which will be recorded in the token node. When a new node requests joining the network, there are three possible scenarios to be discussed. 1. Current backbone is a balanced de Bruijn graph. Figure 6 illustrates an example. The token is held by the master node with the smallest label among all master nodes that have less than same-prefix slaves. In this status, the master node with the token has some free slot for newly joined node and so do all master nodes with larger labels. 2. Current backbone is pseudo-balanced de Bruijn graph under expanding status, i.e., many nodes join the scatternet. Figure 7 illustrates an example. The token is held by the first master node with less than same-prefix slaves in level if it exists, otherwise the first master node in level holds the token. In this status, all master nodes in level 26 token i-1 i i+1 Figure 6: Token in balanced de Bruijn graph. and do not have free slots except the last two master nodes in level . In other words, at most two master nodes have free slots. level m token i i-1 i+1 level m+1 Figure 7: Token in pseudo-balanced de Bruijn graph under expanding status. 3. Current backbone is a pseudo-balanced de Bruijn graph under shrinking status, i.e., many nodes leave the scatternet. Figure 8 illustrates an example. The token is held by the master node in level with the smallest label. In this status, each master node in level and level has and same-prefix slave nodes respectively. level m+1 i+1 i-1 i token level m Figure 8: Token in pseudo-balanced de Bruijn graph under shrinking status. Those statuses balanced, expanding, shrinking will be recorded in the token data structure. 3.2 Node Joining When a new node joins the network, there are three cases. 1. Token status is balanced, that is to say, current backbone is a balanced de Bruijn graph. See Figure 6 for an illustration. (a) The token node has less than slaves. Then it simply adds the joining node into its slave set and assigns it a label , where is one of the un-assigned MAC address in . If the token node now has slaves, then it passes the token to its successor. (b) The token node is fully occupied by slaves. This could happen only when all master nodes in the scatternet have slaves. Then the token is passed back to node if it is not at node . Change the token status to expanding and call Method 1 to split the current piconet mastered by the token node into two parts and add the joining node as a new pure slave with label . 2. Token status is expanding, that is to say, current backbone is a pseudo-balanced de Bruijn graph under expanding status. See Figure 7 for an illustration. (a) If the token node is in level , i.e., with bits label , the it must has less than slaves. It simply adds the joining node into its slave set and assigns it a label , where is one of the un-assigned labels in . If the token node now has slaves, then passes the token to its successor. (b) If the token node is in level , i.e., with -bits label . This could happen only when all master nodes in the scatternet has been fully occupied by slaves. Call Method 1 to split the current piconet mastered by this token node into two piconets, and add the joining node as a new slave with label . 3. Token status is shrinking, that is to say, current backbone is a pseudo-balanced de Bruijn graph under shrinking status . See Figure 8 for an illustration. In this case, token node surely has exactly four slaves (see node leaving for more details ). We first add the joining node as the slave of the token node and assign it one of the un-assigned MAC addresses in . Call Method 1 to split current piconet into two piconets, and pass token to the successor in level . If the current token node is , then set token status to balanced and pass the token to master node . In other words, we basically undo the updating (piconets merging) caused by the previous node leaving event. We then present our algorithm that split one piconet mastered by node to two new piconets mastered by nodes and respectively. M ETHOD 1. Piconet split due to node joining 1. Token node promotes its slave node as the master for a new piconet. We change the label of a pure slave node or a out-neighbor bridge slave node by simply appending in the MAC address, i.e., the new label is . Two new piconets have master node with labels and respectively. The detail of labeling and role updating is as follows: (a) , which assumes master role in first piconet. (b) , which assumes a bridge slave role in first piconet. (c) , which assumes a bridge slave role in first piconet. (d) , which assumes master role in second piconet. (e) , which assumes a bridge slave role in second piconet. (f) , which assumes a bridge slave role in second piconet. 27 Notice this label extension still preserves their prefix. Thus, after the piconet splitting, the message delivery will not be interrupted at all because old addresses are still reachable since the new label has same prefix. In addition, the nodes with new labels with the corresponding MAC addresses will serve the bridge slave role in the two newly created piconets. Figure 9 illustrates the change while piconet splitting. m ... 0,101 x 1 x m ... 1,010 x 1 x m ... 0,010 x 1 x m ... 1,101 x 1 x m ... 0,001 x 1 x m ... 0 x 1 x m ... 1 x 1 x m ... x 1 x m ... ,001 x 1 x m ... ,010 x 1 x m ... ,100 x 1 x m ... ,101 x 1 x m ... ,110 Joining v v u x u 1 x Figure 9: Piconet splits due to node joining. 2. Then, both and need reselect the bridge slaves to connect with its in-neighbors and out-neighbors if needed. Simultaneously , both and 's neighbors need reselect its same-prefix bridge slaves to connect with and . The selection still follows the rule described in Section 2.2, Figure 3 illustrates all possible scenarios. Since the master nodes in the new piconets are in level , each of them has at most out-neighbors in the pseudo-balanced de Bruijn graph . Thus, we have enough bridge slave nodes for each new piconet. The in-neighbor master nodes , where or , of node and in the de Bruijn graph have to change one of its pure slave to bridge slave to connect with node or . Notice this update is only restricted to local regions, so the update is totally localized. 3. Finally, the token is still kept by the master node , whose previous label is . 3.3 Node Leaving If a node leaves elegantly, it should first notify the token node before leaving. If a master/slave node leaves because unexpected reason such as power off, all of its neighborhood will detect it soon and notify the token node. Our method does not consider the detail of the exception detection process, we assume the token node can detect the node leaving in short time. When the token node detects the node leaving, then there are three cases to be addressed again: 1. Token status is balanced, that is to say, current backbone is a balanced de Bruijn graph. Here two cases need be discussed: (a) If the token node does have pure slave node, then the token node requests one pure slave to replace the position of the leaving node, including the label; (b) If the token node has no pure slave nodes, then it passes the token to its predecessor, say node . There are two scenarios also, which as discussed as follows. i. If node has pure slaves, then it requests one pure slave to replace the position of the leaving node. ii. If node also has no pure slaves. This could happen only when , and all master nodes have only slaves serving bridge slave role. Token node changes the token status to shrinking, and call Method 2 to merge its corresponding critical pair, then ask one pure slave to replace the position of the leaving node. 2. Token status is expanding, that is to say, current backbone is a pseudo-balanced de Bruijn graph under expanding status. (a) If the token node is in level , i.e., with -bits label . This could happen only when all master nodes in the scatternet has been fully occupied by slaves. The token need be passed the predecessor, which will ask one pure slave node to replace the position of the leaving node. (b) If the token node is in level , i.e., with bits label . If the token node does have pure slave node, then the token node requests one pure slave to replace the position of the leaving node, otherwise two cases need be discussed here: i. The least significant bit of the token node's label is . The token will be passed to be passed the predecessor, which will ask one pure slave node to replace the position of the leaving node. ii. The least significant bit of the token node's label is . It first merges its corresponding critical pair by calling Method 2, then requests one pure slave to replace the position of the leaving node. Now if the current token node is , then it changes the token status to balanced and passes the token to its predecessor . 3. Token status is shrinking, that is to say, current backbone is a pseudo-balanced de Bruijn graph under shrinking status. (a) If the token node is not , then it passes the token to its second predecessor with least significant bit in level , which will call Method 2 to merge its critical pair piconet and ask one pure slave to replace the position of the leaving node. (b) If the token node is , then it changes the token status to balanced and passes the token to node , which will ask one pure slave to replace the position of leaving node. One special case is that token node leaves. In this case, the token node will promote one of its pure slaves to replace it, i.e., be the master node and the new token node. If no new pure slave exits, similarly, we have to ask some pure slave node from its predecessor to replace its role. When the token node did not leave elegantly, it is more complicated and we need fault tolerance about the token node, which is out of the scope of this paper. We then describe our method to merge two piconets that are mastered by a critical pair. M ETHOD 2. Piconet merge due to node leaving 1. Assume that token node requests merging with its sibling master node . The new piconet has master node with label . Notice that node and node each has at most out-neighbors in the de Bruijn graph. The label change will be achieved by simply deleting the least significant bit as follows: (a) , which is the master node in the new piconet. 28 (b) , which is a pure slave node or the bridge slave node to connect master node if it exists. (c) , which is the bridge slave node to connect master node , whichever exists. (d) moves to replace the leaving node position. (e) , which is the bridge slave node to connect master node , whichever exists. (f) , which is a pure slave node or the bridge slave node to connect master node if it exists. Notice this label shrink still preserves the label prefix. Thus, after the piconets merging, the message delivery will not be affected at all because de Bruijn graph uses prefix based routing, old addresses are still reachable by the same prefix . The piconets mergence will not cause any routing problem although the node label shrink is not acknowledged by other nodes. At the same time, the sibling master node leaves to replace the position of leaving node. To continue the message delivery for node , the new master node will keep the new label of for a period of time and forwards the message targeted to accordingly. More detail is discussed in Section 4. Figure 10 illustrates the change of labels by merging piconets. m ... x 1 x m ... ,001 x 1 x m ... ,101 x 1 x m ... ,110 x 1 x m ... 0,101 x 1 x m ... 1,010 x 1 x m ... 0,010 x 1 x m ... 1,101 x 1 x m ... 0 x 1 x m ... 1 x 1 x m ... ,010 x 1 x m ... 1 u v u v x replace leaving node 1 x Figure 10: Piconets merge due to node leaving. 2. Then, node need reselect the bridge slaves to connect with in-neighbors and out-neighbors if needed. Simultaneously, the neighboring master nodes of and need reselect their same-prefix bridge slaves to connect with . The selection still follows the same rule described in Section 2.2, please see Figure 3 for an illustration for all possible scenarios. Notice this update is totally localized. 3. The token is now kept by the master node . It is not difficult to prove the following theorem. T HEOREM 2. Our method locally updates the dBBlue scatternet using at most communications when a node joins or leaves the network. In most cases, the cost of updating the scatternet is actually since the node can leave and join without affecting the remaining scatternet. The number of nodes affected when a node leaves or joins the network is always bounded from above by a constant. Our method can construct the structure incrementally when the nodes join the network one by one. 3.4 Bounded Network Size The method described so far can incrementally construct the scatternet when the nodes join the network one by one and can update the scatternet structure efficiently when nodes leave or join the network even frequently without affecting the worst case properties of the scatternet. This method is efficient in most cases, however, it could generate lots of merging and splitting of piconets in the worst case: a node joins the scatternet which causes the splitting of a piconet , then a node leaves which in turn causes the merging of two piconets, and repeat joining, leaving. In most applications, the size of the bluetooth network is often stable, for example, within for a small constant . If this is the case, we can apply the following approach to build the scatternet. First, we use Algorithm 1 to build a scatternet with nodes. When a new node joins the network, we first tries to find an empty pure slave slot for this node from the current token node. If no empty slot, we then pass the token to the successor of the current token node. When all master nodes in the scatternet have slaves, we will start to create another piconet to connect to the current backbone. In other words, instead of having pure slave nodes, a master node from the scatternet backbone will replace the pure slave nodes by piconets (at maximum). We call such piconets associated with the master node of the backbone. Clearly, a backbone based on a balanced de Bruijn graph could support from nodes to nodes without associating piconets . By associating piconets to the master nodes of backbone, the number of nodes it can support is increased to since we can replace each pure slave node by a piconet of nodes. One disadvantage of associating piconets to master nodes is that every master node in the backbone will have to forward more messages than the scatternet created by the method described previously . The other disadvantage is that when the network size goes beyond its supported scope, the updating of the scatternet is more costly than before. See the full version of the paper for more detail. ROUTING IN SCATTERNET We first describe the routing in the dBBlue scatternet with balanced backbone. If both source and target nodes are masters, we assume the source master node has label and the target master node has label . According to the routing mechanism described in Section 2.1, node simply forwards the message to its neighbor master node , relayed by their common bridge node if or by if . Then forwards the message again to its neighbor master node accordingly. Clearly, the message is guaranteed to reach the target in at most steps. If the source node is a slave, it first sends the messages to its master node. Notice that pure slave node has only one master node and the bridge slave node has two master nodes. Then bridge slave node just randomly picks one master node. Similarly if the target node is a slave, the message will be first forwarded to its master node. The procedure of routing message between these two master nodes is same as the previous description. Clearly, the routing path from one master node to another master node is at most hops. The longest path between two nodes happens from a slave node to another slave node, which is at most hops. From , we have . Thus, the diameter of the de Bruijn-based scatternet is . T HEOREM 3. For any two nodes in dBBlue scatternet, there is a path with at most hops and such path can be found locally based only on the labels of the source and target. 29 Notice that, two assumptions are made in our routing scheme described above: (1) the source node knows the label of the target node, and (2) the backbone of the scatternet is based on a balanced de Bruijn graph. We will not try to resolve the first assumption in this paper, but discuss it briefly here. The labels of a node can be broadcasted to the whole network if the nodes leaving and joining is not frequent, i.e., the labels of nodes do not change frequently. Or we can adopt a mechanism similar to the Domain Name Service (DNS): the labels are stored in a hierarchical manner and a node can query the label servers to get the labels of the target nodes and then cache them locally. Here, we discuss briefly how to perform broadcast in de Bruijn graph such that it guarantees to reach each node exactly once. We initiate the broadcast from node . Each node with label continues forwarding the message to its out-neighbors. The nodes whose most significant bit is will not forward the message. The broadcast basically works same as the breadth first search (BFS) in a binary tree. Clearly, a node will only forward the message to nodes with larger labels. Thus, a node receives the message exactly once. The communication cost of such broadcasting is exactly messages. We then discuss in detail how to route the packets when the scatternet backbone is pseudo-balanced. Assume the source master node has label and the target master node has label , where . Node will forward the packet to its out-neighbor master node with label , or , or . Notice that since the labels of all nodes are a universal prefix set, we know that exactly one of these three labels does exist. Consequently, the diameter of pseudo-balanced de Bruijn graph is still . The bridge slave node from to has MAC (1) if a master node with label exists; or (2) if a master node with label exists; or (3) if a master node with label exists. Review Section 2.2 for more detail about the rules of labeling nodes and assigning MAC addresses in a piconet. A shorter route is obtained by looking for the longest sequence that is suffix of and prefix of . For the purpose of illustration, let's see how we route packets from master node to master node in the scatternet based on the de Bruijn graph illustrated in Figure 2. First, the master node checks the labels of all out-neighbor master nodes and finds that master node with label exists. Then it forwards the packet to master node via the bridge slave node with MAC . Similarly, master node forwards the packet to master node with label via the bridge slave with MAC . Finally , the master node forwards the packet to node via the bridge slave with MAC . Notice that the last step it takes a shorter path other than via another master node . At last, we discuss how to route the messages while the scatternet is on updating due to nodes leaving or joining the network. When a node joins the network, the piconet mastered by the token node may be split into two piconets. Clearly, the message still can be routed since the labels of the two newly created piconets are the children of this token node. Similarly, when two piconets are merged to create a new piconet, the label-based routing still successfully route the packets. The remaining case is that when a node leaves, we may need find a pure slave node from the current token node to fill the space emptied by this left node. When a message targeted to node reaches the piconet mastered by the token node , node has already been moved. To remedy this, we apply a mechanism similar to the mail-forwarding service provided by the post-office: the master node will keep a record of the nodes moved to other piconets and its new label within a time window. When a message targeted for reaches, the master node forwards the message to the new destination and also acknowledges the source node of the new label of . The source node will then cache the label of node if it is frequently used. To decrease messages forwarding, every master node could record the frequency that a slave node receives messages from other node. When a pure slave node is visited frequently by other nodes, then we switch its role with one of the bridge slaves with same prefix and broadcast the new labels of these two nodes to the network. When we have to move a pure slave node to other piconet to make the scatternet compact, the pure slave node is the least frequently visited nodes among the current piconet. RELATED WORK Zaruba, Basagni and Chlamtac [15] proposed two protocols for forming connected scatternet. In both cases, the resulting topology is termed a bluetree. The number of roles each node can assume is limited to two or three. The first protocol is initiated by a single node, called the blueroot, which will be the root of the bluetree. A rooted spanning tree is built as follows. The root will be assigned the role of master. Every one hop neighbor of the root will be its slave. The children of the root will be now assigned an additional master role, and all their neighbors that are not assigned any roles yet will become slaves of these newly created masters. This procedure is repeated recursively till all nodes are assigned. Each node is slave for only one master, the one that paged it first. Each internal node of the tree is a master on one piconet, and slave of another master (its parent in the initial tree). In order to limit the number of slaves, they [15] observed that if a node in unit disk graph has more than five neighbors, then at least two of them must be connected. This observation is used to re-configure the tree so that each master node has no more than slaves. If a master node has more than slaves, it selects its two slaves and that are connected and instructs to be master of , and then disconnects from itself. Such branch reorganization is carried throughout the network . However, whether this approach will terminate is not proved in [15]. Tan et al. [14] proposed a similar method for single-hop network. In the second protocol [15], several roots are initially selected . Each of them then creates its own scatternet as in the first protocol. In the second phase, sub-tree scatternets are connected into one scatternet spanning the entire network. Notice that the tree topology suffers from a major drawback: the root is a communication bottleneck as it will be overloaded by communications between the different parts of the tree. Obviously, the root node in the tree-based scatternet is the bottleneck of the network and its congestion is , assuming that total traffic demand is a unit and is uniformly distributed. In addition, dynamic updating that preserves correct routing is not discussed in these protocols. Law, Mehta and Siu [9] described an algorithm that creates connected degree bounded scatternet in single-hop networks. The final structure is a tree like scatternet, which limits efficiency and robust-ness . A single-hop Bluetooth scatternet formation scheme based on 1-factors is described in [1]. However, piconets are not degree limited in that scheme. Salonidis et al. [13] proposed another topology construction algorithm recently. It first collects neighborhood information using an inquiry procedure, where senders search for receivers on randomly chosen frequencies, and the detected receivers reply after random backoff delay. Leader is elected in the process, one for each connected component. Leader then collects the information about the whole network, decides the roles for each node, and distributes back the roles. In other words, basically, it is a centralized approach. Thus, the solution is not scalable, and not localized. 30 Moreover, how to assign the roles is not elaborated in [13]. They also assume up to nodes in the network. Another centralized solution for single-hop networks, where the traffic between any pair of nodes is known a priori, is described in [10]. Sun, Chang and Lai [11] described a self-routing topology for single-hop Bluetooth networks. Nodes are organized and maintained in a search tree structure, with Bluetooth ID's as keys (these keys are also used for routing). It relies on a sophisticated scatternet merge procedure with significant communication overhead for creation and maintenance. Bluerings as scatternets are proposed in [4]. Ring structure for Bluetooth has simplicity and easy creation as advantage, but it suffers large diameter (i.e., the maximum number of hops between any two devices) and large number of piconets. The works are most related to our dBBlue scatternet construction method is [2] and [7]. Barriere, Fraigniaud, Narajanan, and Opatrny [2] described a connected degree limited and distributed scatternet formation solution based on projective geometry for single-hop networks. They assume that only slave nodes can act as bridges. They described procedures for adding and deleting nodes from the networks and claimed that its communication cost is and the computation cost is , where is the number of nodes in the network. The degree of the scatternet can be fixed to any , where is a power of a prime number. However, in their method, every node need hold information of the projective plane and the master node who has the "token" needs to know the information of the projective scatternet (which label should be used for the new coming master and which existing nodes need to be connected to it). However, the authors did not discuss in detail how to compute the labels for the new master and its slaves, and what will happen when the number of nodes reaches the number of nodes of a complete projective scatternets. Notice that our dBBlue scatternet can be easily transformed to support a Bluetooth network in which a piconet has any number of slaves, while the method in [2] can only support the piconet with slaves where is a power of a prime number. Moreover , the dynamic updating cost of dBBlue is at most . The construction of dBBlue scatternet is inspired by the method proposed by Fraigniaud and Gauron [7] for constructing a network topology for P2P environment based on de Bruijn graph. When a node joins the P2P network, it [7] randomly selects a node in the de Bruijn graph and then creates two children nodes of : one for and one for . This random selection of node cannot be applied to Bluetooth scatternet since it may create a de Bruijn graph with node whose degree is large than . It is not difficult to show that for Bluetooth scatternet, we can only afford the de Bruijn graph whose node label lengths differ by at most . In this paper, we proposed a novel method for assigning MAC addresses to nodes such that a self-routing is still possible during the updating procedures when node leaves or joins the network. The de Bruijn graph is used as backbone of the scatternet in our dBBlue structure. CONCLUSION In this paper, we addressed the problem of scatternet formation for single-hop Bluetooth based ad hoc networks, with minimal communication overhead. We adopted the well-known structure de Bruijn graph to form the backbone of the dBBlue scatternet. The diameter of the scatternet dBBlue is and we can find a path with at most hops between every pair of nodes without using any routing table. Moreover, the congestion of every node is at most . We discussed in detail the method to locally update the structure dBBlue using at most communications when a node joins or leaves the network. In most cases, the cost of updating the scatternet is actually . Our method can construct the structure dBBlue incrementally when the nodes join the network one by one. Previously no method can guarantee all these properties although some methods can achieve some of the properties. The dBBlue scatternet has lower dynamic updating cost than the structure proposed in [2]. Notice that, instead of having three statuses for the token, we can require that the scatternet is always in the status of expanding . Then the scenarios for updating the scatternet become simpler when nodes join or leave the network, but with a possible high cost of updating: more merging and splitting of piconets will occur. We are currently investigating the tradeoffs of the three approaches described in this paper by conducting simulations on different models of node joining and leaving the network. We are also investigating the scatternet formed based on butterfly structure [6] and compare their performance with the one described here. Notice that the butterfly structure has node degree at most , which maps exactly to the degree requirement by bluetooth piconet. REFERENCES [1] S. Baatz, S. Bieschke, M. Frank, P. Martini, C. Scholz, and C. Kuhl. Building efficient bluetooth scatternet topologies from 1-factors. In Proc. IASTED Wireless and Optical Communications WOC, 2002. [2] L. Barriere, P Fraigniaud, L. Narajanan, and J. Opatrny. Dynamic construction of bluetooth scatternets of fixed degree and low diameter. In 14th ACM-SIAM Symp. on Discrete Algorithms (SODA), pages 781790, 2003. [3] S. Basagni, R. Bruno, and C. Petrioli. Device discovery in bluetooth networks: A scatternet perspective. In Proc. IFIP-TC6 Networking Conference, Networking 2002, 2002. [4] F. Cgun-Choong and C. Kee-Chaing. Bluerings - bluetooth scatternets with ring structure. In Proc. IASTED Wireless and Optical Communications WOC, 2002. [5] N. de Bruijn. A combinatorial problem. In Koninklijke Nederlandse Academie van Wetenschappen, 49, pages 758764, 1946. [6] D.Malkhi, M.Naor, and D.Ratajczak. Viceroy: a scalable and dynamic lookup network. In Proceedings of the 21st ACM Symposium on Principles of Distributed Computing(PODC), 2002. [7] Pierre Fraigniaud and Philippe Gauron. The content-addressable network d2b. Technical Report Technical Report TR-LRI-1349 (also appeared in 22nd ACM Symp. on Principles of Distributed Computing (PODC)), 2003. [8] Jaap C. Haartsen. The bluetooth radio system. IEEE Personal Communications, 7:2836, 2000. [9] C. Law, A.K. Mehta, and K.Y. Siu. Performance of a new bluetooth scatternet formation protocol. In Proc. ACM Symposium on Mobile Ad Hoc Networking and Computing MobiHoc, pages 183192, 2001. [10] D. Miorandi and A. Zanella. On the optimal topology of bluetooth piconets: Roles swapping algorithms. In Proc. Mediterranean Conference on Ad Hoc Networks MedHoc, 2002. [11] C.K. Chang M.T. Sun and T.H. Lai. A self-routing topology for bluetooth scatternets. In 2002 International Symposium on Parallel Architectures, Algorithms and Networks (ISPAN '02), 2002. [12] C. Petrioli and S. Basagni. Degree-constrained multihop scatternet formation for bluetooth networks. In Proc. IEEE GLOBECOM, 2002. [13] T. Salonidis, P. Bhagwat, L. Tassiulas, and R. LaMaire. Distributed topology construction of bluetooth personal area networks. In Proc. IEEE INFOCOM, 2001. [14] G. Tan, A. Miu, J. Guttag, and H. Balakrishnan. Forming scatternets from bluetooth personal area networks. Technical Report MIT-LCS-TR-826, MIT, 2001. [15] G.V. Zaruba, S. Basagni, and I. Chlamtac. Bluetrees - scatternet formation to enable bluetooth based ad hoc networks. In Proc. IEEE International Conference on Communications(ICC), 2001. 31

scalable MAC assignment;scatternet formation;Low Diameter;ad hoc networks;self-routing;const updating cost;de Bruijn graph;Bluetooth;Network topology;Self-routing Scatternet;Bluetooth networks;Bruijn graph;equal traffic;easy updating;low diameter;single-hop

66

Development of E-commerce Statistics and the Implications

This text has analyzed the development of E-commerce in some developed countries such as Canada, U.S.A., Japan, etc and put forward several suggestions on how to set up the system of E-commerce in our country taking the national conditions of our country into account.

INTRODUCTION Since the 1990s, the rapid development of e-commerce has brought extensive, enormous and far-reaching influence on the economy of the countries all over the world. E-commerce has already become the contemporary chief trend of economic and social development. As representatives of advanced productivity of new economic period, the level of its development has already become important signs of measuring the modernization level and comprehensive strength of countries and cities; it has become important means to make changeover in the economic system and reform the style of economic, promote the upgrading of the industrial structure, promote the modernized level of the city and strengthen international competitiveness. So, the governments all over the world have paid close attention to the development of E-commerce Statistics. Though the development of informationization in our country is very quick, it still has great disparity with the developed countries for relatively late start. Our country is still in the interim facing the double task of informationization and industrialization at present. So, in order to carry on an instant, accurate Statistics to the development level of E-commerce and set up perfect E-commerce Statistics system, we must understand, absorb and bring in the theories and methods of E-commerce Statistics from the main foreign countries to make E-commerce Statistics become effective guarantee in leading and promoting e-commerce in a healthy way, combining social source and promoting national power. DEVELOOPMENT STATES OF E-COMMERCE STATISTICS IN THE WORLD We have chosen some representative countries in the world and analyzed the development of E-commerce Statistics in these countries. 2.1 Definitions of e-commerce in main Developed countries. The definition of e-commerce is the standard of carrying on E-commerce Statistics, but there are various kinds of definition of E-commerce in the world because of visual angles are different. So, it is necessary for each country to make a distinct, standard, practical, wide meaningful and measurable definition which should be suitable for each field and can be amenable to time. 2.1.1 Definition of e-commerce in OECD (Organization for Economic Cooperation and Development) There are broadly-defined e-commerce and the narrowly-defined e-commerce. The broadly-defined e-commerce means the activity of electronic transaction on item and service, no matter the transaction occurred between enterprise, family, government and other public or individual organizations. It uses network as a intermediary. Goods and service should be ordered on the network but the payment and goods service don't need to carry on the net; the narrowly-defined e-commerce is only referred to trade activity carrying on through Internet. It is worth pointing out that the source of OECD definition of e-commerce is the Canada official Statistical department. 2.1.2 Definition of e-commerce in Canada The e-commerce definition of Canada official Statistical department is: E-commerce is the transaction which based on the computer network, including the transformation of ownership, transformation of tangible and intangible assets right to use. It consists of B2B (business to business), B2C (business to consumer), B2G (business to government) and G2C (government to consumer). But the transaction taking place inside enterprises will not be included in the e-commerce Statistics. 2.1.3 Definition of e-commerce in the U.S.A. Definition of e-commerce in the U.S.A. is defined by the U.S.A. general survey bureau who divides e-commerce into three parts from angle of the overall situation: e-commerce infrastructure; electronic affairs; e-commerce. E-commerce infrastructure is the economic facility or equipment which is used for supporting electronic affairs or electronic transaction activities. Electronic affairs include the affairs managed by computer network in company, government and other non- profit organization. E-commerce refers to goods or service transaction activity completed on computer network. 70 2.2 Overview and Characteristic of Main Country's e-commerce Statistics 2.2.1 Overviews of e-commerce Statistical Surveys 2.2.1.1 Overviews of Canadian e-commerce Statistical Survey The e-commerce Statistics in Canada is an official activity that was presiding over by government and implemented concretely by State Statistics Bureau of Canada. Up till now, Canada has implemented four pieces of different e-commerce Statistics. a) "Net-banking operation and bank service Statistics survey on internet and e-commerce application in financial department&quot;, this investigation is an irregular survey; its respondents are enterprises of the financial field and its nature is a separate investigation; b) "Annually Statistical survey on internet application in family &quot;, is a fixed annual Statistical survey. It is a supplemented investigation and its respondents are families. c) "The Statistics survey on communication technology and e- commerce&quot;, is an irregular Statistical survey; it is a supplement investigation and the respondents are enterprises in &quot;the standard industry of North America classifies" d) "Annually Statistical survey on e-commerce and relevant technology &quot;, is a fixed annual Statistical survey; it is a supplemental investigation and the respondents are enterprises in "the standard industry of North America classifies" 2.2.1.2 Overviews of e-commerce Statistical survey in U.S.A. U.S.A. is one of the countries that e-commerce and e-commerce Statistical survey launched earliest in the world. The U.S.A. general survey bureau is the principal organ responsible for e-commerce Statistical survey. The annually Statistical survey adopted by U.S.A. general survey bureau is consisted of annual sample investigation of commerce, annual sample investigation of manufacturing industry, annual sample investigation of retailing business and annual sample investigation of service trade. The method taken in these investigations is dividing layer and sampling. The concrete method in e-commerce Statistical survey is joining the questions of e-commerce into the existing questionnaire except the annually sample investigation of manufacturing industry which is joining the supplementary questionnaire. These respondents investigating are enterprises and the enterprise e-commerce activity, business procedure and sales amount are investigated on the foundation of existing investigation. 2.2.1.3 Overview of e-commerce Statistical survey in other countries 2.2.1.3.1 Overview of e-commerce Statistical survey in Japan In Japan, the departments in chare of e-commerce Statistical survey is Statistics Bureau, Ministry of Internal Affairs and Communication, Japan, but other departments participate in the e-commerce Statistical survey such as Statistics Bureau of Cabinet, Statistics Bureau of Ministry of Economics and Industry, etc. So, there are more than forty kinds of official investigation on e-commerce which involve every aspect of e-commerce but have great differences in purpose, frequency and content. These investigations launch around three departments including enterprises, governments and families. 2.2.1.3.2 Overview of e-commerce Statistical Survey in S.Korean Statistics bureau of the S.Korean began the official e-commerce Statistical survey since April of 2000. The investigation mainly concentrates on B2C (business to consumers) and B2B (business to business). The investigation on B2G (business to government) lags behind slightly, which began since the first quarter of 2001. 2.2.2 Characteristics of the e-commerce Statistical Survey in each country. a) The organizers carrying on e-commerce Statistical survey in the above-mentioned countries are all official departments, or implemented by cooperating with other relevant government departments (as Japan). The Statistical survey presided over by the government can not only strengthen its Fairness and dependability, but also give the survey authoritativeness. b) The investigations almost are not specially but supplementary. The main reasons are high cost of special investigations and not perfect e-commerce Statistical systems of each country which have not reach the level of special investigation. c) The above-mentioned countries confirm the content of investigation not only consulting the content that OECD recommends, but also considering the development level and characteristics of the national e-commerce. It is worth pointing out those indexes of Statistical survey in Singapore, Canada and U.S.A. are comprehensive and have involved the preferential investigation content that OECD recommends. d) Most investigations take the annually survey as the core, but there also are monthly, quarter, general survey and irregular surveys. The industries included in monthly and quarterly investigation are not more than on generally, such as &quot;monthly trade sample investigation of retail business&quot; in U.S.A. and &quot;the investigation on family consumption trend&quot;, etc. e) Most countries adopt the sample investigations, but other method as census and census combine with sample investigation are also adopted. The method of sampling is mainly used and the following two kinds are used less. IMPLICATIONS OF E-COMMERCE STATISTICAL SURVEY IN OUR COUNTRY The e-commerce in our country is still in the elementary stage and the e-commerce Statistical survey is just start too. There are just some semi-official or unofficial departments and organization trying to carry on e-commerce Statistical survey but not a formal, overall, official survey on e-commerce in our country. For instance: &quot;Statistical Reports on the Internet Development in China&quot;, &quot;CII research and calculating on e-commerce total index system in China&quot;, &quot;Statistical survey on intranet and e-commerce development level&quot;, &quot;investigation on e-commerce developing in enterprises &quot;, etc. Most of above mentioned investigations are irregular, even once only, lack unified consideration and can't form a system except &quot;Statistical Reports on the Internet Development in China&quot; which hold regularly and establishes its own system to a certain 71 extent. Meanwhile, the unofficial survey is very apt to the systemic deviation and utility nature for it is not mandatory; even affect the fairness, accuracy and representativeness of the investigation result. 3.2 Implications of e-commerce Statistical survey in China According to the experience of some foreign countries that carrying on e-commerce Statistical and the development of e-commerce Statistical in our country, we consider that if we want to set up a comparatively perfect e-commerce Statistical survey system, we should accomplish the following several points at least: 3.2.1 Attach importance to the definition of e-commerce The kind, range and respondents are all fixed according to the definition of e-commerce which is prerequisite of e-commerce Statistical survey. There is not an authoritative definition of e-commerce in our country so that the key problem we met is how to define e-commerce when we carrying on the e-commerce Statistical survey. We consider that open principle should be followed when defining e-commerce according to its characteristic of appearing late and excessive growth, in order to perfect it constantly with the development of e-commerce. 3.2.2 The government should take charge of e-commerce Statistical survey We could understand the development of e-commerce prompt and accurate, find the questions existing in e-commerce and predict the development trend according to the e-commerce Statistical. It is obvious that e-commerce Statistical is important to the sound development of e-commerce. E-commerce could be promoted by just and accurate Statistical survey but the unilateral and utilitarian Statistical survey will mislead even hamper it. However, the e-commerce Statistical survey of our country lacks the authoritativeness and mandatory at present even affected the fairness and accuracy of Statistical survey. So, the Statistical survey of e-commerce in our country should be included in the official Statistical development plan as early as possible and we should set up the official survey system of e-commerce in order to make the e-commerce Statistical survey authoritative and promote the development of it. 3.2.3 Accelerate the research of the e-commerce Statistical theory The problem which should be considered first in research on Statistical theory of e-commerce is to keep the continuity with traditional Statistical. The e-commerce Statistical is not produced without foundation after all but is the extension on the network of traditional Statistical so that the basic theories of traditional Statistical are still suitable for the e-commerce Statistical survey. Secondly, we should make further research on Statistical method, Statistical caliber and Statistical range of e-commerce, and then set up the index system of e-commerce Statistical as soon as possible. Moreover, e-commerce has the characteristics of crossing over the limit of region. We should try our best to keep the harmony with the world on research in e-commerce Statistical theories for the overall and perfect system of e-commerce Statistical need the joint efforts of countries all over the world. 3.2.4 The service of e-commerce Statistical survey should be comprehensively and pointed. The e-commerce Statistical survey should serve not only for the macroscopically strategic policies of countries but also for the micro operation of enterprises. Meanwhile, there should be different surveys to conform to the different respondents in order to offer the personalized service of the Statistical survey. Only in that way can we offer the good development environment for e-commerce and reflect the value of e-commerce Statistical survey. REFERENCES [1]. Seminar of &quot;research on e-commerce Statistical survey and application&quot;. &quot;Statistical Surveys and Application of E-Commerce in Canada&quot; [J]. China Statistics, 2003, 3 2003, 4 [2]. Seminar of &quot;research on e-commerce Statistical survey and application&quot;. &quot;Survey of e-commerce development in S.Korean&quot;. [J]. China Statistics, 2003, 5 [3]. Seminar of &quot;research on e-commerce Statistical survey and application&quot;. "Statistical Surveys and Application of E-Commerce in Japan&quot; [J]. China Statistics, 2003, 6 [4]. Seminar of &quot;research on e-commerce Statistical survey and application&quot;. "Statistical Surveys and Application of E-Commerce in the U.S.A.&quot; [J]. China Statistics, 2003, 7 [5]. &quot;Research paper of e-commerce development all over the world&quot;, translated by Juanying Zhu, Bingzhi Yang United Nations Trade and Development Board [M].2003. [6]. &quot;The application of IT in Statisticals&quot; Feng Cui, [M] Lixin Accounting Publishing House,2003 72

Development stage;Definition of E-commerce;Survey;Authority;Statistics;Statistical methods;Measurement;Implications;E-commerce Statistics;Statistical Survey;China;E-commerce

67

Development through communicative action and information system design: a case study from South Africa

Many authors have recognised the importance of structure in shaping information system (IS) design and use. Structuration theory has been used in IS research and design to assist with the identification and understanding of the structures in which the IS is situated. From a critical theoretical perspective, focusing on the Habermas' theory of communicative action, a community based child health information system was designed and implemented in a municipality in rural South Africa. The structures which shaped and influenced the design of this IS (the restructured health services and social tradition) are explored and discussed. From this case study the implications of using IS design as a developmental tool are raised: namely the development of a shared understanding, the participation of key players and the agreement on joint action.

INTRODUCTION Many authors [Walsham and Sahay 1996; Walsham and Han 1991; Jones 1997; Rose 1999; Orlikowski 1992; Orlikowski and Baroudi 1991; Orlikowski and Robey 1991] have recognised the importance of structure in shaping information system (IS) design and use. Structuration theory has been used in IS research and design to assist with the identification of the structures in which they are situated. Using this meta-analysis tool, information systems have been used to redefine and/or reinforce some of these structures. The IS design process is particularly important, not just in shaping the structures, but also in terms of understanding what structures exist and how they were formed. Critical approaches to IS examine those structures with the perspective of questioning and changing some of them. Critical social researchers seek to emancipate people by finding alternatives to existing social conditions as well as challenging taken-for-granted conditions. In particular, Habermas [1987] examines communication and how through striving for an ideal speech situation these structures can be challenged. In the process of IS design communication is especially important, as is who participates, and how. In this paper the author explores the existing structures which have contributed to the accessibility, or as the case may be inaccessibility, of the health services in the Okhahlamba municipality, KwaZulu-Natal, South Africa. Through the design of the community-based child health information system these structures were explored and addressed throughout the design process. Communication and participation were integral to the process, as well as the recognition of the importance of the context in which the system is designed. The rest of this paper is structured in the following manner. The following section looks at what is meant by structure, the process of structuration and its application to IS design. The third section looks at critical social theory in IS design, in particular Habermas' notion of communicative action. The fourth section outlines the existing structures in a community in KwaZulu-Natal that were important in shaping the IS design process. The fifth section explores how the process of IS design acknowledged and challenged these structures and the last section discusses the implications for IS design as a developmental tool. Author Addresses: Elaine Byrne, School of Public Health, University of the Western Cape, PBag X17, Bellville, 7535, South Africa, elainebyrne@telkomsa.net 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 the copies are not made or distributed for profit or commercial advantage, that the copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than SAICSIT or the ACM must be honoured. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, or to redistribute to lists, requires prior specific permission and/or a fee. c 2003 SAICSIT Proceedings of SAICSIT 2003, Pages 8392 84 Elaine Byrne Figure 1. Dimensions of duality of structure. IS DESIGN AND STRUCTURATION In this paper structure is regarded as 'Rules and resources, recursively implicated in the reproduction of social systems. Structure exists only as memory traces, the organic basis of human knowledgeability, and as instantiated in action' [Giddens 1993] p377. That is, through action, based on rules and resources in peoples' minds, structures in society are produced and reproduced. The rules and resources drawn upon in the production and reproduction of action are simultaneously the means of system reproduction (this is what Giddens refers to as the 'duality of structure'). The rules can be viewed as generalised procedures of action and human agents are aware and knowledgeable of these rules, but may not know what the outcome of that action will be because action can have both intended and unintended consequences. The resources are both authoritative (coordination of the activity of human agents) and allocative (control of material aspects of the natural world), so both human and material resources are included. The process of structuration involves knowledgeable actions of human agents discursively and recursively forming the sets of rules, practices and routines which, over time and space constitute structure. Thus agents and structures are not seen as independent, but as a duality whereby structure is relied upon in human actions, and in so doing structures are produced or reproduced. Over time these social practices becomes reasonably stable and routines develop. Giddens [1993] p29 breaks down social structure and human interaction into three dimensions which are interlinked by three modalities as illustrated in Figure 1. When human actors communicate, they draw on interpretative schemes to help make sense of interactions. At the same time those interactions reproduce and modify those interpretative schemes which are embedded in social structure as meaning or signification. Similarly the human actors allocate resources through use of power, and produce and reproduce social structures of domination. Moral codes or norms help determine what human agents can sanction and thus produce and reproduce social structures of legitimation. It is useful to separate structure and interaction into these three dimensions for analysis of structure, but the dimensions are interlinked. [Rose 1999] The design and use of information systems are shaped by the very structures within which they are situated, but IS can also be used to help define and redefine these structures. By exploring each of the above dimensions in the process of IS design, IS design can be used as a tool for development by refining the structures to include the views and values of those currently disadvantaged by the existing structures. Through a participative and reflective process in IS design, cultural and traditional norms which influences human action can be explained, understood and addressed. The design process, and the IS itself, can improve communication and encourage reflection and change interpretative schemes. Through the process of IS design and reflecting on the situation the excluded can be empowered, which redefines the power and resource structures. In summary IS design can define and refine structures by understanding and incorporating all the dimensions of the duality of structure in the design process. Structuration theory has been used quite widely in IS. Rose [1998] conceptualises the use of the theory in IS for three different purposes: analyse; theorise and operationalise. Walsham and Han [1991] analyse literature under topics of operational studies, meta-theory and specific concepts used, as well as outlining structuration Proceedings of SAICSIT 2003 Development through communicative action and information system design 85 theory. Jones [1997] analyses the use of structuration theory in an attempt to reconstruct theory to accommodate technology. He further explores the application of the theory as an analytical tool, the use of theory as a meta-theory , and use of concepts from the theory. In an attempt to theorise aspects of the IS field using structuration theory Orlikowski and Robey [1991], apply the fundamentals of structuration theory to help understand the relationship between information technology and organisations. In a later article Orlikowski [1992] developed her structurational model of technology to understand the relationship between information technology and institutions. She recognises that technology cannot determine social practices, but can condition them and that technology in conditioning social practices is both facilitating and constraining. In terms of empirical studies Walsham [1993] provides a number of case study analysis which cover issues of IS strategy, development, implementation and evaluation in three different organisations. Walsham and Sahay [1996] use structuration theory, with actor-network theory, to investigate problems in developing Geographical Information Systems in an Indian government department. In a similar manner this paper, from a critical social perspective, uses structuration theory to highlight two key aspects of existing structure which were addressed in and affected the process of designing the IS. The meaning of a critical social perspective is provided in the next section before section 4 decribes the key structural aspects of the case study. CRITICAL SOCIAL THEORY AND IS DESIGN Critical social researchers by their very presence influence and are influenced by the social and technological systems they are studying. 'For critical social theorists, the responsibility of a researcher in a social situation does not end with the development of sound explanations and understandings of it, but must extend to a critique of unjust and inequitable conditions of the situation from which people require emancipation'[Ngwenyama and Lee 1997]p151. Critical social theorists seek to emancipate people; they are concerned with finding alternatives to existing social conditions as well as challenging taken-for-granted conditions. Critical social theorists view people, not as passive receptacles of whatever data or information that is transported to them, but as intelligent actors who assess the truthfulness, completeness, sincerity, and contextuality of the messages they receive. Adopting a critical social theoretical perspective to IS design is not new. In relation to IS research Ngwenyama [1991] gives an in-depth treatment of critical social theory. Ngwenyama and Lee [1997] approach research on communication richness in computer mediated communication from a critical social theoretical perspective. Hirschheim and Klein [1994] deal with a critical approach to qualitative research. Habermas [1987] suggests that critical social theorists should initiate a process of self-reflection among human actors, but it is only participants in the community that can select the appropriate political action. His theory of communicative action notes that all social action assumes a basic set of norms. These norms allow all actors to express themselves fully and openly. They also imply that all actors accept the outcome of open rational argument. According to the theory of communicative action, breakdowns in communication occurs when actors fail to adhere to these norms. There have been numerous studies which refer in particular to the theory of Habermas. Lyytinen [1992] has explored the theory of Habermas to analyse systems development. Hirschheim et al. [1996]using Habermas' theory of communicative action propose a framework for the intellectual trends in IS development research. In this study Habermas' theory of communicative action and the notion of 'the ideal speech situation' is used to explore how effective striving for its attainment is as a transformation strategy. My study uses aspects of critical social theory to examine how community action can be strengthened or changed by exploring the structures which enable or constrain that action. Communication, power and norms are key in trying to grasp an understanding of that action. Fundamental to this exploration is the belief that as intelligent and knowledgeable agents, human actors can, within limits, choose to act in accordance with or against societal norms. SITUATION IN OKHAHLAMBA MUNICIPALITY, UTHUKELA DISTRICT, KWAZULU-NATAL, SOUTH AFRICA The existing district health information system in South Africa excludes children and adults that cannot, and/or do not, access the services at the health facilities (clinics, community centres, mobiles and hospitals). Those who are most vulnerable and socially excluded, and need the health support systems the greatest, are the very ones not accessing the health services. Policies are formed and resources allocated to the community based on the information they recieve. Since the vulnerable are excluded from the formal IS they are further and systematically excluded from these policy and resource decisions. With the impact of HIV/AIDS children have increasingly become an excluded and more vulnerable group. This exclusion and vulnerability of children can be tackled on two interconnected levels. The first is through the creation of awareness of the situation of children and the second through the commitment and action of government and society to address this situation. The first can be supported by designing an information system Proceedings of SAICSIT 2003 86 Elaine Byrne for action - an information system that can be used for advocating and influencing decisions and policies for the rights of these children. So IS design can be used as a developmental tool. Since protecting and improving the health of the children of the entire district is the aim of the district health system, research was conducted on how to develop a community-based health information system that could support a comprehensive district health information system. The research was conducted in Okhahlamba as a component of the child health programme of the uThukela district child survival project and the department of health. Okhahlamba is a municipality of the uThukela district lying in KwaZulu Natal on the eastern coast of South Africa. The primary objective of developing a community-based information system is to assist community members in their decision-making regarding the health of their children. On a secondary level it aims to establish interfaces with the formal health facility information system to enable district managers to use information from the whole district to make informed decisions and policy changes. After a review of the district's health information system and a community meeting on monitoring and evaluation community members, as well as district government staff, recognised their need for a community-based child health information system. To understand what the information needs were, who should be involved in the information system and the format the information should be communicated in a total of 10 interviews, 16 focus group discussions and 1 meeting took place between July and September 2002. From the field work there was a greater understanding around the meaning of 'well-being' and 'at-risk' for a child, what factors/practices contribute to these situations, how the situations can be measured and, based on what action can be taken, who the information should go to. Consequently a community-based child health information system has been integrated into the district health information system. In this section two key aspects of structures which address, or have contributed to, the exclusion of children are outlined, namely restructuring of health services and status of child health, and social traditions. The first aspect provided an opportunity for change and reflection on the current role and function of the IS whilst also providing an understanding of the exclusion of segments of the population. The second aspect again provides an understanding of the position of women and children in society which impacts on IS design as well as presents some challenges in the design process. [For more details of the child health programme and the research see [uThukela District Child Survival Project 2002; 2000a; 2000b; 1999a]] 4.1 Restructuring of health services and status of child health After 1994 the national Health Plan for South Africa and the Reconstruction and Development Programme outlined that a Primary Health Care (PHC) approach is the underlying philosophy for the restructuring of the health system. Crucial to this is the role of the community in the development of a district health system emphasising the movement from a traditionally vertical curative based health system to a newer client centred and preventive based health system. In addition, more recently, there has been the move towards the decentralisation of health service delivery (along with other basic social services) to local authorities from the department of health. The newly established structures, such as the community health committees and community health forums, have meant a renegotiation of roles and responsibilities at the district level. This requires active communication between the parties involved to ensure consent on the new roles and responsibilities of all local government staff [uThukela District Child Survival Project 2000b; 2000a]. Since 1994 children have benefited from the move to PHC. However the free health care policy is not without its fair share of problems. Due to the emphasis on PHC, there has been a 30% increase in clinic attendance and a 2% increase in hospital attendance in the province of KwaZulu-Natal. The additional drugs and personnel needed for the increased attendance at the clinics was not budgeted correctly. As a result the quality of services in terms of shortages of personnel and drugs has been compromised as well as putting severe strain on the budget. Clinics in particular have struggled to accommodate the increased number of clients. Clients also complain that hospital-based health workers are often unsympathetic to their needs. [uThukela District Child Survival Project 1999b] Poorer children living in rural areas have poorer access to PHC facilities than children living in the wealthier more urbanised areas. They have greater distances to walk and fewer health personnel to cater for them. KwaZulu-Natal is one of two provinces with especially poor client-to-clinic ratios (23,000 clients per clinic) and in 1995 only 54.3% of households in KwaZulu-Natal were within 5 kms of medical care, the second lowest in the country.[Crisp and Ntuli 1999] Child health indicators point to the lingering effects of apartheid's racial, geographic and socio-economic policies. Just over half of all children aged 12-23 months in KwaZulu-Natal are not immunised, though 62.2% have their road to health cards. This indicates at least one contact with the health services, but this contact was not sustained as the immunisation schedule has not been completed. The infant mortality rate for KwaZulu-Natal has been estimated at 52.1/1000 and the under-five mortality rate at 74.5/1000.[Crisp and Ntuli 1999] Proceedings of SAICSIT 2003 Development through communicative action and information system design 87 This situation is exacerbated by disparities in access to basic infrastructure. Access to potable (drinkable) water and sanitation are often critical to improving child health outcomes. The government has however committed itself to increasing access to water and sanitation. In spite of two major dams and several springs in the area, a serious shortage of water for agriculture and clean drinking water has impacted nearly every household, and influenced the health status of the area. The cholera epidemic in 2001 is evident of this poor access. A situational analysis for the Okhahlamba municipality completed in July 1998 estimates that only 25% of the population live within 15 minutes walking distance of safe water, and only 25% have adequate sanitary facilities. Transport remains poor, particularly during rains when rivers become impassable. [uThukela District Child Survival Project 1999b] 4.2 Social traditions Strong Zulu cultural and traditional values exist in the Okhahlamba municipality. Traditional leaders are highly respected, though there is some controversy over the roles and powers being eroded with the formation of the new local government structures. Grandmothers and traditional healers are often the first persons to be consulted in times of illness and many locally available remedies and treatments are used and practiced. Grandmothers can have quite a powerful decision-making influence at household level. However, women in general tend to be dependent on males for income and have very little access to independent means of livelihood. Household responsibilities also make women subject to 'time poverty' that is, it is not uncommon for most women in this rural area to work ten hours a day, making it a hardship to travel to seek health care for themselves or their children. Much of each day involves several hours of strenuous manual labour, hauling water and firewood, and performing agricultural work. Women, including mothers, grandmothers and older 'girl children', are predominantly responsible for childcare [uThukela District Child Survival Project 1999b]. However if the health-seeking or care decision involves any financial decisions the head of the household, which is usually a man, will need to be consulted in order to make the final decision. This process often causes a delay in a child attending a clinic as money for transport and alternative child care for the siblings would need to be sourced. Through the existing patriarchal social system women are particularly at risk from HIV/AIDS. These factors include: sexual subservience to men, higher risk of transmission with the migrant labour of partners to cities; differential access to information and resources for prevention, and; women often remain with spouses who are HIV positive rather than vice-versa. Women in their twenties have the highest rate of HIV infection nationally, but between 1997 and 1998 the HIV prevalence among teens attending antenatal clinics jumped over 65%, from 12.7% to 21%. With high teenage fertility rates this picture is unlikely to change in the near future. In 1998, the provincial fertility rate was 3.3%, and the provincial teenage pregnancy rate was 13.8%. In Okhahlamba/ Mtshezi municipalities the average teenage pregnancy rate for young women delivering in facilities in 1999 was 22.9%, significantly higher than the provincial rate [uThukela District Child Survival Project 1999b]. Children are particularly susceptible to the ravages of the HIV/AIDS epidemic through high rates of mother to child transmission and an increasing number of AIDS orphans and consequent child headed households. ASPECTS OF THE PROCESS OF DESIGNING A COMMUNITY-BASED INFORMATION SYSTEM IN OKHAHLAMBA, KWAZULU-NATAL, SOUTH AFRICA One of the fundamental steps that needed to be addressed before addressing the situation of children, and how this was reflected in information systems, was a paradigm shift. It required a shift from the older focus on curative centre based service delivery to the newer health services approach which focuses on prevention, clients and quality. To support this paradigm shift the project adapted a new approach of transformational thinking, or future focussed approach, developed in the business sector, but which is also being integrated in health systems. The approach focuses on working towards holistic well-being for all, rather than just solving health associated problems. Through community meetings and discussions the community determined a vision for their children: 'To achieve optimal health, growth, development and well-being of children within the family and community in the uThukela Health district'. The implications of the paradigm shift for IS was that though it was important to measure children's physical condition, it was also important to measure how far towards our vision we are. So instead of saying 80% of our children are immunised, we would say that we still need to immunise 20% of our children. This approach reflects what we still need to do to attain our vision and thus, hopefully, stimulate action. Adopting a forward looking perspective also stresses the importance of the context we are presently in and the importance of measuring changes in that context. Monitoring the context and acting based on that information, should lead to a situation where most children in the future would find themselves in a state of 'well-being'. Proceedings of SAICSIT 2003 88 Elaine Byrne 5.2 Sharing of information with key actors If people are to act or reflect on information received that information needs to be relevant and communicated in a culturally sensitive and appropriate manner. In terms of a community-based information system for children an important step in the process of the IS design was who should participate in the process. The main role players and duty bearers 1 need to be included as it is them who are in the best position to change or influence the context in which the child is placed. In the case study these key people were: the community health workers, parents, family members, early childhood and creche teachers, home based carers, caretakers, social workers, health facility staff, clinic health committees, councillors, government officials and staff from external organisations. This indicates that a multi-leveled and multi-sectoral group affects the situation of children at community level. What was also important was a common understanding by all parties on what was meant by 'well-being' and 'at-risk' as the monitoring of these situations and conditions would be important if we were to measure whether we were on the right track to attaining our vision. Meanings of 'well-being' and 'at-risk' were gathered through focus-group discussions, interviews and meetings with all the role players and duty bearers. This common understanding was translated into common data definitions in the community-based, as well as in the health facility, information system. A review of the existing data sources and flows was conducted based on the assumption that information flows are a key element of dialogue between providers and consumers of health services. One important conclusion from this review was that some of the data collected through the current district health information system is valid and useful, but is not getting to the people who can act upon or use it. As one project leader mentioned we need to look at how data is flowing and the possibility of establishing 'feedback pathways' for this data. There are many of these pathways at different levels, but the one between community based workers and community forums is core for a community-based health information system. This level of feedback was entirely absent from the district health information system in Okhahlamba. It is also interesting to note what was absent from existing data sets, yet what key role players and duty bearers felt were important in monitoring the situation of their children. Data items relating to the context in which a child is being reared are mostly excluded. Many of the current indicators focus on the condition the child is currently in, such as having immunisation or not, and not the context that caused the child to be in that situation, such as no caregiver to take the child to the clinic. But exclusion is a process and to prevent the child becoming excluded requires analysing the situation of the child throughout that process. Measures for context, such as happiness, playfulness and communication are more intangible and therefore difficult to develop as data items. However through the new observation tools developed by and for the community health worker these measures are now included. So data items on the presence of a caregiver, drug and alcohol abuse, cleanliness of the household for example, are now included as indicators of 'at-risk'. This observation tool is used as part of the dialogue between the health worker, who is a trusted and respected family advisor, and the household. The results from the aggregated monthly data is shared through role play, song, dance, drawings and histograms in the community quarterly meetings. The act of sharing information establishes networks of people at community level who are responsible for the care of the children. These networks form the basis for communication. 5.3 The communication loop In terms of capacity to act, or to make decisions, most respondents, in the research undertaken, felt that they could act if given appropriate information and if key role players were included in the communication loop with one another. The visioning exercise started a communication process, but this needed to be developed into more formal communication structures. Communication was needed with other levels of government. Building on the recent development of clinic health committees and the governments' appointment of the community health workers in the KwaZulu-Natal province, communication loops were developed. These loops are described below at three levels: household, community, and district. --Household level: Following on from a discussion on how to measure the more intangible measures a standardised observation checklist was developed. The checklist is used as a communication tool with household members. Based on the community health worker's assessment a number of choices or options to solve any of the problems identified is given to the household. The community health worker could facilitate the choices, such as contact with certain services, if requested to do so, but the final decision lies with the household. The assessment is used as an empowering tool, rather than as a means of inspection. These visits assist the child caregivers in 1 Role players have a role to play in children's lifes, but duty bearers are those people who are responsible and obligated to fulfill childrens' rights Proceedings of SAICSIT 2003 Development through communicative action and information system design 89 terms of their knowledge of child care and health seeking behaviour within their household. The visits also provide the mother or caregiver with a mediator between them and health facilities as well as a mediator between them and their family. Therefore issues of access to basic social services could be addressed. --Community level: The community health workers, with the assistance of their supervisors (community health facilitators), conduct village health days for discussion of broader issues affecting the community served by the clinic. Bar graphs, role-plays, song, poetry and dance are used as these methods seem to work very well. These meetings form the quarterly community health meetings, that were suggested in the course of the field work. Members of the community and the clinic health committee, health facility staff, community health worker, school children and other key people attend the meetings. More people have now access to the information they requested and in a format that is easy to understand. The village health days also provide a forum for reflection and discussion. --District level: Communication and information flows between community and district involves combining data from various sources to provide a comprehensive database for the district. Important for the collation of this data is the use of the same data definitions in the different data sources. This collation is done through the district information officer as her office already receives this data from the different sectors. A summary of the district data is distributed every quarter. The content of the summary sheet is regularly determined in consultation with the clinic health committees and through feedback on the village health days. Existing local government structures, community and clinic committees, have already established clear communication channels with higher levels of local government. The feedback from these meetings would be sent through these structures when needed. Thus a comprehensive picture of child health in the district is achieved. In summary, with the restructuring of the health services there was the need for a paradigm shift, before addressing the review and design of a community-based child health information system. This shift was from the older more curative health service approach to the newer client and service focused approach of primary health care. With the newly established local clinic and community health committees this offered an opportunity of new people coming into the health services with a new vision and who were also willing to be involved in the IS design process. Furthermore the newly formed local government structures have established clear communication channels with higher levels of local government. The feedback from quarterly community health meetings could be sent through these channels and forms part of the health information flow and communication loop. Challenges around the position of women in society impacted on decisions regarding participation. However as women are the main carers of children they were involved in the process without any question. Furthermore as some of the key positions in the community are occupied by men it was also felt that they needed to participate in the design process as their positions were influential in terms of the situation of children in the community. The dialogue initiated in the design process continues through the community health quarterly meetings which provide an opportunity for dialogue to take place at community level. At the household level the community health workers role is to empower the household in its health seeking and caring practices. This is done through household visits and providing the appropriate education at the appropriate time, for example if a child has diarrhoea the conversation would be around what to do for the child with diarhoea. The community health worker also plays the role of mediator - mediator between households and the community forum and also between the caregiver and the rest of the family. Through the supportive role of the community health worker the position of women and children will not change in society, but their views on health and the care of children will be supported and heard. The process of IS design in the case study supported and questioned two key aspects of structure, namely the restructuring of health services and the status of child health and social traditions and their implications on the process of the design. The next section explores what implications the use of such an approach has for IS design. DISCUSSION IMPLICATIONS FOR IS DESIGN The implications for IS design have been categorised into three main areas: the need for a shared understanding, the need for participation of key people and the need for agreement on joint action. 6.1 Shared understanding If health IS design is to be used in a developmental context their needs to be agreement reached between health care deliverers and those who receive the services on the design and the purpose of the health service. From our case study the importance of having a common vision for the health services was seen as an important first step in this direction, especially given the restructuring of the health services and the adoption of a PHC approach. Creation of this vision and shared understanding necessitates communication between the designers of the system, the users of the IS as well as the users of the health system. The process of IS design is important for Proceedings of SAICSIT 2003 90 Elaine Byrne establishing the relationship between the users and providers of health care, as reaching agreement on subsequent action that needs to take place involves both parties working together. The objective of communicative action is to achieve mutual understanding. In this case study mutual understanding on a vision for the children, how to measure our progress to this vision and who needs to be involved in that process was made. IS design should be ' . . . concerned with achieving and maintaining mutual understanding . . . among all those who are involved in a coordinated organizational situation . . . Organizational actors involved in communicative action depend on a common language and a shared understanding of the organizational context in order to enact meaning from each other's communicative actions.' [Ngwenyama and Lee 1997]p158/9 IS use in developmental contexts can go beyond communicative action and be an enabler of discursive action. Discursive action is intended to achieve or restore agreement for collective action. It is ' oriented toward achieving or restoring agreement and redeeming validity claims. Discursive action is initiated when organizational actors need to achieve agreement for joint action. In such a situation, the individuals would generally engage each other in a debate of the issues until they agree on a course of action' [Ngwenyama and Lee 1997]p155. However there needs to be a common medium of communication, agreement on roles and responsibilities and terms and conditions set for means of discourse. A common understanding was needed in this case on what was meant by 'at-risk' and 'well-being' children and how to measure the situation of the child. The process of IS design can create an environment where people can express themselves, where understanding on various roles can be agreed to, where responsibility can be taken and where action using available information occurs. However unless we explore and change the structures in which a person operates, e.g. the position of children and women in society it is difficult for an actor to be able to engage either in reflection or in discursive action. 6.2 Participation of key players Reaching a common understanding between the users and providers of the health services is impossible without their joint participation. Participation of the excluded increases transparency and opens officials and other responsible parties to dialogue and wider scrutiny by the citizens they serve. Underlying power differences between different actors influences the interaction and negotiation between them (both within the community and between the community and outside groups) and this can influence whose 'interests' are explored and served in information systems. The social dynamics and power relationships that underlie and constitute the actual practice of the information system needs to be explicit. In this research the unequal nature of social relationships and positions between different actors and also institutions was recognised from the outset. Forums were established that suited the needs of the various groups. Discussions were also facilitated from people who were familiar with the area and who also had an understanding of the norms and values of that society. In the initial stages because of these differentials in status and roles within the community, groups comprising, for example, mothers, councillors, facility staff, met separately to discuss what they wanted for their children. These meetings were held in the local language and near the homes of the individuals. The community health worker formed the essential mediation role between the service providers and the clients. At a later stage representatives from the various groups met jointly to share the findings from the research and to discuss the way forward. Even with community participation communication does not always work smoothly, or in favour of children. Communication provides the means for exploring, affirming or denying norms, debating policies and practices, and discussing old experiences and new ideas. The situation of children will change only when action to improve that situation is taken. So the next step was to explore what will happen once the information has been shared. 6.3 Agreement on joint action:a multi-leveled and multi-sectoral approach Once the vision is formulated then the necessary action to attain that vision needs to be agreed to. Often this involves a multi-leveled and multi-sectoral approach. It also needs all key role players to be in communication with one another. It is not easy to challenge or change the institutions and systems established that support the status quo. In Okhahlamba the key role players that could act to change the situation of children were all identified. The most difficult task was achieving agreement by these role players on their action. Most of the confusion was over formal roles and responsibilities which had changed with the recent moves to decentralization of basic social services to local authorities, rather than an unwillingness to support one another. With this move the community health workers had also recently moved from a local non-governmental organization to the Department of Health and were confused over their reporting structures. The district department of health needs to hand over delivery of health services to local authorities, but the local authority does not have the human nor financial capacity to carry out this function. The volunteer clinic health committees are enthusiastic to support initiatives that Proceedings of SAICSIT 2003 Development through communicative action and information system design 91 will improve the situation of their children, but have only been formed recently. It was only after groups met one another and agreement was reached on their roles and responsibilities that agreement on the action, and who was responsible for that action, took place. The recent changes have provided an opportunity for inclusion of children on the agenda as many of the structures and systems are not, or have only recently, been formed. What was encouraging from the field work was that most people felt that they had the capability to act if they received the information. CONCLUSION It is increasingly recognised that globalisation also produces marginalisation. Castells [2000b; 2000a] argues that processes of globalisation are extremely selective, and various parts of the globe in both the developing and developed world run the potential of being excluded from this process. He uses the term 'fourth world' to describe this segment of society. Conditions of history and geography shape the access that groups and societies have to new information and communication technologies. Lack of such access can be exclusionary. Castells describes these processes to be systematic and can lead to further marginalisation and exclusion of societies. In information systems, and not just health information systems, the voices of communities - in particular women, children and youth - are not often heard, both within communities, between communities and between the other levels of society. When, where and how do they get the opportunity to express their needs and aspirations? How do they have the chance to identify and develop the skills and resources they need to address their problems? Where do they get the opportunity to express themselves or to exchange ideas and experience? In a sub-district in KwaZulu-Natal these questions have been addressed through a holistic approach to health information systems development. Some of the challenges for IS design is the need to focus both on the output, as well as the process. Attaining a common vision is fundamental if the system is to be used, but this involves the participation of different sectors and different levels of actors from the outset. It also offers some opportunities. Clarification over roles and responsibilities allows recognition and acceptance by duty bearers of the tasks they need to perform. This is a first step towards action. There was great enthusiasm by these key role players in the design process and a desire to work together. The community monitoring system in Okhahlamba was based on an understanding that people are intelligent and know what affects their children's and their own development. There is the need to co-design systems, processes and tools in IS design and obtain clarity on what we need to measure. IS design should be about facilitating a journey of development, rather than measuring the destination. Communication and participation, as well as the capacity to do so, are needed to strive towards Habermas' ideal speech situation. This is no easy task, as exclusion is built upon a system of norms, interpretative schema and facilities that systematically excludes segments of the population and country from the network society. Communication will not simply be improved by introducing a new or improved health information system. Even so, a process from visioning, developing skills and capacity and constructing a conducive environment can mean that IS design can be viewed as a development tool, as striving towards this 'ideal speech situation', even if this situation is not attained. ACKNOWLEDGMENTS I wish to thank all the people from uThukela district who assisted with the research. In particular thanks must go to the staff from the uThukela District Child Survival Project and the Department of Health, who assisted with the carrying out of the field research, the data analysis and the implementation of the information system. For support on the formatting and editing of this paper I am grateful for the assistance of Bob Jolliffe. I have also benefitted from the insightful comments from Sundeep Sahay on the various drafts of this paper. Financial support for the research was provided by a World Vision/USAID grant to uThukela District Child Survival Project. REFERENCES Castells, M. 2000a. The Information Age: Economy, Society and Culture: The End of the Millenium, 2 ed. Vol. 2. Blackwell Publishers. Castells, M. 2000b. The Information Age: Economy, Society and Culture: The Network Society, 2 ed. Vol. 1. Blackwell Publishers. Crisp, N. and Ntuli, A. , Eds. 1999. South African Health Review. Health Systems Trust, South Africa. Giddens, A. 1993. The constitution of society. Outline of the theory of structuration. Polity Press, Oxford. Habermas, J. 1987. The Theory of Communicative Action. MIT Press. Hirschheim, R. and Klein, H. 1994. Realizing emancipatory principles in information systems development: The case for ethics. MIS Quarterly 18, 1, 83109. Proceedings of SAICSIT 2003 92 Elaine Byrne Hirschheim, R. , Klein, H. K., and Lyytinen, K. 1996. Exploring the intellectual structures of information systems development: A social action theoretical analysis. Accounting, Management and Information Technology 6, 1/2. Jones, M. 1997. Re-Thinking Management Information Systems. Oxford University Press, Chapter structuration and IS. Lyytinen, K. 1992. Critical Management Studies. Sage Publications, London, Chapter Information systems and critical theory, 159180. Ngwenyama, O. 1991. Information Systems Research: Contemporary Approaches and Emergent Traditions . North Holland, Amsterdam, Chapter The Critical Social Theory Approach to Information Systems: Problems and Challenges. Ngwenyama, O. K. and Lee, A. S. 1997. Communciation richness in electronic mail: Critical social theory and the contextuality of meaning. MIS Quarterly 21, 2 (June), 145167. Orlikowski, W. 1992. The duality of technology: rethinking the concept of technology in organisations. Organisation Science 3, 3 (August). Orlikowski, W. and Baroudi, J. J. 1991. Studying information technology in organisations: Research approaches and assumptions. Information Systems Research 2 , 128. Orlikowski, W. and Robey, D. 1991. It and the structuring of organisations. Information Systems Research 2, 2, 143169. Rose, J. 1998. Evaluating the contribution of structuration theory to the is discipline. In Proceedings of the European Conference on Information Systems . Rose, J. 1999. Towards a structurational theory of is - theory development and case study illustrations. In Proceedings of the 7th European Conference on Information Systems . Copenhagen. uThukela District Child Survival Project . 1999a. Final evaluation report. uThukela District, KwaZulu-Natal, South Africa, unpublished. uThukela District Child Survival Project . 1999b. Knowledge, practice and coverage survey. UThukela District Child Survival Project, KwaZulu-Natal, South Africa. uThukela District Child Survival Project . 2000a. Cs xv detailed implementation plan. uThukela District, KwaZulu-Natal, South Africa, unpublished. uThukela District Child Survival Project . 2000b. Integrated managment of childhood illness situational analysis. uThukela District, KwaZulu-Natal, South Africa, unpublished. uThukela District Child Survival Project . 2002. Mid term evaluation report. uThukela District, KwaZulu-Natal, South Africa, unpublished. Walsham, G. 1993. Interpreting Information Systems in Organisations. Chichester, John Wiley. Walsham, G. and Han, C. K. 1991. Structuration theory and information systems research. Journal of Applied Systems Analysis 17, 7785. Walsham, G. and Sahay, S. 1996. Gis for district-level administration in india: Problems and opportunities. International Journal of Geographical Informaiton Systems 10 , 385404. Proceedings of SAICSIT 2003

communicative action;critical social theory;moral codes or norms;community information systems;information system design;Structuration theory;interpretative schemes;critical social theory in IS design;conducive environment;community monitoring system;marginalisation;health information systems;duality of structure;structuration theory;the ideal speech situation

68

Diagnosis of TCP Overlay Connection Failures using Bayesian Networks

When failures occur in Internet overlay connections today, it is difficult for users to determine the root cause of failure. An overlay connection may require TCP connections between a series of overlay nodes to succeed, but accurately determining which of these connections has failed is difficult for users without access to the internal workings of the overlay. Diagnosis using active probing is costly and may be inaccurate if probe packets are filtered or blocked. To address this problem, we develop a passive diagnosis approach that infers the most likely cause of failure using a Bayesian network modeling the conditional probability of TCP failures given the IP addresses of the hosts along the overlay path. We collect TCP failure data for 28.3 million TCP connections using data from the new Planetseer overlay monitoring system and train a Bayesian network for the diagnosis of overlay connection failures . We evaluate the accuracy of diagnosis using this Bayesian network on a set of overlay connections generated from observations of CoDeeN traffic patterns and find that our approach can accurately diagnose failures.

INTRODUCTION When failures occur in Internet overlay connections today, it is difficult for users to determine the root cause of failure. The proliferation of TCP overlays such as content distribution networks and HTTP proxies means that frequently network communication requires a series of TCP connections between overlay nodes to succeed . For example, an HTTP request using the CoDeeN[9] content distribution network first requires a TCP connection to a CoDeeN node and then a connection from a CoDeeN node to a server or another CoDeeN node. A failure in any one of the TCP connections along the overlay path causes the user's HTTP request to fail. If the user knows which TCP connection failed, then they can take appropriate action to repair or circumvent the failure. For instance, 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. SIGCOMM'06 Workshops September 11-15, 2006, Pisa, Italy. Copyright 2006 ACM 1-59593-417-0/06/0009 ... $ 5.00. if they know that the connection from the proxy to the server failed, then they complain to the web server administrator. On the other hand, if the user/proxy connection fails, perhaps they can try connecting to the proxy using a different ISP. If multiple overlay paths exist between the source and destination, nodes and applications may also use this type of diagnostic information to automatically recover or route around failures[1]. Unfortunately, accurately determining which TCP connection in an overlay connection has failed is difficult for end users, who typically do not have access to the internal workings of the overlay. Commercial overlay networks such as Akamai typically do not reveal details of connection failures to users, and the diagnostic tools available to users today are frequently inadequate. Active probing techniques such as tulip[7] and Planetseer[11] frequently cannot provide accurate information due to firewalls and packet filtering. Furthermore, active probing can be costly both in terms of network resources and time, and cannot diagnose the many transient TCP failures that begin and end before one can complete a probe[11]. Additionally, one must take care when using active probing for diagnosis because they may concentrate network traffic at points of failure and trigger intrusion detection systems. Instead, in our research we consider a passive approach to diagnosis in which intelligent diagnostic agents use probabilistic inference to determine the root cause of failure. The reliability of IP links in the Internet varies widely and hence we expect the probability of TCP failure to differ between different sets of hosts. Diagnostic agents in the Internet learn the probability of such failures for different regions in the Internet based on observations of TCP traffic. When users or network administrators detect network failures , they request diagnosis from such diagnostic agents. Agents then use information about the relative probability of failure of the TCP connections that make up an overlay connection to identify the most likely cause of failure when an overlay connection occurs without conducting any additional probes. In addition, diagnostic agents can also use this Bayesian network to predict the probability of overlay and TCP connection failure given information about the path of an overlay connection. We collect data on TCP failure probabilities in order to determine whether this data enables diagnostic agents data to accurately diagnose overlay failures in the Internet. To learn the probability of failure for TCP connections between different points in the network , we observe TCP traffic on the content distribution network CoDeeN using an updated version of Planetseer[11]. Next we construct a Bayesian network for diagnosis using these probabilities. We then use Bayesian inference to infer the most probable cause of failure for TCP-based applications. To evaluate the effectiveness of this approach, we test this Bayesian network on an artificial set of overlay connections based on the 305 traffic observed on CoDeeN. We find that when a failure occurs, knowing only the AS numbers of the source, proxy, and destination , we can determine which TCP connection has failed with over 80% probability. In addition, the probability of failure between ASes stays relatively constant over time, and data learned can be accurately used for diagnosis for many hours into the future. This suggests that the TCP failure probabilities we learn may be useful in the diagnosis of future failures as well. The contribution of this research is to show how inter-AS TCP failure probabilities can be used for probabilistic diagnosis of failures in overlay networks such as CoDeeN using Bayesian inference . We also demonstrate a variety of clustering methods to address the problem of dataset sparsity for learning TCP failure probabilities . In this paper we evaluate our system on CoDeeN overlay connections, but our Bayesian model generalizes to the diagnosis of other TCP-based applications as well. RELATED WORK There has been previous work in passive diagnosis of failures in the Internet. Padmanabhan, Ramabhadran, and Padhye developed Netprofiler, which collects network measurements from a set of end hosts and attempts to identify cause of failure by examining the shared dependencies among hosts that experience failures[8]. They show that this approach can provide information useful for diagnosis, but their paper only provides some preliminary results and do not provide details of how their system might diagnose real-world failures in practice. Shrink probabilistically diagnoses IP link failures based on the observed status of IP links that share resources[4]. Similarly in our work we diagnose failures in overlay connections where an overlay depends on several underlying TCP connections which may share IP hops. Shrink assumes that one can accurately determine the status of all IP links at any point in time. This allows one to identify the shared cause of failure of the failed IP links. Theoretically, we can also use this approach to diagnose overlay failures. That is, we can identify the TCP connections that share common IP hops and observe which overlay connections have failed at any point in time to identify the failed TCP connections. Unfortunately, in real-world diagnosis of TCP connections many of the assumptions made by systems such as Shrink do not hold for the following reasons. 1. The status of overlay connections may change rapidly, making it difficult to correlate failures in different overlay connections over time. 2. In order to construct a Bayesian network that accurately models the IP hops shared among different TCP connections we need an accurate IP level map of the Internet. As the Skitter 1 project demonstrates, accurately constructing such a map is difficult because routes may change and frequently tools such as traceroute do not provide accurate information. 3. Determining the status of an inactive overlay connection or a TCP connection is costly and takes time because it requires an active probe such as a ping, traceroute, or HTTP connection . Furthermore such probes are frequently inaccurate because of the prevalence of packet filtering, network address translation (NAT), and firewalls in the Internet[3]. 4. TCP and IP failures are frequently so transient that by the time one can test the status of a link, the failure no longer exists [11]. 1 http://www.caida.org/tools/measurement/skitter/ Therefore in this paper we present an alternative passive diagnosis approach that does not require simultaneously knowing the status of all overlay connections. Instead, we cluster TCP failures based on the Internet autonomous systems (ASes) of their endpoints and use information about the distribution of TCP failures to infer the cause of failure. An agent first learns a probabilistic model of failures based on a training set of observed TCP connections , and then it uses this model to diagnose future failures when it does not know the connection status. Other researchers have developed methods for diagnosing specific TCP-based applications. Ward, et al. infer the presence of TCP performance failures based on the rate of requests processed at an HTTP proxy server and TCP connection state [10]. Unlike such specialized diagnostic systems, our Bayesian approach to diagnosis can generalize to other applications that rely on TCP connections. Most previous research in probabilistic diagnosis of Internet failures evaluate their work on simulated failures. Steinder and Sethi model network faults using a bipartite causality graph in which the failure of individual links cause the failure of end-to-end connec-tivity , and then perform fault localization using a belief network[6]. In contrast, in our research we evaluate our approach on real-world TCP failures using actual data collected on the Internet. DIAGNOSING OVERLAY CONNECTION FAILURES In this paper we consider the diagnosis of overlay networks in which an overlay network connection requires a series of TCP connections between overlay nodes between the source and destination hosts. For example, Akamai is a content distribution network in which retrieving a resource from a web server may requ ire communication among multiple Akamai nodes along multiple TCP connections. Another example is the content distribution network CoDeeN on Planetlab, in which overlay nodes act as HTTP proxies . An request on CoDeeN[9] first requires a TCP connection to a CoDeeN node and then a connection from a CoDeeN node to server or another CoDeeN node. A failure in any one of these TCP connections causes the user's HTTP connection to fail. The challenge is to determine which of these TCP connections has failed. Sometimes users can determine whether a failure has occurred along the first TCP connection along the overlay path using information provided by their local TCP stack, but if a failure occurs beyond the first connection users cannot tell where a failure occurs without cooperation from the overlay. Depending on the type of overlay, users may have different amounts of information about the overlay path. For example, in an HTTP proxy connection, users know that the proxy is the first hop along the path and that if the connection is not cached, the web server is the last hop along the path. As a first step, in our research we examine a special case of diagnosis in order to gain insight into how well our approach might generalize to other types of diagnosis. The question we wish to answer is, if a two hop overlay connection fails due to a TCP failure, which TCP connection failed? In this paper we define a TCP failure as three consecutive TCP retransmits without a response. We assume that the diagnostic agent only knows that the overlay connection has failed and does not know which of the TCP connections has failed. We want to answer this question knowing only the IP addresses of the source, IP address of the first hop overlay node, and the IP address of the ultimate overlay destination host. Our model for probabilistic diagnosis generalizes to overlay connections with any number of hops, but as a starting point in this paper we only consider overlay connections with two hops. 306 TCP Conn. B C Hour Dst AS Src AS Overlay Conn. A C TCP Conn. A B Hour Dst AS Src AS 0 Failed OK 1 OK OK 0 OK Failed Failed B C 0 Failed P(Status =OK) A B ... 1 1 Hour ... ... ... 0.87 2 1 1 Dst AS 0.99 1 P(Status =OK) Src AS Figure 1: A Bayesian network for TCP overlay path diagnosis 3.1 Probabilistic Diagnosis The reliability of IP links in the Internet varies widely and hence we expect the probability of TCP failure to differ between different sets of hosts. Thus if we have knowledge of the relative probability of failure of the TCP connections that make up an overlay connection , we can then infer the most likely cause of failure when an overlay connection occurs without conducting any additional probes. In this paper we show we can use Bayesian networks both to learn a model of TCP failures and to perform diagnosis. Bayesian networks compactly represent the conditional probability of related events and enable efficient inference based on available evidence[5]. A Bayesian network is a directed acyclic graph in which nodes represent variables, and edges from parent nodes to children nodes represent dependence relations. Each node X has a conditional probability table (CPT) P (X|parents(X)) that encodes the conditional probability of X given evidence about its parents. Bayesian networks have several important features that make them especially suitable for reasoning about failures in the Internet . Firstly, Bayesian networks can model both deterministic and probabilistic dependencies among many types of Internet components and diagnostic tests. For example, an HTTP proxy connection functions if and only if the user/proxy TCP connection functions and the proxy/provider TCP connection functions. The probability that a TCP connection functions depends on the source and destination IP addresses and the time of the connection. To improve accuracy, we cluster IP addresses by AS and connection time by hour (see section 3.2). Figure 1 illustrates a Bayesian network that encodes the conditional probabilities for diagnosing an overlay connection from A to B to C. To diagnose an overlay connection failure from A to C, one can use this Bayesian network to infer the most probable status of the underlying TCP connections from A to B and B to C given information about the AS numbers and hour the connections were made. The variables in the Bayesian network represent the functional status of TCP connections and overlay connections. A node in this Bayesian network represents the functional status of a connection : OK if functioning, Failed if malfunctioning. Malfunctioning means that a connection failure occurs along the path, functioning means that no connection failure occurs. Edges in the Bayesian network represent dependencies among connections. The CPT for an overlay connection node represents the probability that it is functioning given the status of its underlying TCP paths. The CPT for a TCP path represents the probability that the TCP path functions given information about the path. In our Bayesian network we assume that the conditional probability of a TCP connection failure depends only on the source and destination IP addresses and the time of failure for each hop of the overlay, and not on which hop of the overlay connection it is (user/proxy or proxy/server). We represent this by using parameter tying in this Bayesian network so that both TCP paths share the same CPT. We also assume that a diagnostic agent can identify the intermediate hops in the overlay connection, either through active probing or because it has knowledge of the overlay topology. An advantage of modeling network components in terms of Bayesian networks is that a Bayesian network provides an abstract high-level representation for diagnostic data suitable for reasoning. Representing diagnostic data in terms of variables, evidence, and dependencies rather than passing around low-level measurements such as packet traces allows an agent to reason about the causes and consequences of failures without any deep knowledge of the behavior and characteristics of components and diagnostic tests. In addition, the conditional independence assumptions of Bayesian inference reduce the amount of data a diagnostic agent needs to consider for diagnosis. 3.2 Clustering To perform diagnosis using this Bayesian network, we need to learn the conditional probability of failure of a TCP connection given the properties of a connection. Learning the conditional probability of failure for each pair of IP addresses is impractical because it is infeasible to store the probability of failure for the 2 64 combinations of source and destination IP addresses. More importantly, for each pair of IP addresses we only have a limited amount of data with which to train the Bayesian network. For more effective diagnosis , diagnostic agents need a way to diagnose failures involving IP addresses it has not previously observed. Therefore to reduce the size of the conditional probability tables and to improve the accuracy of the learned probabilities, we cluster together IP addresses in a way that facilitates learning and diagnosis. Our hypothesis is that TCP connections that share many IP links with one another will have similar probabilities of failure . Thus two TCP connections with topologically nearby sources and nearby destinations will likely have similar failure probabilities . Therefore we clustered source and destination IP addresses in three ways: by the first eight bits of the IP address, the AS number, and by country. We also cluster TCP connections based on time. We hypothesize that the probability of failure changes over multiple time scales. For instance, if an IP routing change occurs, the probability of failure for affected TCP connections may change from low to high and back to low within a few minutes. On the other hand, the average rate of routing failure over several days may remain relatively constant. We show how different methods for clustering affect the accuracy of diagnosis in section 5. COLLECTING TCP FAILURE DATA It is difficult to obtain accurate information about the distribution of TCP failures in the Internet because failed connections make up only a small percentage of overall TCP traffic and the number 307 of possible source and destination IP addresses is enormous. To collect accurate failure probabilities, we need a way to observe the status of large quantities of TCP connections from many different source and destination hosts. In order to obtain such data, we used an updated version of Planetseer to collect data on TCP connection failures. The new Planetseer monitors TCP connections in the CoDeeN content distribution network and provides notifications when TCP sessions begin, end, and when TCP failures occur. Planetseer runs on over 320 Planetlab [2] nodes distributed around the world. We used Planetseer to monitor all the TCP connections made by 196 CoDeeN nodes. We observed 28.3 million TCP connections and 249,000 TCP failures over a ten hour period. We observed TCP connections to approximately 17,000 distinct IP addresses per hour on average. In our dataset, we observed TCP connections to hosts in 2116 unique Internet autonomous systems. CoDeeN overlay nodes act as HTTP proxies and establish TCP connections with web clients, web servers, and other CoDeeN nodes. In a typical CoDeeN session, a user initiates a TCP connection with the CoDeeN proxy, the proxy connects to a web server and retrieves the requested resource, and finally the proxy sends the requested data back to the user. Note that many requests are cached, and so the destination of the second hop in the overlay is a CoDeeN node and not the web server specified in the HTTP request. We found that 0.28% of user/proxy connections and 0.65% of proxy/server connections experienced TCP failures. Since Planetseer monitors TCP connections from the vantage point of the proxy, we cannot detect those TCP failures in which a user is unable to establish a TCP connection to the proxy. Therefore the lower percentage of user/proxy failures may be partly explained by the fact that all failures between the proxy and user occur after the user successfully establishes a TCP connection to the proxy. We believe that the failure probabilities learned through Planetseer are representative of typical TCP connections in the Internet . CoDeeN nodes operate as HTTP proxies, so the pattern of TCP connections resembles typical web traffic. Though caching at CoDeeN nodes reduces the number of connections to web servers we observe, we believe that the average failure probability to web servers we observe using Planetseer reflects typical failure rates for HTTP related TCP connections. We are currently examining other types of overlay connections to determine how well this TCP data generalizes for the diagnosis of other overlays. We learn the conditional probability table for TCP connection failure using the data collected from Planetseer. We cluster source and destination IP addresses by AS using the Oregon Route Views BGP tables 2 . EVALUATION Our hypothesis is that Bayesian inference using the conditional probability of failure for TCP connections given the AS numbers of the source and destination can accurately diagnose failures in overlay connections. In order to test this hypothesis, we constructed a Bayesian network using the probabilities learned from Planetseer and used it to diagnose failures in CoDeeN connections. We wanted to answer the following questions in our experiments: 1. Which clustering method produces the most accurate diagnosis : AS, IP/8 prefix, or country? We expect that clustering based on AS will produce the most accurate results since it is most closely correlated with the Internet routing topology. 2 http://www.routeviews.org/ 2. How does diagnostic accuracy change as we increase the time interval over which we cluster TCP connections? We expect that as the clustering interval increases, accuracy will increase at first, but then decrease as the learned probabilities less accurately reflect the probabilities of new failures. 3. How does the age of the training set affect diagnostic accuracy ? We expect that as the distribution of TCP failures in the Internet changes over time, diagnostic accuracy will also decrease. 5.1 Experimental Setup We train a Bayesian network using the Bayes Net Toolbox (BNT) for Matlab 3 . In order to diagnose TCP connections between regions we did not observe in the training set, we initialize the prior probabilities of failure according to a uniform Dirichlet distribution, which is equivalent to adding an additional element to the training set for each combination of source cluster, destination cluster, and connection status. We test this Bayesian network on an artificial dataset generated based on the distribution of TCP connections observed on Planetseer. Since Planetseer does not provide information about which TCP connections are associated with each CoDeeN request, we construct a dataset based on the TCP connections we observed. First we identify user/proxy, proxy/proxy, and proxy/server connections based on IP address and port number . Then for each proxy, we count the number of TCP connections to each server and to each proxy. We assume that the number of cached requests equals the number of user/proxy connections minus the number of proxy/server and proxy/proxy connections . We assign each user/proxy TCP connection a corresponding proxy/provider connection, where the provider may either be a web server (if the resource is not cached), another proxy (if the resource is cached at another proxy), or the same proxy (if the resource is cached locally). We make these provider assignments according to the observed distribution of proxy/server and proxy/proxy connections . Of the 19,700 failures in this dataset, approximately 82% of requests are cached locally, 7.9% are cached at other CoDeeN nodes, and 10.6% are uncached. For each CoDeeN request failure our Bayesian network makes two diagnoses: one for the status of the user/proxy connection, and one for the status of the proxy/provider connection. We measure accuracy in terms of the fraction of correct diagnoses. To evaluate the accuracy of diagnosis, we compute the most probable explanation for a TCP failure given evidence that the overlay connection has failed and the AS numbers of the source, proxy, and destination , and then compare this diagnosis with the actual status of the source/proxy and proxy/provider connections. In our experiments we perform diagnosis without evidence about whether a resource is cached at a proxy. Of the CoDeeN requests that failed in the first hour of our dataset, we found that 62% failed at the user/proxy connection, 31% failed at the proxy/server connection, and 7% failed at a the proxy/proxy connection. Therefore knowing only the overall distribution of TCP failures between users and servers, without using information about the IP addresses of the user, proxy, and server, one could diagnose failures with 62% accuracy by diagnosing every failure as a user/proxy failure. In our experiments we wish to determine if our Bayesian approach to diagnosis can achieve significantly better accuracy. In order to properly compute the accuracy of diagnosis, we sepa-rated the set of TCP connections with which we trained the Bayesian network from the set of TCP connections associated with the failed 3 http://www.cs.ubc.ca/ murphyk/Software/BNT 308 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% AS Country IP Baseline Accuracy Figure 2: Clustering Method Comparison 75% 76% 77% 78% 79% 80% 81% 82% 83% 1 2 3 4 5 6 7 8 9 Training Interval Length (hours) Accuracy Figure 3: Accuracy vs. Training Interval Length overlay connections under diagnosis. We collected ten hours of TCP connection data from Planetseer. In our initial experiments we choose to learn the average probability of failure over one hour because we find that clustering over shorter time scales does not provide enough data for accurate diagnosis. 5.2 Experimental Results First we compare the accuracy of three IP clustering methods: by Internet autonomous system number (AS), by the first eight bits of the IP address (IP), and by the country in which a host resides (Country). We determine the country of a host using the hostip.info database 4 , which maps the first 24 bits of an IP address to a country using location information contributed by Internet users. We train three Bayesian networks corresponding to the three clustering methods using data from hour 1. Then we test these Bayesian networks on the proxy connection failures constructed using data from hours 210 and averaged the results. We use a junction tree inference engine to compute the most likely status for each TCP connection and compare the inferred status with the actual status from the data. Since the Bayesian network we use for inference has no cycles, we can perform Bayesian learning and junction tree inference rapidly; in our experiments, inference for a single connection requires approximately 5 ms. 4 http://www.hostip.info/ 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 1 2 3 4 5 6 7 8 9 Training Set Age (hours) Accuracy Figure 4: Accuracy vs. Training Set Age Figure 2 compares the diagnostic accuracy of these three clustering approaches. We define accuracy as the fraction of correct status inferences. As a baseline, we also plot the accuracy of simply guessing that every failure is due to a user/proxy connection failure. Figure 2 shows that all three clustering methods provide similar degrees of accuracy. Our hypothesis was that clustering based on AS would produce the most accurate results, but our experiments show that clustering based on the first 8 bits of the IP address yields higher accuracy for short time intervals. This may be because one hour is not enough time to accurately learn inter-AS TCP failure probabilities, or due to inaccuracies in the Route Views BGP table. Next we computed the accuracy of diagnosis as we increase the time interval over which we cluster TCP connections. If the interval over which we train is too short, then we will not have enough data to accurately learn failure probabilities. If the interval is too long, then it may not accurately reflect changing network conditions. We train a Bayesian network using AS clustering on x hours before hour 10 for values of x from 1 to 9. We then test each Bayesian network on the data from hour 10. Figure 3 shows how accuracy changes as the training time interval changes. This plot shows that accuracy increases as the clustering time interval increases, suggesting that the training value of incorporating additional data outweighs the inaccuracy introduced by using older data. Finally, we compute the accuracy of diagnosis as we increase the age of the data on which we trained the Bayesian network. We train a Bayesian network using AS clustering on data from hour 1 and test it on overlay failures observed during each of the hours from 2 to 10. Figure 4 plots the accuracy of diagnosis over time. Average accuracy changes over time because the distribution of failures we observe using Planetseer varies from hour to hour, but overall diagnostic accuracy diminishes only slightly after nine hours, suggesting that the distribution of TCP failure probabilities remains relatively stationary over time. We also compare the false positive and false negative rates for each clustering method. The false positive rate is the fraction of functioning connections that are incorrectly diagnosed as having failed, while the false negative rate is the fraction of failed connections that are incorrectly diagnosed as functioning. Table 1 lists the false positive and false negative rates for each clustering method. 5.3 Analysis These experiments show that we can diagnose overlay connection failures knowing only the AS numbers of its TCP endpoints. 309 AS Country IP Baseline user/proxy false pos. 0.174 0.358 0.426 1.000 user/proxy false neg. 0.219 0.050 0.060 0.000 proxy/server false pos. 0.219 0.101 0.265 0.000 proxy/server false neg. 0.171 0.128 0.100 1.000 Table 1: Diagnosis error rates by type One reason our approach to diagnosis works is due to the heavy-tailed distribution of TCP connection failure probability. The majority of TCP failures occur among a small number of AS pairs. Therefore most CoDeeN connection failures involve one TCP connection with low failure probability and another TCP connection with high failure probability, so probabilistic inference produces the correct diagnosis. For example, we find that TCP connections from hosts in China to hosts in the USA tend to have a much higher probability of failure than connections within the USA. If an CoDeeN user in China accesses a proxy in the USA to retrieve content from a web server in the USA and experiences a failure, then it is very likely that the failure occurred on the connection between the user and the CoDeeN node. If the probability of failure for every pair of ASes were equal, then our probabilistic approach to diagnosis would not work as well. Another interesting result is that the accuracy of diagnosis diminishes relatively slowly over time, implying that the distribution of TCP failures in the Internet stays relatively stationary over time. This suggests that diagnostic agents can perform accurate diagnosis using inter-AS TCP failure probabilities without having to con-stantly collect the latest TCP failure data. CONCLUSION AND FUTURE WORK Our initial experimental results indicate that our passive probabilistic approach to diagnosing TCP overlay connection failures can provide useful diagnostic information. In this paper we show that Bayesian inference provides a useful framework for diagnosing two hop overlay connection failures on CoDeeN, but our approach can generalize to the diagnosis of other overlay connection failures as well. We view our approach to diagnosing TCP overlay connection failures as just one example of a more general probabilistic approach for Internet fault diagnosis. In this paper we show how to use inter-AS TCP failure probabilities to diagnose failures in overlay networks, but the technique we used to diagnose failures in CoDeeN can be extended to the diagnosis of other overlays as well. We can apply the knowledge we learned from Planetseer to diagnose other classes of network components and applications by adding new nodes and edges to the Bayesian network we use for diagnosis. In this paper we only considered diagnosis without using any additional evidence about a failure. Typically, however, when failures occur users may already know the status of certain network components and can perform diagnostic probes to collect additional evidence for diagnosing failures. We can improve the accuracy of our approach by adding variables and edges to the Bayesian network to take into account this information. For instance, if we know the IP paths that TCP connections traverse, we can incorporate evidence of IP link failures into the Bayesian network. We intend to explore how agents can incorporate such additional evidence into a Bayesian network to improve diagnostic accuracy. In future work we will also examine more accurate models for Internet fault diagnosis that take into account failures at both short and long time scales. In this paper we only evaluated our algorithm on ten hours of data from Planetseer; we would like to conduct additional experiments to more accurately determine the effectiveness of diagnosis using data from other time periods as well. In addition we would like to explore other clustering methods, including dy-namically choosing the prefix length on which to cluster based on how much data an agent has about TCP connections to a particular IP range. Finally, though our paper describes a centralized diagnosis approach , this approach can easily be adapted for distributed diagnosis . Knowledge of the overlay topology and the conditional probabilities in the CPTs can be distributed among multiple agents in the Internet, allowing different agents to collect failure data from different points in the network. We are currently developing such a distributed system for the diagnosis of TCP application failures in the Internet. REFERENCES [1] D. G. Andersen, H. Balakrishnan, M. F. Kaashoek, and R. Morris. Resilient overlay networks. In Proceedings of the 18th ACM Symposium on Operating System Principles (SOSP), 2001. [2] B. Chun, D. Culler, T. Roscoe, A. Bavier, L. Peterson, M. Wawrzoniak, and M. Bowman. Planetlab: an overlay testbed for broad-coverage services. SIGCOMM Comput. Commun. Rev., 33(3):312, 2003. [3] S. Guha and P. Francis. Characterization and measurement of tcp traversal through nats and firewalls. In Internet Measurement Conference 2005 (IMC '05), 2005. [4] S. Kandula, D. Katabi, and J.-P. Vasseur. Shrink: A Tool for Failure Diagnosis in IP Networks. In ACM SIGCOMM Workshop on mining network data (MineNet-05), Philadelphia, PA, August 2005. [5] U. Lerner, R. Parr, D. Koller, and G. Biswas. Bayesian fault detection and diagnosis in dynamic systems. In Proceedings of the Seventeenth National Conference on Artificial Intelligence (AAAI-00), pages 531537, Austin, Texas, August 2000. [6] A. S. M Steinder. Increasing robustness of fault localization through analysis of lost, spurious, and positive symptoms. In Proceedings of INFOCOM, 2002. [7] R. Mahajan, N. Spring, D. Wetherall, and T. Anderson. User-level internet path diagnosis. In Proceedings of ACM SOSP, 2003. [8] V. N. Padmanabhan, S. Ramabhadran, and J. Padhye. Netprofiler: Profiling wide-area networks using peer cooperation. In Proceedings of the Fourth International Workshop on Peer-to-Peer Systems (IPTPS), February 2005. [9] L. Wang, K. Park, R. Pang, V. Pai, and L. Peterson. Reliability and security in the codeen content distribution network. In Proceedings of the USENIX 2004 Annual Technical Conference, 2004. [10] A. Ward, P. Glynn, and K. Richardson. Internet service performance failure detection. SIGMETRICS Perform. Eval. Rev., 26(3):3843, 1998. [11] M. Zhang, C. Zhang, V. Pai, L. Peterson, and R. Wang. Planetseer: Internet path failure monitoring and characterization in wide-area services. In Proceedings of Sixth Symposium on Operating Systems Design and Implementation (OSDI '04), 2004. 310

fault diagnosis;passive diagnosis;NAT;Bayesian networks;Planetseer overlay monitoring system;active probing for diagnosis;inter-AS TCP failure probabilities;TCP overlay connections;Bayesian networks modelling;CoDeeN traffic patterns;TCP overlay path diagnosis;Planetseer;clustering;network address translation

69

Digital Asset Management Using A Native XML Database Implementation

Digital Asset Management (DAM), the management of digital content so that it can be cataloged, searched and re-purposed, is extremely challenging for organizations that rely on image handling and expect to gain business value from these assets. Metadata plays a crucial role in their management, and XML, with its inherent support for structural representation, is an ideal technology for this. This paper analyzes the capabilities of a native XML database solution via the development of a "proof of concept" and describes implementation requirements, strategy, and advantages and disadvantages of this solution.

INTRODUCTION Digital asset creation and management evolved in the late 1990s. Companies have created massive digital assets in the form of images, video and audio files, streaming media, Power Point templates, web pages, and PDF files containing engineering specs, legal documents, internal memos and more. The World Wide Web has drastically increased the need for digital information and its exchange. Ille [8] identifies a digital asset as a strategic asset like any other form of working capital, and states that its efficient management is being increasingly recognized as a competitive lever for companies all over the world. Development of the model for storing any form of digital object into a structured format requires a deft combination of asset analysis, strategic thinking, business planning and appropriate technology. Companies can achieve early strategic advantage by implementing management systems for digital assets that can be repurposed or customized, providing process efficiencies in collaborative work. Digital Asset Management (DAM) can deliver competitive advantage for advertising agencies, technical or engineering documentation departments, designers, producers and others by reducing time spent locating creative assets. Enterprises often require reusing or sharing their digital assets. It is indispensable to store content in an organized way to reuse or process it for future needs. Global enterprises are facing the daunting challenge of figuring out how best to address the growing complexity of creating digital assets and managing the flow of assets through a single infrastructure [11]. Exacerbating the challenge is the fact that companies are creating a massive volume of digital assets but are rarely examining their organized storage and retrieval methods. SIGNIFICANCE OF THE PROBLEM DAM systems are still relatively new, but organizations have started realizing the importance and need for digital asset management. The Gartner Group affirms that only a limited number of technically advanced commercial content providers use DAM systems today to digitally construct, store and distribute rich media content in single medium forms [7]. The systems also have limited corporate use in advertising firms and marketing departments. Gartner predicts that by 2005 more than 25% of all the enterprises with commercial internet operations will use DAM systems. By 2010, more than 45% of all the enterprises with commercial internet operations will use DAM systems. Recently reported cases provide evidence that companies have started investing in technology for DAM. For example, the Coca Cola company has bought technology from IBM for its digital advertisement archives, which contain 9,000 graphical images, 7,000 scanned documents and more than 25,000 corporate videos and television advertisements [13]. The technology includes search tools for retrieving, updating, managing and disseminating historical records online, including the company's famous marketing and advertising icons. Another case is that of the Shoah Foundation. Steven Spielberg established the Shoah Foundation with the mission to videotape 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. CITC4'03, October 1618, 2003, Lafayette, Indiana, USA. Copyright 2003 ACM 1-58113-770-2/03/0010...$5.00. 237 and preserve the testimonies of the Holocaust survivors and witnesses. The foundation has collected more than 50,000 eyewitness testimonies in 57 countries and 32 languages [12]. The challenge now is to manage, catalog and disseminate the video and digital collection of testimonies of those survivors. These digital assets of the Shoah Foundation are cataloged with lists of keywords, text summaries describing the survivors, related documentaries focusing on topics such as the ghettos or labor camps they lived in, and past and present photos of them and their families. The following summarizes the major challenges that are faced during any wide adoption of DAM. Storage: One of the fundamental problems is physical deterioration of storage devices that store digital information. Magnetic media are particularly subject to data loss and errors. There is also the important question of hardware and software obsolescence and the future availability of today's technologies. Procedural Issues: Technical problems are further complicated when resources to be preserved need to be transformed into digital form. Digitization of paper analogs for access and preservation is a time consuming and labor intensive task. Beyond this technical problem, there is a host of financial, legal, and policy problems. Security: Securing digital assets against misuse, theft, or damage is an ongoing concern of organizations. Copyright: One of the major legal issues associated with electronic assets is the actual scanning or digitizing of materials. Copyright holders have the exclusive right to reproduce, distribute, and display their work. Ferullo [3] points out that digitizing and posting infringes on these exclusive rights. Distribution: Digital assets will be utilized to the fullest only when they can be distributed via proper communication channels to the intended users. Infrastructure: DAM requires a robust IT infrastructure to support creation and distribution. Human Factors: While the purpose of DAM is to provide greater efficiency, getting users to adapt to the new environment can be a challenge. This is an important issue because many DAM solutions require a change in work processes before the users see any benefits [6]. The requirement to manage and deliver digital assets is becoming critical to almost all content-related applications, due to the evolution of the Internet and growth of digital assets. Enterprises need plans to benefit from the new world of rich, valuable digital assets that will affect everything from their internal processes and customer relations to their web site and telecommunications infrastructures. AN XML SOLUTION Widespread use of rich media has spurred the growth of the DAM market. Frost and Sullivan, a market analysis firm, claims the average user in a media enterprise looks for a media file 83 times a week and fails to find it 35% of the time [5]. Canto Software research predicts that DAM solutions will drop that figure to 5% [9]. With the growth of internet publication, document metadata is increasingly important because it provides contextual information about digital assets necessary for customization of the content. In response, Adobe Systems plans to unveil new metadata technology designed to ease the process of applying content to multiple types of media. The XMP (Extensible Metadata Platform) provides a framework for standardizing the creation and processing of document metadata across publishing workflows, according to Adobe officials [10]. A review by Doering [2] provides the insight into methods for digital asset management. According to Doering, a DAM solution must include the following critical features: Indexing: As content is generated and stored, it is indexed according to various possible criteria. Metadata does not merely describe by whom, when, and how a piece of content was created; it can also be used to interpret and catalog the content itself. Rights Management: Includes handling rights to the content or restricting the use of the content by the purchaser/end-user. This might occur, for example, with corporate information or licensed images from a third party incorporated into the content. Reuse: With a viable DAM system in place, the internal content developer can research and select appropriate content for reuse in new content. This represents a significant savings potential for the companies. Review: A final benefit of an online catalog with a DAM system is the ability to review older content more easily. 3.2 The XML/Metadata Approach By incorporating a DAM system, a company gains both the savings from reusing content as well as revenue from continued sales of the same elements. According to Fontana [4], XML databases serve in a complementary role to traditional databases especially as XML becomes prevalent. Nearly 85% of large corporations are expected to move all their web-based content to XML over the next three years. Fundamentally, two high-level approaches may be adopted for implementing XML databases. 1) XML-enabled database: In an XML-enabled database, the documents are stored in constituent fragments. Here the XML data is stored in object relational form and one can use an XML SQL Utility (XSU) or SQL functions and packages to generate the whole document from its object relational instances. 2) Native XML database: In a native XML database approach, XML is not fragmented but rather is stored as a whole in the native XML database. This means that documents are stored, indexed, and retrieved in their original format, with all their content, tags, attributes, entity references, and ordering preserved. In this technique, the XML Database is designed to handle XML in its native format, thereby speeding access to data by eliminating the need to transform it into the rows and columns of a standard relational database. 238 Biggs [1] suggests there are three principal reasons to implement a native XML database: 1) enterprises today use a mix of data, such as the data housed in object-oriented databases and unstructured data that needs to be exchanged with partners (native XML databases can leverage all of these disparate sources); 2) XML databases can boost processing performance due to XML-based transactions; and 3) digital assets described in the XML format can be used by other applications. 3.3 A Technique Using Native XML When developing a native XML solution, certain steps should be followed: 1) Identify the need for DAM 2) Know your assets: Identify various assets, define and understand their use in the organization 3) Define search needs and key attributes of your assets 4) Capture the objects (digitized) and data about the objects (metadata) 5) Process: Generate and store XML files associated with each object The basis of this technique lies in creation and usage of semi-structured metadata stored in an XML format to manage digital assets efficiently. When the structure of the data is irregular and implicitly given by the document itself, not by a global scheme, they are often referred to as semi-structured data. The ability of an XML to handle dynamic data provides leeway for the applications to use semi-structured data to describe their digital assets. XML is becoming the de facto data exchange format for the Web. The XML enclosing metadata can be either stored in the database for future retrieval or be easily transferred over the Internet to any other systems. The Oracle 9i DBMS (used for the Proof of Concept) provides an "XMLType" data type that can be used to store and query XML data in the database. The XMLType has member functions one can use to access, extract, and query the XML data using XPath expressions, a standard developed by the World Wide Web Consortium to traverse XML documents. 3.4 Proof of Concept This study analyzed existing approaches for non-text based digital asset management and implemented a DAM solution by applying a native XML Database technique. Meta-data of digital assets is often semi-structured and the digital files are of varied types. XML databases are most appropriate for managing diverse types of semi-structured digital assets. For this project, the Proof of Concept (POC) was developed using facilities and resources of the University's Digital Enterprise Center (DEC) at the School of Technology. The Product Lifecycle Management (PLM) lab at DEC creates, simulates and shares digital information related to a company's products, processes and resources. These digital assets encompass graphical images, presentations, and video/audio clips of manufacturing processes representing manufacturing models. A digital asset produced in the PLM lab is the intellectual property of the DEC and needs to be managed for future use or research. Though the demonstration focused on the manufacturing process models in the PLM lab, the XML Database technique can be applied to any form of digital asset. Since the potential file population is very vast and is of varied types, the POC restricted the sampling to the following categories of digital assets: Audio files, Video files, Images, and Text based files (presentation slides, Word documents, Acrobat files). The POC confined the sample to a limited number of files describing assembly line parts available from the PLM lab. Metadata was stored for each of the digital objects of the assembly line. The global parameters and sub-parameters used to describe the digital files included the following: File Name, Type (Image, Audio, Video, Word etc), Author, Creation Date, General Description, Keywords, and a Comment. A keyword search capability was added for searching within these and other parameters. The Proof of Concept application was developed to provide a storage, search and retrieval facility to manage the digital assets of the PLM lab using the School of Technology's software and hardware resources. The application provides a web-based interface for convenient management of the digital models and has an "n-tiered" architecture. The backend database is Oracle 9i with native XML support. The web interface was developed using JSP and the middle tier was incorporated using Java Servlets. Analysis was conducted with the following steps: Various types of data (heterogeneous data) were collected for demonstration purposes. The validity of metadata was checked before entering it into the system. Upon validation, data was entered in the system (Oracle 9i database) through the web-interface. With the appropriate search parameters, the data entered could be searched for and retrieved. Depending on the requirement, this retrieved data could be viewed, reused or repurposed. Providing different search criteria tested consistency and reliability in the retrieval of the asset. The following technologies were used to develop the POC: Apache Webserver 1.3; Apache JServ 1.1; JDK 1.1; Oracle 9i (9.0.0.2); XML; Java, JSP, Servlets. Tools used for development purposes included: ERwin database design software; Borland Jbuilder; Microsoft Front Page; and Oracle Enterprise Manager. The POC was developed with a three-tier architecture as shown in Figure 1. The first tier of the architecture presents an interface to the user to facilitate access to digital files. The interface is web-enabled and was developed using Java Server Pages. This tier provides a user-friendly navigation through the site for managing digital files, including screens for inserting, deleting and searching on file data. 239 XML Generation Database XML Files (Metadata for Digital Assets) XML Files (Metadata for Digital Assets) User Interface Web & Application Server Digital Asset Storage XML Validation Metadata for Digital Assets Metadata for Digital Assets Digital Asset Digital Asset Tier 3 Tier 2 Tier 1 The POC Architecture Figure 1: 3-Tier Architecture for POC The middle tier of the architecture consists of Servlets and Java helper classes. Servlets direct the flow of the control and provide database connectivity for the application. In addition, the business logic is implemented in Servlets. Helper classes are used by Servlets to create and manage user-defined Java objects. Servlets and JSPs are deployed on an Apache Web server and Apache Jserv. The third tier of the architecture is the database layer itself. As noted previously, the POC uses the XML capabilities of the Oracle 9i database management system. OBSERVATIONS AND EVALUATION The XML database technique stores attributes of digital assets in the database in the form of an XML file. While the XML resides in the database system, digital assets might be files in a file system or a BLOB in a relational database. This demonstration of the technique stores digital assets in a file system. XML databases provide some features that are analogous to conventional text databases. Common to all native XML databases are indexes, which allow the query engine to jump to a particular point in an XML document. This gives such databases a tremendous speed advantage when retrieving entire documents or document fragments. This is because the database can perform a single index lookup and retrieve the entire document or fragment in a single read. Normalizing data for an XML database is largely the same as normalizing it for a relational database. The programmer needs to design the documents so that no data is repeated. One difference between native XML databases and relational databases is that XML supports multi-valued properties while (most) relational databases do not. This makes it possible to "normalize" data in a native XML database in a way that is not possible in a relational database. 4.2 Native XML vs. XML-Enabled A native XML database defines a (logical) model for an XML document -- as opposed to the data in that document -- and stores and retrieves documents according to that model. A native XML database has an XML document as its fundamental unit of (logical) storage, just as a relational database has a row in a table as its fundamental unit of (logical) storage. It is not required to have any particular underlying physical storage model. For example, it can be built on a relational, hierarchical, or object-oriented database, or use a proprietary storage format such as indexed, compressed files. An XML-enabled database has an added XML mapping layer provided either by the database vendor or a third party. This mapping layer manages the storage and retrieval of XML data. Data that is mapped into the database is mapped into application specific formats and the original XML meta-data and structure may be lost. Data retrieved as XML is NOT guaranteed to have originated in XML form. Data manipulation may occur via either XML specific technologies (e.g. XPath) or other database technologies (e.g. SQL). The fundamental unit of storage in an XML-enabled database is implementation dependent. The XML solutions from Oracle and Microsoft, as well as many third party tools, fall into this category. 4.2.1 Advantages of Native XML Native XML databases have several advantages over relational databases. Since native XML databases do not have database schemas, one can store similar documents with more than one schema in the database at the same time. While one may still need to redesign queries and convert the existing documents -- a non-trivial process -- this may ease the transition process. XML databases make search within the XML documents very efficient. If the data is parsed into Document Object Model (DOM) format, XPATH can be used. XML database solutions usually add a text indexing system so that query performance is improved. 4.2.2 Disadvantages of Native XML Currently, only a few native XML databases enforce referential integrity. The reason for this is that most native XML databases do not currently support linking mechanisms, so there are no references for integrity checking. Therefore, applications that rely on referential integrity mechanisms of databases must enforce these constraints themselves for XML databases. In the future, many native XML databases will probably support linking mechanisms and referential integrity. Another disadvantage of XML databases is that while query performance is better, update performance suffers. The reason is that the index entries for a document must be deleted and new entries created whenever a document is inserted or updated. In general, an XML Database approach is better because it supports the full power of XML. However, a major drawback is performance degradation, as data must be constantly reparsed into a DOM tree, wasting cycles and memory. Additionally, update capabilities are weak, and finally, automated enforcement of 240 integrity constraints needlessly places unreasonable burden upon application programmers, increasing risks and costs. CONCLUSION Digital Asset Management (DAM) is an evolving field with a great potential. With the evolution of computers and the Internet, companies have been creating an enormous volume of digital content in various formats. Managing this content, so that it can be cataloged, searched and re-purposed, is extremely challenging for organizations. XML is a commonly used standard in internet applications. Hence, representing the metadata of a digital content in an XML format is, on the surface, a good design decision. XML, by its very nature, provides for a complex, well-defined and yet extensible structural representation of the metadata and interoperability between various applications dealing with digital assets. A major advantage of having XML natively in the database is that one can perform the relational manipulation operations such as insert, update, and delete on the whole or partial XML and can also perform XML specific operations like XPATH search and node modification using the power of SQL. This, and other advantages give native XML databases an edge over systems that don't use XML or that manage XML externally. REFERENCES [1] Biggs, M. (2001, December 3). Two ways to bring XML to your databases. InfoWorld, Framingham, 23 (49), 20. [2] Doering, D. (2001, August). Defining the DAM thing: how digital asset management works. Emedia Magazine, Wilton, 14(8), 28-32. Retrieved February 27, 2002, from http://proquest.umi.com/pqdweb?Did=000000077222944&F mt=4&Deli=1&Mtd=1&Idx=5&Sid=12&RQT=309 [3] Ferullo, D. L. (2002). The challenge of e-reserves. Net Connect, 48 (8), 33-35. [4] Fontana, J. (2001, November 5). Upgraded database makes the most of XML. Network World, Framingham, 18 (45), 29. Retrieved February 27, 2002, from http://proquest.umi.com/pqdweb?Did=000000088274369&F mt=4&Deli=1&Mtd=1&Idx=7&Sid=3&RQT=309 [5] Frost & Sullivan.s (n.d.). U.S. Digital Asset Management Markets. Retrieved March 23, 2002, from http://www.frost.com/prod/servlet/fcom?ActionName=Displa yReport&id=A192-01-00-00-00&ed=1&fcmseq=1043213084248 [6] Garcia, K. (2001, November). Broadcasters starting to ease into digital workflow. TVB Europe, 10(11), 22-24. [7] Gilbert, M., Landers, G., Latham, L. & Lundy, J. (2001, November 26). What's cool, what's hot: content technology hype cycle. Gartner Advisory. Retrieved February 9, 2002, from http://gartner.adpc.purdue.edu/rasdquest/RESEARCH/RAS/ 102700/102760/102760.html [8] Ille, C. (2001, February 26). Market definitions: digital document imaging. Gartner Group. Retrieved February 27, 2002, from http://proquest.umi.com/pqdweb?Did=000000093522513&F mt=3&Deli=1&Mtd=1&Idx=4&Sid=3&RQT=309 [9] Martin, N. (2001, September). DAM right! Artesia technologies focuses on digital asset management. EContent, Wilton, 24(7), 60-62. [10] Moore, C. (2001, September 24). Adobe touts images. InfoWorld, Framingham, 23(39), 36. Retrieved February 27, 2002, from http://proquest.umi.com/pqdweb?Did=000000081956628&F mt=3&Deli=1&Mtd=1&Idx=2&Sid=12&RQT=309 [11] Moore, C. (2001, September 24). Content management plays role in cross-media publishing. Computer World. Retrieved February 9, 2002, from http://www.computerworld.com/storyba/0,4125,NAV47_ST O64339,00.html [12] Solomon, M. (2002, January 14). Managing the memories. Computer World. Retrieved April 16, 2002, from http://www.computerworld.com/storyba/0,4125,NAV47_ST O67304,00.html [13] Weiss, T. (2001, December 10). Coca-Cola ad legacy gets help from IBM. Computer World. Retrieved February 9, 2002, from http://www.computerworld.com/storyba/0,4125,NAV47_ST O66495,00.html 241

keyword search;metadata;multimedia;digital asset management;semi structured data;database;digital images;Digital Asset Management;XML database;storage and retrieval;native XML;DAM;heterogenous data;proof of concept

7

A Fair and Traffic Dependent Scheduling Algorithm for Bluetooth Scatternets

mechanisms and algorithms necessary to set up and maintain them. The operation of a scatternet requires some Bluetooth units to be inter-piconet units (gateways), which need to time-division multiplex their presence among their piconets. This requires a scatternet-scheduling algorithm that can schedule the presence of these units in an efficient manner. In this paper, we propose a distributed scatternet-scheduling scheme that is implemented using the HOLD mode of Bluetooth and adapts to non-uniform and changing traffic. Another attribute of the scheme is that it results in fair allocation of bandwidth to each Bluetooth unit. This scheme provides an integrated solution for both intra-and inter-piconet scheduling, i.e., for polling of slaves and scheduling of gateways.

Introduction The Bluetooth [10] technology was developed as a replacement of cables between electronic devices and this is perhaps its most obvious use. But, it is the ability of Bluetooth devices to form small networks called piconets that opens up a whole new arena for applications where information may be exchanged seamlessly among the devices in the piconet. Typ-ically , such a network, referred to as a PAN (Personal Area Network), consists of a mobile phone, laptop, palmtop, headset , and other electronic devices that a person carries around in his every day life. The PAN may, from time to time, also include devices that are not carried along with the user, e.g., an access point for Internet access or sensors located in a room. Moreover, devices from other PANs can also be interconnected to enable sharing of information. The networking capabilities of Bluetooth can be further enhanced by interconnecting piconets to form scatternets. This requires that some units be present in more than one piconet . These units, called gateways, need to time-division their presence among the piconets. An important issue with the gateways is that their presence in different piconets needs to be scheduled in an efficient manner. Moreover, since the gateway cannot receive information from more than one piconet at a time, there is a need to co-ordinate the presence of masters and gateways. Some previous work has looked at scheduling in a piconet [2,5] and also in a scatternet. In [4], the authors define a Rendezvous-Point based architecture for scheduling in a scatternet , which results in the gateway spending a fixed fraction of its time in each piconet. Such a fixed time-division of the gateway may clearly be inefficient since traffic is dynamic. In [9], the authors propose the Pseudo-Random Coordinated Scatternet Scheduling (PCSS) scheme in which Bluetooth nodes assign meeting points with their peers. The sequence of meeting points follows a pseudo-random process that leads to unique meeting points for different peers of a node. The intensity of these meeting points may be increased or decreased according to the traffic intensity. This work presents performance results for various cases. In [11], a scatternet-scheduling algorithm based on the concept of a switch table, which can be dynamically adjusted based on traffic load, is presented. In [1], the authors present a credit-based scheduling scheme based on the SNIFF mode of Bluetooth, where credits may be reallocated to cater to changing traffic. Our scheduling scheme addresses the issues of fairness and utilization of bandwidth. Since Bluetooth is a low-bandwidth environment, it is important that bandwidth should be effi-ciently utilized. Also, since a low bandwidth can easily lead to starvation of flows, another metric we focus on is fairness. We propose a distributed scatternet-scheduling algorithm that is implemented using the HOLD mode [10] of Bluetooth and adapts to non-uniform and changing traffic. This algorithm provides an integrated solution for both intra- and inter-piconet scheduling, i.e., for polling of slaves and scheduling of gateways. The algorithm leads to a high bandwidth utilization and results in a fair division of (a) the piconet bandwidth between the slaves of a piconet and (b) the gateway presence among different piconets. In section 2, we discuss the Bluetooth technology. In section 3, we present a definition of fairness in the context of Bluetooth scatternets, which takes into account intra- and inter-piconet max-min fairness. Section 4 describes the algorithm and proves its fairness property. Section 5 presents simulation results and section 6 presents the conclusions. Bluetooth technology The Bluetooth system [3] operates in the worldwide unlicensed 2.4 GHz IndustrialScientificMedical (ISM) frequency band. To make the link robust to interference, it uses a Frequency Hopping (FH) technique with 79 radio carriers. It allows a raw data transmission rate of 1 Mbit/s. Two or more Bluetooth units sharing the same channel form a piconet. Each piconet consists of a master unit and up to seven active slave units. The master unit polls the slave units according to a polling algorithm and a slave is only allowed to transmit after the master has polled it. The piconet capacity is thus, shared among the slave units according to the polling algorithm. Furthermore, two or more piconets can be interconnected, forming a scatternet. This requires a unit, called an inter-piconet unit (gateway), to be a part of more than one piconet. Such a unit can simultaneously be a slave member of multiple piconets, but a master in only one, and can transmit and receive data in only one piconet at a time; so participation in multiple piconets has to be on a time-division multiplex basis . The time of the gateway is, thus, also shared among the piconets it belongs to. In this work, we assume that the gateway can only be a slave in its piconets. If a gateway were to be a master in a piconet, it would lead to the stoppage of all transmission in the piconet when the gateway visits some other piconet. Thus, we believe that the use of the gateway as a slave is the most efficient method of scatternetting. Fair allocation of bandwidth As introduced in the previous section, units belonging to a piconet share the piconet capacity according to the polling algorithm used by the master. In an analogous manner, gateways in a scatternet divide their time among their different piconets, according to the "master-listening" algorithm they use. It can be noted that there is a duality in this architecture . On the one hand, a master divides its capacity among the units of its piconet by using a polling algorithm. On the other hand, a gateway shares its capacity among the piconets it belongs to, on the basis of a scheduling algorithm it uses for listening to the masters. The gateway, can, then be viewed as a "virtual master" and its masters can be viewed as "virtual slaves" forming a "virtual piconet", in which the polling cycle is, actually, the "listening cycle" of the gateway. A graphical interpretation of this duality is given in figure 1, in which the solid line shows the actual piconets, and the dotted line shows the virtual piconet. Due to this duality, we design our scheduling scheme such that the same scheduling algorithm is used for fair sharing of both (a) the piconet capacity among slaves and (b) the gateway time among piconets. We now give a definition of max-min fairness [7]. We then go on to define max-min fairness in the context of Bluetooth scatternets, by considering (a) intra-piconet fairness, i.e., fairness in division of piconet bandwidth among slaves (both gateway and non-gateway) of a piconet and (b) inter-piconet fairness, i.e., fairness in division of the gateway's presence among its piconets. We first define a `feasible' rate distribution since this is used in the definition of max-min fairness. Definition 1 (Feasible). A rate distribution is feasible if rates are non-negative, the aggregate rate is not greater than one, and no unit receives a higher rate than required. Definition 2 (Max-min fairness). An allocation of rates 1 , 2 , . . . , s among s units is max-min fair if it is feasible, and for each unit i, i cannot be increased (while maintaining fea-sibility ) without decreasing j for some other unit j for which j i . The distribution of max-min fair rates depends upon the set of rate demands (traffic generated) of the units. In the following subsections, we discuss factors that determine the max-min "fair share" of a slave (gateway or non-gateway). We call these factors the Piconet Presence Fraction and the Scatternet Presence Fraction and show how they may be used to calculate the "fair share" for a slave in a scatternet. 3.1. Piconet presence fraction Consider a piconet consisting of gateway and non-gateway slaves in which the master has complete knowledge of the rate demands of all slaves (an ideal master). Using this knowledge , the master polls the slaves in a max-min fair manner such that each slave gets its "fair share" of the master's polling. We refer to the "fair share" received by a slave as the "piconet presence fraction" (PPF) of the slave. The gateway has a PPF for each piconet it belongs to. Consider the piconets shown in figures 2(a) and 2(b), each consisting of one gateway and two slaves, with the traffic rates of each slave as shown. In figure 2(a) (Piconet I), the PPF of each non-gateway slave is 0.2, while the PPF of the gateway is 0.6. In figure 2(b) (Piconet II), the PPFs of the slaves are 0.2 and 0.4, while the PPF of the gateway is 0.4. A FAIR AND TRAFFIC DEPENDENT SCHEDULING 11 Figure 2. Piconets with traffic rates between master and each slave shownn. 3.2. Scatternet presence fraction A gateway will, in general, be a slave in multiple piconets and may have different amounts of traffic to exchange with each piconet. Consider an ideal gateway that has complete knowledge of the rate demands of all its masters. The gateway can then divide its presence among its piconets in a max-min fair manner, giving each piconet a "fair share" of its presence. We call this fair share the "scatternet presence fraction" (SPF) of the gateway for the piconet. The importance of the SPF is that a fair division of the gateway's presence among its piconets can be achieved based on the SPF. Consider the piconets of figure 2 again, but the gateway of each of the piconets now connects them to form a scatternet, as shown in figure 3. The traffic requirements are the same as shown in figure 2. The SPF of the gateway is 0.5 in Piconet I and 0.5 in Piconet II. 3.3. Fair share We see that for a gateway to be fair, there are two kinds of fairness it has to achieve: that dictated by the PPFs, which achieves fairness between the gateway and the other slaves of a piconet, and that of the SPFs, which distributes the presence of the gateway between its piconets in a fair manner. Both these kinds of fairness may not always be completely achiev-able and this can lead to a change in the values of PPF and SPF, as we now discuss. We observe that an ideal master (as in section 3.1) does not give a gateway more than the PPF of its polling. Thus, if the SPF of a gateway is greater than its PPF for a piconet, the gateway spends a fraction of its time equal to the PPF in the piconet. The gateway cannot stay for a fraction equal to its SPF in the piconet since it is limited by its PPF. Thus, the extra scatternet presence fraction (the difference of the SPF and the PPF) is redistributed in a fair manner among the gateway's other piconets for which the SPF is less than the PPF. This may increase the SPF of the gateway in the other piconets. In other words, the gateway behaves as if its SPF in a particular piconet is reduced to the PPF and thus, its SPF in the other piconets increases. We refer to this changed SPF as the "updated SPF" of the gateway in a piconet . Similarly, an ideal gateway does not stay a fraction of time more than the SPF in a piconet. Thus, if the PPF of the gate-Table 1 Calculation of fair share of the gateway in the two piconets of figure 3. Piconet I Piconet II Actual traffic rate 0.7 0.6 PPF 0.6 0.4 SPF 0.5 0.5 Updated PPF 0.6 0.4 Updated SPF 0.6 0.4 Fair share 0.6 0.4 Figure 3. Gateway shared between two piconets; traffic rates between slaves and the master are shown. way in the piconet is greater than the SPF, the gateway spends a fraction of time equal to the SPF in the piconet. The remaining PPF of the gateway (the difference of the PPF and the SPF) is redistributed in a fair manner among the other slaves of the piconet (if this other slave is a gateway, it is redistributed to it if its SPF is greater than its PPF in the piconet). This may increase the PPF of these slaves. We refer to this changed PPF as the "updated PPF" of the slave in the piconet. In case there is no such redistribution, the updated PPF is equal to the PPF and the updated SPF is equal to the SPF. The fair share can now be calculated from the "updated PPF" and the "updated SPF" as the minimum of these two quantities. Note that all these quantities PPF, SPF, updated PPF, updated SPF and fair shareare dependent on the traffic. Any change in traffic demand of a unit may lead to a change in some of these quantities. We explain the calculation of the fair share using some examples. An example is given in table 1, which shows the actual traffic rate, PPF, SPF, Updated PPF, Updated SPF and fair share of the gateway in the two piconets of figure 3. In Piconet II, the gateway has a PPF of 0.4, which is less than the SPF. In Piconet I, the gateway has a PPF of 0.6 and an SPF of 0.5. Thus, the extra scatternet presence fraction of the gateway in Piconet II (the difference between the SPF and the PPF) is given to Piconet I, which has a higher traffic rate than may be allowed by the SPF. This is reflected in the "updated SPF" values. Thus, the "fair share" of the gateway in Piconet I is 0.6 and in Piconet II is 0.4. The fair shares of the non-gateway slaves are equal to their PPF. As another example, consider the scatternet consisting of 5 piconets with the traffic rates shown as in figure 4. As shown in table 2, gateway G2 has a PPF of 0.5 and an SPF of 0.4 in Piconet B. Thus, the "updated PPF" of G2 in Piconet B is 0.4. The extra PPF ( = PPF - SPF) is added to the PPF of gateway 12 R. KAPOOR ET AL. Figure 4. Scatternet with two gateways. Table 2 Calculation of fair share of the gateways G1 and G2 in the scatternet of figure 4. Gateway G1 Piconet A Piconet B Piconet C Actual traffic rate 0.4 0.6 0.1 PPF 0.25 0.5 0.1 SPF 0.4 0.5 0.1 Updated PPF 0.25 0.6 0.1 Updated SPF 0.25 0.65 0.1 Fair share 0.25 0.6 0.1 Gateway G2 Piconet B Piconet D Piconet E Actual traffic rate 0.7 0.2 0.4 PPF 0.5 0.2 0.4 SPF 0.4 0.2 0.4 Updated PPF 0.4 0.2 0.4 Updated SPF 0.4 0.2 0.4 Fair share 0.4 0.2 0.4 G1 in Piconet B. The "updated PPF" of G1 in Piconet B is, thus, 0.6. Also, gateway G1 has a PPF of 0.25 and an SPF of 0.4 in Piconet A. Thus, the "updated SPF" of G1 in Piconet A is 0.25. The extra SPF ( = SPF-PPF) is added to the SPF of G1 in Piconet B. The "updated SPF" of G1 in Piconet B, is thus, equal to 0.65. The fair shares can now be easily calculated. A division of the master's polling and the gateway's presence based on PPF and SPF as described in this section takes into account the traffic demands of the slaves and the gateways and leads to fairness in the scatternet. In the next section , we introduce and describe an algorithm that aims to achieve such a fair distribution of bandwidth. Description of algorithm We first explain how the algorithm works in the case of a single piconet with no gateway. We then extend the algorithm to the case of a scatternet and explain how the coordination between the master and the gateways is achieved. We then prove the fairness of the algorithm. 4.1. Single piconet with no gateways The polling algorithm is based on the master estimating the traffic rate between each slave and itself. This traffic rate is the sum of the traffic rates from the master to a slave and in the reverse direction. We assume, in order to simplify the explanation of the algorithm, that traffic flows only from slaves to master; masters generate no traffic to slaves. The same algorithm also applies with little change when traffic flows in both directions (explained later). The master uses a Round Robin polling scheme, with the modification that a slave is skipped if it does not belong to the "active list" of the master. The slaves are moved in and out of the active list on the basis of two variables that the master maintains for each slave. These two variables are: r estimate of the rate of traffic generated by the slave; N estimate of the queue length of the slave. When a slave is polled, the masterslave pair gets a chance to exchange a maximum amount of data in each direction, denoted by M. After each such polling phase, the master updates the values of N and r in the following manner: For the slave just polled: N = N + r - x, (1) r = r + (1 - ) xT , x &lt; M , r + (1 - ) xT + , x = M. (2) For other slaves: N = N + r, (3) where is the time elapsed since the last update, x is the amount of data exchanged during the poll phase, T is the total time elapsed since the last poll of the same slave, is a parameter used to smooth the rate estimation and is a parameter used to probe for more bandwidth. Note that x is the actual amount of data exchanged, which may be less than or equal to M, depending upon the number of packets in the slave's queue. Since N is an estimate of the slave's queue length and r is an estimate of the rate at which traffic is generated, N is increased at the rate of r (as in equations (1) and (3)). Also, when a slave is polled, N is decreased by the amount of data exchanged ((equation 1)). After updating these values, the master determines the changes to be made to the active list. A slave is added or deleted from the active list depending upon whether its value of N is greater or smaller than a "threshold". The value of this threshold is the minimum amount of data that the master would like the slave to have in order to poll it. We choose a value equal to a multiple of a DH5 packet for the threshold since this packet incurs least overhead (the selection of the value of the threshold is discussed further in the next subsection ). Thus, a slave is present in the active list if the master's estimate of the value of N for the slave is greater than the threshold. This makes the simple Round Robin polling strategy adaptive to traffic and enables it to utilize bandwidth ef-ficiently , even when slaves have different rates of traffic. The maximum amount of data that can be exchanged at each poll, M , is also set equal to the threshold. Note that if the amount of data, x, in the slave's queue is less than the threshold, the polling of the slave ends after this data has been exchanged. A FAIR AND TRAFFIC DEPENDENT SCHEDULING 13 If the value of N is less than the threshold for all the slaves, then the slave whose value of N is estimated to take the smallest time to reach the threshold is polled, i.e., the slave for which the value of (Threshold - N)/r is the smallest. The master now goes to the next slave according to the Round Robin ordering of slaves. If the slave is present in the active list, it is polled. Else, the procedure is repeated for the next slave in the Round Robin ordering. Also, note that if the amount of data sent by the slave x is equal to M, r is increased by a small amount, . This is basically an attempt by the slave to probe for more bandwidth if it is able to send data at the present rate. The usefulness of this increase is evident in the proof of fairness in the next section. The value of chosen is 0.15 and that of is 0.65. We also discuss the rationale behind choosing these values in the proof of fairness. If traffic flows in both directions, i.e., from the slaves to the master and in the reverse direction, x is the average of the amount of data exchanged in the two directions, r refers to the average of the rate-estimations of the two directions and N refers to the average of the queue length estimates of the two directions. Also, if the number of packets in either direction is less than the threshold, the polling of the slave continues till in both directions, (a) there is no more data to send or (b) amount of data equal to the threshold has been exchanged. The initial value of N is set to the threshold (to ensure that slaves get polled at the beginning) and that of r is set to 0.25 (as a reasonable value). Note that the algorithm converges to the fair share, but a careful selection of initial values makes the initial convergence faster. Another advantage of such a scheme is that it may allow the master to go into a power-saving mode if it realizes that no slave has sufficient packets to send, i.e., if N is smaller than the threshold for all slaves. Though we do not explore this option in this paper, it may be useful since Bluetooth devices are expected to work in power-constrained environments. To improve the algorithm, we add a heuristic to it. The maximum number of polling cycles that a slave is not polled is bounded. If a slave generates a large burst of data occasionally and then does not generate any data for a long time, the value of r for the slave may be very low. This may cause the value of N for the slave to be lower than the threshold for a long time. By limiting the maximum number of cycles missed by the slave, we make sure that such a behavior of the slave does not lead to its starvation. In the experiments, this value is taken to be equal to 5 cycles. We now explain how the above algorithm works in a scatternet. 4.2. Scatternet Scheduling of gateways using Rendezvous Points. Before describing how the algorithm works in a scatternet, we briefly discuss the notion of Rendezvous Points (RPs) described in [4]. A RP is a slot at which a master and a gateway have agreed to meet, i.e., at this slot, the master will poll the gateway and the gateway will listen to the master. In [4], RPs are implemented using the SNIFF mode of Bluetooth, but we implement RPs using the HOLD mode [10]. In the HOLD mode, the slave does not have to listen to the master for a certain time period and may use this time to visit other piconets. Prior to entering the HOLD mode, the master and the slave agree on the time duration the slave remains in the HOLD mode. We implement our algorithm using RPs as described below. The working of the algorithm in a scatternet is very similar to its operation in a piconet. The master continues to poll the non-gateway slaves in the same manner as described in the previous section with the modification that a gateway is polled at a Rendezvous Point. Each RP is a slot at which a particular gateway is polled and a master has different RPs for each of its gateways. These RPs are always unique (i.e., a master cannot have the same RP with more than one gateway). Since the gateway must be polled at the RP, this has implications in the polling of the other slaves (discussed later). Once a gateway has been polled, the master continues with the polling of the other slaves in the same manner as described in the previous section, i.e., it checks its active list to see if the next slave in the polling cycle is to be polled and so on. In order to divide its time among different piconets in a fair manner, the gateway performs similar calculations as described in the earlier section for the master. The gateway maintains values of N and r for each piconet it belongs to and these values are updated each time a gateway is polled (i.e., at each RP). Thus, the calculations performed by a gateway at each RP are: For the piconet in which the gateway was just polled: N = N + r - x, (4) r = r + (1 - ) xT , x &lt; M , r + (1 - ) xT + , x = M. (5) For other piconets: N = N + r, (6) where is the time elapsed since the last update, x is the amount of data exchanged during the poll phase, T is the total time elapsed since the gateway was polled in the same piconet , and and are as defined earlier. Moreover, at each RP, the gateway and the master negotiate the next RP between them. The assignment of this next RP takes into account the fairness between (a) the gateway and other slaves in a piconet and (b) the presence of the gateway in different piconets. Also, we again employ a heuristic that improves the algorithm. When the next RP is being nego-tiated , we keep a bound on the maximum value this can take. This prevents a piconet from not being visited by a gateway for a long time. The maximum value of this next RP used in our experiments is 400 slots. We now see how the master and the gateway use the information that they have to achieve fairness in the scatternet. When a gateway is polled at a RP, the gateway and the master do the following. 14 R. KAPOOR ET AL. (i) Gateway. The gateway calculates the number of slots, N thresh after which N for the piconet will become greater than the threshold; N thresh = (threshold - N)/r, where threshold is as explained in the previous section, N and r are values maintained by the gateway for the piconet. The gateway makes use of this value and does not visit a piconet till its estimate of N for the piconet becomes greater than the threshold. This is similar to the algorithm used by the master in which a slave is not polled till the master's estimate of N for the slave becomes greater than the threshold. Thus, the gateway tries to divide its time between the piconets in a fair manner, i.e., according to the SPFs. Note that N thresh may be negative if N is greater than the threshold. Also, N thresh is allowed to have a maximum value of 400. Moreover, each time a gateway visits a piconet, it knows the RPs for the other piconets it belongs to (except right at the beginning or when the gateway is added to another piconet). (ii) Master. The master calculates the number of slots after which the gateway can be polled such that the fairness with other slaves is maintained. It adopts the following procedure to achieve this: It maintains a counter, num_slots (which is initialized to 0) and checks the value of N for each slave, in a cyclic order, starting from the slave after the current gateway in the cyclic order to the slave before the current gateway. The master checks if the value of N for the slave will be greater than the threshold after num_slots slots. If this condition is true, num_slots is incremented by twice the value of the threshold. After incrementing num_slots, the master also checks to see if it has a RP with any gateway whose value is equal to num_slots and increments num_slots by twice the value of the threshold if this is true. This ensures that the master has a unique RP for each of its gateways. Note that num_slots is incremented by twice the value of the threshold since the master expects to exchange threshold slots of data with a slave in each direction. The master uses the above procedure to estimate the number of slaves who will have their value of N greater than the threshold when the master polls the slaves in their cyclic order starting from the gateway just polled. The value of num_slots determines the number of slots which the master expects to use in polling the other slaves in one cycle before polling the gateway again and is thus, used by the master to maintain fairness between the gateway and the other slaves in the piconet. Again, note that num_slots is allowed to have a maximum value of 400. The master and the gateway now exchange the information they have to calculate their next RP. This exchange takes place using the LMP_hold_req PDU of the LMP (Link Manager Protocol) layer. This PDU carries a hold instant and a hold time, which are used to specify the instant at which the hold will become effective and the hold time, respectively. When the master is sending a packet to a gateway, the value of num_slots can be sent after hold instant and hold time in the packet. The master also sends the values of its RPs with its other gateways in the packet. Similarly, the gateway sends the master the values of its RPs with other piconets and the value of N thresh also in an LMP_hold_req PDU. The master now knows all the RPs of the gateway; similarly, the gateway knows all the RPs of the master. Note that the above information exchange requires a minimal change in the Bluetooth specifications that the contents of the LMP_hold_req PDU need to be enhanced. This PDU is 1-slot in length; thus, some bandwidth of the master is wasted in sending these PDUs. This wasted bandwidth can be reduced by increasing the value of threshold, i.e., the maximum data that a slave and a master may exchange in each direction during one poll of the slave. On the other hand, a large value of the threshold will lead to larger delays for packets. Thus, we have a tradeoff here. We choose a threshold value equal to three times a DH5 packet. The effect of this wasted bandwidth can be seen in the experiments section where the piconet capacity used is slightly less than 1. Note that we pay a small price here to get perfect coordination between the master and the gateway and also to get a high degree of fairness in the system, as the experiments later demonstrate. Now, the master and the gateway both have complete information . So, each of them calculates the next RP in the following manner: They take the maximum value out of num_slots and N thresh and as long as this value is the same as one of the RPs (note that all relevant RPs are known to both the master and the gateway), the value is incremented by 2 threshold. The value at the end of this small procedure is the next RP between the gateway and the master. Since this value takes into account both N thresh and num_slots, it incorporates both the fairness of the master's polling and the gateway's presence. Note that the value of num_slots calculated by the master is just an estimate (the master assumes that each slave included in the calculation of num_slots will exchange threshold slots of data with the master in each direction, but this may not be true). Thus, the master may have polled all the slaves that had to be polled before the RP of the gateway (according to the estimate in the calculation of num_slots) and still be left with some slots before the RP. In this case, the master just continues polling the slaves in their cyclic order and polls the gateway when the time for the RP arrives. Note that this means that the master may have to force a slave to send a packet smaller than a certain length. For example, if two slots are left for the RP, then the master will send a 1-slot packet and ask the slave being polled to do the same. Note that the Bluetooth header has 4 bits to represent the packet type and these can represent 16 packet types. For ACL links, 10 (7 data, 3 control packets) of the packet types are defined. We use 2 of the remaining bit sequences to send packets that force the slave to send packets smaller than or equal to a certain length. This is shown in table 3. From table 3, we see that this procedure is adopted if the number of slots left for the RP is less than 10 (if the number of slots left for the RP is greater than or equal to 10, then the A FAIR AND TRAFFIC DEPENDENT SCHEDULING 15 Table 3 Procedure adopted by the master if slots left for the RP is less than 10. Slots left for RP Maximum Maximum length of packet length of packet sent by master sent by slave 2 1 1 4 1 1 6 3 3 8 3 3 slave's packet length does not have to be restricted). Thus, if the slots left for the RP is 2, the master can send a packet of maximum length = 1 and the gateway can send a packet of maximum length = 1 and so on. Note that for reasons of fairness, the maximum packet length for the master and the gateway is the same. Since the master needs to restrict the maximum length of the gateway's packet to either 1 or 3 (as shown in table 3), we need 2 packet types to achieve this. This procedure effectively suspends the polling of a slave to honor a RP with a gateway. The polling of the slave continues after the gateway has been polled. In addition, a gateway may lose a slot in switching from one piconet to another. This loss is unavoidable since piconets are in general, not synchronized in time. In the experiments in the paper, we set the value of the threshold to three times the payload of a DH5 packet, which can give a switching loss of about 3% at heavy loads (every 2 threshold slots, the gateway loses about one slot in switching). At light loads, this switching loss does not lead to inefficiency since the sum of the fair shares of the gateway in all its piconets is less than 1 and even after the switching loss, the gateway is able to obtain its fair share. The simulations in the next section do not take this switching loss into account and thus, the bandwidth received by the gateway under heavy loads will be a little smaller than the one shown in the results. 4.3. Proof of fairness We now prove that the above algorithm leads to a max-min fair distribution of the bandwidth of a scatternet among units. We start by proving this in the case of a piconet. In the next step, we will extend the proof to the general case of a scatternet . 4.3.1. Fairness in a piconet Let us introduce the following notation: S : number of slave units in the piconet; g i : rate-demand of the ith unit; i : rate achieved by the ith unit; r i : rate-estimation of the ith unit (as defined in equation (2)), where i and r i are average values. Slave unit i is referred to as "satisfied", if it achieves it rate demand, i.e., i = g i ; else, the slave unit is referred to as "unsatisfied". Also, in the proof that follows, "slot" refers to "Bluetooth slot"; "unit" and "slave unit" may be used inter-changeably . If there is one slave unit in a piconet, then it will always get polled and hence, the algorithm is fair. We prove the fairness when there are two or more slave units. We first make the following observations: (a) If a unit has a rate-estimation, r 0.25, it will never achieve a lesser rate than any other unit. r is an estimation of the average number of slots of traffic that a masterslave pair will generate per slot in each direction . Thus, a rate of 0.25 means that a masterslave pair generates, on the average, "threshold" slots of traffic in each direction in every 4 threshold slots. Suppose a piconet has two slaves, and the first has a rate-estimation, r 0.25, then the first slave will be polled at least once in every 4 threshold slots, i.e., will get on the average at least threshold polling slots out of every 2 threshold, regardless of the r of the other slave (since N increases at the rate of r, N will increase by at least 0.25 4 threshold = threshold; thus, the slave will enter into the "active list" in 4 threshold slots). Thus, it will never achieve a lesser rate than another unit. It is easy to see that this property would be true if there were more than two slaves (two slaves is the worst case). (b) For 0.1 and 0.6, an unsatisfied slave will tend to a rate-estimation of at least 0.25. For an unsatisfied slave, the second part of equation (2) (when x = M) is always used for updating the rate. Thus, if r i is the ith rate-estimation: r n +1 = r n + (1 - ) M T n + . This leads to (as n becomes very large): r = (1 - )M k =0 n -k T k + 1 1 - . Thus, for 0.1 and 0.6, for any value of T , the value of r tends to at least 0.25. (c) As long as there is an unsatisfied unit, the utilization of the system capacity is 1 (for 0.15 and 0.65). Consider a piconet consisting of seven slave units, in which the first unit, unit 1 is unsatisfied. From (a) and (b), unit 1 will never achieve a lesser rate than any other unit; this means that it will be polled at least once for each time the other slaves are polled. The value of T (as in equation (2)) for unit 1 is thus, at most, 14 threshold. For this value of T and for = 0.15 and = 0.65, r for unit 1 tends to at least 0.5. A value of r = 0.5 for a slave unit means that it can be polled all the time (since N increases at the rate of r, N will increase by at least 0.5 2 threshold = threshold; thus, the slave will enter into the "active list" in 2 threshold slots, which is also the time of its polling). Thus, the system capacity is totally utilized. If there were less than 7 slave units, the value of T would be smaller (than 14 threshold), and r would tend to a higher value (than 0.5). 16 R. KAPOOR ET AL. We choose values of and to satisfy the above properties , i.e., = 0.15 and = 0.65. The following statements hold. (i) Units with the same rate-demand achieve the same average rate: g i = g j i = j . We prove this by contradiction. Suppose there are two units, unit 1 and unit 2 with rate demands g 1 and g 2 , respectively, such that g 1 = g 2 . Also, suppose one unit achieves a higher average rate than the other, 1 &gt; 2 . Now, unit 2 does not achieve its rate-demand (since 1 &gt; 2 ) . Unit 1 may or may not achieve its rate demand. From property (b), unit 2 will always tend to a value at least equal to 0.25, since it is an unsatisfied slave. Using property (a), this implies that 2 cannot be less than 1 . This is a contradiction. (ii) Units with a higher rate-demand achieve an average rate at least equal to that achieved by units with a lower rate-demand : g i &gt; g j i j . This can be proved by contradiction in the same manner as in part (i). Now, without loss of generality, let us partition the slave units into two sets, S1 and S2, in such a way that units in S1 are satisfied, while units in S2 are not. If the set S2 is empty, than all the units achieve their rate-demand and the system is fair. If the set S2 is not empty, then using statements (i) and (ii), all units share the bandwidth in a fair manner. Moreover, since S2 contains at least one unit, the total system capacity is utilized. Hence, it is not possible to increase the rate of a unit in S2 without decreasing the rate of some other unit. 4.3.2. Fairness in a scatternet The proof of fairness for a scatternet follows trivially from that for a piconet. We make the following two observations: (1) The gateway visits a piconet only after the estimation of N for the piconet becomes greater than the threshold (it calculates N thresh while determining the next RP). In other words, the "virtual master" (gateway) does not poll (visit) its "virtual slave" (master) till the estimate of N becomes greater than the threshold. This is similar to the algorithm used by the master to poll the slaves in which a slave is not polled till its estimate of N becomes greater than the threshold. Thus, the gateway divides its presence among its piconets in a fair manner, i.e., according to the SPF. Note that if the PPF for a gateway in a piconet is less than its SPF, the master does not poll the gateway for more than the PPF. Thus, the apparent rate demand and SPF for the gateway in the piconet are reduced. This may increase the SPF of the gateway in other piconets. In this case, the gateway divides its presence according to the updated SPFs. (2) While calculating the next RP for a gateway, the master calculates the num_slots value which estimates the number of slaves in one polling cycle (starting from the slave after the gateway in the polling cycle) who will have their values of N greater than the threshold at the estimated time of their poll. This achieves fairness between the gateway and the non-gateway slaves. Also, the master continues to use the same algorithm for polling non-gateway slaves in a scatternet as described for a piconet in section 4.1. This maintains fairness between non-gateway slaves, i.e., the division is done according to the PPFs (or the updated PPFs). 4.4. Overhead/limitations of the algorithm The rate calculations will lead to a higher load on the system. Also, the algorithm does not take into account SCO links. We believe (and as has been shown in [6]) that ACL links are capable of carrying voice with small delays. The controlled channel access in Bluetooth can ensure good support of voice using ACL links. Also, scheduling in a scatternet where SCO links are allowed may not be feasible. Since SCO links require a periodic reservation of two slots every two, four or six slots, meeting the demands of such a link with a gateway may be impossible when the gateway is visiting some other piconet. Experiments and results In this section, we present simulation results, which show that the algorithm satisfies the fairness criteria described earlier. We start with simple topologies that illustrate the behavior of the algorithm and then show that it also works well in more complex topologies. There are three topologies that the experiments focus on and these demonstrate the behavior of the algorithm a topology with (a) a gateway belonging to two piconets , (b) a gateway belonging to three piconets and (c) a piconet having two gateways. The experiments also show the adaptivity of the algorithm, i.e., how quickly the algorithm adapts to changing traffic demands of slaves. In the experiments, we specify the "rate of a slave", which refers to the sum of the rates at which a slave generates data for a master (i.e., the rate demand of a slave) and the master generates data for the slave. Moreover, unless mentioned otherwise, we assume that the traffic rate from a slave to a master is equal to that from the master to the slave. Thus, a slave having a rate of 0.4 means that the slave generates data at the rate of 0.2 Bluetooth slots per slot and the master also has a rate demand of 0.2 towards the slave. As we show in the section on asymmetric traffic, the algorithm works well even if these two rates are not the same. The simulation environment used in our experiments is NS-2 [8]. We have augmented NS-2 with the Bluetooth model. The simulator models the Bluetooth baseband, LMP and L2CAP layers and enables the creation of piconets and scatternets. The model contains most of the standard features of Bluetooth like Frequency Hopping, Multi-Slot Packets, Fast ARQ (Automatic Retransmission Query). Note that as mentioned earlier, in our simulator, the switching loss asso-ciated with the gateway moving from one piconet to another A FAIR AND TRAFFIC DEPENDENT SCHEDULING 17 Figure 5. Example scatternet. is not taken into account. This effect can lead to the gateway losing up to 3% of slots at heavy loads. The experiment results are thus, a slight overestimate. In the experiments, all traffic generated is CBR. Each experiment is run for a system time of 32 sec. In the experiments , the term "slave" refers to a non-gateway slave; a gateway slave is referred to as "gateway". Also, in experiments where the PPF and the SPF values (and not the updated PPF and the updated SPF) are shown, the PPF and the updated PPF are equal and the SPF and the updated SPF are also equal. In the graphs, "BW" in the index stands for bandwidth, "GW" stands for gateway. 5.1. Single gateway in two piconets We first consider the simple topology shown in figure 5, which consists of two piconets, numbered I and II, connected by a single gateway. We consider various cases by changing the traffic and the number of slaves in the piconets. Experiment 1. Adaptation between gateway and non-gateway slave traffic Each piconet has one non-gateway slave that generates very high traffic, with rate equal to 1, to the master. The gateway has equal traffic to both masters. We vary the gateway traffic to show the fair sharing of the piconet bandwidth between the gateway and the slave. We show the results for one piconet since the two piconets are exactly symmetric. Figure 6(a) shows the sharing of bandwidth between the gateway and slave for different values of gateway traffic. It also shows the fair share of the slave and the total fraction of the bandwidth obtained by the gateway and the slave in the piconet. It can be seen that the slave obtains a bandwidth equal to its fair share for different values of gateway traffic. Moreover, the sum of the bandwidths obtained by the slave and the gateway is nearly equal to 1. The reason for this to be slightly less than 1 is that some of the piconet capacity is used in sending LMP_hold_req PDUs of the LMP layer. In figure 6(b), the comparison of the fraction of the bandwidth obtained by the gateway to its SPF (PPF and SPF are equal) is shown. Figure 6(b) shows that the gateway gets almost equal to its fair share of the bandwidth for all values of traffic. Again, the reason that the gateway obtains slightly less than its fair share is because some of the slots are used for LMP PDUs. This also explains why the gateway obtains slightly less than the slave in figure 6(a). Figure 6. (a) Sharing of bandwidth between gateway and slave. (b) Comparison of fraction of bandwidth obtained to SPF for the gateway. Experiment 2. Different traffic to piconets The same topology as in the previous case, but each slave has a traffic rate of 0.3 to the master. The gateway has a fixed traffic rate of 0.2 to the master of Piconet I and variable traffic to the other master. The PPF and SPF of the gateway in the first piconet are, thus, both equal to 0.2. The traffic in Piconet I does not change and the gateway and the slave get a constant fraction of 0.2 and 0.3 of the piconet bandwidth, respectively. Figure 7(a) shows the sharing of bandwidth between the gateway and slave for different values of gateway traffic, while figure 7(b) shows the comparison of the fraction of the bandwidth obtained by the gateway in Piconet II to its SPF and PPF. From the graphs, we can see that when the gateway has different traffic to piconets, it divides its presence among the piconets according to the traffic offered and in a fair manner (again, the gateway obtains slightly less than its fair share due to the LMP PDUs). Also, the gateway makes use of the lower traffic offered by the slave in Piconet II to obtain a higher share of the bandwidth in Piconet II. Experiment 3. Different number of slaves Piconet I has 3 slaves, while the number of slaves in Piconet II is variable. Each slave generates traffic to the master at the rate of 0.2. The gateway has a traffic rate of 0.3 to Piconet I and 0.8 to Piconet II. The PPF and SPF of the gateway in Piconet I are, thus, 0.2 and 0.3, respectively. In Piconet II, the value of PPF changes depending upon the number of slaves. In Piconet I, the slaves get a bandwidth fraction of 0.2 and the gateway gets 0.3. Figure 8(a) shows the sharing 18 R. KAPOOR ET AL. Figure 7. (a) Sharing of bandwidth between gateway and slave in Piconet II. (b) Comparison of fraction of bandwidth obtained by the gateway to SPF and PPF in Piconet II. of bandwidth between the gateway and each slave in Piconet II. Figure 8(b) shows the comparison of the fraction of the bandwidth obtained by the gateway in Piconet II to the SPF and PPF. The gateway receives a fraction of the bandwidth almost equal to its fair share. Also, as the number of slaves increases, the fraction of the bandwidth received by the gateway (and each slave) reduces in a fair manner. Experiment 4. Asymmetric traffic We now consider a case where the traffic rates from Master to Slave and Slave to Master are different (asymmetric traffic ). We consider the same topology as in experiment 2 of the current section, with the non-gateway slaves having the same rate as in experiment 2. The gateway has a fixed traffic rate of 0.2 to the master of Piconet I and variable traffic to the other master. The variable traffic is such that traffic from Master to Slave has a rate of 0.1 and traffic from Slave to Master varies. Figure 9 shows the comparison of bandwidth fraction obtained by the gateway in this experiment versus that obtained by the gateway in experiment 2 in Piconet II for different values of gateway traffic (which is the sum of master to gateway and gateway to master traffic rates). We see that the fraction is slightly lower than the fraction obtained in experiment 2. Asymmetric traffic leads to wastage of slots, since an empty slot is returned in one direction where there is no data to send. It can be seen though, that the gateway still behaves in an ap-Figure 8. (a) Sharing of bandwidth between gateway and slave in Piconet II. (b) Comparison of fraction of bandwidth obtained by the gateway to SPF and PPF in Piconet II. Figure 9. Comparison of fraction of bandwidth obtained by gateway in this experiment with that in experiment 2 in Piconet II. proximately fair manner. All other bandwidth fractions for slaves and the gateway are the same as in experiment 2. 5.2. Single gateway shared between three piconets We now consider a topology, where a gateway is shared between 3 piconets, numbered I, II and III. Piconet I has 5, Piconet II has 1 and Piconet III has 4 slaves. Each slave has a traffic rate of 0.2. The gateway has a traffic rate of 0.2 to Piconet I, 0.3 to Piconet III and a variable rate to Piconet II. All traffic is symmetric (same from master to slave and from slave to master). A FAIR AND TRAFFIC DEPENDENT SCHEDULING 19 Figure 10. Bandwidth fraction received by gateway in the three piconets. Figure 11. Example scatternet topology. Figure 10 shows the fraction of bandwidth obtained by the gateway in each piconet with increasing gateway traffic rate to Piconet II. It also shows the PPF and the Updated SPF of the gateway in Piconet II. We do not show the fair shares of the gateway in Piconet I and III since they are constant (0.16 and 0.2, respectively). It can be seen that the gateway manages to get close to its fair share in the 3 piconets. The slaves in Piconet I get a bandwidth fraction of 0.16 and the slaves in Piconet II and III get a bandwidth fraction of 0.2 (all these are equal to their fair shares). 5.3. Piconet with two gateways We now show the working of the algorithm in a piconet having 2 gateways, as shown in figure 11. Piconets I, II and III have 6, 2 and 4 non-gateway slaves, respectively. There are two gateways, GW 1 between Piconets I and II; and GW 2 between Piconets II and III. All slaves have a traffic rate of 0.2. GW 1 has a traffic rate of 0.2 in Piconet I and 0.5 in Piconet II. GW 2 has a traffic rate of 0.2 in Piconet III. We vary the traffic rate of GW 2 in Piconet II and show the fair sharing of bandwidth. Figure 12 shows the fraction of bandwidth obtained by GW 1 and GW 2 in Piconet II compared to their fair shares. The x-axis denotes GW 2 traffic in Piconet II. It can be seen that the bandwidth fractions obtained are very close to the fair value. The non-gateway slaves of Piconet II receive a bandwidth fraction of 0.2, which is equal to their fair share (not shown in the figure). The bandwidth fraction received by slaves in Piconets I and III does not change for different values of GW2 traffic in Piconet II. The fair share of each slave (including the gateway) in Piconet I is 0.14 and in Piconet III Figure 12. Fraction of bandwidth and fair share of GW1 and GW2 in Piconet II. Figure 13. Actual rate estimation of the gateway and its ideal value. is 0.2; the bandwidth fraction received by each slave is very close to these fair shares. 5.4. Adaptivity to changing traffic demands We now show how quickly the algorithm is able to adapt to changing traffic. We again consider the scenario of experiment 1 of section 5.1, consisting of two piconets, each having a non-gateway slave, connected by a single gateway. The non-gateway slaves have a traffic rate of 1; the gateway has equal traffic to both the masters. We vary the traffic rate of the gateway as time progresses: for the first 2.5 seconds, the gateway's rate is 0.1, for the next 2.5 seconds, it is 0.5 and for the remaining time, it is 0.3. Figure 13 shows the actual rate estimation of the gateway (and its ideal value) versus time. It can be seen that the rate estimation adapts very quickly to the new rate. For example, when the rate changes from 0.1 to 0.5, the rate estimation reaches a value of 0.45 in about half a second after 2.5 sec. Thus, the algorithm adapts to quickly changing traffic. Conclusions This paper proposed a distributed scatternet-scheduling algorithm that adapts to non-uniform and changing traffic. This 20 R. KAPOOR ET AL. algorithm provides an integrated solution for both intra- and inter-piconet scheduling and can be implemented using the HOLD mode of Bluetooth. Through analysis and simulations, we showed that the algorithm is traffic-adaptive and results in a fair allocation of bandwidth to units. We explained earlier that the algorithm may allow a unit to go into a power-saving mode. In future, we would like to explore this option, which also assumes importance since Bluetooth devices will most likely operate in a power-constrained environment. As future work, we would also like to evaluate the performance of TCP and other kinds of traffic on our algorithm. We are also working towards interfacing the algorithm with requirements of higher layers. In this respect, we are working towards providing QoS support using the algorithm. References [1] S. Baatz, M. Frank, C. Kehl, P. Martini and C. Scholz, Adaptive scatternet support for Bluetooth using sniff mode, in: Proc. of IEEE LCN (2001). [2] 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: Proc. of IEEE INFOCOM'2001 (2001). [3] J. Haartsen, BLUETOOTH the universal radio interface for ad hoc wireless connectivity, Ericsson Review 3 (1998) 110117. [4] P. Johansson, M. Kazantzidis, R. Kapoor and M. Gerla, Bluetooth an enabler for personal area networking, IEEE Network Magazine, Wireless Personal Area Network (September 2001). [5] M. Kalia, D. Bansal and R. Shorey, MAC scheduling and SAR policies for Bluetooth: A master driven TDD pico-cellular wireless system, in: Proc. of 6th IEEE International Workshop on Mobile Multimedia Communications (MOMUC) (1999). [6] R. Kapoor, L. Chen, Y. Lee and M. Gerla, Bluetooth: carrying voice over ACL links, in: Proc. of MWCN (2002). [7] A. Mayer, Y. Ofek and M. Yung, Approximating max-min fair rates via distributed local scheduling with partial information, in: Proc. of IEEE INFOCOM (1996). [8] NS-2 simulator, http://www.isi.edu/nsnam/ns/ [9] A. Racz, G. Miklos, F. Kubinszky and A. Valko, A pseudo-random coordinated scheduling algorithm for Bluetooth scatternets, in: Proc. of MobiHoc (2001). [10] Specifications of the Bluetooth System core, Vol. 1, v. 1.1, www. Bluetooth.com [11] W. Zhang and G. Cao, A flexible scatternet-wide scheduling algorithm for Bluetooth networks, in: Proc. of IEEE IPCCC (2002). Rohit Kapoor received his Bachelor degree in computer science in 1999 from the University of Roor-kee , India. He is currently a Ph.D. candidate at the University of California, Los Angeles (UCLA). His research focuses on Bluetooth-based personal area networks. He is a member of the Network Research Lab at UCLA. E-mail: rohitk@cs.ucla.edu Andrea Zanella received the Ph.D. degree in telecommunication engineering from the University of Padova, Italy, in 2002. Prior to that he received the Dr. Ing. degree (comparable to Master degree) in computer engineering in 1998, still from the University of Padova. He spent nine months, in 2001, as post-doc researcher at the Department of Computer Science of the University of California, Los Angeles (UCLA), where he was engaged in research on Wireless Networks and Wireless Access to Internet under the supervision of Prof. Mario Gerla. Currently, he is a research fellow in the Department of Information Engineering of the University of Padova, Italy. His research interests are mainly focused on topics related to wireless and mobile networking. In particular, in the last period, he has been working on the performance aspects of wireless personal area networks based on the Bluetooth standard. E-mail: zanella@dei.unipd.it Mario Gerla is a professor in the Computer Science Department at UCLA. He received his graduate degree in engineering from the Politecnico di Milano in 1966, and his M.S. and Ph.D. degrees in engineering from UCLA in 1970 and 1973, respectively. He joined the faculty of the UCLA Computer Science Department in 1977. His current research is in the area of analysis, design and control of communication networks. Ongoing projects include the design and evaluation of QoS routing and multicast algorithms for IP domains, the design and evaluation of all-optical network topologies and access protocols, the design of wireless mobile, multimedia networks for mobile computing applications, and the development of measurement methods and tools for evaluating the performance of high-speed networks and applications. E-mail: gerla@cs.ucla.edu

scheduling scheme;Round Robin;Distributed algorithm;scheduling;traffic adaptive;Scatternet presence fraction;Fairness;traffic rate;scatternets;Scatternet;Bluetooth;Rendezvous Points;Scheduling algorithm;Information exchange;heuristic;Gateway;Scheduling of gateways;Slaves;Non-uniform traffic;changing traffic;Efficiency;fair share;virtual slave;Rendezvous Point;Blueooth;Piconet presence fraction;HOLD mode;Slave unit;polling algorithm;Gateway slave traffic;Scatternets;Master unit;Allocation of bandwidth;Rendezvous point;fairness;Piconet;piconet;slave;Traffic Dependent Scheduling;Time-division multiplex;scatternet;Fair share;Bluetooth technology;Non-gateway slave traffic;bandwidth utilization;allocation of bandwidth;gateway;master;Round Robin polling

70

DirectoryRank: Ordering Pages in Web Directories

ABSTRACT Web Directories are repositories of Web pages organized in a hierarchy of topics and sub-topics. In this paper, we present DirectoryRank , a ranking framework that orders the pages within a given topic according to how informative they are about the topic. Our method works in three steps: first, it processes Web pages within a topic in order to extract structures that are called lexical chains, which are then used for measuring how informative a page is for a particular topic. Then, it measures the relative semantic similarity of the pages within a topic. Finally, the two metrics are combined for ranking all the pages within a topic before presenting them to the users.

INTRODUCTION A Web Directory is a repository of Web pages that are organized in a topic hierarchy. Typically, Directory users locate the information sought simply by browsing through the topic hierarchy, identifying the relevant topics and finally examining the pages listed under the relevant topics. Given the current size and the high growth rate of the Web [10], a comprehensive Web Directory may contain thousands of pages within a particular category. In such a case, it might be impossible for a user to look through all the relevant pages within a particular topic in order to identify the ones that best represent the current topic. Practically, it would be more time-efficient for a user to view the Web pages in order of importance for a particular topic, rather than go through a large list of pages. One way to alleviate this problem is to use a ranking function which will order the pages according to how "informative" they are of the topic that they belong to. Currently, the Open Directory Project [3] lists the pages within a category alphabetically, while the Google Directory [1] orders the pages within a category according to their PageRank [11] value on the Web. While these rankings can work well in some cases, they do not directly capture the closeness of the pages to the topic that they belong to. In this paper, we present DirectoryRank, a new ranking framework that we have developed in order to alleviate the problem of ranking the pages within a topic based on how "informative" these pages are to the topic. DirectoryRank is based on the intuition that the quality (or informativeness) of a Web page with respect to a particular topic is determined by the amount of information that the page communicates about the given topic, relative to the other pages that are categorized in the same topic. Our method takes as input a collection of Web pages that we would like to rank along with a Web Directory's topic hierarchy that we would like to use. At a high level, our method proceeds as follows: first, we identify the most important words inside every page and we link them together , creating "lexical chains". We then use the topic hierarchy and the pages' lexical chains to compute the "relatedness" (or importance ) of the pages to each of their corresponding topics. Having determined the pages' topic importance, we measure the relative semantic similarity among the pages that relate to the same topic. The semantic similarity indicates the amount of content that important pages in some topic share with each other. Finally, we employ our DirectoryRank algorithm that uses the topic importance scores in conjunction with the semantic similarities of the pages in order to compute the ranking order of the pages within a Directory topic. In order to study the effectiveness of DirectoryRank in identifying the most informative pages within a particular topic, we applied our method to the ranking of 318,296 Web pages listed in 156 topics in the Google Directory. We have compared the rankings induced by DirectoryRank to the rankings induced by PageRank for the pages listed in those 156 topics. Our comparison reveals that the two rankings have different merits and thus they are useful in different tasks. To delve into the two rankings' effectiveness and investigate which is more useful for ordering pages in Directories' topics, we conducted a user study, where we asked a group of individuals to compare the rankings delivered by PageRank to the rankings delivered by DirectoryRank, and indicate which of the two is deemed as more useful. Our results show that, in most cases, the users perceived DirectoryRank to be more topic-informative than PageRank. The rest of the paper is organized as follows: We start our discussion in Section 2 with a brief introduction to PageRank, which is currently employed by the Google Directory in order to rank pages. In Section 3, we briefly present the topic hierarchy that we use in our study as well as the process we follow for representing Web pages into lexical chains. We also show how we explore the topic 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. WIDM'05, November 5, 2005, Bremen, Germany Copyright 2005 ACM 1-59593-194-5/05/0011...$5.00. 17 hierarchy and the pagers' lexical chains for measuring the pages' topic-importance and semantic similarities values. Finally, we present how our DirectoryRank metric employs the above values for measuring how informative Web pages are with respect to some topics and rank them accordingly. In Section 4, we experimentally study the effectiveness of DirectoryRank, by comparing its performance to PageRank. We revise related work in Section 5 and we conclude the paper in Section 6. OVERVIEW OF PAGERANK In this section, we briefly explain the main intuition of PageRank, a metric that was primarily invented for ranking pages within the Google Search Engine and that is currently used within the Google Directory for ordering Web pages. For a more elaborate overview on PageRank, we refer the reader to the work of [11]. The intuition of PageRank metric is that a page on the Web is important if there are a lot of other important pages pointing to it. That is, if a page p has many links from other important pages, we may conclude that this page is interesting to many people and that it should be considered as being "important" or of "good" quality. Similarly, if an important page has links to other pages, we expect that part of its quality is transferred to the pages it points to, which in turn become of increased significance/quality. Roughly, PageRank PR(p) defines the importance of page p to be the sum of the importance of the pages that endorse p. At a high level, PageRank is calculating the probability that a "random surfer" is looking at a given page at a given point of time. The "random surfer" is a mathematical model that emulates the behavior of a user that, given a page, follows an outgoing link from that page at random. Formally, given a page p i that has incoming links from the pages p 1 , ..., p n and let c j be the number of out-links from p j , the PageRank of p i is given by: 1 1 ( ) (1 ) ( ) / ... ( ) / i n n PR p d d PR p c PR p c = + + + where d corresponds to the probability that the random surfer will get bored and his next visit will be a completely random page, and 1-d corresponds to the probability that the page the random surfer will pick for his next visit is an outgoing link of the current page. DIRECTORY RANK The ranking of a Web page within a particular topic intuitively depends on two criteria: (i) the importance of the page for the underlying topic. This criterion helps us identify the most important pages out of the several ones that may lie within a particular topic. (ii) the semantic correlation of a page relative to other important pages in the same topic. This criterion helps us rank pages relative to each other within a topic. For measuring the importance of a Web page in some topic, we explore a subject hierarchy that we have built in the course of an earlier study [13] and we use the lexical chaining technique for identifying the most important words inside the page. We start our discussion with a presentation of the topic hierarchy that we use in our work (Section 3.1) and we describe the process we follow for representing Web pages into lexical chains (Section 3.2). We also explain how we utilize the topic hierarchy and the pages' lexical chains for measuring the pages' importance to the hierarchy's topics (Section 3.2.1). The contribution of our work lies in the exploitation of the topic hierarchy and the lexical chains that we generate for representing Web pages in order to compute the semantic similarities between the pages that are important in some topics. Moreover, we have developed a novel framework, which employs the pages' topic importance and semantic similarity measures for ranking pages inside Directory topics. 3.1 The Topic Hierarchy The main intuition in our DirectoryRank metric is that topic relevance estimation of a Web page relies on the page's lexical coherence , i.e. having a substantial portion of words associated with the same topic. To capture this property, we adopt the lexical chaining approach: for every Web page we generate a sequence of semantically related terms, known as lexical chain. In our approach of representing Web pages into lexical chains, we adopt the method reported in [6], which uses WordNet [5] as the knowledge base for providing information about the semantic relations that exist between the words in a text. A detailed description of the lexical chains' generation process is given in Section 3.2. Before that, we present the topic hierarchy that we use for determining the topics that are associated with the contents (i.e. words) of Web pages. Since we are mainly interested in measuring the Web pages' importance in the context of Web Directories, we decided to demonstrate the usefulness of our DirectoryRank metric in ordering Web pages in the topics currently used in a real Web Directory. To that end, we applied DirectoryRank to the main topics used in the Google Directory . Google Directory provides a hierarchical listing of Web pages categorized by topic and reuses the data maintained by the Open Directory Project. Moreover, since DirectoryRank relies on the Web pages' lexical chains rather than their entire contents for measuring the pages' importance to particular topics and since lexical chain generation is dependent on WordNet, we decided to enrich the top level (main) topics of the Google Directory with their respective WordNet lexical hierarchies. The first step we took for enriching the Google topics with WordNet data was to examine the compatibility between these topics and the topics used to annotate WordNet's concepts with domain information . Note that the topic information that exists in the labels of WordNet's contents is taken from the freely available Suggested Upper Merged Ontology (SUMO) [4] and the MultiWordNet Domains (MWND) [2]. Due to space limitations, here we present a summary of our approach into enriching the Google topics with WordNet hierarchies. A detailed description of the process we followed for appending to the Google top level topics their corresponding WordNet hierarchies is given in [13]. In brief, we located the Google's top level topics among the topics used in either SUMO or MWND for annotating WordNet concepts. Out of the 17 Google topics, 13 topics (shown in Table 1) are used for labeling WordNet concepts with topic information. To each of those 13 topics, we integrated their corresponding sub-topics that we ac-quired from either SUMO or MWND. The sub-topic integration was performed automatically, simply by following WordNet's hyper/hyponymy links. At the end of this process, we came down to a hierarchy of 489 sub-topics, which are organized into the 13 top level topics that we used from Google Directory. Table 1. The Hierarchy's First Level Topics First Level Topics Arts News Sports Society Games Computers Home Reference Shopping Recreation Business Science Health In Section 3.4, we will demonstrate how to use our topic hierarchy for automating the task of ranking pages within topical categories. 18 3.2 Measuring Web Pages' Topic Importance The computational model that we adopted for generating lexical chains is presented in the work of Barzilay [6] and it generates lexical chains in a three step approach: (i) it selects a set of candidate terms 1 from a page, (ii) for each candidate term, it finds an appropriate chain relying on a relatedness criterion among members of the chains, and (iii) if such a chain is found, it inserts the term in the chain. The relatedness factor in the second step is determined by the type of WordNet links that connect the candidate term to the terms stored in existing lexical chains. Figure 1 illustrates an example of the lexical chain generated for a text containing the candidate terms: system, network, sensor, weapon, missile, surface and net. The subscript si denotes the id of the word's sense within WordNet. Lexical chain system s6 network s4 system s6 sensor s1 system s6 weapon s2 missile s1 system s6 surface s1 net s2 Figure 1. An example of a lexical chain. Having generated lexical chains, we disambiguate the sense of the words inside every chain by employing the scoring function f introduced in [12], which indicates the probability that a word relation is a correct one. Given two words, w 1 and w 2 , their scoring function f via a relation r, depends on the words' association score, their depth in WordNet and their respective relation weight. The association score (Assoc) of the word pair (w 1 , w 2 ) is determined by the words' co-occurrence frequency in a corpus that has been previously collected. In practice , the greater the association score between a word pair w 1 and w 2 is, the greater the likelihood that w 1 and w 2 refer to the same topic. Formally, the (Assoc) score of the word pair (w 1 , w 2 ) is given by: 1 2 1 2 1 2 log ( ( , ) 1) ( , ) ( ) ( ) + = s s p w w Assoc w w N w N w where p(w 1 ,w 2 ) is the corpus co-occurrence probability of the word pair (w 1 ,w 2 ) and N s (w) is a normalization factor, which indicates the number of WordNet senses that a word w has. Given a word pair (w 1 , w 2 ) their DepthScore expresses the words' position in WordNet hierarchy and is defined as: 2 2 1 2 1 2 ( , ) ( ) ( ) DepthScore w w Depth w Depth w = , where Depth (w) is the depth of word w in WordNet. Semantic relation weights (RelationWeight) have been experimentally fixed to 1 for reiteration, 0.2 for synonymy and hyper/hyponymy, 0.3 for antonymy, 0.4 for mero/holonymy and 0.005 for siblings. The scoring function f of w 1 and w 2 is defined as: 1 2 1 2 1 2 ( , , ) ( , ) ( , ) Re ( ) s f w w r Assoc w w DepthScore w w lationWeight r = The value of the function f represents the probability that the relation type r is the correct one between words w 1 and w 2 . In order to disambiguate the senses of the words within lexical chain C i we calculate its score, by summing up the f s scores of all the words w j1 w j2 (where w j1 and w j2 are successive words) within the chain C i . Formally, the score of lexical chain C i , is expressed as the sum of the score of each relation r j in C i . 1 2 ( ) ( , , ) i s j j j r in C j j Score C f w w r = 1 As candidate terms, we use nouns and proper names because they convey the vast majority of conceptual information in texts. Eventually, in order to disambiguate we will pick the relations and senses that maximize the Score (C i ) for that particular chain. In estimating the importance of a Web page p i in some Directory's topic T k our first step is to identify which node within the hierarchy (see Section 3.1) corresponds to topic T k of the page. 3.2.1 Topic-Importance Scoring Once the topic of a page is located among the hierarchy's topics, we map the words in the page's lexical chain to the WordNet nodes under that particular topic. Recall that upon lexical chain generation , words are disambiguated ensuring that every word inside a page is mapped to a single word within the WordNet hierarchy. We then determine the importance of a page p i to topic T k by counting the number of words in the lexical chain of p i that are subsumed by T k in the hierarchy's graph. The topic importance of a page is given by a Relatedness Score (RScore), which indicates how relevant a page is for a given topic. Formally, the relatedness score of a page p i (represented by the lexical chain C i ) to the hierarchy's topic T k is defined as the product of the page's chain Score (C i ) and the fraction of words in the page's chain that are descendants of T k . Formally , the RScore is given by: RScore (C i , T k ) = k i i i Score(C ) common C and T elements C elements The denominator is used to remove any effect the length of a lexical chain might have on RScore and ensures that the final score is normalized so that all values are between 0 and 1, with 0 corresponding to no relatedness at all and 1 indicating the page that is highly expressive of the page's topic. The RScore of a page to a specific topic captures the importance of the page in the given topic. 3.3 Semantic Similarity Scoring The relatedness score metric that we have just presented can serve as a good indicator for identifying the most important pages within a topic. However, the RScore metric does not capture the amount of common content that is shared between the Web pages in a topic. This is important in the cases where our topic-importance scoring gives a low score for some pages but, at the same time, these pages are very similar to other pages with high topic-importance scores. In order to accommodate for this scenario, we now show how to compute the semantic similarities among the pages that are listed in the same Directory topic. Semantic similarity is indicative of the pages' semantic correlation and helps in determining the ordering of the pages that are deemed important in some topic. Our DirectoryRank metric employs the Web page's topic-importance scores and their semantic similarities to determine their ranking order inside some Directory topics and is presented in the next section. In order to estimate the Web pages' semantic similarity, we compare the elements in a page's lexical chain to the elements in the lexical chains of the other pages in a Directory topic. We assume that if the chains of two Web pages have a large number of elements in common, then the pages are correlated to each other. To compute similarities between pages, p i and p j that are categorized in the same topic, we first need to identify the common elements between their lexical chains, represented as PC i and PC j, respectively. First, we use WordNet to augment the elements of the lexical chains PC i and PC j with their synonyms. Chain augmentation ensures that pages of comparable content are not regarded unrelated, if their lexical chains contain distinct, but semantically equivalent elements . The augmented elements of PC i and PC j are defined as: 19 ( ) ( ) i i i AugElements PC C Synonyms C = U ( ) ( ) j j j AugElements PC C Synonyms C = U where, Synonyms (C i ) denotes the set of the hierarchy's concepts that are synonyms to any of the elements in C i and Synonyms (C j ) denotes the set of the hierarchy's concepts that are synonyms to any of the elements in C j . The common elements between the augmented lexical chains PC i and PC j are determined as: ( , ) ( ) ( ) i j i j ComElements PC PC AugElements PC AugElements PC = I We formally define the problem of computing pages' semantic similarities as follows: if the lexical chains of pages p i and p j share elements in common, we produce the correlation look up table with tuples of the form &lt; AugElements (PC i ), AugElements (PC j ), ComE-lements&gt ;. The similarity measurement between the lexical chains PC i , PC j of the pages P i and P j is given by: ) ) 2 ( , ) ( , ) ( ( i j i j i j ComElements PC PC PC PC s AugElements PC AugElements PC = + where, the degree of semantic similarity is normalized so that all values are between zero and one, with 0 indicating that the two pages are totally different and 1 indicating that the two pages talk about the same thing. 3.4 DirectoryRank Scoring Pages are sorted in Directory topics on the basis of a DirectoryRank metric, which defines the importance of the pages with respect to the particular topics in the Directory. DirectoryRank ( DR) measures the quality of a page in some topic by the degree to which the page correlates to other informative/qualitative pages in the given topic. Intuitively, an informative page in a topic, is a page that has a high relatedness score to the Directory's topic and that is semantically close (similar) to many other pages in that topic. DR defines the quality of a page to be the sum of its topic relatedness score and its overall similarity to the fraction of pages with which it correlates in the given topic. This way, if a page is highly related to topic D and also correlates highly with many informative pages in D, its DR score will be high. Formally, consider that page p i is indexed in Directory topic T k with some RScore (p i , T k ) and let p 1 , p 2 , ..., p n be pages in T k with which p i semantically correlates with scores of s (PC 1 , PC i ), s (PC 2 , PC i ),..., s (PC n , PC i ), respectively. Then, the DirectoryRank ( DR) of p i is given by: 2 1 ( ) ( , ) [ ( , ) ( , ) ...... ( , )] , i k i k s i s i s n DR p T RScore p T PC PC PC PC PC PC n i = + + + + / where n corresponds to the total number of pages in topic T k with which p i semantically correlates. EXPERIMENTAL SETUP To measure the potential of our DirectoryRank metric in delivering topic-informative rankings, we conducted an experiment, where we studied the effectiveness of DR in prioritizing the most informative pages in some Directory's topics. To obtain perceptible evidence of DirectoryRank's efficiency in a practical setting, we applied our DR metric to a set of Web pages listed in a number of topics in Google Directory and we compared the rankings induced by DirectoryRank to the rankings that Google Directory delivers for the same set of pages and topics. In Section 4.1 we explain how we selected the pages for our study, while in Section 4.2 we present the similarity measure that we used for comparing the rankings induced by DirectoryRank to the rankings delivered by PageRank, and we give obtained results. Moreover, to delve into the behavior of DirectoryRank we carried out a user study, presented in Section 4.3. 4.1 Experimental Dataset In selecting our experimental data, we picked pages that are categorized in those topics in Google Directory, which are also present in our hierarchy. Recall that Google Directory is a replica of the Dmoz Directory, from which we borrowed our hierarchy's 13 top-level topics. Out of all the sub-topics organized in those 13 top-level topics in Google Directory, 156 were represented in our hierarchy. Having determined the topics, whose set of ranked pages would be compared, we downloaded a total number of 318,296 pages, categorized in one of the 156 selected topics, which in turn are organized into the 13 top-level topics. Table 2 shows the distribution of the experimental pages in the top level topics in Google Directory. Table 2. Statistics on the experimental data Category # of documents # of sub-topics Arts 28,342 18 Sports 20,662 26 Games 11,062 6 Home 6,262 7 Shopping 52,342 15 Business 60,982 7 Health 23,222 7 News 9,462 4 Society 28,662 14 Computers 35,382 13 Reference 13,712 10 Recreation 8,182 20 Science 20,022 9 Total 318,296 156 Since we were interested in comparing DirectoryRank with PageRank , in the context of ranking Web pages in Directory topics, we recorded for the downloaded Web pages their relative ranking order in Google Directory in each of the 156 selected topics. We then stored the downloaded pages in a secondary index, maintaining their relative PageRank rankings. To compute the DR values for every experimental page, we initially processed the downloaded pages in order to generate and score their lexical chains. For every page, we first computed its RScore to the topic in which it is assigned in Google Directory, and then we computed the semantic similarity ( s ) for every pair of pages listed in each topic. Lastly, using the above two scores (i.e. semantic similarity and topic relatedness ), we computed for every Web page its DirectoryRank ( DR) value and we sorted the pages listed within each of the topics, so that pages with higher DR scores in some topic are prioritized among the set of topic related pages. Using the above data, we evaluated the effectiveness of our DirectoryRank metric in ordering Web pages inside the Directory's topics. 4.2 Overlap of DirectoryRank and PageRank To investigate whether there is any similarity between the rankings induced by DirectoryRank and the rankings delivered by PageRank for our experimental pages in the 156 topics in Google Directory, we used the OSim measure, reported in the work of [9], which indicates the degree of overlap between the top n URLs of the two rankings. Formally, the overlap of two ranked lists A and B (each of size n) is given by: ( ) , / OSim DR PR A B n = I Using the above formula, we computed for each of the 156 topics the overlap between the pages ranked in the top n=10 positions for that topic by DR and PR respectively. Afterwards, we first com-20 puted the average similarity between the two induced rankings for each of the 156 selected topics, and then the average similarity between the two induced rankings for each of the 13 top-level topics. To compute the average similarity between DR and PR for a top level topic T, we summed the average similarity of all sub-topics in T and we divided by the number of sub-topics that T has. Table 3 gives the average similarity scores between DR and PR for each of the 13 top-level topics examined in our experiment. Table 3. Average similarity of rankings for the top level topics Category OSim Arts 0.038 Sports 0.019 Games 0.030 Home 0.057 Shopping 0.013 Business 0.028 Health 0.057 News 0.100 Society 0.043 Computers 0.046 Reference 0.020 Recreation 0.025 Science 0.044 Obtained results demonstrate that there is little average overlap between the top 10 results for the two rankings. Note that for some topics we compared the overlap between DR and PR for a larger set of pages (e.g. n=20 and n=30) and we found that the OSim score of the two rankings increases, albeit slightly, as the size of n grows. For example in the topic Sports, the OSim between DR and PR for n=10 is 0.019, whereas for n=20 the OSim score is 0.023 and for n=30, OSim is 0.028. Our results show that even the pairs with the greatest similarity among all pairs examined (e.g. the rankings delivered for the News topic), according to the OSim measure, have little in common. Despite the usefulness of the OSim measure for making rough estimations about the ability of the two ranking schemes in identifying the same top pages with respect to some topics, it cannot directly capture which ranking is more useful for ordering pages in Directory topics. This is because OSim does not indicate the degree to which the relative orderings of the top n pages of two rankings are in agreement. Having established that PageRank and DirectoryRank order Web pages substantially differ-ently , we proceed to investigate which of these rankings is better for ordering Web pages in Directory topics. To that end, we carried out a user study, reported next. 4.3 DirectoryRank Performance To determine which of the two ranking measures, namely DR and PR, is perceived as more useful by Web users for organizing pages in Web Directories, we carried out a user study. From our sample data, we picked the top 10 pages listed in 7 randomly selected topics (out of the 156 topics examined) and we recruited 15 postgraduate volunteers from our school. Table 4 lists the 7 topics selected. For each topic, the volunteer was shown 2 result rankings; one consisted of the top 10 pages for the topic ranked with DR, and the other consisted of the top 10 pages for the topic ranked with PR. For each topic, the volunteer was asked to read the pages in both lists and indicate which of the two rankings, in their opinion, is more "useful" overall for communicating information about the topic. Volunteers were not told anything about how either of the rankings was generated. In order to avoid misinterpretations while analyzing the user's selection preferences, we asked from the users to indicate their descriptive selections directly. More specifically, we presented to our participants the following choices and we asked them to indicate for which of the following reasons they selected one ranking over the other for each of the topics examined. Table 4. Experimental Topics Experimental Topics T 1 Crime T 2 Photography T 3 Water Sports T 4 Radiology T 5 Mechanics T 6 Econometrics T 7 Collecting Reason R1. "I prefer this ranking because I obtained significant information about the topic from most of the pages". In our analysis , we interpret the ranking preferences established on this reason as "topic-informative" rankings. Reason R2: "I prefer this ranking because I have seen most of the pages before and I liked them". We interpret the ranking preferences established on this reason as "popular" rankings. We then compared the participants' descriptive selections for every topic with the final DR/ PR choices. This way we ensure that users' preferences would be accurately evaluated even if two volunteers had exactly the same descriptive selection, but they ended up casting that selection into different DR, PR rankings. As a final note, we also asked our volunteers to indicate their familiarity with the experimental topics, by characterizing as "familiar" or "unfamiliar" each of the topics examined. In our evaluation, we considered that one ranking was better than the other if at least 50% of the users selected it as more "useful". Table 5 shows the rankings selected by our subjects as more useful for each of the 7 examined topics. Every row corresponds to a separate user. The columns marked as T i show what the preference of the user was for the particular topic. Under the T i columns the keyword DR means that the user considered DirectoryRank as more useful for that topic, while PR means that the user deemed PageRank as more useful. The column marked as R on the right of a T i column indicates the reason for which the user voted over the specified ranking. Table 6 summarizes the rankings preferred by the majority of the users for each of the topics. Table 5. Rankings selected as more useful for each topic User T 1 R T 2 R T 3 R T 4 R T 5 R T 6 R T 7 R #1 DR 1 DR 1 DR 1 DR 1 PR 2 DR 1 PR 2 #2 PR 2 DR 2 PR 2 DR 1 DR 1 DR 1 PR 2 #3 DR 1 DR 1 DR 1 DR 1 DR 2 DR 1 PR 2 #4 PR 1 PR 1 PR 2 DR 2 PR 2 PR 2 PR 1 #5 DR 1 PR 1 PR 2 PR 2 PR 2 DR 2 DR 1 #6 PR 2 DR 1 PR 2 DR 1 DR 1 DR 2 DR 1 #7 DR 2 PR 2 PR 1 DR 1 PR 2 DR 1 DR 1 #8 DR 1 DR 2 DR 1 DR 1 PR 1 DR 1 PR 2 #9 PR 2 DR 1 PR 2 PR 2 PR 2 DR 1 DR 2 #10 DR 1 DR 1 DR 1 DR 1 DR 1 DR 2 DR 2 #11 DR 1 DR 1 DR 1 DR 2 PR 2 PR 2 PR 2 #12 DR 1 DR 1 DR 1 PR 1 PR 2 DR 1 DR 1 #13 DR 2 PR 2 PR 1 DR 1 PR 2 DR 1 DR 1 #14 PR 2 DR 1 PR 2 DR 1 DR 1 DR 1 PR 2 #15 DR 1 DR 2 DR 1 DR 1 PR 1 DR 1 DR 1 Our survey results demonstrate that the majority of the users perceived in overall DirectoryRank as more useful in comparison to PageRank for ordering Web pages in the Directory's topics. This is attested by the fact that for most of the topics examined (for 5 out of the 7 topics), the majority of our subjects preferred DR over PR. A closer look at the obtained results indicates that the reason on which 21 our participants' based most of their DR selections, is Reason 1, which implies that the rankings delivered by DR are perceived as more topic-informative. Conversely, most of the users who liked better the rankings induced by PR, established their selection on Reason 2. This suggests that the usefulness of PR is not implied mainly by how informative a page is about a topic, but rather that it is substantially influenced by the page's popularity. Table 6. Rankings preferred by the majority of users Topic Preferred by majority T 1 Crime DirectoryRank T 2 Photography DirectoryRank T 3 Water Sports PageRank T 4 Radiology DirectoryRank T 5 Mechanics PageRank T 6 Econometrics DirectoryRank T 7 Collecting DirectoryRank Moreover, although not reported here due to space limit, our survey results show that our participants' answers were not generally influenced by their familiarity or not with the underlying topics. This implies that our survey does not entail "topic-bias", since both rankings compared are applied to pages listed in the same topic. RELATED WORK There have been a number of studies trying to identify the best ranking order of the Web pages that are deemed to be relevant to a given query/topic. The most successful of these studies [8, 11] suggest the exploitation of the pages' links connectivity on the Web graph for measuring the pages' importance and rank them accordingly . The most widely known ranking metric that explores the pages' links structure for measuring their importance on the Web is PageRank. Currently, PageRank and its variations are used by most major Web Search Engines to rank the results that they return to Web users in response to their search requests. Despite PageRank's usefulness for ordering pages in the context of Search Engines, it is designed to measure the global importance of the pages on the Web, independent of any particular topics. However, the overall importance of the pages may be not a sufficient measure for ordering the pages inside Directories' topics, essentially because pages that are important in some topics may not be important in others, regardless of the number and structure of the links that may appear in those pages. To alleviate some of the inherent limitations of PageRank , a number of researchers designed new ranking metrics, which mainly rely on modifications of PageRank and are tailored for specific tasks. For example, [9] studies personalization of the PageRank metric by giving different weights to pages, [14] examine the local and the inter-site link structure in order to compute a global PageRank for Web pages, [7] introduce Hilltop, an algorithm which generates query-specific authority scores for improving rankings for popular queries. While most of these works mainly focus on improving the rankings delivered to Web users by measuring the Web pages' overall importance, in this paper we are more concerned about the topic importance of Web pages by measuring the pages' informativeness with respect to particular topics. In this scope, we perceive our work to be complementary to previous studies on personalized rankings [9]. Moreover, there exists prior work that explores the lexical chaining technique as a means for representing documents' contents [6, 12]. Recently, we employed the lexical chaining technique for the automatic classification of Web documents in topic hierarchies [13]. Our findings indicated the potential of lexical chains in successfully capturing the thematic content of Web pages. This motivated our work to use the lexical chains generated for a number of Web pages as a means for ordering pages within Directory topics. In the future we plan to investigate how our approach could benefit from other linguistic approaches , besides lexical chains. CONCLUDING REMARKS In this paper, we introduced DirectoryRank, a practical metric for determining how informative Web pages are for particular topics and ranking them accordingly. To evaluate the potential of DirectoryRank in ordering Web pages inside Directory topics, we conducted an experiment where we applied our DirectoryRank metric to order a set of pages listed within 156 topics in Google Directory and we compared the rankings induced by DirectoryRank to the rankings that PageRank delivers in Google Directory for the same set of pages and topics. In our study, we relied on the judgments made by 15 users to determine which ranking is perceived as more useful for Web Directories' users. Obtained results indicate that in overall users preferred DirectoryRank over PageRank for ordering Web pages inside the Directory's topics. Although it would probably require additional studies in order to evaluate the applicability of our method to Web Directories other than Google and assess DirectoryRank's usefulness to a larger user and categories base, we believe that our work can serve as the first step towards a topic-informative ranking metric within directories. REFERENCES [1] Google Directory http://dir.google.com/. [2] MultiWordNet Domains http://wndomains.itc.it/. [3] Open Directory Project http://dmoz.com/. [4] Sumo Ontology http://ontology.teknowledge.com/. [5] WordNet 2.0 http://www.cogsci.princeton.edu/~wn/. [6] Barzilay R Lexical chains for text summarization. Master's Thesis, Ben-Gurion University, 1997. [7] Bharat K and Mihaila G. Hilltop: a search engine based on expert documents: http://www.cs.toronto.edu/~georgem/ hilltop /. [8] Kleinberg J. Authoritative sources in a hyperlinked environment . In Journal of the ACM, 46(5), 1999, 604-632. [9] Haveliwala T. Topic sensitive PageRank. In Proceedings of the 11 th WWW Conference, 2002, 517-526. [10] Ntoulas A., Cho J. and Olston Ch. What's new on the web? The evolution of the web from a search engine perspective. In Proceedings of the 13 th WWW Conference, 2004, 1-12. [11] Page L., Brin S., Motwani R. and Winograd T. The PageRank citation ranking: Bringing order to the web. Available at http://dbpubs.stanford.edu:8090/pub/1999-66. [12] Song Y.I., Han K.S. and Rim H.C. A term weighting method based on lexical chain for automatic summarization. In Proceedings of the 5 th CICLing Conference, 2004, 636-639. [13] Stamou S., Krikos V., Kokosis P., Ntoulas A. and Christodoulakis D. Web directory construction using lexical chains. In Proceedings of the 10 th NLDB Conference 2005, 138-149. [14] Wang Y. and DeWitt D. Computing PageRank in a distributed internet search system. In Proc. of the 30 th VLDB Conf., 2004. 22

topic hierachy;semantic similarity;ranking metric;scoring;web directory;ranking;lexical chains;DirectoryRank;topic importance;PageRank;information retrieval;Web Directory;semantic similarities

71

Discovering and Ranking Web Services with BASIL: A Personalized Approach with Biased Focus

In this paper we present a personalized web service discovery and ranking technique for discovering and ranking relevant data-intensive web services. Our first prototype called BASIL supports a personalized view of data-intensive web services through source-biased focus. BASIL provides service discovery and ranking through source-biased probing and source-biased relevance metrics. Concretely, the BASIL approach has three unique features: (1) It is able to determine in very few interactions whether a target service is relevant to the given source service by probing the target with very precise probes; (2) It can evaluate and rank the relevant services discovered based on a set of source-biased relevance metrics; and (3) It can identify interesting types of relationships for each source service with respect to other discovered services, which can be used as value-added metadata for each service. We also introduce a performance optimization technique called source-biased probing with focal terms to further improve the effectiveness of the basic source-biased service discovery algorithm. The paper concludes with a set of initial experiments showing the effectiveness of the BASIL system.

INTRODUCTION Most web services today are web-enabled applications that can be accessed and invoked using a messaging system, typically relying on standards such as XML, WSDL, and SOAP [29]. Many companies have latched onto the web services mantra, including major software developers, business exchanges, eCom-merce sites, and search engines [15, 9, 2, 1, 7]. A large and growing portion of the web services today can be categorized as data-intensive web services. This research is partially supported by NSF CNS CCR, NSF ITR, DoE SciDAC, DARPA, CERCS Research Grant, IBM Faculty Award, IBM SUR grant, HP Equipment Grant, and LLNL LDRD. 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. ICSOC'04, November 1519, 2004, New York, New York, USA. Copyright 2004 ACM 1-58113-871-7/04/0011 ... $ 5.00. Data-intensive web services provide access to huge and growing data stores and support tools for searching, manipulating, and analyzing those data stores. For example, both Amazon [1] and Google [7] now provide XML- and SOAP-based web service interfaces to their underlying data repositories with support for advanced search operators over, collectively, billions of items. In the life sciences domain, many bioinformatics services are transitioning from human-in-the-loop web interfaces to the web services model [9], providing direct access to unprecedented amounts of raw data and specialized research tools to provide high-level analysis and search over these data services. With the increasing visibility of web services and the Service-Oriented Computing paradigm [18], there is a growing need for efficient mechanisms for discovering and ranking services. Effective mechanisms for web service discovery and ranking are critical for organizations to take advantage of the tremendous opportunities offered by web services, to engage in business collaborations and service compositions, to identify potential service partners, and to understand service competitors and increase the competitive edge of their service offerings. Current web service discovery techniques can be classified into two types: categorization-based discovery and personalized relevance-based discovery. The former discovers web services by clustering and categorizing a collection of web services into different groups based on certain common properties of the services. Most of the existing UDDI [28] registry-based service discovery methods are of this type. They typically discover relevant services by querying metadata maintained in the common registries (like the ones offered by Microsoft [16] and IBM [10]). A typical question is "Which bioinformatics web services offer BLAST capability" or "Which commercial services offer on-line auctions". The second type of discovery mechanisms uses personalized relevance reasoning and support questions such as "Which services offer the same type of content as NCBI", and "Find the top-ten web services that offer more coverage than the BLAST services at NCBI". These two types of service discovery techniques offer different focus and complementary capabilities. Consider the following examples: A bioinformatics researcher may be interested in finding all services similar to NCBI's BLAST service for searching DNA and protein sequence libraries [17]. Current service registries may provide pointers to other BLAST services, but they do not describe how these other sites relate specifically to NCBI's BLAST service. Which services provide the most similar coverage with respect to NCBI (e.g. of similar proteins or organisms )? Which services are complementary in their coverage (e.g. of other sequence libraries)? How best should the BLAST services be ranked relative to the NCBI service? A health science researcher familiar with the PubMed med-153 ical literature service may be interested in discovering other related medical digital library services. Given his prior knowledge with PubMed, he may want to ask certain personalized (source-biased) discovery requests, which are not supported by the conventional service-registry-based discovery model. Examples include: Find and rank all PubMed-related medical literature sites. Which services have more general coverage than PubMed? Which medical literature services are more specialized than PubMed? These examples highlight two fundamental differences between the categorization-based and the personalization-based discovery model: (1) Categorization of web services based on general descriptions maintained at the service registries are insufficient and inadequate when a user is interested in discovery based on a particular service (or a set of services) with which she has prior experience or knowledge (e.g. NCBI BLAST or PubMed); and (2) There is a need for service ranking metrics that capture the relative scope and coverage of the data offered with respect to a previously known service. Although these two types of discovery mechanisms are complementary, most existing proposals on web service discovery fall into the first type. Surprisingly, there are, to our knowledge, no effective means to provide such personalized and biased discovery and ranking support without relying on significant human intervention. In this paper we present algorithms for discovering and ranking relevant data-intensive web services. Our first prototype called BASIL 1 -- supports a personalized view of web services through source-biased probing and source-biased relevance detection and ranking metrics. Concretely, our approach is capable of discovering and ranking web services by focusing on the nature and degree of the data relevance of the source service to others. Given a service like NCBI's BLAST called the source - the BASIL source-biased probing technique leverages the summary information of the source to generate a series of biased probes to other services called the targets. This source-biased probing allows us to determine whether a target service is relevant to the source by probing the target with very few focused probes. We introduce the biased focus metric to discover and rank highly relevant data services and measure relevance between services. Our initial results on both simulation and web experiments show that the BASIL system supports efficient discovery and ranking of data-intensive web services. MODEL AND PROBLEM STATEMENT We consider a universe of discourse W consisting of D data-intensive web services: W = {S 1 , S 2 , . . . , S D } where each service produces one or more XML documents in response to a particular service request. Hence, we describe each web service S i as a set of M i documents: S i = {doc 1 , doc 2 , , doc M i }. For example, the documents corresponding to the NCBI BLAST service would consist of genetic sequences and documentation generated in response to a service requests. Similarly, the documents corresponding to PubMed would consist of the set of medical journal articles in the PubMed data repository. There are N terms (t 1 , t 2 , ..., t N ) in the universe of discourse W including both the tags and content of the XML documents where common stopwords (like `a', `the', and so on) have been eliminated. Optionally, the set of N terms may be further refined by stemming [19] to remove prefixes and suffixes. Adopting a vector-space model [22, 23] of the service data repository, we describe each service S i as a vector consisting of the terms in the service along with a corresponding weight: Summary(S i ) = {(t 1 , w i 1 ), (t 2 , w i 2 ), , (t N , w iN ) } 1 BASIL: BiAsed Service dIscovery aLgorithm A term that does not occur in any documents served by a service S i will have weight 0. Typically, for any particular service S i , only a fraction of the N terms will have non-zero weight. We refer to the number of non-zero weighted terms in S i as N i . We call the vector Summary(S i ) a service summary for the data-intensive web service S i . A service summary is a single aggregate vector that summarizes the overall distribution of terms in the set of documents produced by the service. In this first prototype of BASIL, we rely on a bag-of-words model that is indifferent to the structure inherent in the XML documents. As we demonstrate in the experiments section, this bag-of-words approach is quite powerful without the added burden of structural comparisons. We anticipate augmenting future versions of BASIL to incorporate structural components (to support schema matching, leverage existing ontologies, etc.). To find Summary(S i ), we must first represent each document doc j (1 j M) as a vector of terms and the frequency of each term in the document: doc j = {(t 1 , f req j 1 ), (t 2 , f req j 2 ), , (t N , f req jN ) } where f req jk is the frequency of occurrence of term t k in document j. The initial weight for each term may be based on the raw frequency of the term in the document and it can be refined using alternative occurrence-based metrics like the normalized frequency of the term and the term-frequency inverse document-frequency (TFIDF ) weight. TFIDF weights the terms in each document vector based on the characteristics of all documents in the set of documents. Given a particular encoding for each document, we may generate the overall service summary in a number of ways. Initially , the weight for each term in the service summary may be based on the overall frequency of the term across all the documents in the service (called the service frequency, or servFreq ): w ik = servF req ik = M j =1 f req jk . Alternatively, we can also define the weight for each term based on the number of documents in which each term occurs (called the document count frequency, or docCount). Once we have chosen our service model, to effectively compare two data-intensive web services and determine the relevance of one service to another, we need two technical components : (1) a technique for generating a service summary; and (2) a metric for measuring the relevance between the two. 2.1 Estimating Service Summaries Ideally, we would have access to the complete set of documents belonging to a data-intensive web service. We call a service summary for S i built on these documents an actual service summary or ASummary(S i ). However, the enormous size of the underlying repositories for many data-intensive web services coupled with the non-trivial costs of collecting documents (through repeated service requests and individual document transfer) make it unreasonable to generate an actual service summary for every service available. As a result, previous researchers in the context of distributed databases have introduced several probing techniques for generating representative summaries based on small samples of a document-based collections [3, 4]. We call such a representative summary an estimated service summary, or ESummary(S i ): ESummary(S i ) = {(t 1 , w i 1 ), (t 2 , w i 2 ), , (t N , w iN ) } The number of occurring terms (i.e. those terms that have non-zero weight) in the estimated summary is denoted by N i . Typically, N i will be much less than the number of non-zero weighted terms N i in the actual service summary since only 154 a fraction of the total documents in a service will be examined . The goal of a prober is typically to find ESummary(S i ) such that the relative distribution of terms closely matches the distribution of terms in ASummary(S i ), even though only a fraction of the total service documents will be examined. Current probing techniques for estimating service summaries aim at estimating the overall summary of the data served by a web service. We classify them into two categories: random sampling and query-based sampling. Random Sampling - N o Bias If we had unfettered access to a data-intensive web service, we could randomly select terms from the service to generate the estimated service summary ESummary(S i ). Barring that, we could randomly select documents with which to base the estimated service summary. We will call such a random selection mechanism an unbiased prober since all terms (or documents ) are equally likely to be selected. In practice, an unbiased prober is unrealistic since most services only provide a request-response mechanism for extracting documents. Query-based Sampling - Query Bias As a good approximation to unbiased probing, Callan et al. [3, 4] have introduced a query-based sampling technique for generating accurate estimates of document-based collections by examining only a fraction of the total documents. The Callan technique has been shown to provide accurate estimates using very few documents (e.g. several hundred). Adapting the Callan technique to the web services context requires repeat-edly requesting documents from a service using a limited set of service requests. Since the documents extracted are not chosen randomly, but are biased by the service request mechanism through the ranking of returned documents and by providing incomplete access to the entire data service repository, we say that the Callan technique displays query bias. There are several ways to define the limited set of queries, including random selection from a general dictionary and random selection augmented by terms drawn from the extracted documents from the service. In the rest of the paper, when we refer to an estimated service summary ESummary(S i ), we mean one that has been produced by a query-biased prober. 2.2 Comparing Service Summaries In order to determine the relevance of one service S i to another service S j and to assess the nature of their relationship, we require an appropriate relevance metric. There are a number of possible relevance metrics to compare two service summaries . A fairly simple and straightforward approach is based on a count of the number of common terms in the two services S i and S j : rel(S i , S j ) = |ESummary(S i ) ESummary(S j ) | max( |ESummary(S j ) |, |ESummary(S i ) |) Two services with exactly the same terms represented in their estimated summaries will have rel(S i , S j ) = 1, indicating the highest possible degree of relevance. Conversely, two services with no terms in common will have rel(S i , S j ) = 0, indicating the lowest possible degree of relevance. We now use an example to illustrate why the existing service summary estimation techniques are inadequate for effectively discovering relevant services, especially in terms of the data coverage of one (target) in the context of the other (source). Example: We collected fifty documents from the Google web service, the PubMed web service, and ESPN's search site, respectively , using a query-based sampling technique for service summary estimation. Using the service summaries constructed, we find that rel(Google, P ubM ed) = 0.05 and rel(ESP N , P ubM ed) = 0.06. In both cases the service summaries share very few terms in common and hence both Google and ESPN appear to be irrelevant with respect to PubMed, even though Google provides considerable health-related content. Based on these figures, we could incorrectly conclude that: (1) Google is irrelevant to PubMed; and (2) Relatively speaking, ESPN is more relevant to PubMed than Google. This example underlines two critical problems with current techniques for probing and comparing service summaries: First, current service summary estimation techniques are concerned with generating overall (or global ) summaries of the underlying data repositories. The goal is to generate essentially an unbiased estimate of the actual service summary. Second, the current relevance comparison metric fails to serve as a valuable ranking metric or indicator of interesting relationships between target services in terms of the data coverage of a target service with respect to the source. THE BASIL SYSTEM Bearing these issues in mind, we now introduce BASIL an efficient web service discovery and ranking prototype that relies on a biased perspective of services rather than on a single global perspective. BASIL relies on three fundamental steps: (1) source-biased probing for web service discovery; (2) evaluation and ranking of discovered services with the biased focus metric; and (3) leveraging the biased perspective of service sources and targets to discover interesting relationships. 3.1 Source-Biased Probing Given a data-intensive web service the source the source-biased probing technique leverages the summary information of the source to generate a series of biased probes for analyzing another service the target. This source-biased probing allows us to determine in very few interactions whether a target service is relevant to the source by probing the target with focused probes. To help differentiate the source-biased approach from others discussed in Section 2, in this section we use to denote the source service and to denote the target service instead of S i and S j . Given two services and , the output of the source-biased probing is a subjective service summary for that is biased towards . We define the source-biased summary of the target service, denoted by ESummary ( ), as follows: ESummary ( ) = {(t 1 , w 1 ), (t 2 , w 2 ), , (t N , w N ) } N is the total number of terms used in analyzing the set of data-intensive web services. w i (1 i N) is the weight of term t i , defined using one of the weight function introduced in Section 2. To distinguish the term weight w j from the corresponding term weight in the biased target summary, we denote the bias by w j . It is important to note that typically the inequality w j = w j does hold. Concretely, the source-biased probing algorithm generates a source-biased summary for a target as follows: It uses the estimated service summary of the source , denoted by ESummary(), as a dictionary of candidate probe terms and sends a series of query requests parameterized by probe terms, selected from ESummary (), to the target service ; for each probe term, it retrieves the top m matched documents from , generates summary terms and updates ESummary ( ). This process repeats until a stopping condition is met. Figure 1 illustrates the source-biased probing process. Note that in this first prototype of BASIL the service requests are constrained to keyword-based probes. Note that the source-biased approach can also be applied to UDDI-directory-based discovery by restricting 155 SourceBiasedProbing(Source , Target ) For target service , initialize ESummary ( ) = . repeat Invoke the probe term selection algorithm to select a one-term query probe q from the source of bias ESummary(). Send the query q to the target service . Retrieve the top-m documents from . Update ESummary ( ) with the terms and frequencies from the top-m documents. until Stop probing condition is met. return ESummary ( ) Figure 1: Source-Biased Probing Algorithm the source summary to be generated from the meta-data description maintained at the registries rather than performing the source-biased probing directly. However, the quality of the discovery results will be lower due to the lack of richness in the metadata maintained at the service registries for many services. Now we use a simple example to illustrate the power of source-biased probing. For presentation brevity, we are considering a simplistic world of only very few terms per service summary. In reality, each service summary would consist of orders of magnitude more terms: Example: Suppose that our goal is to understand the relevance of Google to PubMed. Suppose, ESummary(P ubMed) = {arthritis, bacteria, cancer} (where for simplicity we have dropped the term weights from the summary). Again for simplicity suppose that Google provides access to only three types of information: health, animals, and cars: ASummary(Google) = {arthritis, bacteria, cancer, dog, elephant, frog, garage, helmet, indycar }. An unbiased prober could result in ESummary (Google) = {arthritis, frog, helmet}, whereas a source-biased prober could result in ESummary P ubM ed (Google) = {arthritis, bacteria, cancer }. This simple example illustrates the essence of the source-biased probing and how it accentuates the commonality between the two services. The performance and effectiveness of the source-biased probing algorithm depends upon a number of factors, including the selection criterion used for choosing source-specific candidate probe terms, and the type of stop condition used to terminate the probing process. Mechanisms to Select Probe Terms There are several possible ways to select the probes based on the statistics stored with each service summary, including uniform random selection and selection based on top-weighted terms. In general, the selection criterion will recommend a query term drawn from the set N of all non-zero weighted terms in the unbiased source summary ESummary(). Uniform Random Selection: In this simplest of selection techniques , each term that occurs in ESummary() has an equal probability of being selected, i.e. P rob(selecting term j) = 1 N . Weight-Based Selection: Rather than randomly selecting query terms, we could instead rely on a ranking of the terms by one of the statistics that are recorded with each service summary. For example, all terms in ESummary() could be ranked according to the weight of each term. Terms would then be selected in descending order of weight. Depending on the type of weight cataloged (e.g. servF req, docCount, etc.), several flavors of weight-based selection may be considered. Different Types of Stop Probing Conditions The stop probing condition is the second critical component in the source-biased probing algorithm. We consider four different types of conditions that might be used in practice: Number of Queries: After some fixed number of query probes (M axP robes), end the probing. This condition is agnostic to the number of documents that are examined for each service. Documents Returned: In contrast to the first technique, the second condition considers not the number of queries, but the total number of documents (M axDocs) returned by the service . Since some queries may return no documents, this stopping condition will require more query probes than the first alternative when M axP robes = M axDocs. Document Thresholding: Rather than treating each document the same, this third alternative applies a threshold value to each document to determine if it should be counted toward M axDocs. For each document, we may calculate the relevance of the document to the source of bias ESummary(). If the document relevance is greater than some threshold value, then the document is counted. Otherwise, the document is discarded. Steady-State: Rather than relying on a count of queries or documents , this final stopping condition alternative instead relies on the estimated summary reaching a steady-state. After each probe, we calculate the difference between the new value of ESummary ( ) and the old value. If the difference (which may be calculated in a number of ways) is less than some small value , then we consider the summary stable and stop the probing. Due to the space limitation, we refer readers to our technical report [5] for detailed experiments on the impact of these two parameters. 3.2 Evaluating and Ranking Services Given a source and a target service, once we generate the source-biased summary for the target service, we need an efficient mechanism to evaluate the source-biased relevance of a target service with respect to the source. Once a set of target services have been evaluated with the source-biased relevance metric, we can then rank the target services with respect to the source of bias. We begin by discussing the necessary components of the source-biased metric. Let denote a source service modeled by an estimated summary and denote a target service with a -biased summary, and let f ocus ( ) denote the source-biased focus measure. We define f ocus ( ) to be a measure of the topical focus of the target service with respect to the source of bias . The focus metric ranges from 0 to 1, with lower values indicating less focus and higher values indicating more focus. In general, f ocus is not a symmetric relation. We may describe any two data-intensive web services and with the focus in terms of by f ocus ( ) or in terms of by f ocus (). We propose to use the well-known cosine similarity (or normalized inner product) to approximate the source-biased focus measure. We define the cosine-based focus as follows: Cosine f ocus ( ) = N k =1 w k w k N k =1 (w k ) 2 N k =1 (w k ) 2 where w k is the weight for term k in ESummary() and w k is the -biased weight for term k in ESummary ( ). The cosine ranges from 0 to 1, with higher scores indicating a higher degree of similarity. In contrast, the cosine between orthogonal vectors is 0, indicating that they are completely dissimilar. The cosine measures the angle between two vectors, regardless of the length of each vector. Intuitively, the cosine-based biased focus is appealing since it reasonably captures the relevance between two data-intensive web services. Ranking Relevant Services Given the biased focus measure, we may probe a group of tar-156 get services to identify the most relevant services to the source of bias. For a single source of bias S 1 from our universe of discourse W, we may evaluate multiple target services S 2 , S 3 , ..., S d . For each target service, we may evaluate the appropriate focus measure for each source-target pair (i.e. f ocus S 1 (S 2 ), f ocus S 1 (S 3 ), etc.). We may then rank the target services in descending order in terms of their source-biased focus with respect to S 1 . As we will show in our experiments section, source-biased probing results in the identification of relevant services that existing approaches may overlook. We also show that source-biased probing can generate source-biased summaries of good quality using far fewer documents than existing approaches, placing significantly less burden on the target services. 3.3 Identifying Interesting Relationships The critical third component of the BASIL system consists of the techniques for exploiting and understanding interesting relationships between services using a source-biased lens. By analyzing the nature of the relationships between data-intensive web services, we will provide support for understanding the relative scope and coverage of one service with respect to another. The source-biased probing framework and biased focus measure provide the flexible building blocks for automated identification of interesting relationships between services, especially since the framework promotes an asymmetric source-biased view for any two services. Our relationship discovery module creates a flexible organization of services, where each service is annotated with a list of relationship sets. The two typical relationship types we have identified are similarity-based and hierarchical-based. Similarity-Based Relationships Given the universe of discourse W = {S 1 , S 2 , . . . , S D }, we identify three similarity-based relationship sets for a particular service S i . These relationship sets are defined in terms of threshold values high and low , where 0 low high &lt; 1. - equivalent: The first relationship says that if both focus S i (S j ) &gt; high and f ocus S j (S i ) &gt; high hold, then we may conclude that S i is sufficiently focused on S j and S j is sufficiently focused on S i . Hence, the two services are approximately the same in terms of their data coverage. We call this approximate equality -equivalence. It indicates that the equivalence is not absolute but is a function of the parameter high . Formally , -equivalent(S i ) = {S j W | focus S i (S j ) &gt; high f ocus S j (S i ) &gt; high }. - complement: If both focus S i (S j ) &lt; low and f ocus S j (S i ) &lt; low hold, then we can conclude that S i and S j are sufficiently concerned with different topics since neither one is very focused on the other. We annotate this approximate complementary nature with the prefix. Formally, -complement(S i ) = {S j W | focus S i (S j ) &lt; low focus S j (S i ) &lt; low }. - overlap: When two services S i and S j are neither equivalent nor -complementary, we say that the two services -overlap. Formally, -overlap(S i ) = {S j W | S j / complement (S i ) S j / -equivalent(S i ) }. Hierarchical Relationships In addition to similarity-based relationship sets, we also define hierarchical relationship sets by measuring the relative coverage of target services in W with respect to a particular text service S i (source). These hierarchical relationship sets are defined in terms of a parameter dif f , where 0 dif f 1. - superset: If focus S i (S j ) - focus S j (S i ) &gt; dif f , then a relatively significant portion of S i is contained in S j , indicating that S j has a -superset relationship with S j . We use the prefix to indicate that S j is not a strict superset of S i , but rather that the relationship is parameterized by dif f . Formally, superset (S i ) = {S j W | focus S i (S j ) - focus S j (S i ) &gt; dif f }. - subset: Conversely, If focus S j (S i ) - focus S i (S j ) &gt; dif f , then a relatively significant portion of S j is contained in S i , indicating that S j has a -subset relationship with S i . Similarly , S j is not a strict subset of S i , but rather the relationship is parameterized by dif f . Formally, -subset(S i ) = {S j W | focus S j (S i ) - focus S i (S j ) &gt; dif f }. We note that the determination of the appropriate -values is critical for the correct assignation of services to each relationship set. In our experiments section, we illustrate how these relationship sets may be created, but, for now, we leave the optimization of -values as future work. Using Relationship Sets Both similarity based and hierarchy-based inter-service relationships can be generated automati-cally , and used as metadata annotation to each of the services. These source-biased relevance data provide a flexible foundation for relationship analysis among services. For any service S i , we need only consult the appropriate relationship set. The three similarity-based relationship sets provide the basis for answering queries of the form: "What other services are most like X? Somewhat like X? Or complementary to X?". The two hierarchical-based sets provide the basis for answering queries of the form: "What other services are more general than X? Or more specialized than X?". In addition, these relationship sets are useful for routing service requests to the appropriate services. For example, a user interested in BLAST data may choose to use both NCBI's BLAST service and all of the services that have a -equivalence relationship with NCBI BLAST. Alternatively, a user interested in maximizing coverage of multiple topically-distinct services , may choose to query both the source service she knows about and any members in the complementary set of the source service. The hierarchical relationship sets are particularly helpful in cases where a user may refine a service request to more specialized services, or alternatively, may choose to generalize the scope of the service request by considering services further up the hierarchy to get more matching answers. FOCAL TERM PROBING One of the critical parameters to the success of BASIL's source-biased probing is the choice of probe terms from the source of bias . We have discussed several selection techniques as well as different ways to define stop-probing conditions. In this section we introduce a refinement over these simple selection techniques whereby the source summary is segmented into k groups of co-occurring terms. The main idea is to it-eratively select one term from each of the k groups to probe the target. We call these terms the focal terms of the corresponding group. When used in conjunction with the general source-biased probing algorithm, we have an enhanced version called source-biased probing with focal terms. A unique advantage of using focal terms is that the biased summaries of target services can be generated in fewer queries with higher quality. 4.1 Focal Terms and Focal Term Groups Let denote a source service with its unbiased service summary ESummary . We denote the set of terms with non-zero weight in ESummary (i.e. the terms that actually occur in the service ) as T erms(), where T erms() consists of n terms t 1 , t 2 , ..., t n . A focal term group is a subset of terms in the set T erms() that co-occur in the documents of . We denote a focal term 157 Table 1: Example Focal Terms for PubMed 1 care, education, family, management, ... 2 brain, gene, protein, nucleotide, ... 3 clinical, noteworthy, taxonomy, ... 4 experimental, molecular, therapy, ... 5 aids, evidence, research, winter, ... group i as F T erms i . The main idea behind source-biased probing with focal terms is to partition the set T erms() into k disjoint term groups such that the terms within each term group co-occur in documents of more frequently than they do with terms from other term groups. Formally, we need an algorithm that can find a partition of T erms() into k focal term groups: T erms() = {F T erms 1 , . . . , F T erms i , . . . , F T erms k | k i =1 F T erms i = {t 1 , ..., t n } and F T erms i F T erms j = } In Table 1, we show an example of five focal term groups for a collection of 100 PubMed documents. Note that k is intended to be very small since the focal term groups are meant to be very coarse. Given k focal term groups, by selecting a focal term from each term group F T erms i as a probing query, we hope to retrieve documents that also contain many of the other words in that focal term group. For example, suppose we are using a frequency-based measure for query probe selection from PubMed. The top four query terms may be "brain", "gene", "protein", and "nucleotide". Suppose these four terms tend to co-occur with each other as indicated in Table 1. By sending the first query "brain" to a target service, we could reasonably expect to find the other three terms since our analysis of the source indicates that these four terms tend to co-occur. A naive source-biased prober would ignore this co-occurrence information and, instead, send the other three queries "gene", "protein", and "nucleotide", even though we might reasonably expect for those queries to generate documents similar to those generated by the first query "brain". In essence, we will have used four queries when a single query would have sufficed at adequately exploring the term space of the target. It is important to note that, unlike previous research in grouping terms for query-expansion [31, 21] or finding similar terms [24] our goal is not to find close semantic relationships between terms, but rather to find very coarse co-occurrence associations among terms to support a more efficient and effective biased service summary estimation. For example, though we may discover that "brain" and "protein" tend to co-occur, we do not claim that there is a close semantic relationship between the two terms. 4.2 Finding Focal Terms In this section, we discuss how we may adapt a popular clustering technique to the problem of focal term discovery. Recall that in Section 2, we view a service S i as a set of documents, each of which is described by a vector of terms and weights. We now invert our view of a service using the same set of information . We consider a service S i as a collection of terms, each of which is described by a vector of the documents in which the term occurs and a weight describing the occurrence frequency of the term in the corresponding document. Hence, we have: T erms(S i ) = {term 1 , term 2 , , term N }. For the N terms in the service, each term j (1 j N) is a vector of documents and weights: term j = {(doc 1 , w j 1 ), (doc 2 , w j 2 ), , (doc M , w jM ) } We can define a segmentation technique for finding focal term groups by clustering the set T erms(S i ) into k clusters. Given the term vectors and the similarity function, a number of clus-FocalTerms (Num Clusters k, Input Vectors D) Let D = {d 1 , ..., d n } denote the set of n term vectors Let M denote the total number of documents in D Let d j = &lt; (doc 1 , w j 1), . . . , (doc M , w jM ) &gt; denote a term vector of M elements, w jl is TFIDF weight of the doc l in term j (l = 1, . . . , M ) Let C = {C 1 , ..., C k } denote a clustering of D into k clusters. Let i denote the center of cluster C i foreach cluster C i Randomly pick a term vector, say d j from D Initialize a cluster center i = d j , where d j D repeat foreach input term vector d j D foreach cluster C i C i = 1, . . . , k compute i = sim(d j , mu i ) if h is the smallest among 1 , 2 , . . . , k mu h is the nearest cluster center to d j Assign d j to the cluster C h // refine cluster centers using centroids foreach cluster C i C foreach doc l in d j (l = 1, . . . , M )) cw ij 1 |C i | M l =1 w jl i &lt; (doc 1 , cw i 1 ), . . . , (doc M , cw iM ) &gt; until cluster centers no longer change return C Figure 2: Focal Term Clustering Algorithm tering algorithms can be applied to partition the set T erms(S i ) of N terms into k clusters. We choose Simple K-Means since it is conceptually simple and computationally efficient. The algorithm starts by generating k random cluster centers. Each term is assigned to the cluster with the most similar (or least distant) center. The similarity is computed based on the close-ness of the term and each of the cluster centers. Then the algorithm refines the k cluster centers based on the centroid of each cluster. Terms are then re-assigned to the cluster with the most similar center. The cycle of calculating centroids and assigning terms in T erms(S i ) to k clusters repeats until the cluster centroids stabilize. Let C denote a cluster in the form of a set of terms in the cluster. The centroid of cluster C is: centroid C = (doc 1 , 1 |C| jC w j 1 ) (doc 2 , 1 |C| jC w j 2 ) (doc M , 1 |C| jC w jM ) where w jl is the weight of term j in document l, and the formula 1 |C| l C w jl denotes the average weight of the document l in the cluster C. A sketch of the K-Means term clustering based on term-vector of a service is provided in Figure 2. The similarity function used in Figure 2 can be defined using a number of functions. In this paper, we use the cosine similarity function. Given a set of N terms and a set of M documents, where w ik denotes the weight for term k in document i (1 k N, 1 i M), the cosine function prescribes: sim(term i , term j ) = N k =1 w ik w jk N k =1 (w ik ) 2 N k =1 (w jk ) 2 In Section 5 we report the initial experiments on effectiveness of using focal terms to optimize the source-biased probing algorithm, showing that the source-biased algorithm with focal terms results in more efficient probing for varying numbers of focal-term groups. 158 4.3 Selecting Focal-Based Probes Once the k focal term groups have been constructed for a source, the remaining problem is how to select the best terms for probing a target service. We propose a simple round-robin selection technique whereby a single term is selected from each focal term group in turn. Once a single term has been selected from each group, the cycle repeats by selecting a second term from each group, a third term, and so on. Given this basic strategy, we may use a number of techniques for determining the order by which to select terms from the k groups and for selecting probe terms from each focal term group. One way to determine the order of focal term groups is based upon the size of each group. We begin with the group with the most terms and end each cycle with the group that has the smallest number of terms. For each focal term group, we may decide which term to select for each cycle by using one of the selection criteria discussed in Section 3. EXPERIMENTS In this section, we describe four sets of experiments designed to evaluate the benefits and costs of BASIL. The first set intends to show the effectiveness of our source-biased probing algorithm and compare its performance with query-biased probing and unbiased probing. The second set evaluates the biased focus measure as an effective tool for ranking services. The third set shows the efficiency of the biased focus measure in identifying interesting inter-service relationships. The fourth set evaluates the efficacy of source-biased probing with focal terms by comparing the basic source-biased probing versus source-biased probing with varying number of groups of focal terms. Our experiments show that focal term probing can achieve about ten percent performance improvement over the basic algorithm for source-biased probing. Since there are no large data-intensive web service collections for experimentation, we rely on: (1) a large collection of newsgroups designed to emulate the diversity and scope of real-world data-intensive web services; and (2) a modest collection of real-world web sources. Since the services in the web collection change frequently and are beyond our control, and in an effort not to overload any one site, we relied on the newsgroup dataset for rigorous experimental validation. Newsgroup Collection: We collected articles from 1,000 randomly selected usenet newsgroups over the period June to July 2003. We eliminated overly small newsgroups containing fewer than 100 articles, heavily spammed newsgroups, and newsgroups with primarily binary data. After filtering out these groups, we were left with 590 single topic newsgroups, ranging in size from 100 to 16,000 articles. In an effort to match the heterogeneity and scope inherent in many real-world services , we constructed 135 additional groups of mixed topics by randomly selecting articles from anywhere from 4 to 80 single topic newsgroups, and 55 aggregate topic newsgroups by combining articles from related newsgroups (e.g. by selecting random documents from all the subgroups in comp.unix.* into a single aggregate group). In total, the newsgroup collection consists of over 2.5GB worth of articles in 780 groups. Web Collection: For the second collection, we randomly selected 50 sites from the ProFusion [20] directory of web sites that support queries, in addition to Google and PubMed. We queried each site with a randomized set of single-word probes drawn from the standard Unix dictionary, and collected a maximum of 50 documents per site. Probing Framework: We built a probing engine in Java 1.4 for use in all of our experiments. For each group in both 0.00 0.10 0.20 0.30 0.40 0 20 40 60 80 100 Documents Examined Average Source Similarity Source Bias Query Bias 2 Query Bias 1 No Bias Figure 3: Probing Efficiency for 100 Pairs datasets, we constructed the estimated service summary based on the overall term frequency of each term (servF req). We eliminated a set of common stopwords (e.g. "a", "the", and so on) as well as collection-specific stopwords (e.g. "wrote", "said", and so on for the newsgroup collection). 5.1 Effectiveness of Source-Biased Probing The goal of our first set of experiments is to compare source-biased probing with existing probing techniques and to evaluate the efficiency and quality of source-biased probing. The source-biased probing show significant gain in terms of the percentage of documents probed that are similar to the source. We first evaluate the efficiency of source-biased probing in terms of the number of documents required to be extracted from each target and the percentage of the documents extracted that are similar to the source. The higher percentage of documents similar (relevant) to the source, the more effective a probing algorithm is. We selected 100 random source-target pairs from the newsgroup collection. For each pair, we evaluated four probing techniques a source-biased prober (Source Bias) that selects probe terms from the source summary in decreasing order of servF req; a query-biased prober (Query Bias 1 ) that randomly selects probes from the standard Unix dictionary of English terms; a query-biased prober (Query Bias 2 ) that selects its initial probe from the Unix dictionary, but once the first document has been retrieved from the target, all subsequent probes are selected based on the estimated servF req of the target's service summary; and an unbiased prober (No Bias) that selects documents at random from each target. For each pair, we evaluated each of the four probing techniques for up to 100 total documents extracted from each target, collecting a maximum of 5 documents per probe query from each target. In Figure 3, we show the average percentage of documents similar (relevant) to the source (Cosine f ocus ( )) over all 100 source-target pairs as a function of the number of documents examined in each target. The percentage of the documents extracted that are similar to the source (biased f ocus measure) indicates the quality of document being extracted from each target. We see that the source-biased probing outperforms the No Bias prober and the Query Bias 1 prober by about 10% and outperforms the Query Bias 2 prober by about 15%. Clearly, the higher focus value means the higher success for a probing algorithm. Figure 4 shows another experiment where we also identified, in our set of 100 source-target pairs, all of those pairs that were a priori similar (e.g. mac.apps and mac.system) or dissimilar (e.g. textiles.sewing and perl.misc). We show the relative performance of the Source Bias, Query Bias 1, and No Bias 159 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0 20 40 60 80 100 Documents Examined Average Source Similarity Source Bias Query Bias No Bias Similar Dissimilar Figure 4: Probing Efficiency Breakdown 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0 20 40 60 80 100 Documents Examined % Docs Similar to Source Source Bias Query Bias 1 Query Bias 2 No Bias Figure 5: Average Document Quality for 100 Pairs probers against these similar and dissimilar pairs. The source-biased prober requires fewer documents to achieve the same relevance level as the other probers for all 100 source-target pairs and for the similar and dissimilar pairs. For example, for the similar source-target pairs in Figure 4, the source-biased prober identifies target documents with 0.8 focus after extracting fewer than 30 documents. In contrast, the other probers require between two and three times as many documents to achieve the same quality. The third experiment is shown in Figure 5. Here we want to show how quickly a source-biased prober can hone on the most source-relevant documents in a target by plotting the percentage of the documents extracted that are similar (relevant ) to the source for each of the four probers. As shown in Figure 5, the source-biased prober performs nearly two-times better than other probers: over 70% of the first 10 documents extracted from a target are source-relevant, whereas the other probers identify between 25% and 45% source-relevant documents . As more documents are examined for each target, the source-biased prober continues to maintain an advantage over the other probers. 5.2 Ranking Effectiveness with Biased Focus The second set of experiments intends to evaluate how well source-biased probing compares with the alternative techniques when it comes to evaluating and ranking collection of target services. We use PubMed as the source and examine all 50 web sites as targets. We computed the biased focus score using Cosine f ocus ( ) and then ranked all targets relative to PubMed using the biased focus measure. Since the web sites do not support random document selection, we are unable to evaluate an unbiased prober. So this experiment only compares the source-biased prober with query biased prober 1. Table 2 shows Table 2: Identifying Web Sources Relevant to PubMed Query Bias Source Bias 1. AMA 1. Open Directory (13) 2. WebMD 2. Google (27) 3. Linux Journal 3. About (11) 4. HealthAtoZ 4. WebMD (2) 5. DevGuru 5. AMA (1) 6. FamilyTree Magazine 6. HealthAtoZ (4) 7. Mayo Clinic 7. Monster (22) 8. Novell Support 8. Mayo Clinic (7) 9. Random House 9. Random House (9) 10. January Magazine 10. BBC News (12) 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 comp.sys.mac.system comp.unix.misc gnu.emacs.help rec.aviation.owning rec.games.chess.misc rec.org.sca rec.pets.cats.misc sci.physics.research soc.culture.hawaii talk.religion.misc Relevance Precision No Bias Query Bias Source Bias Figure 6: Precision for 10 Source Newsgroups the top-10 ranked sites relative to PubMed. In the Source Bias column we also list in parenthesis the rank of each site assigned by the Query Bias prober. The query-biased prober identifies several health-related sites in the web collection, but it mistakenly lists Linux Journal ahead of HealthAtoZ, as well as listing a web development site (DevGuru) and a genealogical magazine (FamilyTree) ahead of the health-related Mayo Clinic. Overall, only four of the top-ten sites could be considered topically relevant to PubMed. In contrast, the source-biased prober's top-eight sites are all relevant to PubMed. In addition to the health-related sites, the source-biased prober also identifies three general sites that offer access to medical literature (Open Directory, Google, and About) that are ranked significantly lower by the query-biased prober. Interestingly, the source-biased prober identifies a fair number of scientific and bioinformatics-related job descriptions in the Monster jobs site, resulting in its high relevance (similarity ) score to PubMed (high biased focus value). To validate the quality of source-biased service evaluation, we next randomly selected 10 sources from the newsgroup collection to evaluate against the entire set of 780 newsgroups. We compared the three probers Source Bias, Query Bias 1, and No Bias. For each of the 10 sources, we measured relevance (similarity) precision as the percentage of the top-10 ranked target services that are considered relevant to the source using Cosine f ocus ( ). Relevance judgments were determined by the consensus opinion of three volunteers. Figure 6 shows the precision for the three probers after extracting 40 documents per target service. Source Bias results in the highest precision in nine of ten cases, tying with the next best prober in only two cases. For the lone failure, Source Bias does succeed after extracting 80 documents, indicating that the mistake may be attributable to the error inherent in probing very few documents . In general, the average precision of the source-biased prober is nearly double that of the next best prober. In Figure 7 we show the average precision for the ten sources when increasingly more documents are extracted per target. The source-biased approach displays higher precision than both 160 0.00 0.10 0.20 0.30 0.40 0.50 20 40 60 80 100 Documents Examined Relevance Precision No Bias Query Bias Source Bias Figure 7: Average Relevance Precision the query-biased and unbiased probers in all cases considered, especially when based on very few documents. 5.3 Identifying Interesting Relationships The third set of experiments is designed to evaluate the effectiveness of using the source-biased framework to support the identification of interesting inter-service relationships that the alternative schemes do not. Unlike the query-biased and unbiased probers, the asymmetric nature of source-biased probing allows us to characterize the nature of the relationship beyond the single relevance ranking using biased focus measure. We first illustrate relationship sets for PubMed over the web collection. In Table 3 we show four classes of relationship sets for high = 0.15, low = 0.05, and dif f = 0.10 using the source-biased prober described above. Again we note, that our interest here is to illustrate the power of the -formalation; we leave the optimization of -values to future work. In contrast to the simple relevance ranking in Table 2, we see how the source-biased framework can differentiate between the very similar services (the -equivalent sites) and the more general services (the -superset sites) relative to PubMed. In addition, we can identify sites with some common data (the -overlap sites) and sites concerned with significantly different topics (the -complement sites). Similarly, we show in Table 4 several interesting relationships derived from the newsgroup collection for high = 0.70, low = 0.40, and dif f = 0.30 using the Source Bias prober discussed before. Again, by relying on BASIL's source-biased analysis we may characterize relationships sets for each source. As an example, we identify sci.physics.particle as a member of the -subset relationship set of the mixed topic newsgroup mixed11, which consists of 25% physics-related articles in addition to articles on backgammon, juggling, and telecommunications . Interestingly, we can see that there are several overlapping relationships between newsgroups in related but slightly different fields (e.g. volleyball and cricket). Finally , we also identify several unrelated newsgroups, including comp.sys.mac.system relative to misc.immigration.usa. 5.4 Probing with Focal Terms In our final set of experiments, we consider the impact of focal term probing on the success rate of source-biased probing. We evaluate four flavors of focal term probing with 2, 3, 5, and 10 focal term groups from which to draw source-biased probes. In our initial experiments with focal term probing, we discovered that there was little impact on either the efficiency of probing or the quality of target service evaluation when considering sources from the single-topic newsgroup collection. [Due to space limitations, we omit these results here]. 0.2 0.3 0.4 0.5 0.6 0 20 40 60 80 100 Documents Examined Average Source Similarity Original Focal - 2 Focal - 3 Focal - 5 Focal - 10 Figure 8: Impact of Focal Term Probing In contrast, we discovered that focal term probing had a significant impact when used on mixed topic newsgroups, in which there are documents from several unrelated single topic newsgroups. In Figure 8, we show the probing efficiency for the four focal term source-biased probers relative to the best basic source-biased prober for 10 source-target pairs from the newsgroup collection. In each case, the sources were drawn exclusively from the mixed topic newsgroups. All of the focal term techniques resulted in more efficient probing versus basic source-biased probing and only minor differences in ranking precision and relationship set generation quality, indicating that focal term probing can be advantageous in certain circumstances. Our intuition is that identifying focal terms is considerably more important in cases in which there are clear distinctions in term distributions as would be reflected in the mixed topic newsgroups in which several groups of documents are concerned with different topics. RELATED WORK Researchers have previously explored different aspects of the service discovery problem, ranging from discovery in a federated environment [25], to identifying services that meet certain quality-of-service guarantees [13], to evaluating services based on a distributed reputation metric [26], to other quality metrics like in [32]. In contrast, we focus on the data relationships between services to power efficient discovery and ranking. Other researchers have previously studied the problem of re-peatedly querying an unknown database in an effort to generate a summary of the database internals [11, 3, 30, 4, 8, 27, 14, 6]. The main purpose of these techniques is to generate a representative content summary of the underlying database. Querying methods suggested include the use of random queries, queries learned from a classifier, and queries based on a feedback cycle between the query and the response. More recently, Gravano et al. [12] have introduced an extension to the Callan-style probing technique that relies on a learned set of queries for database classification. Their probing method is effective for classifying web sites into a pre-determined Yahoo!-style hierarchy, but requires the potentially burdensome and inflexible task of labelling training data for learning the classifier probes in the first place. Additionally, if new categories are added or old categories removed from the hierarchy , new probes must be learned and each source re-probed. Previous research on grouping terms (as in our source-biased probing with focal terms) has focussed on finding terms that are effective for query-expansion [31, 21] or finding similar terms [24]. Our focal term formulation is similar to that used in [21], though their goal is to find close semantic relationships between terms, unlike our coarse-grained groupings. 161 Table 3: Source-Biased Analysis: Identifying Relationships Relative to PubMed Service (S) URL Description focus P M ( S) focus S ( P M) Relationship WebMD www.webmd.com Health/Medical 0.23 0.18 -equivalent AMA www.ama-assn.org Health/Medical 0.19 0.16 -equivalent HealthAtoZ www.healthatoz.com Health/Medical 0.18 0.16 -equivalent Open Directory dmoz.org Web Directory 0.44 0.08 -superset Google www.google.com Web Search Engine 0.37 0.10 -superset About www.about.com Web Channels 0.25 0.08 -superset Monster www.monster.com Jobs 0.14 0.08 -overlap Mayo Clinic www.mayoclinic.com Health/Medical 0.12 0.11 -overlap Silicon Investor www.siliconinvestor.com Finance 0.03 0.04 -complement Usenet Recipes recipes2.alastra.com Recipes 0.02 0.03 -complement Table 4: Source-Biased Analysis: Identifying Relationships in the Newsgroup Collection A B focus A ( B) focus B ( A) Relationship comp.sys.mac.apps comp.sys.mac.system 0.86 0.76 -equivalent comp.sys.mac.system comp.sys.mac.advocacy 0.79 0.74 -equivalent sci.physics.particle sci.physics 0.86 0.80 -equivalent sci.physics.particle mixed45 0.86 0.62 -subset/superset comp.unix.misc mixed120 0.91 0.56 -subset/superset rec.sport.volleyball rec.sport.cricket 0.47 0.46 -overlap rec.games.go rec.games.chess.misc 0.50 0.53 -overlap rec.crafts.textiles.sewing comp.lang.perl.misc 0.35 0.32 -complement comp.sys.mac.system misc.immigration.usa 0.23 0.36 -complement CONCLUSIONS In this paper, we have presented a novel web service discovery and ranking prototype called BASIL that supports a personalized view of data-intensive web services through source-biased focus. BASIL supports personalized discovery requests and relevance reasoning through efficient source-biased probing and source-biased relevance metrics. Concretely, we have shown that BASIL allows us to determine in very few interactions whether a target service is relevant to the source service by probing the target with very precise probes. The biased focus measure allows us to evaluate and rank the services discovered and to identify interesting types of source-biased relationships for a collection of services. Additionally, we have introduced source-biased probing with focal terms as a performance optimization to further improve the effectiveness of the basic source-biased algorithm. REFERENCES [1] Amazon.com. Amazon.com Web Services. http://www.amazon.com/gp/aws/landing.html, 2004. [2] Ariba. http://www.ariba.com, 2003. [3] J. Callan, M. Connell, and A. Du. Automatic discovery of language models for text databases. In SIGMOD '99. [4] J. P. Callan and M. E. Connell. Query-based sampling of text databases. Information Systems, 19(2):97130, 2001. [5] J. Caverlee, L. Liu, and D. Rocco. Discovering and ranking web services with BASIL: A personalized approach with biased focus. Technical report, GIT, 2004. [6] W. W. Cohen and Y. Singer. Learning to query the web. In AAAI Workshop on Internet-Based Information Systems. 1996. [7] Google. Google Web APIs FAQ. http://www.google.com/apis/, 2003. [8] D. Hawking and P. Thistlewaite. Methods for information server selection. ACM Transactions on Information Systems, 17(1):4076, 1999. [9] IBM. Web Services for Life Sciences. http://www.alphaworks.ibm.com/tech/ws4LS/, 2003. [10] IBM. IBM UDDI Business Registry. www.ibm.com/services/uddi/, 2004. [11] P. G. Ipeirotis, L. Gravano, and M. Sahami. Probe, count, and classify: Categorizing hidden-web databases. In SIGMOD '01. [12] P. G. Ipeirotis, L. Gravano, and M. Sahami. QProber: A system for automatic classification of hidden-web databases. ACM TOIS, 21(1):141, 2003. [13] Y. Liu, A. H. Ngu, and L. Zeng. QoScomputation and policing in dynamic web service selection. In WWW '04. [14] W. Meng, C. T. Yu, and K.-L. Liu. Detection of heterogeneities in a multiple text database environment. In CoopIS '99. [15] Microsoft. .NET. http://www.microsoft.com/net/, 2003. [16] Microsoft. Microsoft UDDI Business Registry Node. http://uddi.microsoft.com/, 2004. [17] National Center for Biotechnology Information. NCBI BLAST. http://www.ncbi.nih.gov/BLAST/, 2004. [18] M. P. Papazoglou. Service-oriented computing: Concepts, characteristics and directions. In WISE '03. [19] M. F. Porter. An algorithm for suffix stripping. Program, 14(3):130137, 1980. [20] ProFusion. http://www.profusion.com/, 2004. [21] Y. Qiu and H.-P. Frei. Concept-based query expansion. In SIGIR '93, pages 160169, Pittsburgh, US. [22] G. Salton and C. Buckley. Term-weighting approaches in automatic text retrieval. In Readings in Information Retrieval. Morgan Kauffman, San Francisco, CA, 1997. [23] G. Salton, A. Wong, and C. Yang. A vector space model for automatic indexing. CACM, 18(11):613620, 1971. [24] H. Schutze and J. O. Pedersen. A cooccurrence-based thesaurus and two applications to information retrieval. Information Processing and Management, 33(3), 1997. [25] K. Sivashanmugam, K. Verma, and A. Sheth. Discovery of web services in a federated registry environment. In ICWS '04. [26] R. M. Sreenath and M. P. Singh. Agent-based service selection. Journal on Web Semantics (JWS), 2003. [27] A. Sugiura and O. Etzioni. Query routing for web search engines: Architecture and experiments. In WWW '00. [28] UDDI. http://www.uddi.org/, 2004. [29] W3C Working Group. Web Services Architecture. http://www.w3.org/TR/2004/NOTE-ws-arch-20040211/, February 2004. [30] W. Wang, W. Meng, and C. Yu. Concept hierarchy based text database categorization in a metasearch engine environment. In WISE '00. [31] J. Xu and W. B. Croft. Query expansion using local and global document analysis. In SIGIR '96, pages 411. [32] L. Zeng, , B. Benatallah, M. Dumas, J. Kalagnanam, and Q. Z. Sheng. Quality driven web services composition. In WWW '03. 162

focal terms;biased discovery;ranking;data-intensive web services;data-intensive services;query-biased probing;web service discovery;source-biased probing

72

Distance Measures for MPEG-7-based Retrieval

In visual information retrieval the careful choice of suitable proximity measures is a crucial success factor. The evaluation presented in this paper aims at showing that the distance measures suggested by the MPEG-7 group for the visual descriptors can be beaten by general-purpose measures. Eight visual MPEG-7 descriptors were selected and 38 distance measures implemented. Three media collections were created and assessed, performance indicators developed and more than 22500 tests performed. Additionally, a quantisation model was developed to be able to use predicate-based distance measures on continuous data as well. The evaluation shows that the distance measures recommended in the MPEG-7-standard are among the best but that other measures perform even better.

INTRODUCTION The MPEG-7 standard defines among others a set of descriptors for visual media. Each descriptor consists of a feature extraction mechanism, a description (in binary and XML format) and guidelines that define how to apply the descriptor on different kinds of media (e.g. on temporal media). The MPEG-7 descriptors have been carefully designed to meet partially complementary requirements of different application domains: archival, browsing, retrieval, etc. [9]. In the following, we will exclusively deal with the visual MPEG-7 descriptors in the context of media retrieval. The visual MPEG-7 descriptors fall in five groups: colour, texture, shape, motion and others (e.g. face description) and sum up to 16 basic descriptors. For retrieval applications, a rule for each descriptor is mandatory that defines how to measure the similarity of two descriptions. Common rules are distance functions, like the Euclidean distance and the Mahalanobis distance. Unfortunately, the MPEG-7 standard does not include distance measures in the normative part, because it was not designed to be (and should not exclusively understood to be) retrieval-specific. However, the MPEG-7 authors give recommendations, which distance measure to use on a particular descriptor. These recommendations are based on accurate knowledge of the descriptors' behaviour and the description structures. In the present study a large number of successful distance measures from different areas (statistics, psychology, medicine, social and economic sciences, etc.) were implemented and applied on MPEG-7 data vectors to verify whether or not the recommended MPEG-7 distance measures are really the best for any reasonable class of media objects. From the MPEG-7 tests and the recommendations it does not become clear, how many and which distance measures have been tested on the visual descriptors and the MPEG-7 test datasets. The hypothesis is that analytically derived distance measures may be good in general but only a quantitative analysis is capable to identify the best distance measure for a specific feature extraction method. The paper is organised as follows. Section 2 gives a minimum of background information on the MPEG-7 descriptors and distance measurement in visual information retrieval (VIR, see [3], [16]). Section 3 gives an overview over the implemented distance measures. Section 4 describes the test setup, including the test data and the implemented evaluation methods. Finally, Section 5 presents the results per descriptor and over all descriptors. BACKGROUND The visual part of the MPEG-7 standard defines several descriptors. Not all of them are really descriptors in the sense that they extract properties from visual media. Some of them are just structures for descriptor aggregation or localisation. The basic descriptors are Color Layout, Color Structure, Dominant Color, Scalable Color, Edge Histogram, Homogeneous Texture, Texture Browsing, Region-based Shape, Contour-based Shape, Camera Motion, Parametric Motion and Motion Activity. Other descriptors are based on low-level descriptors or semantic information: Group-of-Frames/Group-of-Pictures Color (based on 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'03, November 7, 2003, Berkeley, California, USA. Copyright 2003 ACM 1-58113-778-8/03/00011...$5.00. 130 Scalable Color), Shape 3D (based on 3D mesh information), Motion Trajectory (based on object segmentation) and Face Recognition (based on face extraction). Descriptors for spatiotemporal aggregation and localisation are: Spatial 2D Coordinates, Grid Layout, Region Locator (spatial), Time Series, Temporal Interpolation (temporal) and SpatioTemporal Locator (combined). Finally, other structures exist for colour spaces, colour quantisation and multiple 2D views of 3D objects. These additional structures allow combining the basic descriptors in multiple ways and on different levels. But they do not change the characteristics of the extracted information. Consequently, structures for aggregation and localisation were not considered in the work described in this paper. 2.2 Similarity measurement on visual data Generally, similarity measurement on visual information aims at imitating human visual similarity perception. Unfortunately, human perception is much more complex than any of the existing similarity models (it includes perception, recognition and subjectivity). The common approach in visual information retrieval is measuring dis-similarity as distance. Both, query object and candidate object are represented by their corresponding feature vectors. The distance between these objects is measured by computing the distance between the two vectors. Consequently, the process is independent of the employed querying paradigm (e.g. query by example). The query object may be natural (e.g. a real object) or artificial (e.g. properties of a group of objects). Goal of the measurement process is to express a relationship between the two objects by their distance. Iteration for multiple candidates allows then to define a partial order over the candidates and to address those in a (to be defined) neighbourhood being similar to the query object. At this point, it has to be mentioned that in a multi-descriptor environment especially in MPEG-7 we are only half way towards a statement on similarity. If multiple descriptors are used (e.g. a descriptor scheme), a rule has to be defined how to combine all distances to a global value for each object. Still, distance measurement is the most important first step in similarity measurement. Obviously, the main task of good distance measures is to reorganise descriptor space in a way that media objects with the highest similarity are nearest to the query object. If distance is defined minimal, the query object is always in the origin of distance space and similar candidates should form clusters around the origin that are as large as possible. Consequently, many well known distance measures are based on geometric assumptions of descriptor space (e.g. Euclidean distance is based on the metric axioms). Unfortunately, these measures do not fit ideally with human similarity perception (e.g. due to human subjectivity). To overcome this shortage, researchers from different areas have developed alternative models that are mostly predicate-based (descriptors are assumed to contain just binary elements, e.g. Tversky's Feature Contrast Model [17]) and fit better with human perception. In the following distance measures of both groups of approaches will be considered. DISTANCE MEASURES The distance measures used in this work have been collected from various areas (Subsection 3.1). Because they work on differently quantised data, Subsection 3.2 sketches a model for unification on the basis of quantitative descriptions. Finally, Subsection 3.3 introduces the distance measures as well as their origin and the idea they implement. 3.1 Sources Distance measurement is used in many research areas such as psychology, sociology (e.g. comparing test results), medicine (e.g. comparing parameters of test persons), economics (e.g. comparing balance sheet ratios), etc. Naturally, the character of data available in these areas differs significantly. Essentially, there are two extreme cases of data vectors (and distance measures): predicate-based (all vector elements are binary, e.g. {0, 1}) and quantitative (all vector elements are continuous, e.g. [0, 1]). Predicates express the existence of properties and represent high-level information while quantitative values can be used to measure and mostly represent low-level information. Predicates are often employed in psychology, sociology and other human-related sciences and most predicate-based distance measures were therefore developed in these areas. Descriptions in visual information retrieval are nearly ever (if they do not integrate semantic information) quantitative. Consequently, mostly quantitative distance measures are used in visual information retrieval. The goal of this work is to compare the MPEG-7 distance measures with the most powerful distance measures developed in other areas. Since MPEG-7 descriptions are purely quantitative but some of the most sophisticated distance measures are defined exclusively on predicates, a model is mandatory that allows the application of predicate-based distance measures on quantitative data. The model developed for this purpose is presented in the next section. 3.2 Quantisation model The goal of the quantisation model is to redefine the set operators that are usually used in predicate-based distance measures on continuous data. The first in visual information retrieval to follow this approach were Santini and Jain, who tried to apply Tversky's Feature Contrast Model [17] to content-based image retrieval [12], [13]. They interpreted continuous data as fuzzy predicates and used fuzzy set operators. Unfortunately, their model suffered from several shortcomings they described in [12], [13] (for example, the quantitative model worked only for one specific version of the original predicate-based measure). The main idea of the presented quantisation model is that set operators are replaced by statistical functions. In [5] the authors could show that this interpretation of set operators is reasonable. The model offers a solution for the descriptors considered in the evaluation. It is not specific to one distance measure, but can be applied to any predicate-based measure. Below, it will be shown that the model does not only work for predicate data but for quantitative data as well. Each measure implementing the model can be used as a substitute for the original predicate-based measure. Generally, binary properties of two objects (e.g. media objects) can exist in both objects (denoted as a), in just one (b, c) or in none of them (d). The operator needed for these relationships are UNION, MINUS and NOT. In the quantisation model they are replaced as follows (see [5] for further details). 131 + + = = = k jk ik jk ik k k j i else x x M if x x s s X X a 0 2 2 , 1 ( ) ( ) + + = = = = = = = = = k jk ik jk ik k k j i k ik jk ik jk k k i j k jk ik jk ik k k j i else x x if x x M s s X X d else x x M if x x s s X X c else x x M if x x s s X X b 0 2 2 , 0 , 0 , 1 2 2 with: ( ) [ ] ( ) { } 0 \ . 0 1 . 0 1 , 2 2 1 min max max min + = = = = = = R p k i x where else p if p M k i x where else p if p M x x M x x x with x X i k ik i k ik ik ik i a selects properties that are present in both data vectors (X i , X j representing media objects), b and c select properties that are present in just one of them and d selects properties that are present in neither of the two data vectors. Every property is selected by the extent to which it is present (a and d: mean, b and c: difference) and only if the amount to which it is present exceeds a certain threshold (depending on the mean and standard deviation over all elements of descriptor space). The implementation of these operators is based on one assumption. It is assumed that vector elements measure on interval scale. That means, each element expresses that the measured property is &quot;more or less&quot; present (&quot;0&quot;: not at all, &quot;M&quot;: fully present). This is true for most visual descriptors and all MPEG-7 descriptors. A natural origin as it is assumed here (&quot;0&quot;) is not needed. Introducing p (called discriminance-defining parameter) for the thresholds 2 1 , has the positive consequence that a, b, c, d can then be controlled through a single parameter. p is an additional criterion for the behaviour of a distance measure and determines the thresholds used in the operators. It expresses how accurate data items are present (quantisation) and consequently, how accurate they should be investigated. p can be set by the user or automatically. Interesting are the limits: 1. M p 2 1 , In this case, all elements (=properties) are assumed to be continuous (high quantisation). In consequence, all properties of a descriptor are used by the operators. Then, the distance measure is not discriminant for properties. 2. 0 , 0 2 1 p In this case, all properties are assumed to be predicates. In consequence, only binary elements (=predicates) are used by the operators (1-bit quantisation). The distance measure is then highly discriminant for properties. Between these limits, a distance measure that uses the quantisation model is depending on p more or less discriminant for properties. This means, it selects a subset of all available description vector elements for distance measurement. For both predicate data and quantitative data it can be shown that the quantisation model is reasonable. If description vectors consist of binary elements only, p should be used as follows (for example, p can easily be set automatically): ( ) , min . . , 0 , 0 2 1 = = p g e p In this case, a, b, c, d measure like the set operators they replace. For example, Table 1 shows their behaviour for two one-dimensional feature vectors X i and X j . As can be seen, the statistical measures work like set operators. Actually, the quantisation model works accurate on predicate data for any p. To show that the model is reasonable for quantitative data the following fact is used. It is easy to show that for predicate data some quantitative distance measures degenerate to predicate-based measures. For example, the L 1 metric (Manhattan metric) degenerates to the Hamming distance (from [9], without weights): distance Hamming c b x x L k jk ik = + = 1 If it can be shown that the quantisation model is able to reconstruct the quantitative measure from the degenerated predicate-based measure, the model is obviously able to extend predicate-based measures to the quantitative domain. This is easy to illustrate. For purely quantitative feature vectors, p should be used as follows (again, p can easily be set automatically): 1 , 2 1 = p Then, a and d become continuous functions: + = = + + = = + k jk ik k k jk ik k jk ik k k jk ik x x M s where s d true M x x x x s where s a true M x x M 2 2 2 2 b and c can be made continuous for the following expressions: ( ) ( ) = = + = = = = k jk ik k k k ik jk ik jk k k ik jk ik jk k jk ik jk ik k k jk ik jk ik x x s where s c b else x x if x x s where s c x x M x x M else x x if x x s where s b x x M x x M 0 0 0 0 0 0 Table 1. Quantisation model on predicate vectors. X i X j a b c d (1) (1) 1 0 0 0 (1) (0) 0 1 0 0 (0) (1) 0 0 1 0 (0) (0) 0 0 0 1 132 = = = = k ik jk k k k jk ik k k x x s where s b c x x s where s c b This means, for sufficiently high p every predicate-based distance measure that is either not using b and c or just as b+c, b-c or c-b, can be transformed into a continuous quantitative distance measure. For example, the Hamming distance (again, without weights): 1 L x x x x s where s c b k jk ik k jk ik k k = = = = + The quantisation model successfully reconstructs the L 1 metric and no distance measure-specific modification has to be made to the model. This demonstrates that the model is reasonable. In the following it will be used to extend successful predicate-based distance measures on the quantitative domain. The major advantages of the quantisation model are: (1) it is application domain independent, (2) the implementation is straightforward, (3) the model is easy to use and finally, (4) the new parameter p allows to control the similarity measurement process in a new way (discriminance on property level). 3.3 Implemented measures For the evaluation described in this work next to predicate-based (based on the quantisation model) and quantitative measures, the distance measures recommended in the MPEG-7 standard were implemented (all together 38 different distance measures). Table 2 summarises those predicate-based measures that performed best in the evaluation (in sum 20 predicate-based measures were investigated). For these measures, K is the number of predicates in the data vectors X i and X j . In P1, the sum is used for Tversky's f() (as Tversky himself does in [17]) and , are weights for element b and c. In [5] the author's investigated Tversky's Feature Contrast Model and found =1, =0 to be the optimum parameters. Some of the predicate-based measures are very simple (e.g. P2, P4) but have been heavily exploited in psychological research. Pattern difference (P6) a very powerful measure is used in the statistics package SPSS for cluster analysis. P7 is a correlation coefficient for predicates developed by Pearson. Table 3 shows the best quantitative distance measures that were used. Q1 and Q2 are metric-based and were implemented as representatives for the entire group of Minkowski distances. The w i are weights. In Q5, i i , are mean and standard deviation for the elements of descriptor X i . In Q6, m is 2M (=0.5). Q3, the Canberra metric, is a normalised form of Q1. Similarly, Q4, Clark's divergence coefficient is a normalised version of Q2. Q6 is a further-developed correlation coefficient that is invariant against sign changes. This measure is used even though its particular properties are of minor importance for this application domain. Finally, Q8 is a measure that takes the differences between adjacent vector elements into account. This makes it structurally different from all other measures. Obviously, one important distance measure is missing. The Mahalanobis distance was not considered, because different descriptors would require different covariance matrices and for some descriptors it is simply impossible to define a covariance matrix. If the identity matrix was used in this case, the Mahalanobis distance would degenerate to a Minkowski distance. Additionally, the recommended MPEG-7 distances were implemented with the following parameters: In the distance measure of the Color Layout descriptor all weights were set to &quot;1&quot; (as in all other implemented measures). In the distance measure of the Dominant Color descriptor the following parameters were used: 20 , 1 , 3 . 0 , 7 . 0 2 1 = = = = d T w w (as recommended). In the Homogeneous Texture descriptor's distance all ( ) k were set to &quot;1&quot; and matching was done rotation- and scale-invariant. Important! Some of the measures presented in this section are distance measures while others are similarity measures. For the tests, it is important to notice, that all similarity measures were inverted to distance measures. TEST SETUP Subsection 4.1 describes the descriptors (including parameters) and the collections (including ground truth information) that were used in the evaluation. Subsection 4.2 discusses the evaluation method that was implemented and Subsection 4.3 sketches the test environment used for the evaluation process. 4.1 Test data For the evaluation eight MPEG-7 descriptors were used. All colour descriptors: Color Layout, Color Structure, Dominant Color, Scalable Color, all texture descriptors: Edge Histogram, Homogeneous Texture, Texture Browsing and one shape descriptor: Region-based Shape. Texture Browsing was used even though the MPEG-7 standard suggests that it is not suitable for retrieval. The other basic shape descriptor, Contour-based Shape, was not used, because it produces structurally different descriptions that cannot be transformed to data vectors with elements measuring on interval-scales. The motion descriptors were not used, because they integrate the temporal dimension of visual media and would only be comparable, if the basic colour, texture and shape descriptors would be aggregated over time. This was not done. Finally, no high-level descriptors were used (Localisation, Face Recognition, etc., see Subsection 2.1), because to the author's opinion the behaviour of the basic descriptors on elementary media objects should be evaluated before conclusions on aggregated structures can be drawn. Table 2. Predicate-based distance measures. No. Measure Comment P1 c b a . . Feature Contrast Model, Tversky 1977 [17] P2 a No. of co-occurrences P3 c b + Hamming distance P4 K a Russel 1940 [14] P5 c b a + Kulczvnski 1927 [14] P6 2 K bc Pattern difference [14] P7 ( )( )( )( ) d c d b c a b a bc ad + + + + Pearson 1926 [11] 133 The Texture Browsing descriptions had to be transformed from five bins to an eight bin representation in order that all elements of the descriptor measure on an interval scale. A Manhattan metric was used to measure proximity (see [6] for details). Descriptor extraction was performed using the MPEG-7 reference implementation. In the extraction process each descriptor was applied on the entire content of each media object and the following extraction parameters were used. Colour in Color Structure was quantised to 32 bins. For Dominant Color colour space was set to YCrCb, 5-bit default quantisation was used and the default value for spatial coherency was used. Homogeneous Texture was quantised to 32 components. Scalable Color values were quantised to sizeof(int)-3 bits and 64 bins were used. Finally, Texture Browsing was used with five components. These descriptors were applied on three media collections with image content: the Brodatz dataset (112 images, 512x512 pixel), a subset of the Corel dataset (260 images, 460x300 pixel, portrait and landscape) and a dataset with coats-of-arms images (426 images, 200x200 pixel). Figure 1 shows examples from the three collections. Designing appropriate test sets for a visual evaluation is a highly difficult task (for example, see the TREC video 2002 report [15]). Of course, for identifying the best distance measure for a descriptor, it should be tested on an infinite number of media objects. But this is not the aim of this study. It is just evaluated if for likely image collections better proximity measures than those suggested by the MPEG-7 group can be found. Collections of this relatively small size were used in the evaluation, because the applied evaluation methods are above a certain minimum size invariant against collection size and for smaller collections it is easier to define a high-quality ground truth. Still, the average ratio of ground truth size to collection size is at least 1:7. Especially, no collection from the MPEG-7 dataset was used in the evaluation because the evaluations should show, how well the descriptors and the recommended distance measures perform on &quot;unknown&quot; material. When the descriptor extraction was finished, the resulting XML descriptions were transformed into a data matrix with 798 lines (media objects) and 314 columns (descriptor elements). To be usable with distance measures that do not integrate domain knowledge, the elements of this data matrix were normalised to [0, 1]. For the distance evaluation next to the normalised data matrix human similarity judgement is needed. In this work, the ground truth is built of twelve groups of similar images (four for each dataset). Group membership was rated by humans based on semantic criterions. Table 4 summarises the twelve groups and the underlying descriptions. It has to be noticed, that some of these groups (especially 5, 7 and 10) are much harder to find with low-level descriptors than others. 4.2 Evaluation method Usually, retrieval evaluation is performed based on a ground truth with recall and precision (see, for example, [3], [16]). In multi-descriptor environments this leads to a problem, because the resulting recall and precision values are strongly influenced by the method used to merge the distance values for one media object. Even though it is nearly impossible to say, how big the influence of a single distance measure was on the resulting recall and precision values, this problem has been almost ignored so far. In Subsection 2.2 it was stated that the major task of a distance measure is to bring the relevant media objects as close to the origin (where the query object lies) as possible. Even in a multi-descriptor environment it is then simple to identify the similar objects in a large distance space. Consequently, it was decided to Table 3. Quantitative distance measures. No. Measure Comment No. Measure Comment Q1 k jk ik i x x w City block distance (L 1 ) Q2 ( ) k jk ik i x x w 2 Euclidean distance (L 2 ) Q3 + k jk ik jk ik x x x x Canberra metric, Lance, Williams 1967 [8] Q4 ( ) + k jk ik jk ik x x x x K 2 1 Divergence coefficient, Clark 1952 [1] Q5 ( ) ( ) ( ) ( ) k k j jk i ik k j jk i ik x x x x 2 2 Correlation coefficient Q6 + + k ik k jk ik k ik k k jk k ik jk ik x m Km x x m Km x x x m Km x x 2 . . 2 2 2 2 2 Cohen 1969 [2] Q7 k k jk ik k jk ik x x x x 2 2 Angular distance, Gower 1967 [7] Q8 ( ) ( ) ( ) + + 1 2 1 1 K k jk jk ik ik x x x x Meehl Index [10] Table 4. Ground truth information. Coll. No Images Description 1 19 Regular, chequered patterns 2 38 Dark white noise 3 33 Moon-like surfaces Brodatz 4 35 Water-like surfaces 5 73 Humans in nature (difficult) 6 17 Images with snow (mountains, skiing) 7 76 Animals in nature (difficult) Corel 8 27 Large coloured flowers 9 12 Bavarian communal arms 10 10 All Bavarian arms (difficult) 11 18 Dark objects / light unsegmented shield Arms 12 14 Major charges on blue or red shield 134 use indicators measuring the distribution in distance space of candidates similar to the query object for this evaluation instead of recall and precision. Identifying clusters of similar objects (based on the given ground truth) is relatively easy, because the resulting distance space for one descriptor and any distance measure is always one-dimensional. Clusters are found by searching from the origin of distance space to the first similar object, grouping all following similar objects in the cluster, breaking off the cluster with the first un-similar object and so forth. For the evaluation two indicators were defined. The first measures the average distance of all cluster means to the origin: distance avg clusters no size cluster distance clusters no i i size cluster j ij d i _ . _ _ _ _ = where distance ij is the distance value of the j-th element in the i-th cluster, = CLUSTERS i i CLUSTERS i size cluster j ij size cluster distance distance avg i _ _ _ , no_clusters is the number of found clusters and cluster_size i is the size of the i-th cluster. The resulting indicator is normalised by the distribution characteristics of the distance measure (avg_distance). Additionally, the standard deviation is used. In the evaluation process this measure turned out to produce valuable results and to be relatively robust against parameter p of the quantisation model. In Subsection 3.2 we noted that p affects the discriminance of a predicate-based distance measure: The smaller p is set the larger are the resulting clusters because the quantisation model is then more discriminant against properties and less elements of the data matrix are used. This causes a side-effect that is measured by the second indicator: more and more un-similar objects come out with exactly the same distance value as similar objects (a problem that does not exist for large p's) and become indiscernible from similar objects. Consequently, they are (false) cluster members. This phenomenon (conceptually similar to the &quot;false negatives&quot; indicator) was named &quot;cluster pollution&quot; and the indicator measures the average cluster pollution over all clusters: clusters no doubles no cp clusters no i size cluster j ij i _ _ _ _ = where no_doubles ij is the number of indiscernible un-similar objects associated with the j-th element of cluster i. Remark: Even though there is a certain influence, it could be proven in [5] that no significant correlation exists between parameter p of the quantisation model and cluster pollution. 4.3 Test environment As pointed out above, to generate the descriptors, the MPEG-7 reference implementation in version 5.6 was used (provided by TU Munich). Image processing was done with Adobe Photoshop and normalisation and all evaluations were done with Perl. The querying process was performed in the following steps: (1) random selection of a ground truth group, (2) random selection of a query object from this group, (3) distance comparison for all other objects in the dataset, (4) clustering of the resulting distance space based on the ground truth and finally, (5) evaluation. For each combination of dataset and distance measure 250 queries were issued and evaluations were aggregated over all datasets and descriptors. The next section shows the partially surprising results. RESULTS In the results presented below the first indicator from Subsection 4.2 was used to evaluate distance measures. In a first step parameter p had to be set in a way that all measures are equally discriminant. Distance measurement is fair if the following condition holds true for any predicate-based measure d P and any continuous measure d C : ( ) ( ) C P d cp p d cp , Then, it is guaranteed that predicate-based measures do not create larger clusters (with a higher number of similar objects) for the price of higher cluster pollution. In more than 1000 test queries the optimum value was found to be p=1. Results are organised as follows: Subsection 5.1 summarises the Figure 1. Test datasets. Left: Brodatz dataset, middle: Corel dataset, right: coats-of-arms dataset. 135 best distance measures per descriptor, Section 5.2 shows the best overall distance measures and Section 5.3 points out other interesting results (for example, distance measures that work particularly good on specific ground truth groups). 5.1 Best measure per descriptor Figure 2 shows the evaluation results for the first indicator. For each descriptor the best measure and the performance of the MPEG-7 recommendation are shown. The results are aggregated over the tested datasets. On first sight, it becomes clear that the MPEG-7 recommendations are mostly relatively good but never the best. For Color Layout the difference between MP7 and the best measure, the Meehl index (Q8), is just 4% and the MPEG-7 measure has a smaller standard deviation. The reason why the Meehl index is better may be that this descriptors generates descriptions with elements that have very similar variance. Statistical analysis confirmed that (see [6]). For Color Structure, Edge Histogram, Homogeneous Texture, Region-based Shape and Scalable Color by far the best measure is pattern difference (P6). Psychological research on human visual perception has revealed that in many situation differences between the query object and a candidate weigh much stronger than common properties. The pattern difference measure implements this insight in the most consequent way. In the author's opinion, the reason why pattern difference performs so extremely well on many descriptors is due to this fact. Additional advantages of pattern difference are that it usually has a very low variance and because it is a predicate-based measure its discriminance (and cluster structure) can be tuned with parameter p. The best measure for Dominant Color turned out to be Clark's Divergence coefficient (Q4). This is a similar measure to pattern difference on the continuous domain. The Texture Browsing descriptor is a special problem. In the MPEG-7 standard it is recommended to use it exclusively for browsing. After testing it for retrieval on various distance measures the author supports this opinion. It is very difficult to find a good distance measure for Texture Browsing. The proposed Manhattan metric, for example, performs very bad. The best measure is predicate-based (P7). It works on common properties (a, d) but produces clusters with very high cluster pollution. For this descriptor the second indicator is up to eight times higher than for predicate-based measures on other descriptors. 5.2 Best overall measures Figure 3 summarises the results over all descriptors and media collections. The diagram should give an indication on the general potential of the investigated distance measures for visual information retrieval. It can be seen that the best overall measure is a predicate-based one. The top performance of pattern difference (P6) proves that the quantisation model is a reasonable method to extend predicate-based distance measures on the continuous domain. The second best group of measures are the MPEG-7 recommendations, which have a slightly higher mean but a lower standard deviation than pattern difference. The third best measure is the Meehl index (Q8), a measure developed for psychological applications but because of its characteristic properties tailor-made for certain (homogeneous) descriptors. Minkowski metrics are also among the best measures: the average mean and variance of the Manhattan metric (Q1) and the Euclidean metric (Q2) are in the range of Q8. Of course, these measures do not perform particularly well for any of the descriptors. Remarkably for a predicate-based measure, Tversky's Feature Contrast Model (P1) is also in the group of very good measures (even though it is not among the best) that ends with Q5, the correlation coefficient. The other measures either have a significantly higher mean or a very large standard deviation. 5.3 Other interesting results Distance measures that perform in average worse than others may in certain situations (e.g. on specific content) still perform better. For Color Layout, for example, Q7 is a very good measure on colour photos. It performs as good as Q8 and has a lower standard deviation. For artificial images the pattern difference and the Hamming distance produce comparable results as well. If colour information is available in media objects, pattern difference performs well on Dominant Color (just 20% worse Q4) and in case of difficult ground truth (group 5, 7, 10) the Meehl index is as strong as P6. 0,000 0,001 0,002 0,003 0,004 0,005 0,006 0,007 0,008 Q8 MP 7 P6 MP 7 Q4 MP 7 P6 MP 7 P6 MP 7 P6 MP 7 P6 MP 7 P7 Q2 Color Layout Color Structure Dominant Color Edge Histogram Homog. Texture Region Shape Scalable Color Texture Browsing Figure 2. Results per measure and descriptor. The horizontal axis shows the best measure and the performance of the MPEG-7 recommendation for each descriptor. The vertical axis shows the values for the first indicator (smaller value = better cluster structure). Shades have the following meaning: black=- (good cases), black + dark grey= (average) and black + dark grey + light grey=+ (bad). 136 CONCLUSION The evaluation presented in this paper aims at testing the recommended distance measures and finding better ones for the basic visual MPEG-7 descriptors. Eight descriptors were selected, 38 distance measures were implemented, media collections were created and assessed, performance indicators were defined and more than 22500 tests were performed. To be able to use predicate-based distance measures next to quantitative measures a quantisation model was defined that allows the application of predicate-based measures on continuous data. In the evaluation the best overall distance measures for visual content as extracted by the visual MPEG-7 descriptors turned out to be the pattern difference measure and the Meehl index (for homogeneous descriptions). Since these two measures perform significantly better than the MPEG-7 recommendations they should be further tested on large collections of image and video content (e.g. from [15]). The choice of the right distance function for similarity measurement depends on the descriptor, the queried media collection and the semantic level of the user's idea of similarity. This work offers suitable distance measures for various situations. In consequence, the distance measures identified as the best will be implemented in the open MPEG-7 based visual information retrieval framework VizIR [4]. ACKNOWLEDGEMENTS The author would like to thank Christian Breiteneder for his valuable comments and suggestions for improvement. The work presented in this paper is part of the VizIR project funded by the Austrian Scientific Research Fund FWF under grant no. P16111. REFERENCES [1] Clark, P.S. An extension of the coefficient of divergence for use with multiple characters. Copeia, 2 (1952), 61-64. [2] Cohen, J. A profile similarity coefficient invariant over variable reflection. Psychological Bulletin, 71 (1969), 281-284 . [3] Del Bimbo, A. Visual information retrieval. Morgan Kaufmann Publishers, San Francisco CA, 1999. [4] Eidenberger, H., and Breiteneder, C. A framework for visual information retrieval. In Proceedings Visual Information Systems Conference (HSinChu Taiwan, March 2002), LNCS 2314, Springer Verlag, 105-116. [5] Eidenberger, H., and Breiteneder, C. Visual similarity measurement with the Feature Contrast Model. In Proceedings SPIE Storage and Retrieval for Media Databases Conference (Santa Clara CA, January 2003), SPIE Vol. 5021, 64-76. [6] Eidenberger, H., How good are the visual MPEG-7 features? In Proceedings SPIE Visual Communications and Image Processing Conference (Lugano Switzerland, July 2003), SPIE Vol. 5150, 476-488. [7] Gower, J.G. Multivariate analysis and multidimensional geometry. The Statistician, 17 (1967),13-25. [8] Lance, G.N., and Williams, W.T. Mixed data classificatory programs. Agglomerative Systems Australian Comp. Journal, 9 (1967), 373-380. [9] Manjunath, B.S., Ohm, J.R., Vasudevan, V.V., and Yamada, A. Color and texture descriptors. In Special Issue on MPEG-7 . IEEE Transactions on Circuits and Systems for Video Technology, 11/6 (June 2001), 703-715. [10] Meehl, P. E. The problem is epistemology, not statistics: Replace significance tests by confidence intervals and quantify accuracy of risky numerical predictions. In Harlow, L.L., Mulaik, S.A., and Steiger, J.H. (Eds.). What if there were no significance tests? Erlbaum, Mahwah NJ, 393-425. [11] Pearson, K. On the coefficients of racial likeness. Biometrica, 18 (1926), 105-117. [12] Santini, S., and Jain, R. Similarity is a geometer. Multimedia Tools and Application, 5/3 (1997), 277-306. [13] Santini, S., and Jain, R. Similarity measures. IEEE Transactions on Pattern Analysis and Machine Intelligence, 21/9 (September 1999), 871-883. [14] Sint, P.P. Similarity structures and similarity measures. Austrian Academy of Sciences Press, Vienna Austria, 1975 (in German). [15] Smeaton, A.F., and Over, P. The TREC-2002 video track report. NIST Special Publication SP 500-251 (March 2003), available from: http://trec.nist.gov/pubs/trec11/papers/ VIDEO.OVER.pdf (last visited: 2003-07-29) [16] Smeulders, A.W.M., Worring, M., Santini, S., Gupta, A., and Jain, R. Content-based image retrieval at the end of the early years. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22/12 (December 2000), 1349-1380. [17] Tversky, A. Features of similarity. Psychological Review, 84/4 (July 1977), 327-351. 0,000 0,002 0,004 0,006 0,008 0,010 0,012 0,014 0,016 0,018 0,020 P6 MP 7 Q8 Q1 Q4 Q2 P2 P4 Q6 Q3 Q7 P1 Q5 P3 P5 P7 Figure 3. Overall results (ordered by the first indicator). The vertical axis shows the values for the first indicator (smaller value = better cluster structure). Shades have the following meaning: black=-, black + dark grey= and black + dark grey + light grey=+. 137

Similarity Perception;MPEG-7 descriptors;Distance Measurement;Content-based Image Retrieval;MPEG-7;distance measure;quantisation;Content-based Video Retrieval;Similarity Measurement;visual information retrieval;Visual Information Retrieval;human similarity perception

73

Dogs or Robots: Why do Children see them as Robotic Pets rather than Canine Machines?

In the not too distant future Intelligent Creatures (robots, smart devices, smart vehicles, smart buildings , etc) will share the everyday living environment of human beings. It is important then to analyze the attitudes humans are to adopt for interaction with morphologically different devices, based on their appearance and behavior. In particular, these devices will become multi-modal interfaces, with computers or networks of computers, for a large and complex universe of applications. Our results show that children are quickly attached to the word `dog' reflecting a conceptualization that robots that look like dogs (in particular SONY Aibo) are closer to living dogs than they are to other devices. By contrast, adults perceive Aibo as having stronger similarities to machines than to dogs (reflected by definitions of robot). Illustration of the characteristics structured in the definition of robot are insufficient to convince children Aibo is closer to a machine than to a dog.

Introduction The play R. U. R. (Rossum's Universal Robots), written by the Czech author Karel Capek, was produced in London in 1923. The term robot entered the English language (in Czech the word `robota' means `heavy labor'). The robot concept remained science fiction until 1961 when Unimation Inc. installed the world's first industrial robot in the US. Unimation Inc. made Australia's first robot, installed in 1974. The emergence of legged autonomous robots and their commercial release (as in Honda's Asimo and Sony's Aibo) contribute to support the hypothesis that mobile robots will soon become common in our everyday environments. The commercial release of the Personal Computer (PC) occurred just a generation ago, yet now it is a common household item. This forecast has prompted some studies into the acceptabil-Copyright c 2004, Australian Computer Society, Inc. This paper appeared at 5th Australasian User Interface Conference (AUIC2004), Dunedin. Conferences in Research and Practice in Information Technology, Vol. 28. A. Cockburn, Ed. Reproduction for academic, not-for profit purposes permitted provided this text is included. This work was supported by a Griffith University Research Grant as part of the project "Robots to guide the blind the application of physical collaborative autonomous intelligent agents". ity and attitudes these artifacts generate among human beings (Fong, Nourbakhsh & Dautenhahn 2003). For example, Kahn et al have recently embarked on a series of studies with children (Kahn Jr., Friedman , Freier & Severson 2003) and adults (Kahn Jr., Friedman & Hagman 2002) investigating matters such as humans attributing social and ethical stance to robots like Sony's Aibo. Earlier Bumby and Dautenhahn (Bumby & Dautenhahn 1999) explored reactions of children as they interacted with a robot. Reports have recently appeared in the media of cases where humans also build emotional attachments to robots that do not look like animals, similar to those they have for home pets. An example is the robotic vacuum cleaners in (Kahney 2003). While some experts in childhood development argue that the use of robot-dolls (machines with realism ) in children's environments `cuts off open-ended imaginative play' there are some studies showing that intelligent children will still explore beyond the limitations of the machine. Similar concerns about other technologies have been also subject of debate. For example, the concerns about children reducing their play with other children, or increasingly aggressive behavior because of computer games, can be attributed more to exposure to the type of game than to the technology of computer games themselves (Lawry, Upitis, Klawe, Anderson, Inkpen, Ndunda, Hsu, Leroux & Sedighian 1995). Researchers have found that children (in particular boys) will entertain and seek more challenging and interesting computer games (not necessarily violent games) and that there is no observable increase in violent behavior or deterioration in social behavior (Lawry et al. 1995). Recent studies have focused on the attitudes gen-erated by Sony's Aibo on humans, we propose here to explore the differences between Dog (as in living animal) and Robot (as in lifeless machine assembled from parts) in the concepts and language formations in children. Naturally, the smaller the difference, the more it is likely that humans will attribute animal characteristics (as high up as rights) to a robot. The question is, `what makes a small or a large difference?' If the difference is very small, perhaps humans will interact with autonomous robots as they do with animals . We suggest that in today's children's world, the issue is not confusion between reality and fantasy (Aylett 2002). To a child, Sony's Aibo is not a fantasy but reality. Identification of what makes children perceive a robot as a dog (an animal) or as a robot is important, especially if one wants to design robots stressing the difference or diluting it. Our research reveals that the look and feel of Sony's Aibo and its body shape go a long way into its acceptability as a dog. Its play-7 ful behavior, tail wagging, legged walk, recovery from falls, sitting and hand shaking are absorbed into the child's mind. Later, illustration of its robotic features are repeatedly insufficient to fully convince the children that this is an artifact (and not a being with feelings). Unless Aibo does something unacceptable for a dog (like speak with a human voice), it remains essentially a dog. Our findings that human speech in Aibo reduces its dog-ness and increases its robot-ness may be attributed to the `uncanny valley' (Scheeff, Pinto, Rahardja, Snibbe & Tow 2000). Although we are not measuring emotional response, we have observed dissatisfaction with Aibo as a dog, since clearly it is only humans that talk (although children accept talking toys and talking animals in fantasy or animated movies). The rest of this paper is organized as follows. Section 2 will describe our research methods. Section 3 will elaborate on the findings. Section 4 will present conclusions and final remarks. Our aim is to explore and contrast the properties currently accepted in the definition of `mobile autonomous robot'. The International Federation of Robotics (IFR) and the Australian Robot Association follow the ISO standard vocabulary (ISO 8373) to describe `manipulating industrial robots operated in a manufacturing environ-ment' . A robot has three essential characteristics: 1. It possesses some form of mobility (formally, a robot must possess at least three programmable axes of motion). 2. It can be programmed to accomplish a large variety of tasks. 3. After being programmed, it operates automati-cally . Mobile robots can move away from a fixed location and come in two varieties: tethered and autonomous; a tethered robot may have its power supply or its control unit overboard, possibly relying on a desktop computer and a wall outlet and a long cord. Autonomous mobile robots bring everything along (control unit, body, actuators, sensors and power supply). The control unit is typically a computer and software. Our research attempts to find out if children do indeed notice these properties or fixate more on the form and behavior of the artifact. The methods This research was performed by a series of demonstrations of Aibo and other robots, toys and models . in particular using Lego Mindstorms 1 constructions (Knudsen 1999), remote control toy cars and autonomous battery toys. The demonstrations were conducted with preschool children as well as those from the first 4 years of primary school across 3 schools, two childcare centers , and a museum in the urban area of Brisbane, Australia. Table 1 summarizes the presentations and the age groups of children. Whenever consent was given, a video/or audio of the session was recorded. Alternatively, a secretary took notes of the statement made by children. Sessions lasted between 30 minutes and approximately one hour. The sessions consisted of five stages. 1. Establishing the language. In the first minutes of the session, children are asked to describe what they see. The goal of this stage is to obtain a vocabulary from the children to discuss the artifacts that will be presented later. 1 A trademark of the LEGO group. 2. Demonstration of Aibo. In the next 5 to 7 minutes , a demonstration of Aibo is performed with one black model without a memory stick, so the default simple `dog-like' behaviors are in place. 3. Demonstration of the concept of robot. This stage illustrates the main features that are common in accepted definitions of a robot. It also ensures that the children observe that Sony's Aibo shares these properties with robots. 4. Investigate animal attributes on Aibo. This stage questions the children for the existence of animal properties on Sony's Aibo and the justification for their belief. 5. Challenge Aibo's animal attributes with the other artifacts used in the session. Children are asked to confirm or justify that Aibo is a robot. Attempts are made to convince them of the artificial nature of Aibo by showing the same property in an artifact accepted as lifeless and to compel them to decide on one side of the Dog vs Robot debate (or generate a new term to describe it). The initial phase starts with the projection of a video from RoboCup-2002 legged league (Veloso, Uther, Fujita, Asada & Kitano 1998) (the video is the match between the University of Newcastle and Team Sweden). Presentations at the Powerhouse Museum in Brisbane consisted of a live match with 8 dogs programmed as the Mi-PAL Team Griffith participation in RoboCup-2003. After two minutes the children are asked to describe what they see in the video. In the video, human manipulators are visible, which contributes to the realization that these are real things and not an animation film. Children are requested to indicate what the `dogs' are doing (if they suggested this word, but if they suggest `puppies' then `puppies' is used). That is, we use the same words the children themselves have introduced to describe what they see. Children are then asked to justify why it is a game of `soccer/rugby' (or whatever word was the most common or immediate reply). The phase finishes by bringing an Aibo that is switched off, placing it on the ground, turning it on, and waiting. Since Aibo requires some seconds to `boot' we observe children's reactions and comments. This phase is obviously different for blind children. It is replaced by allowing children to explore the robot with their hands. Blind children still find and recognize legs, paws, ears, head and tail because of shape, texture, malleability and movement. We then proceed to phase two where we illustrate the default behavior of the Aibo, which includes the following interactions. A couple of fast strokes on its back starts it on a walk while it makes the sounds of a marching tune. Hard strokes on its head produce sounds and a red light on its head LEDs. Soft strokes on its head produce sounds and a green light on its LEDs. Scratching under its chin produces another set of sounds and lights. Placing it on the floor on its side produces some sounds, and then Aibo performs a routine to stand back on its four legs. After getting up, Aibo shakes, wagging his tail and making other sounds. Presenting a pink ball produces sounds and the LED on its tail to go pink. 8 School Level Children's age Group size Boronia Childcare Center pre-school 4-5 10 Carole Park State School pre-school 4-5 17 Holland Park State School 3rd year primary 7-8 25 Holland Park State School 1st year primary 5-6 24 Camp Hill State School 1st year primary 5-6 10 Camp Hill State School 1st year primary 5-6 12 Camp Hill State School 1st year primary 5-6 11 Carole Park State School 1st year primary 5-6 22 Carole Park State School 2nd year primary 6-7 24 Carole Park State School 2nd year primary 6-7 28 Carole Park State School 3rd year primary 7-8 26 Carole Park State School 4th year primary 8-9 20 Powerhouse Museum pre-school 4-5 10 Narbethong State Special School pre-school (blind) 4-5 3 Table 1: Presentations conducted and age groups. Figure 1: A 4-legged walking toy with visible battery, motor and gears. A flexible tail resembles a dog tail. Other behaviors that Aibo produces are not directly triggered by the manipulator. These include Aibo sitting down, Aibo lying on his stomach and waving all fours as a synchronized dance, Aibo waiving one leg, Aibo moving his head from side to side and flapping his ears. Children are then invited to interact with Aibo directly . In particular, to show the pink ball, to produce the green lights or invite him to walk. They are also invited to explain what Aibo is doing in their own words. There are a series of questions that the presenter goes through as the illustration of behaviors is performed. These questions are as follows: What is Aibo doing now? Is he happy or is he sad? Does he like to be touched like this? Do you think he can get up by himself? At the completion of this phase, Aibo is turned off and focus is transferred to other examples of robotics. Because the commonly accepted definitions of mobile robots includes that they have their own control unit, phase three consists of the following: A presentation of a 4-legged toy with a tail (made of a spring) that has a visible battery, motor and gears (Figure 1). It is illustrated that this toy needs a battery to operate it and that it has an on-off-reverse button. Children are asked to carry out the task of setting it off or stopping it by taking the battery out. This illustrates that mobile autonomous robots require power and carry their source of power. Figure 2: A model of a humanoid robot. Figure 3: Remote control car to contrast with the notion of autonomous control. A presentation of a model of a humanoid robot (Figure 2). Although it looks like a robot, it can be seen that it has no motors, no batteries and essentially does nothing. A presentation of a remote control car (radio con-trolled ) (Figure 3). This car is also shown to have batteries on board but all behavior is de-rived from the actions on the two-lever remote control. The first lever produces forward or reverse motion on the back wheels and the second lever produces left/right turns on the front wheels. This illustrates the notion of control (remote and human). A Lego Mindstorm construction extremely similar to `Hank the Bumper Tank' (Knudsen 1999) (Figure 4). This robot is shown to have a behavior that allows it to move around and steer away from obstacles it bumps into (the program is very similar to the one suggested in (Knudsen 1999, Chapter 2)). As part of the interactive nature of the presentation, the children are asked to act 9 Figure 4: A Mindstorm construction with touch and light sensor. Figure 5: A Mindstorm construction with touch and light sensor, that acts on its environment with a mechanical arm, and plays sounds. as obstacles for `adapted' Hank. The presenter points out the sensors behind the robots bumper, and illustrates that disconnecting these sensors makes it `unaware' of obstacles. We also added to Hank a program that used a light sensor to monitor the color of the ground beneath it. This program was similar to the obstacle avoidance previously mentioned, but rather than avoid objects it would move away from dark areas. We presented this behavior as `being afraid of the dark'. By switching between these modes, we illustrate that the behavior of the robot changes with the chosen program. Using Lego's ROBO-Lab (a graphical programming application) to build a very simple program that makes Hank spin in a circle for four seconds, we show the children its programmable nature. The children are taken through the process of building the program, transferring it onto Hank via an infrared interface and finally running it. When the program is running, the children are encouraged to count along, thus verifying that the program is indeed the one just built. This is repeated for at least two other timings (around 10 to 20 seconds). A Lego construction extremely similar to Minerva (Knudsen 1999, Chapter 6) was presented next (Figure 5). The components were shown to be the same as Hank's and the Lego RCX is labeled as the `computer control'. A program similar to the one suggested (Knudsen 1999) produces the behavior illustrated to the children. Minerva moves around a white floor until it finds a black object, uses a robotic arm to pick it up, then turns around and brings it to another position close to where it started. It then releases the object and plays a tune. The presenter en-sured that the children observed that Minerva perceives its environment and can act to change it (thus the notion of actuator is illustrated). A series of pictures (or videos) of autonomous robots were shown to the children. These images demonstrate that robots come in all sorts of shapes and sizes. Among these are pictures of more Aibos, Honda ASIMO, the Sony humanoid SDX, MINERVA (Thrun, Bennewitz, Burgard, Cremers, Dellaert, Fox, Hahnel, Rosenberg, Roy, Schulte & Schulz 1999) and Kismet (Brooks 2002). It was pointed out that robots can produce smiles, walk, and be as small as a cockroach. Pictures of experimental robots were shown to display the wires, gears and components inside their packaging. Aibo was then brought back as the presenter repeated the main concepts, namely: Aibo requires power and carries a battery. Aibo is turned on and off. Also, it is shown that Aibo's behaviors are interrupted and stopped if the battery is removed. Aibo has motors. The gears on Aibo's joints, and wires near its neck are pointed out to the children . Aibo has sensors. The strokes on head and sensitivity to the pink ball are illustrated once more. Using another pink object (perhaps a piece of clothing, or the memory stick of Aibo), we show that the behavior is triggered by the color being noticed by a sensor and not Aibo understanding the concept of a ball. Aibo has actuators and a control program. We install a memory stick and re-start Aibo. With the new program it kicks a ball as in the RoboCup video. We illustrate Aibo's behavior changes with different memory sticks. Once this is completed the next phase commences by the presenter asking one of the following questions. 1. Does Aibo have feelings? 2. Where does Aibo get energy from? 3. Will Aibo have babies? Responses of several children were collected. We expected that this question would have distinct answers depending on whether we were referring to a robot or a `living' dog. We then passed to the final stage. Each time a child made a response that seemed to indicate animal essence or animal agency in the Aibo, we chal-lenged the response. For example, if a child indicated that Aibo had feelings, we next asked what the child thinks happens to the feelings when Aibo is turned off. We found children would continue to support their point of view. Following the previous example, many children followed the path that Aibo was just asleep when turned off. The challenge continued as the presenter requested children to explain what sort of feelings Aibo has or if the feelings fade when the battery runs out. Also, the presenter checked if the other artifacts, shown before, have feelings and asked the children to explain why the others do or do not have those or other feelings. The sequence of challenges for the 3 questions above were as follows. 1. Does Aibo have feelings? What happens to Aibo's feelings when he is turned off? 10 Figure 6: The demonstrator with a class of grade 2 children and 3 of the objects: Aibo, Hank and the 4-legged walking toy. What happens to his feelings when the battery runs out? What happens to his feelings if we re-move/change/replace the memory stick and Aibo`s personality changes? What happens to his feelings if Aibo is broken ? Will the feelings come back if we glue him? What feelings do you know Aibo has? How do you know he is happy/sad? Is it possible to pretend to be happy but not be happy? Do you think Aibo is happy or just pretending? 2. Where does Aibo get his energy from? Where do you get your energy from? Where do the other artifacts get their energy from? Does Aibo work without a battery? Do the others work without a battery? What do you think Aibo eats/drinks? Do you think he needs to visit the toilet? 3. Will Aibo have babies? Is Aibo a baby dog? How do you know? How will Aibo look after (care for) the babies ? Does Aibo need to charge/replace the battery of the babies? These paths of questioning were not all developed ahead of the first presentation. They evolved from one presentation to the next. Their length reflects the resistance of the children to change their opinion, even if all other artifacts have opposite responses to these questions. That is, for each question, before we progressed to the next, we confirmed that the children sustained the notion that a difference remains between Aibo and the other artifacts. For example, in the last question sequence, children would start by confirming that none of the other artifacts can have babies while Aibo can. When progressing to `Is Aibo a baby?' and contrasting this with `Is Hank a baby?', most children realized that Aibo is really like Hank and cannot have babies. The findings After an analysis of the transcript of our videos and notes we summarize the following findings. It is remarkable that when we queried the children for their first impressions of the video we obtained the following results. To the question `What do you see?' all sessions had children responding that they saw `dogs'. This is surprising for two main reasons: firstly the video shows the robots playing soccer, a behavior not commonly attributed to dogs. And secondly, the children would have anticipated seeing robots through prior conversations with parents and teachers. This may explain why a few children claimed that the adults in the video where robots. As the age of the pupils involved in the study in-creased , we noticed that the tendency to regard the robots as `dogs' decreased. The more mature respondents were more likely to label the robots as `robots' or `robotic dogs'. One pupil also gave the more generic answer of `animals', and another thought that they were `cats'. Interestingly, on a few occasions children referred to the humans in the video as the robots. We believe this is due to their anticipation to see robots and perhaps the media culture of humanoid robots. To the question, `What are they doing?' most children identified the activity as a game of `soccer'. This is surprising, since the RoboCup has barriers around the pitch that make the game more similar to ice-hockey , and although the robots are legged, they do not kick the ball with one leg. All robots in this video kick the ball with two legs simultaneously or head-butt the ball. The ball is also bright orange, clearly not a soccer ball. Another point is that although played in Australia, it is not the most commonly played sport. Other suggestions included, `they're fighting', `playing hockey', and one child thought he was watching a game of tennis. Justifications for `why is it a game of soc-cer/football ?' included a rounded ball on the ground, goals, two teams, referees and goalies. When initially presenting the Aibo to the children, rather than give it the label `it', we found that they would usually use `him' or `her'. Once again this was more pronounced with younger subjects. As the presenter went on to explain the attributes of the Aibo and show its operation, the children while probing with questions would begin to lose the gender label. The children generally were of the opinion that the Aibo did have emotions, with a couple of them claiming that this was so because it had a `mind'. This opinion was seemingly an accepted one with many children declaring at certain stages of the proceedings that it was either happy or sad. Upon the exhibition of other robots and robot-like artifacts, the general consensus was that the Aibo did in fact meet the criteria for being a robot. However, the most common term used to describe it was that of `robotic dog', where dog is the noun. This emphasis on the dog nature of the robot suggests that the subjects were still willing to consider it animal-like. The youngest group, however, needed the most convincing. They insisted that the Aibo was a dog, even after repeated demonstrations of its robotic nature , with the presenter even stating in no uncertain terms that it was a robot. They did come around eventually, with one child using the `robotic dog' description , and his peers following suit. We briefly describe some reaction to the other objects . Although initially enthusiastic, children were quickly disappointed by the model of the humanoid; mainly its inaction made it uninteresting. One child said, "it's just a toy, not a robot". The 4-legged walking toy caused some laughs because it bumps into things, but children realized rapidly that it did not offer any `interesting' interaction beside turning it on and off (potentially reversing the direction). The remote control car was appealing and children wanted to play with it even after the presentation. It was clear to them they were controlling it. Hank did 11 cause surprise and children wanted to continue playing with it, or asked about how to program it. Children wanted to interact with it and explored different obstacles for its obstacle-backing behavior. On two occasions we witnessed children convinced that Hank also had feelings because it was "afraid of the dark". The mechanical-arm robot caused amazement. We believe this greater surprise was because children familiar with Lego do not expect the action of a mechanical arm lifting an object. We also performed a variation in our initial approach to confirm some of these findings. We approached a different grade 6 class (12 year-olds) that had been already working with Mindstorm robots and had done some research assignments on the Internet and in the library on topics such as `What is a robot?'. We did a presentation in which the objects were not necessarily the focus, but the properties of a robot were the focus. We also demonstrated different applications of robotics, like using Miranda to assist a blind person to read a WEB page. The method for collecting the children's attitudes was a questionnaire of 25 questions asking children to choose between two positions and to give their reasons for such decisions. We invited them to reflect on their responses, so they were asked to answer the questions over a day at school and at home. The results of 23 answered questionnaires confirmed that a dog-looking robot rapidly acquires animalistic properties and values in the minds of children . In particular, 75% of the children confirmed that Miranda should be called a `robotic dog' rather than a `dog-looking robot'. Note that the preferred noun is dog over robot. The reasons provided in the questionnaire are illustrative of their thinking: "It has more dog features than robot features", "Miranda has characteristics a dog has", "Kinda looks like a real dog" "It is an automatic dog" and "Just doesn't look like a dog, she has a dog personality". And on the question "Does Miranda have feelings?" again 75% responded positively. Some of the reasons were "She just isn't a robot. She's almost a real dog", "She can be happy, unhappy", If you hit her hard, she would make a noise, but she felt it". Note that in this presentation we actually changed programs several times, radically changing the behaviors and personality of Miranda. Also, real dogs do not talk, but our programs had a female voice for instructions to kick a ball and a male voice reading Web pages. Only one child classified Miranda as a robot because dogs do not sing. Discussion and Final Remarks The blurring of the concept of robotic pet or canine machine is of interest to us because of the direct applications of autonomous mobile robots in helping people . In particular, we foresee that people with disabilities , the elderly and other groups in need of assistance , are the first humans that will benefit from autonomous mobile robots. Naturally, the attitudes, acceptability and adequate expectations are to match an effective human-computer interaction. If the person expects smarter behavior of the robot (things like gesture/voice recognition) and the technology does not deliver, then rather than assisting, we will frustrate the person. It is also important that anyone who encounters a person assisted by a robot approaches with attitudes and gestures that allow the robotic assistant to facilitate the approach. The main motivation behind this research is a related project on using Aibo to assist blind people. While it may seem straightforward that a robotic assistant for the vision impaired person should be shaped as a dog, this is not so. Even with guide dogs, other humans find it difficult to approach and assist a blind person. Humans expect a strong bond and loyalty of the animal to its owner, fearing that dogs may misinterpret help as interfering with the bond, causing then to react violently. Our findings agree with those of others (Kahn Jr. et al. 2002) in that there is a progression of attributes that humans ascribe to robots like Aibo. This progression starts from Essence, and advances to Agency, Social Standing and Moral Standing. Our findings are that Aibo fulfills biological animistic underpinnings (children refer to its tail, legs, ears and behaviors in the same way as for living dogs). It also fulfills Agency properties (children attribute intentions, feelings , emotional states, wishes, desires, goals). We left aside social standing in our methodology, but strongly suspect that children attribute an emotional connection and companionship to Aibo. We observed a clear preference among children for `Do you want to pat the dog/puppy?' over `Do you want to touch the robot?'. Many children made unsolicited comments about how similar it was playing with Aibo to playing with their dog at home. Similarly, we refrained from exploring children's attribution of moral standing to Aibo (for example, should Aibo be punished for doing something wrong). Nevertheless, we received unsolicited suggestions that `leaving Aibo alone or not playing with him would make him sad' and that `batteries should always be charged, which may mean more responsibility than for a living dog'. These types of comments do attribute some rights to Aibo and a sense that it also deserves some respect. Our observations indicate that Essence and Agency are maintained in the child's beliefs even in the presence and practical illustration of other machines for which they will not typically attach such biological or animistic properties, nor psychological characteristics (although Bob the Builder's cars and machines talk). In fact, we witnessed arguments and debates among the children which turned the balance the other way around, some managing to convince others that Hank had feelings like `being afraid of the dark, because afraid is a feeling'. Also, we found observations that concur with the writing of anthropologist S. Turkle (Turkle 1999). In particular, although we did not intend to observe adults, we witnessed parents and teachers attempting to convince the children that Aibo was a machine and not a dog. Some child-carers seem to interpret our experiment as a lecture on the living versus the non-living. We believe this reflected some of Turkle's conclusions about the `thinking about alive-ness' with older people interpreting machines and computers through mechanistic/physical interpretations while the newer generation interprets beings in computer games and robots as `sort of alive'. Our best example of this was witnessing a parent selecting a particular physical argument to convince her 5-year old of `the clear difference' why Aibo is not a dog. This also pointed out a difference between Aibo and dogs that we had not observed but that the adult believed made "the difference". We attempt to illustrate it with Figure 7. Aibo has one less joint in the back leg than a dog (the dog, as shown in Figure 7(a), has hip (1), knee (2), ankle (3) and toes (4)). This is one degree of freedom less and also the toes bend back in the dog, while they do not on Aibo. Note that if we were to choose a physical argument it is perhaps more obvious that Aibo does not have two eyes or does not have a wet nose. The point is that a basic minimum of physical structure is enough to engage children in a psychological/conceptual interpretation that then is hard to remove on the basis of physical evidence. We believe our results indicate that children are 12 (a) (b) Figure 7: A dog (a) has one more degree of freedom per leg than Aibo (b) and has more movement in the toes than Aibo. quickly attached to the notion that `robotic dogs' are closer to living dogs. Although we would not go as far as S. Turkle to suggest that `living' has a new meaning for this generation of children, we suggest that they will see them as robotic pets more than canine machines. We expect, therefore, that in the future, humans will adopt more of them as an interface for human-computer interaction. Prof. B. Shneiderman is probably the world's leading authority in Human-Computer Interaction. He has repeatedly been outspoken about reducing `ma-chine intelligence' and `software agents' for building computers that are more effective tools (Beardsley 1999). At first, our research seems to contradict some of his ideas; but, interaction with a robot is interaction with a computer and we agree that it allows for direct manipulation, even more realistic and perhaps more meaningful than on the computer screen. Also, it is now clear that domestic robots will soon be around us and computers will not be restricted to output devices like monitors, nor will computers be confined to fixed locations. Third, we argue that studies such as ours advance the possibilities of having a `controllable, consistent and predictable interaction' with a robot. Thus, we share the vision of interaction facilitated by proper design. Finally, our aim is interaction with people who are blind. In such case, visualization (the coloring of pixels in a monitor) for `insight' cannot be used. Shneiderman also agrees on this point. We argue that properly designed robots will offer a multi-modal interface where insight is com-municated by embodiment and movement as well as sound. Other papers in the literature confirm that people may develop strong attachments, and even affectionate relationships with artificial information systems. Those studies involve human adults on one side and rather simple emulations of human intelligence in the other. In such cases, the interface has been rather simple (or at least not multi-modal), like through a phone conversation. It is interesting that this may have both positive and negative outcomes. For example , as reported in the case of a `Health Behavior Ad-visor System' (Kaplan, Farzanfar & Friedman 1999), some patients felt motivated to follow a healthier life style, while others found it inflicted a sense of guilt that did not motivate healthier habits. We believe that understanding people's expectations for robots is important since these expectations will define the context for the interactions that may result in effective use of robotic technology. An example is the potential attribution of moral standing to robots. This could eventually regard the robot (and not its manufacturer ) as responsible for its actions. Certainly, this would have many implications for our society. Acknowledgments The authors wish to thank the anonymous referees for the constructive feedback provided in their reviews. This work was supported by a Griffith University Research Grant as part of the project "Robots to guide the blind - the application of physical collaborative autonomous intelligent agents". References Aylett, B. (2002), ROBOTS -- Bringing intelligent machines to life?, ABC Books, Sydney NSW, Australia. Beardsley, T. (1999), `Humans unite!', Scientific American March, 3536. Profile Column. Brooks, R. (2002), `Humanoid robots', Communications of the ACM 45(3), 3338. Bumby, K. & Dautenhahn, K. (1999), Investigating children's attitudes towards robots: A case study, in `Proceedings of the Third Cognitive Technology Conference, CT'99', M.I.N.D. Lab, Michigan State University, East Lansing, MI., pp. 391410. Fong, T., Nourbakhsh, I. & Dautenhahn, K. (2003), `A survey of socially interactive robots', Robotics and Autonomous Systems 42, 235243. Kahn Jr., P., Friedman, B. & Hagman, J. (2002), I care about him as a pal: Conceptions of robotic pets in online Aibo discussion forum, in `Proceedings of CHI, Interactive Poster: Fun changing the world, changing ourselves', pp. 632633. Kahn Jr., P. J., Friedman, B., Freier, N. & Severson, R. (2003), Coding manual for children's interactions with Aibo, the robotic dog -- the preschool study, Technical Report UW CSE 03-04-03, Department of Computer Science and Engineering, University of Washington, Seattle, US. Kahney, L. (2003), `The new pet craze: Robovacs', Wired Magazine. June, 16th; visited Septenber 10th, 2003, www.wired.com/news/technology/0,1282,59249,00.html. Kaplan, B., Farzanfar, R. & Friedman, R. (1999), Ethnographic interviews to elicit patients, reactions to an intelligent interactive telephone health behavior advisor system, in M. N.M. Lorenzi, Bethesda, ed., `Proceedings: AMIA Symposiu', American Medical Informatics Association, www.amia.org/pubs/symposia/D005604.PDF. Knudsen, J. (1999), The Unofficial Guide to LEGO MINDSTORM Robots, O'Reilly, Sebastopol, CA. 13 Lawry, J., Upitis, R., Klawe, M., Anderson, A., Inkpen, K., Ndunda, M., Hsu, D., Leroux, S. & Sedighian, K. (1995), `Exploring common conceptions about boys and electronic games', Journal of Computer in Math and Science Teaching 14 (4), 439459. Scheeff, M., Pinto, J., Rahardja, K., Snibbe, S. & Tow, R. (2000), Experiences with Sparky: A social robot, in `Proceedings of the Workshop on Interactive Robot Entertainment'. Thrun, S., Bennewitz, M., Burgard, W., Cremers, A., Dellaert, F., Fox, D., Hahnel, D., Rosenberg, C., Roy, N., Schulte, J. & Schulz, D. (1999), MINERVA : A tour-guide robot that learns, in `KI Kunstliche Intelligenz', pp. 1426. Turkle, S. (1999), What are we thinking about when we are thinking about computers?, in M. Biagi-oli , ed., `The Science Studies Reader', Routledge, New York. Veloso, M., Uther, W., Fujita, M., Asada, M. & Kitano, H. (1998), Playing soccer with legged robots, in `In Proceedings of IROS-98, Intelligent Robots and Systems Conference', Victoria, Canada. 14

intelligent creatures;human attitudes;language;essence;agency;robots;perceived attitude;behavioral science;zoo-morphological autonomous mobile robots;robot attributes;multi-modal interfaces;interaction;feelings;hci

74

Downloading Textual Hidden Web Content Through Keyword Queries

An ever-increasing amount of information on the Web today is available only through search interfaces: the users have to type in a set of keywords in a search form in order to access the pages from certain Web sites. These pages are often referred to as the Hidden Web or the Deep Web. Since there are no static links to the Hidden Web pages, search engines cannot discover and index such pages and thus do not return them in the results. However, according to recent studies, the content provided by many Hidden Web sites is often of very high quality and can be extremely valuable to many users. In this paper, we study how we can build an effective Hidden Web crawler that can autonomously discover and download pages from the Hidden Web. Since the only "entry point" to a Hidden Web site is a query interface, the main challenge that a Hidden Web crawler has to face is how to automatically generate meaningful queries to issue to the site. Here, we provide a theoretical framework to investigate the query generation problem for the Hidden Web and we propose effective policies for generating queries automatically. Our policies proceed iteratively, issuing a different query in every iteration . We experimentally evaluate the effectiveness of these policies on 4 real Hidden Web sites and our results are very promising. For instance, in one experiment, one of our policies downloaded more than 90% of a Hidden Web site (that contains 14 million documents ) after issuing fewer than 100 queries.

INTRODUCTION Recent studies show that a significant fraction of Web content cannot be reached by following links [7, 12]. In particular, a large part of the Web is "hidden" behind search forms and is reachable only when users type in a set of keywords, or queries, to the forms. These pages are often referred to as the Hidden Web [17] or the 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. JCDL'05, June 711, 2005, Denver, Colorado, USA. Copyright 2005 ACM 1-58113-876-8/05/0006 ... $ 5.00. Deep Web [7], because search engines typically cannot index the pages and do not return them in their results (thus, the pages are essentially "hidden" from a typical Web user). According to many studies, the size of the Hidden Web increases rapidly as more organizations put their valuable content online through an easy-to-use Web interface [7]. In [12], Chang et al. estimate that well over 100,000 Hidden-Web sites currently exist on the Web. Moreover, the content provided by many Hidden-Web sites is often of very high quality and can be extremely valuable to many users [7]. For example, PubMed hosts many high-quality papers on medical research that were selected from careful peer-review processes, while the site of the US Patent and Trademarks Office 1 makes existing patent documents available, helping potential inventors examine "prior art." In this paper, we study how we can build a Hidden-Web crawler 2 that can automatically download pages from the Hidden Web, so that search engines can index them. Conventional crawlers rely on the hyperlinks on the Web to discover pages, so current search engines cannot index the Hidden-Web pages (due to the lack of links). We believe that an effective Hidden-Web crawler can have a tremendous impact on how users search information on the Web: Tapping into unexplored information: The Hidden-Web crawler will allow an average Web user to easily explore the vast amount of information that is mostly "hidden" at present. Since a majority of Web users rely on search engines to discover pages, when pages are not indexed by search engines, they are unlikely to be viewed by many Web users. Unless users go directly to Hidden-Web sites and issue queries there, they cannot access the pages at the sites. Improving user experience: Even if a user is aware of a number of Hidden-Web sites, the user still has to waste a significant amount of time and effort, visiting all of the potentially relevant sites, querying each of them and exploring the result. By making the Hidden-Web pages searchable at a central location, we can significantly reduce the user's wasted time and effort in searching the Hidden Web. Reducing potential bias: Due to the heavy reliance of many Web users on search engines for locating information, search engines influence how the users perceive the Web [28]. Users do not necessarily perceive what actually exists on the Web, but what is indexed by search engines [28]. According to a recent article [5], several organizations have recognized the importance of bringing information of their Hidden Web sites onto the surface, and committed considerable resources towards this effort. Our 1 US Patent Office: http://www.uspto.gov 2 Crawlers are the programs that traverse the Web automatically and download pages for search engines. 100 Figure 1: A single-attribute search interface Hidden-Web crawler attempts to automate this process for Hidden Web sites with textual content, thus minimizing the associated costs and effort required. Given that the only "entry" to Hidden Web pages is through querying a search form, there are two core challenges to implementing an effective Hidden Web crawler: (a) The crawler has to be able to understand and model a query interface, and (b) The crawler has to come up with meaningful queries to issue to the query interface. The first challenge was addressed by Raghavan and Garcia-Molina in [29], where a method for learning search interfaces was presented. Here, we present a solution to the second challenge, i.e. how a crawler can automatically generate queries so that it can discover and download the Hidden Web pages. Clearly, when the search forms list all possible values for a query (e.g., through a drop-down list), the solution is straightforward. We exhaustively issue all possible queries, one query at a time. When the query forms have a "free text" input, however, an infinite number of queries are possible, so we cannot exhaustively issue all possible queries. In this case, what queries should we pick? Can the crawler automatically come up with meaningful queries without understanding the semantics of the search form? In this paper, we provide a theoretical framework to investigate the Hidden-Web crawling problem and propose effective ways of generating queries automatically. We also evaluate our proposed solutions through experiments conducted on real Hidden-Web sites. In summary, this paper makes the following contributions: We present a formal framework to study the problem of Hidden-Web crawling. (Section 2). We investigate a number of crawling policies for the Hidden Web, including the optimal policy that can potentially download the maximum number of pages through the minimum number of interactions. Unfortunately, we show that the optimal policy is NP-hard and cannot be implemented in practice (Section 2.2). We propose a new adaptive policy that approximates the optimal policy. Our adaptive policy examines the pages returned from previous queries and adapts its query-selection policy automatically based on them (Section 3). We evaluate various crawling policies through experiments on real Web sites. Our experiments will show the relative advantages of various crawling policies and demonstrate their potential . The results from our experiments are very promising. In one experiment, for example, our adaptive policy downloaded more than 90% of the pages within PubMed (that contains 14 million documents) after it issued fewer than 100 queries. FRAMEWORK In this section, we present a formal framework for the study of the Hidden-Web crawling problem. In Section 2.1, we describe our assumptions on Hidden-Web sites and explain how users interact with the sites. Based on this interaction model, we present a high-level algorithm for a Hidden-Web crawler in Section 2.2. Finally in Section 2.3, we formalize the Hidden-Web crawling problem. 2.1 Hidden-Web database model There exists a variety of Hidden Web sources that provide information on a multitude of topics. Depending on the type of information , we may categorize a Hidden-Web site either as a textual database or a structured database. A textual database is a site that Figure 2: A multi-attribute search interface mainly contains plain-text documents, such as PubMed and Lexis-Nexis (an online database of legal documents [1]). Since plain-text documents do not usually have well-defined structure, most textual databases provide a simple search interface where users type a list of keywords in a single search box (Figure 1). In contrast , a structured database often contains multi-attribute relational data (e.g., a book on the Amazon Web site may have the fields title=`Harry Potter' , author=`J.K. Rowling' and isbn=`0590353403' ) and supports multi-attribute search interfaces (Figure 2). In this paper, we will mainly focus on textual databases that support single-attribute keyword queries. We discuss how we can extend our ideas for the textual databases to multi-attribute structured databases in Section 6.1. Typically, the users need to take the following steps in order to access pages in a Hidden-Web database: 1. Step 1. First, the user issues a query, say "liver," through the search interface provided by the Web site (such as the one shown in Figure 1). 2. Step 2. Shortly after the user issues the query, she is presented with a result index page. That is, the Web site returns a list of links to potentially relevant Web pages, as shown in Figure 3(a). 3. Step 3. From the list in the result index page, the user identifies the pages that look "interesting" and follows the links. Clicking on a link leads the user to the actual Web page, such as the one shown in Figure 3(b), that the user wants to look at. 2.2 A generic Hidden Web crawling algorithm Given that the only "entry" to the pages in a Hidden-Web site is its search from, a Hidden-Web crawler should follow the three steps described in the previous section. That is, the crawler has to generate a query, issue it to the Web site, download the result index page, and follow the links to download the actual pages. In most cases, a crawler has limited time and network resources, so the crawler repeats these steps until it uses up its resources. In Figure 4 we show the generic algorithm for a Hidden-Web crawler. For simplicity, we assume that the Hidden-Web crawler issues single-term queries only. 3 The crawler first decides which query term it is going to use (Step (2)), issues the query, and retrieves the result index page (Step (3)). Finally, based on the links found on the result index page, it downloads the Hidden Web pages from the site (Step (4)). This same process is repeated until all the available resources are used up (Step (1)). Given this algorithm, we can see that the most critical decision that a crawler has to make is what query to issue next. If the crawler can issue successful queries that will return many matching pages, the crawler can finish its crawling early on using minimum resources. In contrast, if the crawler issues completely irrelevant queries that do not return any matching pages, it may waste all of its resources simply issuing queries without ever retrieving actual pages. Therefore, how the crawler selects the next query can greatly affect its effectiveness. In the next section, we formalize this query selection problem. 3 For most Web sites that assume "AND" for multi-keyword queries, single-term queries return the maximum number of results. Extending our work to multi-keyword queries is straightforward. 101 (a) List of matching pages for query "liver". (b) The first matching page for "liver". Figure 3: Pages from the PubMed Web site. A LGORITHM 2.1. Crawling a Hidden Web site Procedure (1) while ( there are available resources ) do // select a term to send to the site (2) q i = SelectTerm() // send query and acquire result index page (3) R(q i ) = QueryWebSite( q i ) // download the pages of interest (4) Download( R(q i ) ) (5) done Figure 4: Algorithm for crawling a Hidden Web site. S q 1 q q q 2 3 4 Figure 5: A set-formalization of the optimal query selection problem. 2.3 Problem formalization Theoretically, the problem of query selection can be formalized as follows: We assume that the crawler downloads pages from a Web site that has a set of pages S (the rectangle in Figure 5). We represent each Web page in S as a point (dots in Figure 5). Every potential query q i that we may issue can be viewed as a subset of S, containing all the points (pages) that are returned when we issue q i to the site. Each subset is associated with a weight that represents the cost of issuing the query. Under this formalization, our goal is to find which subsets (queries) cover the maximum number of points (Web pages) with the minimum total weight (cost). This problem is equivalent to the set-covering problem in graph theory [16]. There are two main difficulties that we need to address in this formalization. First, in a practical situation, the crawler does not know which Web pages will be returned by which queries, so the subsets of S are not known in advance. Without knowing these subsets the crawler cannot decide which queries to pick to maximize the coverage. Second, the set-covering problem is known to be NP-Hard [16], so an efficient algorithm to solve this problem optimally in polynomial time has yet to be found. In this paper, we will present an approximation algorithm that can find a near-optimal solution at a reasonable computational cost. Our algorithm leverages the observation that although we do not know which pages will be returned by each query q i that we issue, we can predict how many pages will be returned. Based on this information our query selection algorithm can then select the "best" queries that cover the content of the Web site. We present our prediction method and our query selection algorithm in Section 3. 2.3.1 Performance Metric Before we present our ideas for the query selection problem, we briefly discuss some of our notation and the cost/performance metrics . Given a query q i , we use P (q i ) to denote the fraction of pages that we will get back if we issue query q i to the site. For example, if a Web site has 10,000 pages in total, and if 3,000 pages are returned for the query q i = "medicine", then P (q i ) = 0.3. We use P (q 1 q 2 ) to represent the fraction of pages that are returned from both q 1 and q 2 (i.e., the intersection of P (q 1 ) and P (q 2 )). Similarly, we use P (q 1 q 2 ) to represent the fraction of pages that are returned from either q 1 or q 2 (i.e., the union of P (q 1 ) and P (q 2 )). We also use Cost (q i ) to represent the cost of issuing the query q i . Depending on the scenario, the cost can be measured either in time, network bandwidth, the number of interactions with the site, or it can be a function of all of these. As we will see later, our proposed algorithms are independent of the exact cost function. In the most common case, the query cost consists of a number of factors, including the cost for submitting the query to the site, retrieving the result index page (Figure 3(a)) and downloading the actual pages (Figure 3(b)). We assume that submitting a query incurs a fixed cost of c q . The cost for downloading the result index page is proportional to the number of matching documents to the query, while the cost c d for downloading a matching document is also fixed. Then the overall cost of query q i is Cost (q i ) = c q + c r P (q i ) + c d P (q i ). (1) In certain cases, some of the documents from q i may have already been downloaded from previous queries. In this case, the crawler may skip downloading these documents and the cost of q i can be Cost (q i ) = c q + c r P (q i ) + c d P new (q i ). (2) Here, we use P new (q i ) to represent the fraction of the new documents from q i that have not been retrieved from previous queries. Later in Section 3.1 we will study how we can estimate P (q i ) and P new (q i ) to estimate the cost of q i . Since our algorithms are independent of the exact cost function, we will assume a generic cost function Cost (q i ) in this paper. When we need a concrete cost function, however, we will use Equation 2. Given the notation, we can formalize the goal of a Hidden-Web crawler as follows: 102 P ROBLEM 1. Find the set of queries q 1 , . . . , q n that maximizes P (q 1 q n ) under the constraint n i=1 Cost (q i ) t. Here, t is the maximum download resource that the crawler has. KEYWORD SELECTION How should a crawler select the queries to issue? Given that the goal is to download the maximum number of unique documents from a textual database, we may consider one of the following options : Random: We select random keywords from, say, an English dictionary and issue them to the database. The hope is that a random query will return a reasonable number of matching documents. Generic-frequency: We analyze a generic document corpus collected elsewhere (say, from the Web) and obtain the generic frequency distribution of each keyword. Based on this generic distribution , we start with the most frequent keyword, issue it to the Hidden-Web database and retrieve the result. We then continue to the second-most frequent keyword and repeat this process until we exhaust all download resources. The hope is that the frequent keywords in a generic corpus will also be frequent in the Hidden-Web database, returning many matching documents. Adaptive: We analyze the documents returned from the previous queries issued to the Hidden-Web database and estimate which keyword is most likely to return the most documents. Based on this analysis, we issue the most "promising" query, and repeat the process. Among these three general policies, we may consider the random policy as the base comparison point since it is expected to perform the worst. Between the generic-frequency and the adaptive policies, both policies may show similar performance if the crawled database has a generic document collection without a specialized topic. The adaptive policy, however, may perform significantly better than the generic-frequency policy if the database has a very specialized collection that is different from the generic corpus. We will experimentally compare these three policies in Section 4. While the first two policies (random and generic-frequency policies ) are easy to implement, we need to understand how we can analyze the downloaded pages to identify the most "promising" query in order to implement the adaptive policy. We address this issue in the rest of this section. 3.1 Estimating the number of matching pages In order to identify the most promising query, we need to estimate how many new documents we will download if we issue the query q i as the next query. That is, assuming that we have issued queries q 1 , . . . , q i-1 we need to estimate P (q 1 q i-1 q i ), for every potential next query q i and compare this value. In estimating this number, we note that we can rewrite P (q 1 q i-1 q i ) as: P ((q 1 q i-1 ) q i ) = P (q 1 q i-1 ) + P (q i ) - P ((q 1 q i-1 ) q i ) = P (q 1 q i-1 ) + P (q i ) - P (q 1 q i-1 )P (q i |q 1 q i-1 ) (3) In the above formula, note that we can precisely measure P (q 1 q i-1 ) and P (q i | q 1 q i-1 ) by analyzing previously-downloaded pages: We know P (q 1 q i-1 ), the union of all pages downloaded from q 1 , . . . , q i-1 , since we have already issued q 1 , . . . , q i-1 and downloaded the matching pages. 4 We can also measure P (q i | q 1 q i-1 ), the probability that q i appears in the pages from q 1 , . . . , q i-1 , by counting how many times q i appears in the pages from q 1 , . . . , q i-1 . Therefore, we only need to estimate P (q i ) to evaluate P (q 1 q i ). We may consider a number of different ways to estimate P (q i ), including the following : 1. Independence estimator: We assume that the appearance of the term q i is independent of the terms q 1 , . . . , q i-1 . That is, we assume that P (q i ) = P (q i |q 1 q i-1 ). 2. Zipf estimator: In [19], Ipeirotis et al. proposed a method to estimate how many times a particular term occurs in the entire corpus based on a subset of documents from the corpus. Their method exploits the fact that the frequency of terms inside text collections follows a power law distribution [30, 25]. That is, if we rank all terms based on their occurrence frequency (with the most frequent term having a rank of 1, second most frequent a rank of 2 etc.), then the frequency f of a term inside the text collection is given by: f = (r + ) (4) where r is the rank of the term and , , and are constants that depend on the text collection. Their main idea is (1) to estimate the three parameters, , and , based on the subset of documents that we have downloaded from previous queries, and (2) use the estimated parameters to predict f given the ranking r of a term within the subset. For a more detailed description on how we can use this method to estimate P (q i ), we refer the reader to the extended version of this paper [27]. After we estimate P (q i ) and P (q i |q 1 q i-1 ) values, we can calculate P (q 1 q i ). In Section 3.3, we explain how we can efficiently compute P (q i |q 1 q i-1 ) by maintaining a succinct summary table. In the next section, we first examine how we can use this value to decide which query we should issue next to the Hidden Web site. 3.2 Query selection algorithm The goal of the Hidden-Web crawler is to download the maximum number of unique documents from a database using its limited download resources. Given this goal, the Hidden-Web crawler has to take two factors into account. (1) the number of new documents that can be obtained from the query q i and (2) the cost of issuing the query q i . For example, if two queries, q i and q j , incur the same cost, but q i returns more new pages than q j , q i is more desirable than q j . Similarly, if q i and q j return the same number of new documents, but q i incurs less cost then q j , q i is more desirable . Based on this observation, the Hidden-Web crawler may use the following efficiency metric to quantify the desirability of the query q i : Efficiency (q i ) = P new (q i ) Cost (q i ) Here, P new (q i ) represents the amount of new documents returned for q i (the pages that have not been returned for previous queries). Cost (q i ) represents the cost of issuing the query q i . Intuitively, the efficiency of q i measures how many new documents are retrieved per unit cost, and can be used as an indicator of 4 For exact estimation, we need to know the total number of pages in the site. However, in order to compare only relative values among queries, this information is not actually needed. 103 A LGORITHM 3.1. Greedy SelectTerm() Parameters: T : The list of potential query keywords Procedure (1) Foreach t k in T do (2) Estimate Efficiency (t k ) = P new (t k ) Cost(t k ) (3) done (4) return t k with maximum Efficiency (t k ) Figure 6: Algorithm for selecting the next query term. how well our resources are spent when issuing q i . Thus, the Hidden Web crawler can estimate the efficiency of every candidate q i , and select the one with the highest value. By using its resources more efficiently, the crawler may eventually download the maximum number of unique documents. In Figure 6, we show the query selection function that uses the concept of efficiency. In principle, this algorithm takes a greedy approach and tries to maximize the "potential gain" in every step. We can estimate the efficiency of every query using the estimation method described in Section 3.1. That is, the size of the new documents from the query q i , P new (q i ), is P new (q i ) = P (q 1 q i-1 q i ) - P (q 1 q i-1 ) = P (q i ) - P (q 1 q i-1 )P (q i |q 1 q i-1 ) from Equation 3, where P (q i ) can be estimated using one of the methods described in section 3. We can also estimate Cost (q i ) sim-ilarly . For example, if Cost (q i ) is Cost (q i ) = c q + c r P (q i ) + c d P new (q i ) (Equation 2), we can estimate Cost (q i ) by estimating P (q i ) and P new (q i ). 3.3 Efficient calculation of query statistics In estimating the efficiency of queries, we found that we need to measure P (q i |q 1 q i-1 ) for every potential query q i . This calculation can be very time-consuming if we repeat it from scratch for every query q i in every iteration of our algorithm. In this section, we explain how we can compute P (q i |q 1 q i-1 ) efficiently by maintaining a small table that we call a query statistics table. The main idea for the query statistics table is that P (q i |q 1 q i-1 ) can be measured by counting how many times the keyword q i appears within the documents downloaded from q 1 , . . . , q i-1 . We record these counts in a table, as shown in Figure 7(a). The left column of the table contains all potential query terms and the right column contains the number of previously-downloaded documents containing the respective term. For example, the table in Figure 7(a) shows that we have downloaded 50 documents so far, and the term model appears in 10 of these documents. Given this number , we can compute that P (model|q 1 q i-1 ) = 10 50 = 0.2. We note that the query statistics table needs to be updated whenever we issue a new query q i and download more documents. This update can be done efficiently as we illustrate in the following example . E XAMPLE 1. After examining the query statistics table of Figure 7(a), we have decided to use the term "computer" as our next query q i . From the new query q i = "computer," we downloaded 20 more new pages. Out of these, 12 contain the keyword "model" Term t k N (t k ) model 10 computer 38 digital 50 Term t k N (t k ) model 12 computer 20 disk 18 Total pages: 50 New pages: 20 (a) After q 1 , . . . , q i-1 (b) New from q i = computer Term t k N (t k ) model 10+12 = 22 computer 38+20 = 58 disk 0+18 = 18 digital 50+0 = 50 Total pages: 50 + 20 = 70 (c) After q 1 , . . . , q i Figure 7: Updating the query statistics table. q i 1 i-1 q \/ ... \/ q q i / S Figure 8: A Web site that does not return all the results. and 18 the keyword "disk." The table in Figure 7(b) shows the frequency of each term in the newly-downloaded pages. We can update the old table (Figure 7(a)) to include this new information by simply adding corresponding entries in Figures 7(a) and (b). The result is shown on Figure 7(c). For example, keyword "model" exists in 10 + 12 = 22 pages within the pages retrieved from q 1 , . . . , q i . According to this new table, P (model|q 1 q i ) is now 22 70 = 0.3. 3.4 Crawling sites that limit the number of results In certain cases, when a query matches a large number of pages, the Hidden Web site returns only a portion of those pages. For example , the Open Directory Project [2] allows the users to see only up to 10, 000 results after they issue a query. Obviously, this kind of limitation has an immediate effect on our Hidden Web crawler. First, since we can only retrieve up to a specific number of pages per query, our crawler will need to issue more queries (and potentially will use up more resources) in order to download all the pages. Second, the query selection method that we presented in Section 3.2 assumes that for every potential query q i , we can find P (q i |q 1 q i-1 ). That is, for every query q i we can find the fraction of documents in the whole text database that contains q i with at least one of q 1 , . . . , q i-1 . However, if the text database returned only a portion of the results for any of the q 1 , . . . , q i-1 then the value P (q i |q 1 q i-1 ) is not accurate and may affect our decision for the next query q i , and potentially the performance of our crawler. Since we cannot retrieve more results per query than the maximum number the Web site allows, our crawler has no other choice besides submitting more queries. However, there is a way to estimate the correct value for P (q i |q 1 q i-1 ) in the case where the Web site returns only a portion of the results. 104 Again, assume that the Hidden Web site we are currently crawling is represented as the rectangle on Figure 8 and its pages as points in the figure. Assume that we have already issued queries q 1 , . . . , q i-1 which returned a number of results less than the maximum number than the site allows, and therefore we have downloaded all the pages for these queries (big circle in Figure 8). That is, at this point, our estimation for P (q i |q 1 q i-1 ) is accurate. Now assume that we submit query q i to the Web site, but due to a limitation in the number of results that we get back, we retrieve the set q i (small circle in Figure 8) instead of the set q i (dashed circle in Figure 8). Now we need to update our query statistics table so that it has accurate information for the next step. That is, although we got the set q i back, for every potential query q i+1 we need to find P (q i+1 |q 1 q i ): P (q i+1 |q 1 q i ) = 1 P (q 1 q i ) [P (q i+1 (q 1 q i-1 ))+ P (q i+1 q i ) - P (q i+1 q i (q 1 q i-1 ))] (5) In the previous equation, we can find P (q 1 q i ) by estimating P (q i ) with the method shown in Section 3. Additionally, we can calculate P (q i+1 (q 1 q i-1 )) and P (q i+1 q i (q 1 q i-1 )) by directly examining the documents that we have downloaded from queries q 1 , . . . , q i-1 . The term P (q i+1 q i ) however is unknown and we need to estimate it. Assuming that q i is a random sample of q i , then: P (q i+1 q i ) P (q i+1 q i ) = P (q i ) P (q i ) (6) From Equation 6 we can calculate P (q i+1 q i ) and after we replace this value to Equation 5 we can find P (q i+1 |q 1 q i ). EXPERIMENTAL EVALUATION In this section we experimentally evaluate the performance of the various algorithms for Hidden Web crawling presented in this paper. Our goal is to validate our theoretical analysis through real-world experiments, by crawling popular Hidden Web sites of textual databases. Since the number of documents that are discovered and downloaded from a textual database depends on the selection of the words that will be issued as queries 5 to the search interface of each site, we compare the various selection policies that were described in section 3, namely the random, generic-frequency, and adaptive algorithms. The adaptive algorithm learns new keywords and terms from the documents that it downloads, and its selection process is driven by a cost model as described in Section 3.2. To keep our experiment and its analysis simple at this point, we will assume that the cost for every query is constant. That is, our goal is to maximize the number of downloaded pages by issuing the least number of queries. Later, in Section 4.4 we will present a comparison of our policies based on a more elaborate cost model. In addition, we use the independence estimator (Section 3.1) to estimate P (q i ) from downloaded pages. Although the independence estimator is a simple estimator, our experiments will show that it can work very well in practice. 6 For the generic-frequency policy, we compute the frequency distribution of words that appear in a 5.5-million-Web-page corpus 5 Throughout our experiments, once an algorithm has submitted a query to a database, we exclude the query from subsequent submissions to the same database from the same algorithm. 6 We defer the reporting of results based on the Zipf estimation to a future work. downloaded from 154 Web sites of various topics [26]. Keywords are selected based on their decreasing frequency with which they appear in this document set, with the most frequent one being selected first, followed by the second-most frequent keyword, etc. 7 Regarding the random policy, we use the same set of words collected from the Web corpus, but in this case, instead of selecting keywords based on their relative frequency, we choose them ran-domly (uniform distribution). In order to further investigate how the quality of the potential query-term list affects the random-based algorithm, we construct two sets: one with the 16, 000 most frequent words of the term collection used in the generic-frequency policy (hereafter, the random policy with the set of 16,000 words will be referred to as random-16K), and another set with the 1 million most frequent words of the same collection as above (hereafter, referred to as random-1M). The former set has frequent words that appear in a large number of documents (at least 10, 000 in our collection ), and therefore can be considered of "high-quality" terms. The latter set though contains a much larger collection of words, among which some might be bogus, and meaningless. The experiments were conducted by employing each one of the aforementioned algorithms (adaptive, generic-frequency, random-16K , and random-1M) to crawl and download contents from three Hidden Web sites: The PubMed Medical Library, 8 Amazon, 9 and the Open Directory Project[2]. According to the information on PubMed's Web site, its collection contains approximately 14 million abstracts of biomedical articles. We consider these abstracts as the "documents" in the site, and in each iteration of the adaptive policy, we use these abstracts as input to the algorithm. Thus our goal is to "discover" as many unique abstracts as possible by repeat-edly querying the Web query interface provided by PubMed. The Hidden Web crawling on the PubMed Web site can be considered as topic-specific, due to the fact that all abstracts within PubMed are related to the fields of medicine and biology. In the case of the Amazon Web site, we are interested in downloading all the hidden pages that contain information on books. The querying to Amazon is performed through the Software De-veloper's Kit that Amazon provides for interfacing to its Web site, and which returns results in XML form. The generic "keyword" field is used for searching, and as input to the adaptive policy we extract the product description and the text of customer reviews when present in the XML reply. Since Amazon does not provide any information on how many books it has in its catalogue, we use random sampling on the 10-digit ISBN number of the books to estimate the size of the collection. Out of the 10, 000 random ISBN numbers queried, 46 are found in the Amazon catalogue, therefore the size of its book collection is estimated to be 46 10000 10 10 = 4.6 million books. It's also worth noting here that Amazon poses an upper limit on the number of results (books in our case) returned by each query, which is set to 32, 000. As for the third Hidden Web site, the Open Directory Project (hereafter also referred to as dmoz), the site maintains the links to 3.8 million sites together with a brief summary of each listed site. The links are searchable through a keyword-search interface. We consider each indexed link together with its brief summary as the document of the dmoz site, and we provide the short summaries to the adaptive algorithm to drive the selection of new keywords for querying. On the dmoz Web site, we perform two Hidden Web crawls: the first is on its generic collection of 3.8-million indexed 7 We did not manually exclude stop words (e.g., the, is, of, etc.) from the keyword list. As it turns out, all Web sites except PubMed return matching documents for the stop words, such as "the." 8 PubMed Medical Library: http://www.pubmed.org 9 Amazon Inc.: http://www.amazon.com 105 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 50 100 150 200 fraction of documents query number Cumulative fraction of unique documents - PubMed website adaptive generic-frequency random-16K random-1M Figure 9: Coverage of policies for Pubmed 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 100 200 300 400 500 600 700 fraction of documents query number Cumulative fraction of unique documents - Amazon website adaptive generic-frequency random-16K random-1M Figure 10: Coverage of policies for Amazon sites, regardless of the category that they fall into. The other crawl is performed specifically on the Arts section of dmoz (http:// dmoz.org/Arts ), which comprises of approximately 429, 000 indexed sites that are relevant to Arts, making this crawl topic-specific , as in PubMed. Like Amazon, dmoz also enforces an upper limit on the number of returned results, which is 10, 000 links with their summaries. 4.1 Comparison of policies The first question that we seek to answer is the evolution of the coverage metric as we submit queries to the sites. That is, what fraction of the collection of documents stored in the Hidden Web site can we download as we continuously query for new words selected using the policies described above? More formally, we are interested in the value of P (q 1 q i-1 q i ), after we submit q 1 , . . . , q i queries, and as i increases. In Figures 9, 10, 11, and 12 we present the coverage metric for each policy, as a function of the query number, for the Web sites of PubMed, Amazon, general dmoz and the art-specific dmoz, respectively . On the y-axis the fraction of the total documents downloaded from the website is plotted, while the x-axis represents the query number. A first observation from these graphs is that in general , the generic-frequency and the adaptive policies perform much better than the random-based algorithms. In all of the figures, the graphs for the random-1M and the random-16K are significantly below those of other policies. Between the generic-frequency and the adaptive policies, we can see that the latter outperforms the former when the site is topic spe-0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 100 200 300 400 500 600 700 fraction of documents query number Cumulative fraction of unique documents - dmoz website adaptive generic-frequency random-16K random-1M Figure 11: Coverage of policies for general dmoz 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 50 100 150 200 250 300 350 400 450 fraction of documents query number Cumulative fraction of unique documents - dmoz/Arts website adaptive generic-frequency random-16K random-1M Figure 12: Coverage of policies for the Arts section of dmoz cific. For example, for the PubMed site (Figure 9), the adaptive algorithm issues only 83 queries to download almost 80% of the documents stored in PubMed, but the generic-frequency algorithm requires 106 queries for the same coverage,. For the dmoz/Arts crawl (Figure 12), the difference is even more substantial: the adaptive policy is able to download 99.98% of the total sites indexed in the Directory by issuing 471 queries, while the frequency-based algorithm is much less effective using the same number of queries, and discovers only 72% of the total number of indexed sites. The adaptive algorithm, by examining the contents of the pages that it downloads at each iteration, is able to identify the topic of the site as expressed by the words that appear most frequently in the result-set. Consequently, it is able to select words for subsequent queries that are more relevant to the site, than those preferred by the generic-frequency policy, which are drawn from a large, generic collection. Table 1 shows a sample of 10 keywords out of 211 chosen and submitted to the PubMed Web site by the adaptive algorithm, but not by the other policies. For each keyword, we present the number of the iteration, along with the number of results that it returned. As one can see from the table, these keywords are highly relevant to the topics of medicine and biology of the Public Medical Library, and match against numerous articles stored in its Web site. In both cases examined in Figures 9, and 12, the random-based policies perform much worse than the adaptive algorithm, and the generic-frequency. It is worthy noting however, that the random-based policy with the small, carefully selected set of 16, 000 "qual-ity" words manages to download a considerable fraction of 42.5% 106 Iteration Keyword Number of Results 23 department 2, 719, 031 34 patients 1, 934, 428 53 clinical 1, 198, 322 67 treatment 4, 034, 565 69 medical 1, 368, 200 70 hospital 503, 307 146 disease 1, 520, 908 172 protein 2, 620, 938 Table 1: Sample of keywords queried to PubMed exclusively by the adaptive policy from the PubMed Web site after 200 queries, while the coverage for the Arts section of dmoz reaches 22.7%, after 471 queried keywords . On the other hand, the random-based approach that makes use of the vast collection of 1 million words, among which a large number is bogus keywords, fails to download even a mere 1% of the total collection, after submitting the same number of query words. For the generic collections of Amazon and the dmoz sites, shown in Figures 10 and 11 respectively, we get mixed results: The generic-frequency policy shows slightly better performance than the adaptive policy for the Amazon site (Figure 10), and the adaptive method clearly outperforms the generic-frequency for the general dmoz site (Figure 11). A closer look at the log files of the two Hidden Web crawlers reveals the main reason: Amazon was functioning in a very flaky way when the adaptive crawler visited it, resulting in a large number of lost results. Thus, we suspect that the slightly poor performance of the adaptive policy is due to this experimental variance. We are currently running another experiment to verify whether this is indeed the case. Aside from this experimental variance, the Amazon result indicates that if the collection and the words that a Hidden Web site contains are generic enough, then the generic-frequency approach may be a good candidate algorithm for effective crawling. As in the case of topic-specific Hidden Web sites, the random-based policies also exhibit poor performance compared to the other two algorithms when crawling generic sites: for the Amazon Web site, random-16K succeeds in downloading almost 36.7% after issuing 775 queries, alas for the generic collection of dmoz, the fraction of the collection of links downloaded is 13.5% after the 770th query. Finally, as expected, random-1M is even worse than random-16K , downloading only 14.5% of Amazon and 0.3% of the generic dmoz. In summary, the adaptive algorithm performs remarkably well in all cases: it is able to discover and download most of the documents stored in Hidden Web sites by issuing the least number of queries. When the collection refers to a specific topic, it is able to identify the keywords most relevant to the topic of the site and consequently ask for terms that is most likely that will return a large number of results . On the other hand, the generic-frequency policy proves to be quite effective too, though less than the adaptive: it is able to retrieve relatively fast a large portion of the collection, and when the site is not topic-specific, its effectiveness can reach that of adaptive (e.g. Amazon). Finally, the random policy performs poorly in general, and should not be preferred. 4.2 Impact of the initial query An interesting issue that deserves further examination is whether the initial choice of the keyword used as the first query issued by the adaptive algorithm affects its effectiveness in subsequent itera-tions . The choice of this keyword is not done by the selection of the 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 fraction of documents query number Convergence of adaptive under different initial queries - PubMed website pubmed data information return Figure 13: Convergence of the adaptive algorithm using different initial queries for crawling the PubMed Web site adaptive algorithm itself and has to be manually set, since its query statistics tables have not been populated yet. Thus, the selection is generally arbitrary, so for purposes of fully automating the whole process, some additional investigation seems necessary. For this reason, we initiated three adaptive Hidden Web crawlers targeting the PubMed Web site with different seed-words: the word "data", which returns 1,344,999 results, the word "information" that reports 308, 474 documents, and the word "return" that retrieves 29, 707 pages, out of 14 million. These keywords represent varying degrees of term popularity in PubMed, with the first one being of high popularity, the second of medium, and the third of low. We also show results for the keyword "pubmed", used in the experiments for coverage of Section 4.1, and which returns 695 articles. As we can see from Figure 13, after a small number of queries, all four crawlers roughly download the same fraction of the collection, regardless of their starting point: Their coverages are roughly equivalent from the 25th query. Eventually, all four crawlers use the same set of terms for their queries, regardless of the initial query. In the specific experiment, from the 36th query onward , all four crawlers use the same terms for their queries in each iteration, or the same terms are used off by one or two query numbers . Our result confirms the observation of [11] that the choice of the initial query has minimal effect on the final performance. We can explain this intuitively as follows: Our algorithm approximates the optimal set of queries to use for a particular Web site. Once the algorithm has issued a significant number of queries, it has an accurate estimation of the content of the Web site, regardless of the initial query. Since this estimation is similar for all runs of the algorithm, the crawlers will use roughly the same queries. 4.3 Impact of the limit in the number of results While the Amazon and dmoz sites have the respective limit of 32,000 and 10,000 in their result sizes, these limits may be larger than those imposed by other Hidden Web sites. In order to investigate how a "tighter" limit in the result size affects the performance of our algorithms, we performed two additional crawls to the generic-dmoz site: we ran the generic-frequency and adaptive policies but we retrieved only up to the top 1,000 results for every query. In Figure 14 we plot the coverage for the two policies as a function of the number of queries. As one might expect, by comparing the new result in Figure 14 to that of Figure 11 where the result limit was 10,000, we conclude that the tighter limit requires a higher number of queries to achieve the same coverage. For example, when the result limit was 10,000, the adaptive pol-107 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 500 1000 1500 2000 2500 3000 3500 Fraction of Unique Pages Query Number Cumulative fraction of unique pages downloaded per query - Dmoz Web site (cap limit 1000) adaptive generic-frequency Figure 14: Coverage of general dmoz after limiting the number of results to 1,000 icy could download 70% of the site after issuing 630 queries, while it had to issue 2,600 queries to download 70% of the site when the limit was 1,000. On the other hand, our new result shows that even with a tight result limit, it is still possible to download most of a Hidden Web site after issuing a reasonable number of queries. The adaptive policy could download more than 85% of the site after issuing 3,500 queries when the limit was 1,000. Finally, our result shows that our adaptive policy consistently outperforms the generic-frequency policy regardless of the result limit. In both Figure 14 and Figure 11, our adaptive policy shows significantly larger coverage than the generic-frequency policy for the same number of queries. 4.4 Incorporating the document download cost For brevity of presentation, the performance evaluation results provided so far assumed a simplified cost-model where every query involved a constant cost. In this section we present results regarding the performance of the adaptive and generic-frequency algorithms using Equation 2 to drive our query selection process. As we discussed in Section 2.3.1, this query cost model includes the cost for submitting the query to the site, retrieving the result index page, and also downloading the actual pages. For these costs, we examined the size of every result in the index page and the sizes of the documents, and we chose c q = 100, c r = 100, and c d = 10000, as values for the parameters of Equation 2, and for the particular experiment that we ran on the PubMed website. The values that we selected imply that the cost for issuing one query and retrieving one result from the result index page are roughly the same, while the cost for downloading an actual page is 100 times larger. We believe that these values are reasonable for the PubMed Web site. Figure 15 shows the coverage of the adaptive and generic-frequency algorithms as a function of the resource units used during the download process. The horizontal axis is the amount of resources used, and the vertical axis is the coverage. As it is evident from the graph, the adaptive policy makes more efficient use of the available resources, as it is able to download more articles than the generic-frequency, using the same amount of resource units. However, the difference in coverage is less dramatic in this case, compared to the graph of Figure 9. The smaller difference is due to the fact that under the current cost metric, the download cost of documents constitutes a significant portion of the cost. Therefore, when both policies downloaded the same number of documents, the saving of the adaptive policy is not as dramatic as before. That 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 5000 10000 15000 20000 25000 30000 Fraction of Unique Pages Total Cost (c q =100, c r =100, c d =10000) Cumulative fraction of unique pages downloaded per cost unit - PubMed Web site adaptive frequency Figure 15: Coverage of PubMed after incorporating the document download cost is, the savings in the query cost and the result index download cost is only a relatively small portion of the overall cost. Still, we observe noticeable savings from the adaptive policy. At the total cost of 8000, for example, the coverage of the adaptive policy is roughly 0.5 while the coverage of the frequency policy is only 0.3. RELATED WORK In a recent study, Raghavan and Garcia-Molina [29] present an architectural model for a Hidden Web crawler. The main focus of this work is to learn Hidden-Web query interfaces, not to generate queries automatically. The potential queries are either provided manually by users or collected from the query interfaces. In contrast , our main focus is to generate queries automatically without any human intervention. The idea of automatically issuing queries to a database and examining the results has been previously used in different contexts. For example, in [10, 11], Callan and Connel try to acquire an accurate language model by collecting a uniform random sample from the database. In [22] Lawrence and Giles issue random queries to a number of Web Search Engines in order to estimate the fraction of the Web that has been indexed by each of them. In a similar fashion, Bharat and Broder [8] issue random queries to a set of Search Engines in order to estimate the relative size and overlap of their indexes. In [6], Barbosa and Freire experimentally evaluate methods for building multi-keyword queries that can return a large fraction of a document collection. Our work differs from the previous studies in two ways. First, it provides a theoretical framework for analyzing the process of generating queries for a database and examining the results, which can help us better understand the effectiveness of the methods presented in the previous work. Second, we apply our framework to the problem of Hidden Web crawling and demonstrate the efficiency of our algorithms. Cope et al. [15] propose a method to automatically detect whether a particular Web page contains a search form. This work is complementary to ours; once we detect search interfaces on the Web using the method in [15], we may use our proposed algorithms to download pages automatically from those Web sites. Reference [4] reports methods to estimate what fraction of a text database can be eventually acquired by issuing queries to the database. In [3] the authors study query-based techniques that can extract relational data from large text databases. Again, these works study orthogonal issues and are complementary to our work. In order to make documents in multiple textual databases searchable at a central place, a number of "harvesting" approaches have 108 been proposed (e.g., OAI [21], DP9 [24]). These approaches essentially assume cooperative document databases that willingly share some of their metadata and/or documents to help a third-party search engine to index the documents. Our approach assumes uncoop-erative databases that do not share their data publicly and whose documents are accessible only through search interfaces. There exists a large body of work studying how to identify the most relevant database given a user query [20, 19, 14, 23, 18]. This body of work is often referred to as meta-searching or database selection problem over the Hidden Web. For example, [19] suggests the use of focused probing to classify databases into a topical category, so that given a query, a relevant database can be selected based on its topical category. Our vision is different from this body of work in that we intend to download and index the Hidden pages at a central location in advance, so that users can access all the information at their convenience from one single location. CONCLUSION AND FUTURE WORK Traditional crawlers normally follow links on the Web to discover and download pages. Therefore they cannot get to the Hidden Web pages which are only accessible through query interfaces. In this paper, we studied how we can build a Hidden Web crawler that can automatically query a Hidden Web site and download pages from it. We proposed three different query generation policies for the Hidden Web: a policy that picks queries at random from a list of keywords, a policy that picks queries based on their frequency in a generic text collection, and a policy which adaptively picks a good query based on the content of the pages downloaded from the Hidden Web site. Experimental evaluation on 4 real Hidden Web sites shows that our policies have a great potential. In particular, in certain cases the adaptive policy can download more than 90% of a Hidden Web site after issuing approximately 100 queries. Given these results, we believe that our work provides a potential mechanism to improve the search-engine coverage of the Web and the user experience of Web search. 6.1 Future Work We briefly discuss some future-research avenues. Multi-attribute Databases We are currently investigating how to extend our ideas to structured multi-attribute databases. While generating queries for multi-attribute databases is clearly a more difficult problem, we may exploit the following observation to address this problem: When a site supports multi-attribute queries, the site often returns pages that contain values for each of the query attributes. For example, when an online bookstore supports queries on title, author and isbn, the pages returned from a query typically contain the title, author and ISBN of corresponding books. Thus, if we can analyze the returned pages and extract the values for each field (e.g, title = `Harry Potter', author = `J.K. Rowling' , etc), we can apply the same idea that we used for the textual database: estimate the frequency of each attribute value and pick the most promising one. The main challenge is to automatically segment the returned pages so that we can identify the sections of the pages that present the values corresponding to each attribute. Since many Web sites follow limited formatting styles in presenting multiple attributes -- for example, most book titles are preceded by the label "Title:" -- we believe we may learn page-segmentation rules automatically from a small set of training examples. Other Practical Issues In addition to the automatic query generation problem, there are many practical issues to be addressed to build a fully automatic Hidden-Web crawler. For example, in this paper we assumed that the crawler already knows all query interfaces for Hidden-Web sites. But how can the crawler discover the query interfaces? The method proposed in [15] may be a good starting point. In addition, some Hidden-Web sites return their results in batches of, say, 20 pages, so the user has to click on a "next" button in order to see more results. In this case, a fully automatic Hidden-Web crawler should know that the first result index page contains only a partial result and "press" the next button automatically . Finally, some Hidden Web sites may contain an infinite number of Hidden Web pages which do not contribute much significant content (e.g. a calendar with links for every day). In this case the Hidden-Web crawler should be able to detect that the site does not have much more new content and stop downloading pages from the site. Page similarity detection algorithms may be useful for this purpose [9, 13]. REFERENCES [1] Lexisnexis http://www.lexisnexis.com. [2] The Open Directory Project, http://www.dmoz.org. [3] E. Agichtein and L. Gravano. Querying text databases for efficient information extraction. In ICDE, 2003. [4] E. Agichtein, P. Ipeirotis, and L. Gravano. Modeling query-based access to text databases. In WebDB, 2003. [5] Article on New York Times. Old Search Engine, the Library, Tries to Fit Into a Google World. Available at: http: //www.nytimes.com/2004/06/21/technology/21LIBR.html , June 2004. [6] L. Barbosa and J. Freire. Siphoning hidden-web data through keyword-based interfaces. In SBBD, 2004. [7] M. K. Bergman. The deep web: Surfacing hidden value,http: //www.press.umich.edu/jep/07-01/bergman.html . [8] K. Bharat and A. Broder. A technique for measuring the relative size and overlap of public web search engines. In WWW, 1998. [9] A. Z. Broder, S. C. Glassman, M. S. Manasse, and G. Zweig. Syntactic clustering of the web. In WWW, 1997. [10] J. Callan, M. Connell, and A. Du. Automatic discovery of language models for text databases. In SIGMOD, 1999. [11] J. P. Callan and M. E. Connell. Query-based sampling of text databases. Information Systems, 19(2):97130, 2001. [12] K. C.-C. Chang, B. He, C. Li, and Z. Zhang. Structured databases on the web: Observations and implications. Technical report, UIUC. [13] J. Cho, N. Shivakumar, and H. Garcia-Molina. Finding replicated web collections. In SIGMOD, 2000. [14] W. Cohen and Y. Singer. Learning to query the web. In AAAI Workshop on Internet-Based Information Systems, 1996. [15] J. Cope, N. Craswell, and D. Hawking. Automated discovery of search interfaces on the web. In 14th Australasian conference on Database technologies, 2003. [16] T. H. Cormen, C. E. Leiserson, and R. L. Rivest. Introduction to Algorithms, 2nd Edition. MIT Press/McGraw Hill, 2001. [17] D. Florescu, A. Y. Levy, and A. O. Mendelzon. Database techniques for the world-wide web: A survey. SIGMOD Record, 27(3):5974, 1998. [18] B. He and K. C.-C. Chang. Statistical schema matching across web query interfaces. In SIGMOD Conference, 2003. [19] P. Ipeirotis and L. Gravano. Distributed search over the hidden web: Hierarchical database sampling and selection. In VLDB, 2002. [20] P. G. Ipeirotis, L. Gravano, and M. Sahami. Probe, count, and classify: Categorizing hidden web databases. In SIGMOD, 2001. [21] C. Lagoze and H. V. Sompel. The Open Archives Initiative: Building a low-barrier interoperability framework In JCDL, 2001. [22] S. Lawrence and C. L. Giles. Searching the World Wide Web. Science, 280(5360):98--100, 1998. [23] V. Z. Liu, J. C. Richard C. Luo and, and W. W. Chu. Dpro: A probabilistic approach for hidden web database selection using dynamic probing. In ICDE, 2004. [24] X. Liu, K. Maly, M. Zubair and M. L. Nelson. DP9-An OAI Gateway Service for Web Crawlers. In JCDL, 2002. [25] B. B. Mandelbrot. Fractal Geometry of Nature. W. H. Freeman & Co. [26] A. Ntoulas, J. Cho, and C. Olston. What's new on the web? the evolution of the web from a search engine perspective. In WWW, 2004. [27] A. Ntoulas, P. Zerfos, and J. Cho. Downloading hidden web content. Technical report, UCLA, 2004. [28] S. Olsen. Does search engine's power threaten web's independence? http://news.com.com/2009-1023-963618.html . [29] S. Raghavan and H. Garcia-Molina. Crawling the hidden web. In VLDB, 2001. [30] G. K. Zipf. Human Behavior and the Principle of Least-Effort. Addison-Wesley, Cambridge, MA, 1949. 109

crawler;deep web;hidden web;Hidden Web crawling;query selection;efficiency;Deep Web crawler;coverage;keyword selection;adaptive algorithm;potential bias;adaptive algorithmn;accurate language model;keyword query;keyword queries

75

Easy Language Extension with Meta-AspectJ

Domain-specific languages hold the potential of automating the software development process. Nevertheless, the adop-tion of a domain-specific language is hindered by the difficulty of transitioning to different language syntax and employing a separate translator in the software build process. We present a methodology that simplifies the development and deployment of small language extensions, in the context of Java. The main language design principle is that of language extension through unobtrusive annotations. The main language implementation idea is to express the language as a generator of customized AspectJ aspects, using our Meta-AspectJ tool. The advantages of the approach are twofold. First, the tool integrates into an existing software application much as a regular API or library, instead of as a language extension. This means that the programmer can remove the language extension at any point and choose to implement the required functionality by hand without needing to rewrite the client code. Second, a mature language implementation is easy to achieve with little effort since AspectJ takes care of the low-level issues of interfacing with the base Java language

INTRODUCTION AND MOTIVATION The idea of extensible languages has fascinated programmers for many decades, as evident by the extensibility fea This material is based upon work supported by the National Science Foundation under Grants No. CCR-0220248 and CCR-0238289. Copyright is held by the author/owner. ICSE'06, May 2028, 2006, Shanghai, China. ACM 1-59593-085-X/06/0005. tures in languages as old as Lisp. From a Software Engineering standpoint, the main advantages of expressing a concept as a language feature, as opposed to a library API, are in terms of conciseness, safety, and performance. A language feature allows expressing the programmer's intent much more concisely--in contrast, libraries are limited to a function- or method-call syntax. A language feature also allow better static error checking--a library can only check the static types of arguments of a function call against the declared types of the formals. Finally, a language feature can take advantage of context information and employ an optimized implementation, while a library routine cannot be customized according to its uses. Despite these advantages, there are excellent reasons why full language extensibility is undesirable. Changing the syntax and semantics of a programming language is confusing and can lead to incomprehensible code. Furthermore, programming languages are complex entities, designed to provide a small number of features but allow them to be combined as generally as possible. A new feature can either increase the complexity of the language implementation significantly (because of interactions with all existing features), or will need to be limited in its interactions, which is a bad language design principle that leads to single-use features and design bloat. In past work [3], we have advocated expressing small language extensions purely through unobtrusive annotations . Indeed, the introduction of user-defined annotations in mainstream programming languages, such as C# and Java, has allowed specialized language extensions (e.g., for distributed computing, persistence, or real-time programming ) to be added without changing the base syntax. We believe that the approach of limited language extension through annotations meshes very well with an implementation technique that uses our Meta-AspectJ (MAJ) tool [4] to express the semantics of the language extension. Specifically, MAJ is a language that allows writing programs that generate aspects in the AspectJ language [1]. The programmer can easily express an extension to the Java language as a program that: a) reads annotations and type information from an existing program using Java reflection; b) outputs a customized AspectJ aspect that transforms the original program according to the information in the annotation ; c) executes the generated aspect by using the standard AspectJ compiler. In other words, our approach uses the AspectJ language as a compiler back-end. AspectJ code is not written by the application programmer but generated by the language 865 extension, for the sole purpose of expressing program transformations easily and generally. This is appropriate, as AspectJ embodies the Aspect-Oriented Programming [2] philosophy of expressing program enhancements orthogonally and independently of the original source code. Our approach has the advantage of simplifying the implementation of the language extension significantly, without encouraging undisciplined language extension (since the only extensions allowed are through annotations). Specifically , the approach leverages the engineering sophistication of the AspectJ compiler implementation and its provisions for dealing correctly with different Java language features. If a programmer were to replicate the same effort by hand, she would likely need to reproduce much of the AspectJ compiler complexity. The purpose of this paper is to support the idea of implementing small language extensions as programs that produce aspects. We have recently implemented a number of such small extensions to Java and they all exhibit a striking simplicity. Specifically, we did not have to implement (or extend) a Java parser, we did not need to deal with syntax tree pattern matching and transformation, and we did not need to provide special handling for many Java complexities. The combined annotations-MAJ approach ensured that our small language extensions were implementable in a few hundreds of lines of code, without sacrificing generality in their conditions of use. We discuss two such extensions in detail, after first introducing the MAJ language. BACKGROUND MAJ SYNTAX MAJ is an extension of Java that allows writing programs that generate AspectJ source code. MAJ offers two operators for creating AspectJ code fragments: `[...] ("quote") and #[...] ("unquote"). The quote operator creates representations of AspectJ code fragments. Parts of these representations can be variable and are desig-nated by the unquote operator (instances of unquote can only occur inside a quoted code fragment). For example , the value of the MAJ expression `[call(* *(..))] is a data structure that represents the abstract syntax tree for the fragment of AspectJ code call(* *(..)). Similarly , the MAJ expression `[!within(#className)] is a quoted pattern with an unquoted part. Its value depends on the value of the variable className. If, for instance, className holds the identifier "SomeClass", the value of `[!within(#className)] is the abstract syntax tree for the expression !within(SomeClass). MAJ also introduces a new keyword infer that can be used in place of a type name when a new variable is being declared and initialized to a quoted expression. For example, we can write: infer pct1 = `[call(* *(..))]; This declares a variable pct1 that can be used just like any other program variable. For instance, we can unquote it: infer adv1 = `[void around() : #pct1 { }]; This creates the abstract syntax tree for a piece of AspectJ code defining (empty) advice for a pointcut. Of course, since AspectJ is an extension of Java, any regular Java program fragment can be generated using MAJ. We can now see a full MAJ method that generates a trivial but complete AspectJ file: void generateTrivialLogging(String classNm) { infer aspectCode = `[ package MyPackage; aspect #[classNm + &quot;Aspect&quot;] { before : call(* #classNm.*(..)) { System.out.println(&quot;Method called&quot;); } } ]; System.out.println(aspectCode.unparse()); } The generated aspect causes a message to be printed before every call of a method in a class. The name of the affected class is a parameter passed to the MAJ routine. This code also shows the unparse method that MAJ supports for creating a text representation of their code. EXAMPLE 1 FILLING INTERFACE METHODS Our first language extension is simple but a good example to our approach, since it can be defined very quickly and it is hard to implement with alternate means. The Java language ensures that a class cannot declare to "implement" an interface unless it provides implementations for all of its methods. Nevertheless, this often results in very tedious code. For instance, it is very common in code dealing with the Swing graphics library to implement an event-listener interface with many methods, yet provide empty implementations for most of them because the application does not care about the corresponding events. The example code below is representative: private class SomeListener implements MouseListener, MouseMotionListener { public void mousePressed (MouseEvent event) { ... // do something } public void mouseDragged (MouseEvent event) { ... // do something } // the rest are not needed. Provide empty bodies. public void mouseReleased (MouseEvent event) {} public void mouseEntered (MouseEvent event) {} public void mouseExited (MouseEvent event) {} public void mouseMoved (MouseEvent event) {} } Of course, the programmer could avoid providing the empty method bodies on a per-interface basis, by associating each interface with a class that by default provides empty implementations of all interface methods. Then a client class can inherit the empty implementations and only provide implementations for the methods it needs. This pattern is indeed supported in Swing code (through library classes called adapters), but it is usually not possible to employ since the listener class may already have another superclass. Instead, it would be nice to provide a simple Java language extension implemented as an annotation. The implementation of the extension would be responsible for finding the unimple-mented methods and supplying empty implementations by default (or implementations that just return a default primitive or null value, in the case of methods that have a return 866 type). In this case, the above class could be written more simply as: @Implements ({&quot;MouseListener&quot;,&quot;MouseMotionListener&quot;}) public class SomeListener { public void mousePressed (MouseEvent event) { ... // do something } public void mouseDragged (MouseEvent event) { ... // do something } } Of course, this extension should be used carefully since it weakens the tests of interface conformance performed by the Java compiler. We implemented the above Java extension using MAJ. The code for the implementation was less than 200 lines long, with most of the complexity in the traversal of Java classes, interfaces, and their methods using reflection. The code processes a set of given Java classes and retrieves the ones with an Implements annotation. It then finds all methods that are in any of the interfaces passed as arguments to the Implements annotation and are not implemented by the current class. For each such method, code is generated in an AspectJ aspect to add an appropriate method implementation to the class. For instance, the code to add the method to the class in the case of a void return type is: infer newMethod = `[ public void #methodName (#formals) {} ]; aspectMembers.add(newMethod); Finally, the class needs to be declared to implement the interfaces specified by the annotation. This is easily done by emitting the appropriate AspectJ code: infer dec = `[declare parents: #[c.getName()] implements #[iface.getName()]; ]; The final aspect (slightly simplified for formatting reasons ) generated for our earlier listener class example is: public aspect SomeListenerImplementsAspect1 { void SomeListener.mouseEntered(MouseEvent e) {} void SomeListener.mouseExited(MouseEvent e) {} void SomeListener.mouseMoved(MouseEvent e) {} void SomeListener.mouseReleased(MouseEvent e) {} declare parents: SomeListener implements MouseListener; declare parents: SomeListener implements MouseMotionListener; } This aspect performs exactly the modifications required to the original class so that it correctly implements the MouseListener and MouseMotionListener interfaces. We invite the reader to consider how else this language extension might be implemented. Our approach of using annotations in combination with MAJ yielded a very simple implementation by letting AspectJ deal with most of the complexities of Java. Specifically, we did not have to deal with the low-level complexities of either Java source syntax or Java bytecode. For instance, we did not have to do any code parsing to find the class body or declaration that needs to be modified. Dealing with Java syntactic sugar, such as inner classes, was automatic. We did not need to do a program transformation to add the implements clauses or the extra methods to the class. Similarly, we did not need to worry about the valid syntax for adding an implemented interface if the class already implements another. EXAMPLE 2 LANGUAGE SUPPORT FOR OBJECT POOLING Our second example language extension addresses a common programming need, especially in server-side programming . Software applications often find the need for pooling frequently-used objects with high instantiation costs. We use the following database connection class as a running example : public class DBConnection { public DBConnection(String dbURI, String userName, String password ) { ... } public void open() { ... } public void close() { ... } } The cost of an open() call is very high for a database connection . In applications concerned with performance, such as high-volume websites with lots of database requests, one often finds the need to pool database connections and keep them open, instead of repeatedly creating new ones and opening them. Making a class such as DBConnection into a "pooled" class involves at the very least creating a pooling manager class that knows how to manage instances of the class being pooled. A different pooling manager class is needed for each class being pooled, since the manager needs to have class-specific information such as how to instantiate a new instance of the class when the pool is running low (e.g., DBConnection objects are created by a constructor call, followed by an open() call), and how to uniquely identify objects of the same class that belong to different pools (e.g., DBConnection objects of different dbURI, userName, and password combinations need to be in different pools, and the pooling manager needs to understand which pool to fetch objects from when a request arrives). We expressed the pooling concept as a language feature that can used transparently with any Java class, as long as some broad properties hold regarding its construction and instantiation interface. The rest of the application will be completely oblivious to the change. This facilitates the application of pooling after a large code base which uses the class in its non-pooled form has been developed. Using our extension, converting a class to a pooled class involves only 4 annotations: @pooled, @constructor, @request, and @release. For example, to convert the DBConnection class into a "pooled" class, and to adapt an existing code base to using the pooled functionality, the user only has the add the following annotations to the code: @pooled(mgr=pooled.PoolTypes.BASIC, max=10, min=2) public class DBConnection { @constructor public DBConnection(String dbURI,String userName, String password ) { ... } 867 @request public void open() { ... } @release public void close() { ... } } The @pooled annotation indicates that class DBConnection should be pooled. It accepts parameters that can be used to customize the pooling policy. @constructor annotates the constructor call whose parameters serve as unique identifiers for different kinds of DBConnection objects. In this example, DBConnection objects with different dbURI, userName, and password combinations should be maintained separately. @request annotates the method that signals for the request of a pooled object, and @release annotates the method call that signals for the return of the pooled object back to the pooling manager. The implementation of this extension using MAJ is less than 400 lines of code. The MAJ program searches for classes annotated with @pooled, and generates two Java classes and one aspect to facilitate converting this class to be pooled. We next describe the generated code in more detail. The reader may want to consider in parallel how the same task could be accomplished through other means. Neither conventional Java facilities (i.e., classes and generics) nor AspectJ alone would be sufficient for expressing the functionality we describe below in a general way, so that it can be applied with little effort to arbitrary unomdified classes. For instance, none of these facilities can be used to create a proxy class with methods with identical signatures as those of an arbitrary Java class. First, a pooling manager class, PoolMgrForDBConnection, is generated for DBConnection. The pooling manager class contains methods for requesting and releasing pooled DBConnection objects, as well as code to manage the expansion of the pool based on the min and max parameters. In order to retrofit an existing code base to use DBConnection as a pooled class, we need to introduce proxy objects that will be used wherever an object of the original class would exist in the application code. This is necessary as different objects from the perspective of the client code will correspond to the same pooled object. We generate a proxy class as a subclass of the pooled class. In our example : DBConnection_Proxy extends DBConnection. All instances of the proxy class share a static reference to an instance of PoolMgrForDBConnection. Each proxy instance holds (non-static) references to the parameters to the @constructor constructor call, and the DBConnection object obtained from the pooling manager. The proxy class rewrites the @request and @release methods: the @request method is rewritten to obtain an object of DBConnection type from the pooling manager, using the unique identifiers kept from the constructor call, and the @release method returns the DBConnection method back to the pool, while setting the reference to this object to null. The MAJ code in the proxy takes care to exactly replicate the signature of the original methods, including modifiers and throws clauses. For instance, the @release method in the proxy is generated as: infer meth = `[ #mods #ret #[m.getName()] (#formals) #throwStmt { m_poolMgr.release(m_uniqueId, m_proxiedObj); m_proxiedObj = null; }]; All other methods simply delegate to the same method in the superclass. The idea is to have variables declared to hold a DBConnection object, now hold a DBConnection_Proxy object . Therefore, to complete the "proxy" pattern, we need to change all the calls of new DBConnection(...) to new DBConnection_Proxy(...). This is the role of our generated aspect: tedious recoding effort is easily replaced by an aspect: the aspect intercepts all the constructor calls of DBConnection, and returns an object instantiated by calling new DBConnection_Proxy(...). In summary, a user can easily turn a class into a pooled class, and retrofit any existing code base to use this class in its new, pooled form. The client code does not need to be hand-edited at all, other than with the introduction of our 4 annotations. FUTURE WORK We believe that the years to come will see the emergence of a healthy ecology of small language extensions based on the annotation features of Java and C#. There are already major examples of such extensions, especially with distribution- and persistence-related annotations, implemented in the context of J2EE Application Servers. Such extensions can be implemented with heavyweight support-e .g., parsing files, or recognizing annotations in a class loader and performing bytecode manipulation. In fact, the JBoss AOP mechanism (in whose early design and implementation we have played an active role) is the foremost example of infrastructure used to implement annotation-based language extensions. Nevertheless, experience from compilers in general-purpose languages has shown that it is beneficial to develop a mature back-end language and implement high-level features by translating to that back-end. Our approach proposes that AspectJ is well-suited as such a back-end language for small Java language extensions and that generating AspectJ code offers significant simplicity benefits. In the future we plan to support this claim with more examples and perform a thorough comparison with competing mechanisms. REFERENCES [1] G. Kiczales, E. Hilsdale, J. Hugunin, M. Kersten, J. Palm, and W. G. Griswold. An overview of AspectJ. In ECOOP '01: Proceedings of the 15th European Conference on Object-Oriented Programming, pages 327353, London, UK, 2001. Springer-Verlag. [2] G. Kiczales, J. Lamping, A. Menhdhekar, C. Maeda, C. Lopes, J.-M. Loingtier, and J. Irwin. Aspect-oriented programming. In M. Ak sit and S. Matsuoka, editors, Proceedings European Conference on Object-Oriented Programming, volume 1241, pages 220242. Springer-Verlag, Berlin, Heidelberg, and New York, 1997. [3] Y. Smaragdakis. A personal outlook on generator research. In C. Lengauer, D. Batory, C. Consel, and M. Odersky, editors, Domain-Specific Program Generation. Springer-Verlag, 2004. LNCS 3016. [4] D. Zook, S. S. Huang, and Y. Smaragdakis. Generating AspectJ programs with meta-AspectJ. In Generative Programming and Component Engineering (GPCE), pages 118. Springer-Verlag, October 2004. 868

language extensions;annotation;domain-specific language;language extension;Meta-AspectJ;Java;simplicity;domain-specific languages

76

Hourly Analysis of a Very Large Topically Categorized Web Query Log

We review a query log of hundreds of millions of queries that constitute the total query traffic for an entire week of a general-purpose commercial web search service. Previously, query logs have been studied from a single, cumulative view. In contrast, our analysis shows changes in popularity and uniqueness of topically categorized queries across the hours of the day. We examine query traffic on an hourly basis by matching it against lists of queries that have been topically pre-categorized by human editors. This represents 13% of the query traffic. We show that query traffic from particular topical categories differs both from the query stream as a whole and from other categories. This analysis provides valuable insight for improving retrieval effectiveness and efficiency. It is also relevant to the development of enhanced query disambiguation, routing, and caching algorithms.

INTRODUCTION Understanding how queries change over time is critical to developing effective, efficient search services. We are unaware of any log analysis that studies differences in the query stream over the hours in a day; much less how those differences are manifested within topical categories. We focus on Circadian changes in popularity and uniqueness of topical categories. Emphasis on changing query stream characteristics over this longitudinal (time) aspect of query logs distinguishes this work from prior static log analysis, surveyed<A href="76.html#7"> in [7]. We began with the hypothesis that there are very different characteristics during peak hours and off-peak hours during a day. After reviewing a week's worth of data hundreds of millions of queries - we have found, not surprisingly, that: The number of queries issued is substantially lower during non-peak hours than peak hours. However, we knew little about how often queries are repeated from one hour of the day to the next. After examining the behavior of millions of queries from one hour of the day to the next we have found the less obvious result: The average number of query repetitions in an hour does not change significantly on an hourly basis throughout the day. Most queries appear no more than several times per hour. These queries consistently account for a large portion of total query volume throughout the course of the day. The queries received during peak hours are more similar to each other than their non-peak hour counterparts. We also analyze the queries representing different topics using a topical categorization of our query stream. These cover approximately 13% of the total query volume. We hypothesized that traffic behavior for some categories would change over time and that others would remain stable. For 16 different categories, we examined their traffic characteristics: Some topical categories vary substantially more in popularity than others as we move through an average day. Some topics are more popular during particular times of the day, while others have a more constant level of interest over time. The query sets for different categories have differing similarity over time. The level of similarity between the actual query sets received within topical categories varies differently according to category. This leads us to believe that predictive algorithms that are able to estimate the likelihood of a query being repeated may well be possible. This could have a significant impact on future cache management and load-balancing algorithms. Such algorithms could improve retrieval effectiveness by assisting in query disambiguation, making it easier to determine what information need is being expressed by a query at a given time. They could also assist research in search efficiency that takes into account query arrival-rate<A href="76.html#7">s [3]. Our analysis covers the entirety of the tens of millions of queries each day in the search log from America Online over a complete week in December. This represents a population of tens of millions of users searching for a wide variety of topics. Section 2 reviews the prior work in query log analysis. Section 3 describes our analysis of overall query traffic. Section 4 describes our analysis of trends in categorized queries. Finally, in Section 5 we present our conclusions and directions for future work. 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 2529, 2004, Sheffield, South Yorkshire, UK. Copyright 2004 ACM 1-58113-881-4/04/0007...$5.00. 321 PRIOR WORK Examinations of search engine evaluation indicate that performance likely varies over time due to differences in query sets and collections<A href="76.html#7"> [6]. Although the change in collections over time has been studied (e.g., the growth of the web)<A href="76.html#7"> [10], analysis of users' queries has been primarily limited to the investigation of a small set of available query logs that provide a snapshot of their query stream over a fixed period of time. Prior work can be partitioned into static query log analysis and some recent disclosures by web search engines. Query log analysis can be partitioned into large-scale log analysis, small-scale log analysis and some other applications of log analysis such as categorization and query clustering. Jansen and Pooch provide a framework for static log analysis, but do not address analysis of changes in a query stream over time<A href="76.html#7"> [7]. Given that most search engines receive on the order of between tens and hundreds of millions of queries a day<A href="76.html#8"> [22], current and future log analysis efforts should use increasingly larger query sets to ensure that prior assumptions still hold. Previous studies measured overall aspects of users' queries from static web query logs. In the only large-scale study (all others involve only a few million queries), Silverstein concludes that users typically view only the top ten search results and that they generally enter short queries from a static analysis of an AltaVista query log from six weeks in 1998 consisting of 575 million non-empty queries<A href="76.html#8"> [16]. He also found that only 13.6% of queries appear more than three times, the top 25 queries represent 1.5% of the total query volume, and in 75% of sessions users do not revise their queries. Additionally, co-occurrence analysis of the most frequent 10,000 queries showed that the most correlated terms are often constituents of phrases. No time-based or topic-based analysis of this query load was reported; it does not provide insight into how or when any usage or topical interest changes occur. Other studies examine the effect of advanced query operators on the search service coverage of Google, MSN, and AOL, finding that in general, they had little effect<A href="76.html#7"> [4]. These overall statistics do not provide any insight into temporal changes in the query log, but do provide some insight into how people use search services. Jansen, et. al, also provide analysis of query frequency<A href="76.html#7"> [7][<A href="76.html#8">19]. Their findings indicate that the majority (57%) of query terms from the Excite log of more than 51,000 queries are used only once, and a large majority (78%) occur three times or less. These studies show that neither queries nor their component terms follow a Zipfian distribution, as the number of rare, infrequently repeated queries and terms is disproportionately large. Other studies have focused on user behavior at the query session level and found varying results, with some estimating reformulated queries constituting 40-52% of queries in a log<A href="76.html#8"> [18][21]. Wang, et. al examined a log of more than 500,000 queries to a university search engine from 1997-2001 <A href="76.html#8">[23]. They find trends in the number of queries received by season, month, and day. We extend upon this work by examining the larger community of general web searchers and analyzing trends corresponding to hour of day. Several studies examine query categories in small, static logs. Spink, et. al analyzed logs totaling more than one million queries submitted to the Excite web search engine during single days in 1997, 1999, and 2<A href="76.html#8">001 [18][19][20]. They classified approximately 2,500 queries from each log into 11 topical categories and found that although search topics have changed over the years, users' behaviors have not. Ross and Wolfram categorized the top 1,000 term pairs from the one million query Excite log into 30 subject areas to show commonalities of terms in categorie<A href="76.html#8">s [13]. Jansen, et. al used lists of terms to identify image, audio, and video queries and measure their presence in the one million query Excite<A href="76.html#7"> log [9]. In order to examine the differences in queries from users in different countries, Spink, et. al, examined a 500,000 query log from the FAST web search engine during 2001, believed to be used largely by Europeans at that time, classifying 2,500 queries from it into the same topical categories. They found differences between FAST and Excite in the topics searched fo<A href="76.html#8">r [17]. Other work manually grouped queries by task. Broder defines queries as informational, navigational or transactional and presents a study of AltaVista users via a popup survey and manual categorization of 200 queries from a log<A href="76.html#7"> [2]. Beitzel, et. al implicitly categorized queries from a search log as navigational by matching them to edited titles in web directories to automatically evaluate navigational web search<A href="76.html#7"> [1]. Xie and Wolfram automatically categorized query terms by using results from web search engines to assign the terms to broad subject categorie<A href="76.html#8">s [25]. Several studies of query caching examine query frequency distributions from a static log, focusing on the average likelihood of an arbitrary query being repeated over the entire, fixed-length log. Lempel and Moran evaluated the performance of caching strategies over a log of seven million queries to AltaVista in 2001 and found that the frequencies of queries in their log followed a power law<A href="76.html#8"> [11]. Eiron and McCurley compared query vocabulary from a log of nearly 1.3 million queries posed to a corporate intranet to the vocabulary of web page anchor text and found that the frequency of queries and query terms follows a tail-heavy power law<A href="76.html#7"> [5]. Xie and O'Hallaron studied query logs from the Vivisimo meta-search engine of 110,881 queries over one month in 2001 in comparison to the Excite log of 1.9 million over one day in 1999 and found that although as in other studies over half of the queries are never repeated, the frequencies of queries that are repeated do follow a Zipfian distribution<A href="76.html#8"> [26]. Saraiva, et. al evaluated a two-level caching scheme on a log of over 100,000 queries to a Brazilian search engine and found that query frequencies follow a Zipf-like distri<A href="76.html#8">bution [15]. Markatos simulated the effect of several types of query caches on an Excite query log of approximately one million queries and found that traditional caching methods provide significant improvements in efficiency<A href="76.html#8"> [12]. Although traditional MRU-style caches obviously enhance throughput by exploiting temporal locality at the minute-to -minute level, these studies do not examine changes in the query stream according to the hour of the day that may be leveraged in enhanced cache design. It is well known that different users represent the same information need with different query terms, making query clustering attractive when examining groups of related queries. However, as Raghavan and Sever have shown, traditional similarity measures are unsuitable for finding query-to-query similarity<A href="76.html#8"> [13]. Wen, et. al, incorporated click-through to cluster users' queries<A href="76.html#8"> [23]. In evaluating their system, they analyzed a random subset of 20,000 queries from a single month of their approximately 1-million queries-per-week traffic. They found 322 that the most popular 22.5% queries represent only 400 clusters of queries using differing sets of query terms. Many web search services have begun to offer views of the most popular and/or changing (becoming drastically more or less popular) queries: AOL Member Trends, Yahoo - Buzz Index, Lycos - The Lycos 50 with Aaron Schatz, Google Zeitgeist, AltaVista - Top Queries, Ask Jeeves, Fast (AllTheWeb). These views necessarily incorporate a temporal aspect, often showing popular queries for the current time period and those that are consistently popular. Some also break down popularity by topical categories. Systems seeking to display changing queries must address the issue of relative versus absolute change in a query's frequency to find queries whose change is "interesting", not simply a query that went from frequency one to two (a 200% jump), or one that went from 10,000 to 11,000 (a 1000 absolute change). OVERALL QUERY TRAFFIC We examine a search log consisting of hundreds of millions of queries from a major commercial search service over the seven-day period from 12/26/03 through 1/1/04. This log represents queries from approximately 50 million users. We preprocess the queries to normalize the query strings by removing any case differences, replacing any punctuation with white space (stripping advanced search operators from the approximately 2% of queries containing them), and compressing white space to single spaces. The average query length is 1.7 terms for popular queries and 2.2 terms over all queries. On average, users view only one page of results 81% of the time, two pages 18% and three or more 1% of the time. First, we examine trends in the query stream as a whole, and then focus on trends related to queries manually categorized into topical categories. We begin our analysis of the overall stream by examining how the volume of query traffic changes as we move from peak to non-peak hours. We show the percentage of the day's total and distinct number of queries for each hour in the day on average over our seven-day peri<A href="76.html#3">od in Figure 1 (all times in our query log are Eastern Standard Time). Only 0.75% of the day's total queries appear from 5-6AM, whereas 6.7% of the day's queries appear from 9-10PM. Perhaps more interestingly, the ratio of distinct to total queries in a given hour is nearly constant throughout the day. This shows that the average number of times a query is repeated is virtually constant over the hours in a day, remaining near 2.14 with only a 0.12 standard deviation. Although the average repetition of queries remains nearly constant, we can examine this in greater detail by measuring the frequency distribution of queries at various hours in the day, as seen<A href="76.html#3"> in Figure 2. From this analysis it is clear that the vast majority of queries in an hour appear only one to five times and that these rare queries consistently account for large portions of the total query volume throughout the course of the day. Figure 1 Although we have shown that the query distribution does not change substantially over the course of a day, this does not provide insight into how the sets of queries vary from one hour to the next. To examine this, we measure the overlap between the sets of queries entered during those hours. We use traditional set and bag overlap measures as gi<A href="76.html#4">ven in Equation 1 and Equation 2, respectively. Distinct overlap measures the similarity between the sets of unique queries from each hour, while overall (bag) overlap measures the similarity of their frequency distributions by incorporating the number of times each query appears in an hour, ) ; ( A q C i . While these measures examine the similarity of the sets of queries received in an hour and the number of times they are entered, they do not incorporate the relative popularity or ranking of queries within the query sets. To examine this, we also measure the Pearson correlation of the queries' frequencies. As can be seen from<A href="76.html#4"> Equation 3 (where ) ; ( A q C is the mean number of query repetitions in period A and ) ; ( A q C s is the standard deviation of all the query frequencies in period A), this measures the degree of linear correlation between the frequencies of the queries in each hour, so two hours that had exactly the same queries with exactly the same frequencies would have a correlation of one. Note that this normalizes for the effect of differing query volume, i.e., the correlation of two hours with exactly the same underlying query distributions simply scaled by a constant would also have a correlation of one. Figure 2 Percentage of Average Daily Query Traffic at Each Hour 0% 1% 2% 3% 4% 5% 6% 7% 8% 0 6 12 18 Hour of Day P e r c en t a g e o f D a i l y Qu er y T r af f i c Average Total Queries Average Distinct Queries Frequency Distribution of Selected Hours from 12/26/03 0% 5% 10% 15% 20% 25% 30% 35% 40% 1, 00 11 0 , 0 0 0 50 11 , 0 0 0 20 15 0 0 10 12 0 0 51 1 0 0 26 5 0 21 2 5 16 2 0 11 1 5 10 9 8 7 6 5 4 3 2 1 Frequency Ranges P erc en t a g e o f T o t a l Q u er i e s 12AM-1AM 6AM-7AM 12PM-1PM 6PM-7PM 323 B A B A B A overlap dist = ) , ( . Equation 1: Distinct Overlap of Query Sets from Hours A and B + = B A q i i B q i A q i B A q i i i i i i B q C A q C B q C A q C B q C A q C B A overlap )) ; ( ), ; ( min( ) ; ( ) ; ( )) ; ( ), ; ( min( ) , ( Equation 2: Overall Overlap of Query Sets from Hours A and B ) ; ( ) ; ( 1 , ) ) ; ( ) ; ( )( ) ; ( ) ; ( ( 1 1 B q C A q C n i i i B A s s B q C B q C A q C A q C n r = = Equation 3: Pearson Correlation of Query Frequencies from Hours A and B Figure 3 In<A href="76.html#4"> Figure 3 we examine the average level of overlap and correlation between the query sets received during the same hour for each day over our week. As measuring overlap over the set of all queries appearing in our week would be computationally expensive, we use the set of all the tens of millions of queries in the day after our seven-day period as an independent sample and measure overlap at each hour in our week of the queries matching those in that sample. Although we previously saw that the frequency distribution of queries does not substantially change across hours of the day,<A href="76.html#4"> Figure 3 shows that the similarity between the actual queries that are received during each hour does in fact change. This trend seems to follow query volume, which is apparent if we sort the same overlap data by query volume as is done<A href="76.html#4"> in Figure 4. Clearly, as query volume increases the queries that compose that traffic are more likely to be similar across samples of those peak time periods. This finding is consistent with prior analyses of web query caches showing they significantly improve performance under heavy load. The more redundancy they are able to detect, the more caching algorithms are able to enhance throughput. Although the prior work primarily measures the effect of this redundancy in cache performance, it is obvious that redundancy must exist and be detected for caching to succeed. By examining the overall query stream by hour we are able to infer the effectiveness of general caching algorithms at those times. Figure 4 QUERY CATEGORIES In Section 3 we analyzed the entire query log. However, this blanket view of the query traffic does not provide insight into the characteristics of particular categories of queries that might be exploited for enhanced efficiency or effectiveness. For example, a search provider who returns specialized results for entertainment queries cannot determine from general query traffic alone whether a given query is more likely to be referring to entertainment related content or how to best process and cache that query. The remainder of our analysis focuses on trends relating to topical category of queries. Our query set is categorized simply by exactly matching queries to one of the lists corresponding to each category. These lists are manually constructed by editors who categorize real users' queries, generate likely queries, and import lists of phrases likely to be queries in a category (e.g., cities in the US for the US Sites category). Queries that match at least one category list comprise 13% of the total query traffic on average. This represents millions of queries per day. Figure 5 To verify that our defined category lists sufficiently cover the topics in the query stream, we manually classified a random sample of queries, assigning them to "Other" if they did not intuitively fit into an existing category, as can be seen<A href="76.html#4"> in Figure 5. To determine the number of queries required to achieve a representative sample, we calculate the necessary sample size in queries, ss = (z 2 2 )/ 2 , where z is the confidence level value, is Sorted Average Overlap Characteristics from 1/2/04 that Matched Each Hour 10% 15% 20% 25% 30% 35% 40% 45% 50% 55% 60% 5 6 4 7 3 2 8 1 9 10 0 11 12 13 14 23 15 16 17 18 22 19 21 20 Hour of Day P e rcen t a g e Overlap Distinct Overlap Pearson Sampled Categorized Query Stream Breakdown Personal Finance 3% Computing 9% Research & Learn 9% Entertainment 13% Games 5% Holidays 1% Home 5% US Sites 3% Porn 10% Shopping 13% Sports 3% T ravel 5% Other 16% Health 5% Average Overlap Characteristics of Matching Queries from 1/2/04 10% 15% 20% 25% 30% 35% 40% 45% 50% 55% 0 6 12 18 Hour of Day Overlap Distinct Overlap Pearson 324 the sample standard deviation, and is the error rate. By setting our confidence level to 99% and error rate to 5%, we require a sample of 600 queries. The relative percentages for each category of the approximately 13% of query volume that match any category list over our week (s<A href="76.html#5">ee Figure 9) are within the error rate of those from our manually categorized sample. This shows that our lists are a reasonable representation of these topical categories. We focus on a subset of these categories and examine music and movies independent of other entertainment queries. The relative size of each category list we used is given in<A href="76.html#5"> Figure 6. Obviously, not all queries listed actually match those entered by users, especially when the category contains large imported lists of phrases. Figure 6 Although we have shown that our lists are a fair representation of the topics in the query stream, this does not indicate what portion of the frequency distribution of that stream they represent. To determine this, we measured the average proportion of queries matching any category list that appear at various frequencies each hour and compared them to the average overall hourly frequency distribution of the query stream (see<A href="76.html#5"> Figure 7). Unsurprisingly, this comparison shows that queries in the category lists represent more popular, repeated queries than average, although the general shape of the distributions is similar. Figure 7 4.1 Trends in Category Popularity We begin our temporal analysis of topical categories by measuring their relative popularity over the hours in a day. First, we examine the percent of total query volume matching a selected group of category lists, as can be seen<A href="76.html#5"> in Figure 8. It is clear that different topical categories are more and less popular at different times of the day. Personal finance, for example, becomes more popular from 7-10AM, while music queries become less popular. Although it is difficult to compare the relative level of popularity shift from one category to another due to the differences in scale of each of their percentages of the query stream, it is clear that some categories' popularity changes more drastically throughout the day than others. Figure 8 In order to quantify this, we calculated the KL-divergence (<A href="76.html#5">Equation 4) between the likelihood of receiving any query at a particular time and the likelihood of receiving a query in a particular category, as can be seen in<A href="76.html#5"> Figure 9. This reveals that the top three categories in terms of popularity are pornography, entertainment, and music. ) , | ( ) | ( log ) | ( )) , | ( ) | ( ( t c q p t q p t q p t c q p t q p D q = Equation 4: KL-Divergence of Query Occurrence Likelihood for Category c and Total Stream at Time t Figure 9 Comparing these divergences to the proportion of categorized queries in each category in<A href="76.html#5"> Figure 6 quickly illustrates that divergence is not correlated with the number of queries categorized in each category. Also shown in<A href="76.html#5"> Figure 9 is the average percentage of the entire query volume and distinct queries that match each category. Although the categories that cover the largest portions of the query stream also have the most relative popularity fluctuation, this correlation does not continue throughout all categories. Relative Percentage of Categorized Queries 0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% Shopp ing Com puti ng Trave l Hom e Hea lth Gov ernm ent Games Rese arch & L earn ing Porn Holiday s Spor ts Mov ies Pers onal Fina nce Ente rtain men t US S ites Music P e r c e n t a ge of C a t e gor i z e d Q u e r i e s Hourly Frequency Distribution of Matching Queries vs. All Queries Averaged over 7 Days and 16 Categories 0% 5% 10% 15% 20% 25% 30% 35% &gt; 1,0 00 201-500 51-1 00 21-2 5 11-1 5 9 7 5 3 1 Frequency Ranges P e r cent a g e of A v er age T o t a M a t c hi ng Q u er i e s Avg. Matching Queries Avg. Queries Category Percentage of Entire Query Stream and Divergence from Likelihood of any Query at Each Hour 0% 1% 2% 3% 4% 5% 6% Com puti ng Spo rts Holi day s Rese arch and L earni ng Hea lth Gam es US S ites Shopp ing Gov ernm ent Mov ies Tra vel Pers ona l Fin ance Home Music Ente rtain men t Porn Category KL-Divergence % of query stream Distinct % of query stream Categorical Percent over Time 0% 1% 2% 3% 4% 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Hour of Day Entertainment Games Health Personal Finance Shopping Music USSites Porn 325 We drilled down into the highly fluctuating categories and examined the behavior of the queries with the most highly fluctuating frequencies in each category. From this we hoped to gain some insight into the reasons why certain categories fluctuate, and the effect of terms and queries with very high flux on those categories. For example, the three most changing queries for the entertainment category on average over our week were: Table 1: Top Three Fluctuating Entertainment Queries gwyneth paltrow paris hilton orlando bloom All three of these queries are specifically related to recent events in US popular culture; the actress Gwyneth Paltrow recently married in secret, and the news of her nuptials broke during the week we analyzed. Hilton Hotel heiress Paris Hilton has been a popular topic recently; she starred in a prime time reality TV show entitled "The Simple Life". Also popular is Orlando Bloom, the actor who portrays a popular character in the "Lord of the Rings" trilogy. As the final installment of the series was released in US theatres during the week prior to our query log, it is no surprise to see his name as a top-changing query. Drilling down further, we pinpointed some of the specific instances where these popular queries jumped the most. For example, in the afternoon of Friday, December 27th, the popularity of the query "gwyneth paltrow" skyrocketed. From 3-4PM , it occurred once, from 4-5PM it occurred 67 times, and from 5PM-6PM it occurred 11,855 times. The top changing (on average) twenty-five queries, after normalization, in the Entertainment and Music categories are shown in<A href="76.html#6"> Table 2. Table 2: Top 25 Fluctuating Queries from Music and Entertainment Music Entertainment lyrics music britney spears furniture love hilary duff good charlotte sloppy seconds jessica simpson b2k eminem christina aguilera simple plan justin timberlake free music linkin park michael jackson beyonce jennifer lopez 50 cent kinky napster chic tupac blink 182 gwyneth paltrow paris hilton orlando bloom espn disney johnny depp much music disney channel hgtv disneychannel com www disneychanel com katie holmes pictures pamela anderson cartoon network hilary duff fake chad michael murray vivica a fox disneychannel care bears sailor moon www cartoonnetwork com days of our lives charmed tom welling We also looked at some of the most frequently changing terms to see how they relate to the change of entire queries containing those terms. Some excellent examples of this behavior in the Entertainment category include the terms "pictures" (the tenth-most changing term) and "duff" (the 17 th -most changing term). We looked at the popularity change (i.e., change in frequency) for queries containing these terms and found that several of them also exhibited large changes over time. For example, on the afternoon of December 28 th from noon to 5PM EST, the query "hilary duff" changed from an initial frequency of 27 from 12-1PM to a peak of 131 (from 3-4PM), and then stabilized around 70 for the rest of the evening; similar spikes in frequency for this query occurred at similar times during other days in our period of study. 4.2 Trends in Uniqueness of Queries Within Categories Although we have shown that different categories have differing trends of popularity over the hours of a day, this does not provide insight into how the sets of queries within those categories change throughout the day. In order to examine this, we return to the overlap measures used in Sec<A href="76.html#3">tion 3. Overlap, distinct overlap, and the Pearson correlation of query frequencies for Personal Finance and Music are show<A href="76.html#6">n in Figure 10 a<A href="76.html#7">nd Figure 11. Figure 10 Although the uniqueness of queries in categories in general appears to be correlated with that of the entire query stream (<A href="76.html#4">Figure 3), that of particular categories appears to be substantially different from one to the next. For example, if we compare the overlap characteristics of personal finance with those of music, we see they are quite different. Not only does personal finance have generally higher overlap, but it has a much higher overall overlap than distinct overlap, whereas they are nearly equal for music. Other categories with generally high overlap and distinct overlap are shopping, computing, and travel. Also, the correlation of frequencies of personal finance queries is very high all day, indicating searchers are entering the same queries roughly the same relative amount of times, this is clearly not true for music. Some categories have a high Pearson correlation. This indicates that a significant portion of the queries in these categories is often ranked similarly by frequency. These categories are: pornography, travel, research and learning, and computing, and their Pearson correlations are illustrate<A href="76.html#7">d in Figure 12. Personal Finance Overlap 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 Overlap Dist. Olap Pearson 326 Figure 11 It is clear that some categories have very similarly ranked queries by frequency throughout the day, while others vary dramatically according to query volume. Referring back to<A href="76.html#5"> Figure 6 and<A href="76.html#5"> Figure 9, uniqueness of queries in particular categories does not appear to be correlated with the number of queries in their respective category lists, the proportion of the query stream they represent, or the number of distinct queries they match. Figure 12 This type of data is potentially of great use to query caching algorithms. For example, if it is known a priori that queries for certain categories are similarly ranked throughout the day, they can be given higher priority in a query-caching scheme. Similarly, queries in categories whose rankings change vastly over time might be given low caching priority. CONCLUSIONS AND FUTURE WORK This study focuses on investigating the nature of changes in the query stream of a very large search service over time. Understanding how users' queries change over time is critical to developing effective, efficient search systems and to engineering representative test sets and evaluations that drive this development. In this study we find trends over time that are stable despite continuing fluctuation in query volume. Although the average query is repeated only twice during any given hour of the day, the total query traffic varies both in magnitude from one hour to the next, and also in degree of overlap and correlation in popularity of the queries that are received. In addition, we also find that the frequency distribution of an hour's worth of queries remains constant throughout the day. Also, at the most general level, we find that query volume is highest and query sets are most stable during peak hours of the day. This study further investigates changes in the query stream over time by examining the nature of changes in popularity of particular topical categories. For this we use a set of topical categories created by human editors that represents approximately 13% of the average query traffic. We show that popularity of some of these categories fluctuates considerably while other categories remain relatively stable over the hours in a day. Additionally, we show that the overlap and correlation in popularity of the queries within each topical category varies quite differently over the course of the day. Extending this analysis to investigate changes in the very rare queries not often matched by our category lists would provide insight into whether those are changing similarly to more popular queries. One method for approaching this might be to incorporate automatic query classification methods to extend our basic lists This study is the gateway to a large and diverse body of future work. Integrating this knowledge of Circadian changes in the query stream by category will likely yield improved query disambiguation, query caching, and load balancing algorithms. BIBLIOGRAPHY [1] Beitzel, S., Jensen, E., Chowdhury, A., and Grossman, D. Using Titles and Category Names from Editor-driven Taxonomies for Automatic Evaluation. In Proceedings of CIKM'03 (New Orleans, LA, November, 2003), ACM Press. [2] Broder, A. A Taxonomy of Web Search. SIGIR Forum 36(2) (Fall, 2002). [3] Chowdhury, A., G. Pass. "Operational Requirements for Scalable Search Systems", In Proceedings of CIKM'03 (New Orleans, LA, November 2003), ACM Press. [4] Eastman, C., B. Jansen, "Coverage, Relevance, and Ranking: The Impact of Query Operators on Web Search Engine Results", ACM Transactions on Information Systems, Vol. 21, No. 4, October 2003, Pages 383411. [5] Eiron, N., K. McCurley. "Analysis of Anchor Text for Web Search", In Proceedings of SIGIR'03 (Toronto, Canada, July 2003), ACM Press. [6] Hawking, D., Craswell, N., and Griffiths, K. Which Search Engine is Best at Finding Online Services? In Proceedings of WWW10 (Hong Kong, May 2001), Posters. Actual poster available as http://pigfish.vic.cmis.csiro.au/~nickc/pubs/www10actualpos ter.pdf [7] Jansen, B. and Pooch, U. A review of Web searching studies and a framework for future research. Journal of the American Society for Information Science and Technology 52(3), 235-246, 2001. [8] Jansen, B., Spink, A., and Saracevic, T. Real life, real users, and real needs: a study and analysis of user queries on the web. Information Processing and Management, 36(2) (2000), 207-227. [9] Jansen, B.J., Goodrum, A., Spink, A. Searching for multimedia: video, audio, and image Web queries. World Wide Web 3(4), 2000. [10] Lawrence, S. and Giles, C.L. Searching the World Wide Web. Science 280(5360), 98-100, 1998. Pearson Correlations of Frequencies for Categories 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 Personal Finance Music Movies Porn Computing Games Entertainment Government Music Overlap 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 Overlap Dist. Olap Pearson 327 [11] Lempel, R. and Moran, S. Predictive caching and prefetching of query results in search engines. In Proceedings of WWW12 (Budapest, May 2003). [12] Markatos, E.P. On Caching Search Engine Query Results. In the Proceedings of the 5th International Web Caching and Content Delivery Workshop, May 2000. [13] Raghavan, V. and Sever, H. On the Reuse of Past Optimal Queries. In Proc. of the 1995 SIGIR Conference, 344-350, Seattle, WA, July 1995. [14] Ross, N. and Wolfram, D. End user searching on the Internet: An analysis of term pair topics submitted to the Excite search engine. Journal of the American Society for Information Science 51(10), 949-958, 2000. [15] Saraiva, P., Moura, E., Ziviani, N., Meira, W., Fonseca, R., Riberio-Neto, B. Rank-preserving two-level caching for scalable search engines. In Proc. of the 24th SIGIR Conference, 51-58, New Orleans, LA, September, 2001. [16] Silverstein, C., Henzinger, M., Marais, H., and Moricz, M. Analysis of a very large web search engine query log. SIGIR Forum 33(1) (Fall, 1999), 6-12. [17] Spink, A., Ozmutlu, S., Ozmutlu, H.C., and Jansen, B.J. U.S. versus European web searching trends. SIGIR Forum 36(2), 32-38, 2002. [18] Spink, A., Jansen, B.J., Wolfram, D., and Saracevic, T. From E-sex to e-commerce: Web search changes. IEEE Computer, 35(3), 107-109, 2002. [19] Spink, A., Wolfram, D., Jansen, B.J. and Saracevic, T. Searching the Web: The Public and Their Queries. Journal of the American Society of Information Science 53(2), 226-234 , 2001. [20] Spink, A., Jansen, B.J., and Saracevic, T. Vox populi: The public searching of the web. Journal of the American Society of Information Science 52 (12), 1073-1074, 2001. [21] Spink, A., Jansen, B.J., and Ozmultu, H.C. Use of query reformulation and relevance feedback by Excite users. Internet Research: Electronic Networking Applications and Policy 10 (4), 2000. [22] Sullivan, D. Searches Per Day. Search Engine Watch, February, 2003. http://searchenginewatch.com/reports/article.php/2156461 [23] Wang, P., Berry, M., and Yang, Y. Mining longitudinal web queries: Trends and patterns. Journal of the American Society for Information Science and Technology 54(8), 743-758, June 2003. [24] J. Wen, J. Nie, H. Zhang "Query Clustering using User Logs" ACM Transactions on Information Systems, Vol. 20, No. 1, January 2002, pp59-81. [25] Wolfram, D., H. Xie, "Subject categorization of query terms for exploring Web users' search interests", Journal of the American Society for Information Science, v.53 n.8, p.617-630 , June 2002. [26] Xie, Y., O'Hallaron, D. Locality in Search Engine Queries and Its Implications for Caching. Infocom 2002. 328

query traffic;query stream;frequency distribution;topical categories;log analysis;query log;Query Log Analysis;Web Search

77

Efficient Multi-way Text Categorization via Generalized Discriminant Analysis

Text categorization is an important research area and has been receiving much attention due to the growth of the on-line information and of Internet. Automated text categorization is generally cast as a multi-class classification problem. Much of previous work focused on binary document classification problems. Support vector machines (SVMs) excel in binary classification, but the elegant theory behind large-margin hyperplane cannot be easily extended to multi-class text classification. In addition, the training time and scaling are also important concerns. On the other hand, other techniques naturally extensible to handle multi-class classification are generally not as accurate as SVM. This paper presents a simple and efficient solution to multi-class text categorization. Classification problems are first formulated as optimization via discriminant analysis . Text categorization is then cast as the problem of finding coordinate transformations that reflects the inherent similarity from the data. While most of the previous approaches decompose a multiclass classification problem into multiple independent binary classification tasks, the proposed approach enables direct multi-class classification. By using Generalized Singular Value Decomposition (GSVD), a coordinate transformation that reflects the inherent class structure indicated by the generalized singular values is identified. Extensive experiments demonstrate the efficiency and effectiveness of the proposed approach.

INTRODUCTION With the ever-increasing growth of the on-line information and the permeation of Internet into daily life, methods that assist users in organizing large volumes of documents are in huge demand. In particular, automatic text categorization has been extensively studied recently. This categorization problem is usually viewed as supervised learning, where the gaol is to assign predefined category labels to unlabeled documents based on the likelihood in-ferred from the training set of labeled documents. Numerous approaches have been applied, including Bayesian probabilistic approaches [20, 31], nearest neighbor [22, 19], neural networks [33], decision trees [2], inductive rule learning [4, 9], support vector machines [18, 14], Maximum Entropy [26], boosting [28], and linear discriminate projection [3] (see [34] for comparative studies of text categorization methods). Although document collections are likely to contain many different categories, most of the previous work was focused on binary document classification. One of the most effective binary classification techniques is the support vector machines (SVMs) [32]. It has been demonstrated that the method performs superbly in binary discriminative text classification [18, 34]. SVMs are accurate and robust, and can quickly adapt to test instances. However, the elegant theory behind the use of large-margin hyperplanes cannot be easily extended to multi-class text categorization problems. A number of techniques for reducing multi-class problems to binary problems have been proposed, including one-versus-the-rest method, pairwise comparison [16] and error-correcting output coding [8, 1]. In these approaches, the original problems are decomposed into a collection of binary problems, where the assertions of the binary classifiers are integrated to produce the final output. In practice, which reduction method is best suited is problem-dependent, so it is a non-trivial task to select the decomposition method. Indeed, each reduction method has its own merits and limitations [1]. In addition, regardless of specific details, these reduction techniques do not appear to be well suited for text categorization tasks with a large number of categories, because training of a single, binary SVM requires O n time for 1 7 2 1 where n is the number of training data [17]. Thus, having to train many classifiers has a significant impact on the overall training time. Also, the use of multiple classifiers slows down prediction. Thus, despite its elegance and superiority, the use of SVM may not be best suited for multi-class document classification. However, there do not appear to exist many alternatives, since many other techniques that can be naturally extended to handle multi-class classification problems, 317 such as neural networks and decision trees, are not so accurate as SVMs [34, 35]. In statistics pattern recognition literature, discriminant analysis approaches are well known to be able to learn discriminative feature transformations (see, e.g., [12]). For example, Fisher discriminant analysis [10] finds a discriminative feature transformation as eigenvectors associated with the largest eigenvalues of matrix T ^ 1 w ^ b , where ^ w is the intra-class covariance matrix and ^ b is the inter-class covariance matrix 1 . Intuitively, T captures not only compactness of individual classes but separations among them. Thus, eigenvectors corresponding to the largest eigenvalues of T are likely to constitute a discriminative feature transform. However, for text categorization, ^ w is usually singular owing to the large number of terms. Simply removing the null space of ^ w would eliminate important discriminant information when the projections of ^ b along those directions are not zeros [12]. This issue has stymied attempts to use traditional discriminant approaches in document analysis. In this paper we resolve this problem. We extend discriminant analysis and present a simple, efficient, but effective solution to text categorization. We propose a new optimization criterion for classification and cast text categorization as the problem of finding transformations to reflect the inherent similarity from the data. In this framework, given a document of unknown class membership, we compare the distance of the new document to the centroid of each category in the transformed space and assign it to the class having the smallest distance to it. We call this method Generalized Discriminant Analysis (GDA), since it uses generalized singular value decomposition to optimize transformation. We show that the transformation derived using GDA is equivalent to optimization via the trace or determinant ratios. GDA has several favorable properties: First, it is simple and can be programed in a few lines in MATLAB. Second, it is efficient. (Most of our experiments only took several seconds.) Third, the algorithm does not involve parameter tuning. Finally, and probably the most importantly, it is very accurate. We have conducted extensive experiments on various datasets to evaluate its performance. The rest of the paper is organized as follows: Section 2 reviews the related work on text categorization. Section 3 introduces our new criterion for discriminant analysis. Section 4 introduces the basics of generalized singular value decomposition and gives the solution of the optimization problem. Section 5 shows that the transformation derived using GDA can also be obtained by optimizing the trace or determinant ratios. Section 6 presents some illustrating examples . Section 7 shows experimental results. Finally, Section 8 provides conclusions and discussions. RELATED WORK Text categorization algorithms can be roughly classified into two types: those algorithms that can be naturally extended to handle multi-class cases and those require decomposition into binary classification problems. The first consists of such algorithms as Naive Bayes [22, 19], neural networks [25, 33], K-Nearest Neighbors [22, 19], Maximum Entropy [26] and decision trees. Naive Bayes uses the joint distributions of words and categorizes to estimate the probabilities that an input document belongs to each document class and 1 This is equivalent to using eigenvectors associated with the smallest eigenvalues of matrix T ^ 1 b ^ w . It indicates that traditional discriminant analysis requires the non-singularity of at least one covariance matrix. Since the rank of ^ w is usually greater than that of ^ b , we will base our discussion on the eigenvalue-decomposition of T ^ 1 w ^ b . then selects the most probable class. K-Nearest Neighbor finds the k nearest neighbors among training documents and uses the categories of the k neighbors to determine the category of the test document . The underlying principle of maximum entropy is that without external knowledge, uniform distribution should be preferred. Based on this principle, it estimate the conditional distribution of the class label given a document. The reduction techniques that are used by the second group include one-versus-the-rest method [29], error-correcting output coding [8], pairwise comparison [16], and multi-class objective functions , where the first two have been applied to text categorization [34, 13]. In the one-versus-the-rest method a classifier separating between from a class and the rest is trained for each class. Multi-class classification is carried out by integrating prediction of these individual classifiers with a strategy for resolving conflicts. The method is sometimes criticizes for solving asymmetric problems in a symmetrical manner and for not considering correlations between classes. Error-correcting output coding (ECOC) [8] partitions the original set of classes into two sets in many different ways. A binary classifier is trained for each partition. The partitions are carefully chosen so that the outputs of these classifiers assign a unique binary codeword for each class (with a large Hamming distance between any pair of them). The class of an input with unknown class membership is chosen by computing the outputs of the classifiers on that input and then finding the class with the codeword closest to the output codeword. Although SVMs are considered to be very effective in binary classification, its large training costs may make it unsuitable for multi-class classification with a large number of classes if the above decomposition techniques are applied. Also, the lack of a clear winner among the above techniques makes the reduction task complicated . Our GDA directly deals with multi-class classification and does not require reduction to binary classification problems. Other techniques for text categorization exist. Godbole et al. [14] propose a new multi-class classification technique that exploits the accuracy of SVMs and the speed of Naive Bayes. It uses a Naive Bayes classifier to compute a confusion matrix quickly. Then it uses this matrix to reduce both the number and the complexity of binary SVMs to be built. Chakrabarti et al. [3] propose a fast text classification technique that uses multiple linear projections. It first projects training instances to low-dimensional space and then builds decision tree classifiers on the projected spaces. Fragoudis et al. [11] propose a new algorithm that targets both feature and instance selection for text categorization. In summary, as pointed out in [34, 26], there is no obvious winner in multi-class classification techniques. For practical problems, the choice of approach will have to be made depending on the constraints , e.g., the desired accuracy level, the time available, and the nature of the problem. NEW CRITERION FOR DISCRIMINANT ANALYSIS Suppose the dataset D has m instances, d 1 d m , having p features each. Then D can be viewed as a subset of R p as well as a member of R m p . Suppose D has L classes, D 1 D L having m 1 m L instances, respectively, where m L i 1 m i . For each i, 1 i L, let J i be the set of all j, 1 j m, such that the j-th instance belongs to the i-th class, and let c i be the centroid of the i-th class, i.e., the component-wise average of the m i vectors in the 318 class. Let c be the centroid of the entire dataset. The intra-class scatter matrix of D, ^ w , is defined by ^ w L i 1 j J i d j c i T d j c i and its inter-class scatter matrix, ^ b , is defined by ^ b L i 1 j J i d j c T d j c Let A w be the m p matrix constructed by stacking D 1 e 1 T c 1 , , D L e L T c L and let A b be the p m matrix whose columns are, from left to right, m 1 c 1 c T m L c L c T . Then ^ w A w A T w and ^ b A b A T b Although there are ways (such as Kernel tricks [24]) for utilizing non-linear transformation, we will focus on linear transformation . Given a linear transformation , the covariance matrices in the transformed space are A b T A b T A T b A b T ^ b and A w T A w T A T w A w T ^ w Fisher's linear discriminant analysis discriminates inter-class distance and intra-class distance by using their corresponding covariance matrices. The optimal projection can be obtained by solving the generalized eigenvalue problem: ^ b ^ w (1) If ^ w is nonsingular, is given by the eigenvectors of matrix ^ 1 w ^ b . As we already pointed out, the approach fails if ^ w is singular which is often the case in document classification 2 . Usually, this problem is overcome by using a nonsingular intermediate space of ^ w obtained by removing the null space of ^ w and then computing eigenvectors. However, the removal of the null space of ^ w possibly eliminates some useful information because some of the most discriminant dimensions may be lost by the removal. In fact, the null space of ^ w is guaranteed to contain useful discriminant information when the projections of ^ b are not zeros along those directions. Thus, simple removal of the null space of ^ w is not an effective resolution [12]. Once the transformation has been determined, classification is performed in the transformed space based on a distance metrics, such as Euclidean distance d x y i x i y i 2 and cosine measure d x y 1 i x i y i i x 2 i i y 2 i A new instance, z, it is classified to argmin k d z x k (2) where x k is the centroid of k-th class. 2 In fact, ^ w is nonsingular only if there are p L samples. This is usually impractical. 3.2 The New Criterion We propose the use of the following criterion for discriminating inter-class and intra-class distances by inter-class and intra-class covariance matrices: min A b I n 2 F A w 2 F (3) where X F is the Frobenius norm of the matrix X , i.e., i j x 2 i j . The criterion does not involve the inverse of the intra-class matrix and is similar to Tikhonov regularization of least squares problems. Intuitively, the first term of (3) is used to minimize the difference between the projection of x i x in a new space and the i-th unit vector of the new space. The second term is used to minimize the intra-class covariance. The equation (3) can be rewritten as min A w A b 0 I n 2 F (4) and this is a least squares problem with the solution A T w A w A T b A b A T b (5) GENERALIZED SINGULAR VALUE DECOMPOSITION Here we will show how to use GSVD to compute efficiently the solution to the optimization problem formulated in Section 3 and show that the solution thus obtained is stable. 4.1 The Basics of GSVD Singular value decomposition (SVD) is a process of decomposing a rectangular matrix into three other matrices of a very special form. It can be viewed as a technique for deriving a set of uncor-related indexing variables or factors [6]. A Generalized Singular Value Decomposition (GSVD) is an SVD of a sequence of matrices . GSVD has played a significant role in signal processing and in signal identification and has been widely used in such problems as source separation, stochastic realization and generalized Gauss-Markov estimation. The diagonal form of GSVD, shown below, was first introduced in [21]. T HEOREM 1. (GSVD Diagonal Form [21]) If A R m p , B R n p , and rank A T B T k, then there exist two orthogonal matrices, U R m m and V R n n , and a non-singular matrix, R p p , such that U T 0 0 V T A B X C S I k 0 (6) where C and S are nonnegative diagonal and of dimension m k and n k, respectively, 1 S 11 S min n k min n k 0, and C T C S T S I k . The generalized singular values are defined to be the component-wise ratios of the diagonal entries of the two diagonal matrices. In signal processing, A is often the signal matrix and B is the noise matrix, in which case the generalized singular values are referred to as signal-noise ratios. 4.2 Stable Solution By plugging the GSVD matrices of A w and A b in (5), we have X I k 0 S T V T . Since V is orthogonal, we can drop it without 319 changing the squared distance. So, we have X I k 0 S T (7) This derivation of holds even if ^ w is singular. Thus, by using GSVD to solve the new criterion, we can avoid removing null space, thereby keeping all the useful information. The degree of linear independence of the original data, rank A T w A T b , is equal to k, Since R p k , rank A w T A b T , the degree of linear independence in the transformed space, is at most k. We now state a theorem that shows that the solution is stable. T HEOREM 2. (GSVD relative perturbation bound [7]) Suppose A and B be matrices with the same number of columns and B is of full column rank. Let A A 1 D 1 and B B 1 D 2 such that D 1 and D 2 have full rank. Let E E 1 D 1 and F F 1 D 2 be perturbations of A and B, respectively, such that for all x there exist some 1 2 1 for which it holds that E 1 x 2 1 A 1 x 2 F 1 x 2 2 B 1 x 2 Let i and ~ i be the i-th generalized singular value of A B and that of A E B F , respectively. Then either i ~ i 0 or i ~ i i 1 2 1 2 The above theorem gives a bound on the relative error of the generalized eigenvalues (C ii and S ii ) if the difference between the estimated covariance matrices and the genuine covariance matrices is small. This guarantees that the relative error of is bounded by the relative error of estimated intra- and inter-class covariance matrices. GSVD also brings some favorable features, which might improve accuracy. In particular, computation of the cross products A T b A b and A T w A w , which causes roundoff errors, is not required. 4.3 The GDA Algorithm The pseudo codes of the training and prediction procedures are described as follows: Algorithm 1 Training procedure = Train (x's) Input: the training data x i 's Output: the transformation ; begin 1. Construct the matrices A w and A b ; 2. Perform GSVD on the matrix pair; 3. Obtain as described in equation 7. 4. Return ; end Algorithm 2 Prediction Procedure T = Predict ( , x) Input: the transformation generated by the training procedure; and a new instance x; Output: the label T of the new instance; begin 1. Perform Prediction as in equation 2; 2. Return T ; end CONNECTIONS Here we show that the above transformation derived using our new criterion can also be obtained by optimizing the trace or determinant ratios. 5.1 Optimizing the determinant ratio Fisher's criterion is to maximize the ratio of the determinant of the inter-class scatter matrix of the projected samples to the determinant of the intra-class scatter matrix of the projected samples: J T ^ b T ^ w (8) One way to overcome the requirements of non-singularity of Fisher's criterion is looking for solutions that simultaneously maximize T ^ b minimize T ^ w . Using GSVD, A b and A w are decomposed as A w UC I k 0 X 1 and A b V S I k 0 X 1 . To maximize J , T ^ b should be increased while decreasing T ^ w . Let C C I k 0 and S S I k 0 . Then we have ^ b A T b A b X S 2 X 1 and ^ w A T w A w XC 2 X 1 . This implies T ^ b T X S 2 X 1 S X 1 2 and T ^ w T XC 2 X 1 C X 1 2 Thus, the matrix satisfying X 1 I k 0 would simultaneously maximize T ^ b and minimize T ^ w (since the diagonal of S is decreasing). So, we have X I k 0 . In the case where we must weight the transformation with the generalized singular , X I k 0 S T is the transformation we want. 5.2 Optimizing the trace ratio The same transformation can also be obtained by optimizing the trace ratio. Using GSVD, we have trace T ^ b trace S S T X 1 T X T trace S S T GG T k i 1 S 2 ii g ii and trace T ^ w trace C C T X 1 T X T trace C C T GG T k i 1 C 2 ii g ii where G X 1 and g ii is the ii-th term of G. Since C T C S T S I k , we have trace T ^ b trace T ^ w k i 1 S 2 ii g ii k i 1 C 2 ii g ii k i 1 g ii If we force that trace T ^ b 1, the optimization is formulated as minimization of trace T ^ w k i 1 g ii 1. Here g ii 's are diagonal elements of a positive semi-definite matrix, so they are nonnegative. Also, for all i, g ii 0 implies that for all j 320 g i j g ji 0. Note that GG T is a p p matrix. Since only the first k diagonal entries, g ii ki 1 , appear in the formula for trace T ^ w k i 1 g ii 1, the quantities of other m k diagonal entries do not affect the optimization. Thus, we may set all of these to 0, thereby obtaining X I k 0 . In the case when we want to weight the transformation with the generalized singular values, we obtain X I k 0 S T . TEXT CLASSIFICATION VIA GDA EX-AMPLES A well-known transformation method in information retrieval is Latent Semantic Indexing (LSI) [6], which applies Singular Value Decomposition (SVD) to the document-term matrix and computes eigenvectors having largest eigenvalues as the directions related to the dominant combinations of the terms occurring in the dataset (latent semantics). A transformation matrix constructed from these eigenvectors projects a document onto the latent semantic space. Although LSI has been proven extremely useful in information retrieval , it is not optimal for text categorization because LSI is com-pletely unsupervised. In other words, LSI deals with the data without paying any particular attention to the underlying class structure . It only aims at optimally transforming the original data into a lower dimensional space with respect to the mean squared error , which has nothing to do with the discrimination of the different classes. Our GDA approach possesses advantages of both discriminant analysis and of latent semantic analysis. By explic-itly taking the intra-class and inter-class covariance matrices into the optimization criterion, GDA deals directly with discrimination between classes. Furthermore, by employing GSVD to solve the optimization problem, GDA tries to identify the latent concepts indicated by the generalized singular values. To illustrate how well GDA can perform, we present here two examples. In the first example, we compare GDA against LDA and LSI. Figure 1 shows a small dataset consisting of nine phrases in three topics: user interaction, graph theory, and distributed systems. No. Class Phrase 1 1 Human interface for user response 2 1 A survey of user opinion of computer system response time 3 1 Relation of user-perceived response time to error measurement 4 2 The generation of random, binary, unordered trees 5 2 The intersection graph of paths in trees 6 2 Graph Minors IV: Widths of trees and well-quasi-ordering 7 3 A survey of distributed shared memory system 8 3 RADAR: A multi-user distributed system 9 3 Management interface tools for distributed computer system Figure 1: Nine example sentences After removing words (terms) that occurs only once, we have the document-term matrix as shown in Figure 2. The first and second samples in each class are used for training . GDA , LDA, and LSI are run on the training data to obtain transformation matrices. Figure 3 shows the plot of the word \ No. 1 2 3 4 5 6 7 8 9 a 1 1 1 computer 1 1 distributed 1 1 1 for 1 1 graph 1 1 interface 1 1 of 2 1 1 1 1 1 response 1 1 1 survey 1 1 system 1 1 1 1 the 1 1 time 1 1 trees 1 1 1 user 1 1 1 1 Figure 2: Document-term Matrix distances/similarities between document pairs in the transformed space using each of the three methods. (a) GDA (b) LDA (c) LSI Figure 3: Pairwise document similarity via GDA , LDA, and LSI. The darker the close is the more similar the documents are. GDA is a clear winner. The second example illustrates differences between GDA and LSI. Distinction among three newsgroups in 20NG are attempted by selecting from each newsgroup twenty training and twenty for testing. Figure 4 shows plots of the the sixty testing articles using the two dominant directions as the axes. GDA has clear separation while the LSI plot shows an L-shaped concentration of the data points. The confusion matrices of these methods are shown in Table 1. GDA clearly performed better than LSI. prediction prediction actual 1 2 3 actual 1 2 3 1 20 0 0 1 20 0 0 2 0 19 1 2 0 3 17 3 0 0 0 3 7 5 8 Table 1: The confusion matrices. Left: GDA . Right: LSI. EXPERIMENTS For our experiments we used a variety of datasets, most of which are frequently used in the information retrieval research. The range of the number of classes is from four to 105 and the range of the number of documents is from 476 to 20,000, which seem varied 321 -1.5 -1 -0.5 0 0.5 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 Group 1 Group 2 Group 3 (a) GDA -0.07 -0.06 -0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 -0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 Group 1 Group 2 Group 3 (b) LSI Figure 4: Document plots. The three groups are separated sig-nificantly better with GDA than with LSI. enough to obtain good insights as to how GDA performs. Table 2 summarizes the characteristics of the datasets. 20Newsgroups The 20Newsgroups (20NG) dataset contains approximately 20,000 articles evenly divided among 20 Usenet newsgroups. The raw text size is 26MB. All words were stemmed using a porter stemming program, all HTML tags were skipped, and all header fields except subject and organization of the posted article were ignored. WebKB The WebKB dataset 3 contains Web pages collected from university computer science departments. There are approximately 8,300 documents in the set and they are divided into seven categories: student, faculty, staff, course, project, department, and other. The raw text size of the dataset is 27MB. Among the seven categories, student, faculty, course, and project are the four most populous. The subset consisting only of these categories is also used here, which is called WebKB4. In neither of the datasets, we used stemming or stop lists. Industry Sector The Industry Section dataset 4 is based on the data made available by Market Guide, Inc. (www.marketguide.com). The set consists of company homepages that are categorized in a hierarchy of industry sectors, but we disregarded the hierarchy. There were 9,637 documents in the dataset, which were divided into 105 classes. We tokened the documents by skipping all MIME and HTML headers and using a standard stop list. We did not perform stemming. Reuters The Reuters-21578 Text Categorization Test Collection contains documents collected from the Reuters newswire in 1987. It is a standard text categorization benchmark and contains 135 categories. We used its subsets: one consisting of the ten most frequent categories, which we call Reuters-Top10, and the other consisting of documents associated with a single topic, which we call Reuters-2. Reuters-2 had approximately 9,000 documents and 50 categories. TDT2 TDT2 is the NIST Topic Detection and Tracking text corpus version 3.2 released in December 6, 1999 [30]. This corpus contains news data collected daily from nine news sources in two languages (American English and Mandarin Chinese), over a period of six months (JanuaryJune in 1998). We used only the English news texts, which were collected from New York Times Newswire Service, Associated Press Worldstream Service, Cable News Network, Voice of America, American Broadcasting Company , and Public Radio International. The documents were manu-ally annotated using 96 target topics. We selected the documents having annotated topics and removed the brief texts. The resulting 3 Both 20NG and WebKB are available at http://www-2 .cs.cmu.edu/afs/cs/project/theo-11/www/wwkb. 4 Available at http://www.cs.cmu.edu/ TextLearning/datasets.html dataset contained 7,980 documents. K-dataset This dataset was obtained from the WebACE project [15]. It contained 2,340 documents consisting of news articles from Reuters News Service made available on the Web in October 1997. These documents were divided into 20 classes. They were processed by eliminating stop words and HTML tags, stemming the remaining words using Porter's suffix-stripping algorithm. CSTR This is the dataset of the abstracts of technical reports published in the Department of Computer Science at the University of Rochester between 1991 and 2002 5 . The dataset contained 476 abstracts, which were divided into four research areas: Symbolic-AI , Spatial-AI, Systems, and Theory. We processed the abstracts by removing stop words and applying stemming operations on the remaining words. Datasets # documents # class 20NG 20,000 20 WebKB4 4,199 4 WebKB 8,280 7 Industry Sector 9,637 105 Reuters-Top10 2,900 10 Reuters-2 9,000 50 CSTR 476 4 K-dataset 2,340 20 TDT2 7,980 96 Table 2: Data Sets Descriptions 7.2 Data Preprocessing In all experiments, we randomly chose 70% of the documents for training and assigned the rest for testing. It is suggested in [35] that information gain is effective for term removal and it can remove up to 90% or more of the unique terms without performance degrade. So, we first selected the top 1,000 words by information gain with class labels. The feature selection is done with the Rainbow package [23]. Here we use classification accuracy for evaluation. Different measures, such as precision-recall graphs and F 1 measure [34], have been used in the literature. However, since the datasets used in our experiments are relatively balanced and single-labeled, and our goal in text categorization is to achieve low misclassification rates and high separation between different classes on a test set, we thought that accuracy is the best measure of performance. All of our experiments were carried out on a P4 2GHz machine with 512M memory running Linux 2.4.9-31. 7.3 Experimental Results Now we present and discuss the experimental results. Here we compare GDA against Naive Bayes (NB for short), K-Nearest Neighbor (KNN for short), Maximum Entropy (ME for short), LDA, and SVM on the same datasets with the same training and testing data. Recall that the first three of the methods we compare against are commonly-used direct methods for multi-class classification (in the sense that they do not require reduction to binary classification problems). For experiments involving SVM we used SVMTorch [5] 6 , which uses the one-versus-the-rest decomposition. Table 3 and Figure 5 show performance comparisons. GDA outperformed all the other five methods on 20NG, WebKB4, WebKB and Industry Sector. SVM performed the best on Reuters-2, 5 The TRs are available at http://www.cs.rochester.edu/trs. 6 Download-able at http://old-www.idiap.ch/learning/SVMTorch.html. 322 K-dataset, and TDT2. GDA outperformed LDA on all the experiments , and the improvement was significant (more than 10%) when the sample size was relatively small (in the case of CSTR, Reuters-Top10 , and K-dataset). On 20NG, the performance of GDA s 95 03%, which is approximately 10% higher than that of NB, 6% higher than that of ME, and 4% higher than that of SVM. On the WebKB4 dataset, GDA beats NB by approximately 5%, and both ME and SVM by approximately 2%. On the WebKB dataset, GDA beats NB by approximately 16% and ME by 6%. The performance of GDA is about 8% higher than that of NB and by 6% than that of ME on the Industry Sector. The results with GDA and with SVM are almost the same on WebKB, Industry Sector, Reuters-Top10, and CSTR. On Reuters-2, K-dataset, and TDT2, SVM performs slightly better than GDA by 3%. ME achieves the best results on the CSTR dataset while NB is the winner on Reuters-top10 in terms of performance On CSTR, the performance of GDA is 2% lower than that of NB and 4% lower than that of ME. On Reuters-Top10, GDA is beaten by NB by approximately 1%. In total, the performance of GDA is always either the winner or very close to the winner: it is ranked the first four times, ranked the second three times, and ranked the third in the remaining two. Although there is no single winner over all datasets, GDA seems to outperform the rest on most counts. We can say that GDA is a viable, competitive algorithm in text categorization. Datasets GDA NB KNN ME LDA SVM 20NG 95.03 85.60 50.70 89.06 93.90 91.07 WebKB4 94.01 85.13 37.29 91.93 90.72 92.04 WebKB 79.02 61.01 44.81 71.30 77.35 78.89 Industry Sector 66.57 56.32 39.48 58.84 66.49 65.96 Reuters-Top10 81.98 83.33 74.07 81.65 71.46 81.13 Reuters-2 89.82 87.88 73.22 88.56 88.65 92.43 CSTR 88.50 90.85 82.53 92.39 68.29 88.71 K-dataset 88.44 86.14 58.26 86.19 77.69 91.90 TDT2 90.54 91.59 86.63 89.18 88.41 93.85 Table 3: Performance comparisons. For KNN we set k to 30. 0 10 20 30 40 50 60 70 80 90 100 20N ewsg roup s Web KB 4 Web KB Ind ustr y S ecto r Reu ters -top 10 Reu ters -2 URCS K-d atas et TDT 2 GDA NB KNN ME LDA SVM Figure 5: Performance Comparison GDA is also very efficient and most experiments are done in several seconds. Table 4 summarizes the running time for all the experiments of GDA and SVM. Figure 6 and Figure 7 present the comparisons of training and prediction time respectively. The time saving of GDA is very obvious. In summary, these experiments have shown that GDA provides an alternate choice for fast and efficient text categorization. GDA GDA SVM SVM Datasets Training Prediction Training Prediction 20NG 171.80 6.86 270.20 64.28 WebKB4 63.4 0.20 114.67 54.72 WebKB 94.64 0.43 1108.17 103.03 Industry Sector 88.23 6.45 423.54 79.82 Reuters-Top10 61.23 0.15 94.28 18.65 Reuters-2 96.19 1.13 566.53 85.10 CSTR 3.65 0.02 7.50 2.77 K-dataset 62.88 0.18 84.56 47.70 TDT2 21.69 5.14 89.91 26.76 Table 4: Time Table in seconds. 0 200 400 600 800 1000 1200 20N ewsg rou ps We bKB 4 We bKB Ind ustry Sect or Reu ters -top 10 Reu ters -2 CST R K-d ata set TDT 2 Training Time GDA SVM Figure 6: Training Time Comparisons 0 20 40 60 80 100 120 20N ewsg rou ps We bKB 4 We bKB Ind ustry Sect or Reu ters -top 10 Reu ters -2 CST R K-d ata set TDT 2 Prediction Time GDA SVM Figure 7: Prediction Time Comparisons DISCUSSIONS AND CONCLUSIONS In this paper, we presented GDA , a simple, efficient, and yet accurate , direct approach to multi-class text categorization. GDA utilizes GSVD to transform the original data into a new space, which could reflect the inherent similarities between classes based on a new optimization criterion. Extensive experiments clearly demonstrate its efficiency and effectiveness. Interestingly enough, although traditional discriminant approaches have been successfully applied in pattern recognition, little work has been reported on document analysis. As we mentioned earlier, this is partly because the intra-class covariance matrix is usually singular for document-term data and hence restrict the usage of discriminant. Our new criterion avoids the problem while still preserving the discriminative power of the covariance matrix. 323 Another big barrier to application of discriminant analysis in document classification is its large computation cost. As we know, traditional discriminant analysis requires a large amount of computation on matrix inversion, SVD, and eigenvalue-analysis. The costs of these operations are extremely large in document analysis because the matrices have thousands of dimension. Our approach makes use of effective feature selection via information gain, with which we can remove up to 90% or more of the unique terms without significant performance degrade [35]. One of our future plans is to explore how the performance correlates with different feature selection methods and the number of words selected. There are also other possible extensions such as using random projection to reduce the dimensionality before applying discriminant analysis [27]. Acknowledgments This work is supported in part by NSF grants EIA-0080124, DUE-9980943 , and EIA-0205061, and NIH grant P30-AG18254. REFERENCES [1] Allwein, E. L., Schapire, R. E., & Singer, Y. (2000). Reducing multiclass to binary: A unifying approach for margin classifiers. ICML-00 (pp. 916). [2] Apte, C., Damerau, F., & Weiss, S. (1998). Text mining with decision rules and decision trees. Proceedings of the Workshop with Conference on Automated Learning and Discovery: Learning from text and the Web. [3] Chakrabarti, S., Roy, S., & Soundalgekar, M. V. (2002). Fast and accurate text classification via multiple linear discriminant projections. Proceedings of the 28th International Conference on Very Large Databases (pp. 658669). [4] Cohen, W. W., & Singer, Y. (1996). Context-sensitive learning methods for text categorization. Proceedings of the 19th Annual International ACM SIGIR Conference on Research and Development in Information (pp. 307315). [5] Collobert, R., & Bengio, S. (2001). SVMTorch: Support vector machines for large-scale regression problems. Journal of Machine Learning Research, 1, 143160. [6] Deerwester, S. C., Dumais, S. T., Landauer, T. K., Furnas, G. W., & Harshman, R. A. (1990). Indexing by latent semantic analysis. Journal of the American Society of Information Science, 41, 391407. [7] Demmel, J., & Veselic, K. (1992). Jacobi's method is more accurate than QP. SIAM Journal on Matrix Analysis and Applications, 13, 1019. [8] Dietterich, T. G., & Bakiri, G. (1995). Solving multiclass learning problems via error-correcting output codes. Journal of Artificial Intelligence Research, 2, 263286. [9] Dumais, S., Platt, J., Heckerman, D., & Sahami, M. (1998). Inductive learning algorithms and representations for text categorization. CIKM-98 (pp. 148155). [10] Fisher, R. (1936). The use of multiple measurements in taxonomic problems. Annals of Eugenics, 7, 179188. [11] Fragoudis, D., Meretakis, D., & Likothanassis, S. (2002). Integrating feature and instance selection for text classification. SIGKDD-02 (pp. 501506). [12] Fukunaga, K. (1990). Introduction to statistical pattern recognition. Academic Press. [13] Ghani, R. (2000). Using error-correcting codes for text classification. ICML-00 (pp. 303310). [14] Godbole, S., Sarawagi, S., & Chakrabarti, S. (2002). Scaling multi-class support vector machine using inter-class confusion. SIGKDD-02 (pp. 513518). [15] Han, E.-H., Boley, D., Gini, M., Gross, R., Hastings, K., Karypis, G., Kumar, V., Mobasher, B., & Moore, J. (1998). WebACE: A web agent for document categorization and exploration. Agents-98 (pp. 408415). [16] Hastie, T., & Tibshirani, R. (1998). Classification by pairwise coupling. Advances in Neural Information Processing Systems. The MIT Press. [17] Joachims, T. (1998). Making large-scale support vector machine learning practical. In Advances in kernel methods: Support vector machines. [18] Joachims, T. (2001). A statistical learning model of text classification with support vector machines. SIGIR-01 (pp. 128136). [19] Lam, W., & Ho., C. (1998). Using a generalized instance set for automatic text categorization. SIGIR-98 (pp. 8189). [20] Lewis, D. D. (1998). Naive (Bayes) at forty: The independence assumption in information retrieval. ECML-98. [21] Loan, C. V. (1976). Generalizing the singular value decomposition. SIAM J. Num. Anal., 13, 7683. [22] Masand, B., Linoff, G., & Waltz., D. (1992). Classifying news stories using memory based reasoning. SIGIR-92 (pp. 5964). [23] McCallum, A. K. (1996). Bow: A toolkit for statistical language modeling, text retrieval, classification and clustering. http://www.cs.cmu.edu/ mccallum/bow. [24] Mika, S., Ratsch, G., Weston, J., Scholkopf, B., & Muller, K.-R. (1999). Fisher discriminant analysis with kernels. Neural Networks for Signal Processing IX (pp. 4148). IEEE. [25] Ng, H. T., Goh, W. B., & Low, K. L. (1997). Feature selection, perceptron learning, and a usability case study for text categorization. Proceedings of the 20th Annual International ACM SIGIR Conference on Research and Development in Information (pp. 6773). [26] Nigam, K., Lafferty, J., & McCallum, A. (1999). Using maximum entropy for text classification. In IJCAI-99 Workshop on Machine Learning for Information Filtering (pp. 6167). [27] Papadimitriou, C. H., Tamaki, H., Raghavan, P., & Vempala, S. (1998). Latent semantic indexing: A probabilistic analysis. Proceedings of the Symposium on Principles of Database Systems (pp. 159168). [28] Schapire, R. E., & Singer, Y. (2000). Boostexter: A boosting-based system for text categorization. Machine Learning, 39, 135168. [29] Scholkopf, B., & J.Smola, A. (2002). Learning with kernels. MIT Press. [30] TDT2 (1998). Nist topic detection and tracking corpus. http://www.nist.gove/speech/tests/tdt/tdt98/index.htm. [31] Tzeras, K., & Hartmann, S. (1993). Automatic indexing based on Bayesian inference networks. SIGIR-93 (pp. 2234). [32] Vapnik, V. N. (1998). Statistical learning theory. Wiley, New York. [33] Wiener, E. D., Pedersen, J. O., & Weigend, A. S. (1995). A neural network approach to topic spotting. 4th Annual Symposium on Document Analysis and Information Retrieval (pp. 317332). [34] Yang, Y., & Liu, X. (1999). A re-examination of text categorization methods. SIGIR-99 (pp. 4249). [35] Yang, Y., & Pederson, J. O. (1997). A comparative study on feature selection in text categorization. ICML-97 (pp. 412420). 324

multi-class classification;text categorization;GSVD;Discriminant Analysis;Multi-class Text Categorization;SVMs;GDA;discriminant analysis

78

Efficient Phrase Querying with an Auxiliary Index

Search engines need to evaluate queries extremely fast, a challenging task given the vast quantities of data being indexed. A significant proportion of the queries posed to search engines involve phrases. In this paper we consider how phrase queries can be efficiently supported with low disk overheads. Previous research has shown that phrase queries can be rapidly evaluated using nextword indexes, but these indexes are twice as large as conventional inverted files. We propose a combination of nextword indexes with inverted files as a solution to this problem. Our experiments show that combined use of an auxiliary nextword index and a conventional inverted file allow evaluation of phrase queries in half the time required to evaluate such queries with an inverted file alone, and the space overhead is only 10% of the size of the inverted file. Further time savings are available with only slight increases in disk requirements. Categories and Subject Descriptors

INTRODUCTION Search engines are used to find data in response to ad hoc queries. On the Web, most queries consist of simple lists of words. However, a significant fraction of the queries include phrases, where the user has indicated that some of the 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'02, August 11-15, 2002, Tampere, Finland. Copyright 2002 ACM 1-58113-561-0/02/0008 ... $ 5.00. query terms must be adjacent, typically by enclosing them in quotation marks. Phrases have the advantage of being unambiguous concept markers and are therefore viewed as a valuable retrieval technique [5, 6, 7, 10] In this paper, we explore new techniques for efficient evaluation of phrase queries. A standard way to evaluate phrase queries is to use an inverted index, in which for each index term there is a list of postings, and each posting includes a document identifier, an in-document frequency, and a list of offsets. These offsets are the ordinal word positions at which the term occurs in the document. Given such a word-level inverted index and a phrase query, it is straightforward to combine the postings lists for the query terms to identify matching documents. This does not mean, however, that the process is fast. Even with an efficient representation of postings [16], the list for a common term can require several megabytes for each gigabyte of indexed text. Worse, heuristics such as frequency-ordering [13] or impact-ordering [1] are not of value, as the frequency of a word in a document does not determine its frequency of participation in a particular phrase. A crude solution is to use stopping, as is done by some widely-used web search engines (the Google search engine, for example, neglects common words in queries), but this approach means that a small number of queries cannot be evaluated, while many more evaluate incorrectly [12]. Another solution is to index phrases directly, but the set of word pairs in a text collection is large and an index on such phrases difficult to manage. In recent work, we proposed nextword indexes as a way of supporting phrase queries and phrase browsing [2, 3, 15]. In a nextword index, for each index term or firstword there is a list of the words or nextwords that follow that term, together with the documents and word positions at which the firstword and nextword occur as a pair. The disadvantage of a nextword index is its size, typically around half that of the indexed collection. Also, as described originally, nextword index processing is not particularly efficient, as the nextwords must be processed linearly and (compared to an standard inverted index) for rare firstwords the overhead of the additional layer of structure may outweigh the benefits. In this paper we propose that phrase queries be evaluated through a combination of an inverted index on rare words and a form of nextword index on common words. We explore the properties of phrase queries and show experimentally that query evaluation time can be halved if just the three most common firstwords are supported through a nextword index. While phrase browsing is not possible 215 with such an arrangement, the disk overheads of the partial nextword index are small and the benefits are substantial. We have observed that many ordinary queries -- those without quotation marks -- nonetheless resolve successfully if processed as a phrase query, a phenomenon that search engine users are familiar with, as the most popular engines highly rank matches in which the query terms are adjacent. This suggests that phrase querying is a potential method for a fast "first cut" evaluation method, as it allows more rapid identification of documents in which the terms occur as a phrase. PROPERTIES OF QUERIES With large web search engines being used daily by millions of users, it has become straightforward to gather large numbers of queries and see how users are choosing to express their information needs. Some search engine companies have made extracts of their query logs freely available. In our research , we have made extensive use of query logs provided by Excite dating to 1997 and 1999, as well as more recent logs from other sources. These logs have similar properties (with regard to our purposes), and we report primarily on the Excite logs in this work. In the Excite log, after sanitizing to remove obscenity there are 1,583,922 queries (including duplicates). Of these, 132,276 or 8.3% are explicit phrase queries, that is, they include a sequence of two or more words enclosed in quotes. Amongst the rest of the queries--those without a phrase-- about 5% contain a word that does not occur at all in the 21.9 gigabytes (Gb) of data we use. However, almost exactly 41% of the remaining non-phrase queries actually match a phrase in the 21.9 Gb dataset we use in our experiments. A surprising proportion of the phrases include a common term. Amongst the explicit phrase queries, 11,103 or 8.4% include one of the three words that are commonest in our dataset, "the", "to", and "of". 14.4% of the phrase queries include one of the 20 commonest terms. In some of these queries the common word has a structural function only, as in tower of london, and can arguably be safely neglected during query evaluation. In other queries, however, common words play an important role, as in the movie title end of days or the band name the who, and evaluation of these queries is difficult with the common words removed, especially when both "the" and "who" happen to be common terms [12]. Taken together, these observations suggest that stopping of common words will have an unpredictable effect. Stopping may yield efficiency gains, but means that a significant number of queries cannot be correctly evaluated. We experimented with a set of 122,438 phrase queries that between them match 309 10 6 documents. Stopping of common words means that a query such as tower of london must be evaluated as tower -- london: the query evaluation engine knows that the two remaining query terms must appear with a single term between them. If the commonest three words are stopped, there are 390 10 6 total matches for all queries extracted from the log. However, these are distributed extremely unevenly amongst the queries: for some queries the great majority of matches are incorrect. The figure rises to 490 10 6 for the commonest 20 words, and 1693 10 6 for the commonest 254 words, while a significant number of queries, containing only stopped words, cannot be evaluated at all. It can be argued that stopwords are often insignificant, and that even a document that is technically a mismatch-due to the wrong stopword being present--may be just as likely to be relevant as a document where the match is correct . However, it is worth emphasising that there are many queries in which stopwords do play an important role. The words "to" and "from" are often stopped, for example, but mismatches to the query flights to london are likely to be incorrect. Another instance is that the word "the" often forms part of a description, thus the moon should not match websites about a moon of Jupiter, Keith Moon, or a book publisher. Amongst the phrase queries, the median number of words in a phrase is 2, and the average is almost 2.5. About 34% of the queries have three words or more, and 1.3% have six words or more. A few queries are much longer, such as titles: the architect of desire beauty and danger in the stanford white family by suzannah lessard. Another point of interest is where in a phrase the common words occur. In English, the common words rarely terminate a phrase query. Only 0.4% of phrase queries with "the", "to", or "of" have these words at the end. Almost all of these queries are short: virtually no queries of four words or more terminate with one of the commonest terms. In the short queries ending in a common term, the other query terms are themselves usually common. We take advantage of these trends in the methods for phrase query evaluation proposed in this paper. INVERTED INDEXES Inverted indexes are the standard method for supporting queries on large text databases; there are no practical alternatives to inverted indexes that provide sufficiently fast ranked query evaluation. An inverted index is a two-level structure. The upper level is all the index terms for the collection . For text databases, the index terms are usually the words occurring in the text, and all words are included. The lower level is a set of postings lists, one per index term. Following the notation of Zobel and Moffat [17], each posting is a triple of the form: d, f d,t , [o 1 , . . . , o f d,t ] where d is the identifier of a document containing term t, the frequency of t in d is f d,t , and the o values are the positions in d at which t is observed. An example inverted file is shown in Figure 1. In this example, there is a vocabulary of five words, each of which has a postings list. It is straightforward to use an inverted index to evaluate a phrase query. Consider the query magdalene sue prentiss. Of these terms, "magdalene" is the rarest, and its inverted list is fetched first. The postings are decoded and a temporary structure is created, recording which documents contain this word and the ordinal word positions in each document at which it occurs. The term "prentiss" is the next rarest, and is processed next. For each document identifier and word offset in the temporary structure created earlier, a posting is sought to see whether "prentiss" is in the document two words later. If the search fails, that word position is discarded from the temporary structure, as is the document identifier if no word positions for that document remain. As both the structure and the postings are sorted, this process is a linear merge. Then the postings list 216 On Disk Vectors 1,(&lt;9,2,[4,1001]&gt;) 53,(&lt;9,3,[3,8,90] ...) 23,(&lt;4,2,[5,34]&gt;, ...) 243,(&lt;5,1,[45]&gt;,&lt;9,1,[7]&gt; ...) In Memory Vocabulary new in historic hampshire railroads 15,(&lt;1,1,[100]&gt;,&lt;9,1,[6]&gt; ...) Figure 1: An inverted file for a collection with a vocabulary of five words. for "sue" is fetched and decoded, and used to further delete entries from the temporary structure. The remaining entries are documents and word positions at which the phrase occurs. Summarizing, phrase queries are evaluated as follows. 1. Sort the query terms from rarest to commonest, keeping note of their original position in the phrase. 2. Fetch the postings list for the first (rarest) query term. Decode this list into a temporary structure of document identifiers and word offset positions. 3. For each remaining query term, decode its postings list, merging it with the temporary data; this merge process discards from the temporary structure all document identifiers and word offsets that do not match any entry in the postings list. In this query evaluation model, processing of the first query term establishes a superset of the possible locations of the complete phrase, which are maintained in a temporary structure; as the subsequent query terms are evaluated, this structure is pruned, never added to. It is thus essential to begin processing with the rarest query term, to avoid creation of an excessively large temporary structure (or of having to process the inverted lists in stages to stay within a memory limit). A simple heuristic to address this problem is to directly merge the inverted lists rather than decode them in turn. On the one hand, merging has the disadvantage that techniques such as skipping [11] cannot be as easily used to reduce processing costs (although as we discuss later skipping does not necessarily yield significant benefits). On the other hand, merging of at least some of the inverted lists is probably the only viable option when all the query terms are moderately common. Whether the lists are merged or processed in turn, the whole of each list needs to be fetched (unless query processing terminates early due to lack of matches). For ranked query processing it is possible to predict which postings in each inverted list are most likely to be of value, and move these to the front of the inverted list; techniques for such list modification include frequency-ordering [13] and impact-ordering [1]. With these techniques, only the first of the inverted lists need be fetched during evaluation of most queries, greatly reducing costs. In contrast, for phrase querying it is not simple to predict which occurrences of the term will be in a query phrase, and thus such reordering is unlikely to be effective. Offsets only Table 1: Size of inverted index (Mb) after stopping of common words. Number of Index size words stopped (Mb) 0 2350 3 2259 6 2195 10 2135 20 2089 254 1708 have to be decoded when there is a document match, but they still have to be retrieved. Other techniques do have the potential to reduce query evaluation time, in particular skipping [11], in which additional information is placed in inverted lists to reduce the decoding required in regions in the list that cannot contain postings that will match documents that have been identified as potential matches. On older machines, on which CPU cycles were relatively scarce, skipping could yield substantial gains. On current machines, however, disk access costs are the more important factor, and in other experiments we have observed that the increase in length of lists required by skipping outweighs the reduction in decoding time. We therefore do not use skipping in our experiments. We have implemented a phrase query evaluator based on inverted lists, using compression techniques similar to those employed in MG [16] to reduce costs, and have used it to test the efficiency of phrase query evaluation. Our test data is 21.9 Gb of HTML containing about 8.3 Gb of text (drawn from the TREC large web track [9]). Table 1 shows the size of the index with a range of levels of stopping. As can be seen, the three commonest words account for around 4% of the index size, and only small space savings are yielded by stopping. However, as Table 2 shows, the impact of stopping on query evaluation time is dramatic. Just removing the three commonest words reduces average time by about 60%, and by a factor of 3 for longer queries. For these longer queries, the savings continue to increase as more common words are stopped. It is the scale of these savings that make stopping attractive, despite the fact that they are at the cost of inaccurate query results. 217 Table 2: Times for phrase query evaluation (seconds ) on an inverted index after stopping of common words. Results are shown for all queries; 2-word queries only; and 5-word queries only. Number of Overall 2-word 5-word words stopped time (sec) queries queries 0 1.56 0.49 6.41 3 0.66 0.30 1.94 6 0.45 0.29 1.07 10 0.40 0.28 0.81 20 0.37 0.28 0.70 254 0.18 0.16 0.26 NEXTWORD INDEXES Inverted indexes allow evaluation of phrase queries, but faster evaluation is possible with phrase-oriented indexes. One possibility is to use a conventional inverted index in which the terms are word pairs. Another way to support phrase based query modes is to index and store phrases directly [8] or simply by using an inverted index and approximating phrases through a ranked query technique [5, 10]. Greater efficiency, with no additional in-memory space overheads, is possible with a special-purpose structure, the nextword index [15], where search structures are used to accelerate processing of word pairs. The nextword index takes the middle ground by indexing pairs of words and, therefore, is particularly good at resolving phrase queries containing two or more words. As noted above and observed elsewhere, the commonest number of words in a phrase is two [14]. A nextword index is a three-level structure. The highest level is of the distinct index terms in the collection, which we call firstwords. At the middle level, for each firstword there is a data structure (such as a front-coded list, or for fast access a structure such as a tree) of nextwords, which are the words observed to follow that firstword in the indexed text. For example, for the firstword "artificial", nextwords include "intelligence", "insemination", and "hip". At the lowest level, for each nextword there is a postings list of the positions at which that firstword-nextword pair occur. An example nextword index is shown in Figure 2. In this example, there are two firstwords, "in" and "new". Some of the nextwords for "in" are "all", "new", and "the". For each firstword-nextword pair, there is a postings list. (A nextword index is of course a form of inverted index, but for consistency with other work we use "inverted index" solely to refer to a standard word-level inverted file.) In nextword indexes, the postings lists are typically short, because most pairs only occur infrequently. For example, the postings list for the firstword-nextword pair "the" "who" is orders of magnitude smaller than the postings lists for these words in an inverted file. It follows that phrase query evaluation can be extremely fast. Nextword indexes also have the benefit of allowing phrase browsing or phrase querying [4, 15]; given a sequence of words, the index can be used to identify which words follow the sequence, thus providing an alternative mechanism for searching text collections. However, we do not consider phrase browsing further in this paper. For phrase queries of more than two words, multiple postings lists must be fetched from the nextword index to resolve the query. Selection of which listings to fetch requires a little care. For example, with the query boulder municipal employees credit union the query can be resolved by fetching the postings lists for the firstword-nextword pairs "boulder" "municipal", "employees" "credit", and "credit""union". Alternatively, it would be possible to get the lists for "boulder" "municipal", "municipal" "employees", and "credit""union". Which is most efficient depends on which is shorter: the list for "employees" "credit" or the list for for "municipal""employees". Unfortunately, establishing which is shorter requires two disk accesses, to retrieve the nextwords for "employees" and "municipal". However, we have observed that the frequency of a firstword closely correlates to the lengths of its nextword lists. Thus in the query historic railroads in new hampshire we can with confidence choose "railroads" "in" in preference to "in" "new", because "railroads" is much less common than "in". We have considered algorithms for choosing order of evaluation elsewhere [3]. An efficient algorithm for evaluating phrase queries with a nextword index is as follows . 1. If the number of query terms n is even, the query can consist of n/2 disjoint firstword-nextword pairs. If the number of query terms n is odd, n/2 firstword-nextword pairs must be chosen. However, in both cases it is more efficient to choose more than the minimum number of pairs, if doing so avoids choice of a common word as a firstword. 2. The method we use is to choose all n - 1 firstword-nextword pairs; then sort them by increasing firstword frequency; then discard from the list the pairs that are completely covered by preceding selections. This approach can lead to processing of more than n/2 pairs, but experimentally was shown to reduces costs overall. 3. The selected word pairs are sorted by increasing frequency of the firstwords, then their postings lists are processed as for inverted file phrase query processing. The nextword index for our Web collection is 4487 Mb in size, almost exactly twice that of an inverted file. For phrase queries, the savings in query evaluation time are dramatic. Average query evaluation time is reduced to 0.06 seconds, faster than inverted files by a factor of 25. For two-word queries, the time falls to 0.01 seconds, which is faster by a factor of 50. The time for 5-word queries is 0.32. An interesting possibility suggested by these results is that--given space for a nextword index--all queries be evaluated as if they were phrases. We observed above that a significant fraction of all queries successfully evaluate, and indeed on browsing the query logs it is obvious that many of the queries without quotation marks are nonetheless intended to be phrases. Spink et al. [14] suggest that most two-word queries should be treated as a phrase query even if they were entered as a ranked query. Given that search engines return as highest matches the pages in which the 218 In Memory Vocabulary On Disk Nextword Lists in new ... age hampshire house ... the all new On Disk Inverted Vectors 15,(&lt;1,15,[100]&gt;,&lt;65,1,[1]&gt;,&lt;74,7,[23,43,54,62,68,114,181,203]&gt;, ...) 1,(&lt;9,1,[6]&gt;) 3,(&lt;1,1,[12]&gt;,&lt;34,3,[23,34,111]&gt;,&lt;77,1,[29]&gt;) 305,(&lt;9,2,[7,199]&gt;,&lt;532,1,[256]&gt;, ...) 2,(&lt;9,1,[423]&gt;,&lt;19,1,[4]&gt;) 2,(&lt;31,3,[21,41,91]&gt;,&lt;44,1,[34)]&gt;) Figure 2: A nextword index with two firstwords. query words appear in sequence, use of a nextword index provides a rapid mechanism for finding these pages. Much of the speed improvement for phrase queries yielded by nextword indexes is for queries involving a non-rare word. Indeed, for queries of rare words there may be little gain, as query processing with nextword indexes involves more complex structures than does processing with inverted indexes. As the two approaches to phrase query processing appear, then, to have complementary advantages, it is attractive to try to combine their strengths. COMBINED QUERY EVALUATION We have observed above that inverted indexes are the least efficient for phrases involving common words, the case where nextword indexes yield the greatest speed advantage. We therefore propose that common words only be used as firstwords in a stripped-down nextword index, and that this new index be used where possible in evaluation of phrase queries. We call this a top frequency based scheme, since only the most frequent words are indexed in the nextword index. We have explored other schemes based on the frequency of words in the indexed collection, or based on the frequency of words in the query log. None of the investi-gated schemes offered a better space and time trade-off, so we report only results from the top frequency scheme. An example of a top frequency combined index is shown in Figure 3. At the left there is a vocabulary of five words. Each word has an inverted list, together constituting a complete inverted file for these words. In addition, the common words "in" and "new" have a nextword index. With a combined index, processing involves postings lists from both the inverted index and the nextword index. Consider again the query: historic railroads in new hampshire Neither "historic" nor "railroads" is a common word, so establishing that these terms occur in the phrase involves fetching their postings lists from the inverted index and processing in the usual way. However, "in" and "new" are both common. The posting list for the firstword-nextword pair "in" "new" from the nextword index must be fetched and processed. Then there is a choice. On the one hand, the nextword index postings list for "new" "hampshire" cannot be longer than the inverted index postings list for "hampshire" and in all likelihood is a great deal shorter. On the other hand, compared to the inverted index, an extra disk access is required to fetch a postings list from the nextword index. In our implementation, we process using the nextword index if possible, and resort to the inverted index only for terms that are not in an indexed firstword-nextword pair. In summary, we use the following process: 1. Identify all pairs in the list in which the first term is an indexed firstword. Sort these terms, and prune the list as for standard evaluation of phrase queries via a nextword index. 2. For all terms not in a firstword-nextword pair, sort. 3. Process the postings lists in increasing order of firstword frequency, so that processing of nextword index lists and of inverted file lists is interleaved. In this model, a common word need only be evaluated via its postings list in the inverted file if it occurs as the last word in a query, which in the Excite query log is a rare event. We have tested other query resolution methods that involved term sorting based on nextword frequency (or NWF, the number of nextwords for a firstword), inverted document frequency (or IDF, the number of documents in which a word occurs), or both. We also experimented with resolving nextword entries of a given query always first, or always last. We found overall that these different resolution methods did not significantly vary in query speed and behaved almost identically to sorting by IDF only. We therefore sort inverted index terms and nextword terms based on IDF since we do not need to keepanother statistical value per index term and sorting is straightforward. EXPERIMENTAL RESULTS All experiments were run on an Intel 700 MHz Pentium III-based server with 2 Gb of memory, running the Linux operating system under light load. In Table 3 we show sizes of nextword indexes in which only the commonest terms are allowed as firstwords. The table shows that a nextword index that contains only the three commonest terms consumes 254 Mb, that is, just over 10% of the space of the inverted index or around 1% of the size of the original HTML collection . Query evaluation time with a combined index is shown in Table 4. (The "0" line is repeated from Table 2.) As can be seen, use of a nextword index allows evaluation of all phrase queries, and much more rapidly than was previously possible. Use of a partial nextword index of 1% of the HTML collection halves query evaluation time; a partial nextword 219 in hampshire historic new railroads In Memory Vocabulary On Disk Nextword Lists ... the all new 15,(&lt;15,1,[100]&gt;,&lt;65,1,[1]&gt;,&lt;74,7,[23,43,54,62,68,114,181]&gt; ...) 251,(&lt;5,1,[45]&gt;,&lt;9,1,[6]&gt; ...) 1,(&lt;9,1,[7]&gt;) 23,(&lt;9,3,[4,8,245]&gt; ...) 2,(&lt;1,1,[53]&gt;,&lt;9,2,[4,1001&gt;]) 23,(&lt;1,2,[65,98]&gt;,&lt;9,4,[7,54,64,69]&gt; ...) age hampshire house ... 15,(&lt;2,1,[100]&gt;,&lt;6,1,[1]&gt;,&lt;9,8,[1,5,54,62,68,114,181,203]&gt; ...) On Disk Inverted Vectors 3,(&lt;1,1,[12]&gt;,&lt;34,3,[23,34,111]&gt;,&lt;77,1,[29]&gt;) 2,(&lt;31,3,[21,41,91]&gt;,&lt;44,1,[34]&gt;) 305,(&lt;9,2,[7,54]&gt;,&lt;532,1,[256]&gt; ...) 2,(&lt;9,1,[423]&gt;,&lt;19,1,[4]&gt;) Figure 3: A combined inverted file and nextword index. Table 3: Size of nextword index (Mb) containing only common firstwords. Number of Index size common words (Mb) 3 254 6 427 10 520 20 657 254 1366 index of less than 3% of the size of the collection cuts time to a third. These are substantial savings at low cost. Phrase query processing time with a nextword index is only slightly greater than with a stopped inverted file, and no answers are lost. Such combined processing can be integrated with other heuristics for phrase query evaluation. For example, a strategy that is likely to be successful in the context of a web search engine is to maintain indexes (perhaps for a limited time only) on phrases, or word pairs from phrases, that are commonly posed as queries. Amongst our 132,276 queries, 72,184 are distinct. The commonest phrase query (thumbnail post) occurs 205 times and involves no common words. The queries themselves contain 92,846 distinct word pairs; the commonest pair occurs 683 times. Indexing of common query pairs has the potential to yield significant further savings. CONCLUSIONS We have proposed that phrase queries on large text collections be supported by use of a small auxiliary index. In this approach, all words in the text are indexed via an inverted file; in addition, the commonest words are indexed via an auxiliary nextword index, which stores postings lists for firstword-nextword pairs. We have shown that the cost of evaluating phrase indexes can be cut by a factor of three, with an auxiliary index that is only 3% of the size of the Table 4: Times for phrase query evaluation (seconds ) on a combined index, with different numbers of common words used in the nextword index. Results are shown for all queries; 2-word queries only; and 5-word queries only. Number of Overall 2-word 5-word common words time (sec) queries queries 0 1.56 0.49 6.41 3 0.76 0.31 2.99 6 0.57 0.31 2.28 10 0.53 0.30 2.10 20 0.50 0.30 1.98 254 0.46 0.27 1.83 indexed data. These results show that there is no need to use stopping in phrases. Indeed, the statistics on the number of matches indicate that such stopping leads to significant error rates. While it can be argued that mistakes in matching due to stopping of common words are in many cases unimportant, there are numerous queries where the stopwords are significant ; moreover, we have demonstrated that there is no reason to make such mistakes at all. Our schemes have scope for improvement. In particular, choosing of pairs during query evaluation requires further exploration, and we are further investigating structures for representing nextword lists. However, our results show that evaluation of phrase queries can be dramatically accelerated with only a small additional index, and that stopping of phrases leads to errors and is not necessary for efficiency. 220 Acknowledgements This research was supported by the Australian Research Council. We thank Amanda Spink, Doug Cutting, Jack Xu, and Excite Inc. for providing the query log. REFERENCES [1] V. N. Anh, O. Kretser, and A. Moffat. Vector-Space ranking with effective early termination. In W. B. Croft, D. J. Harp er, D. H. Kraft, and J. Zobel, editors, Proc. ACM-SIGIRInternational Conference on Research and Development in Information Retrieval, pages 3542, New York, Sept. 2001. [2] D. Bahle, H.E. Williams, and J. Zobel. Compaction techniques for nextword indexes. In Proc. 8th International Symposium on String Processing and Information Retrieval (SPIRE2001), pages 3345, San Rafael, Chile, 2001. [3] D. Bahle, H. E. Williams, and J. Zobel. Optimised phrase querying and browsing in text databases. In M. Oudshoorn, editor, Proc. Australasian Computer Science Conference, pages 1119, Gold Coast, Australia, Jan. 2001. [4] P. Bruza, R. McArthur, and S. Dennis. Interactive internet search: keyword, directory and query reformulation mechanisms compared. In N. J. Belkin, P. Ingwersen, and M.-K. Leong, editors, Proc. ACM-SIGIRInternational Conference on Research and Development in Information Retrieval, pages 280287, Athens, 2000. [5] C. L. Clarke, G. V. Cormack, and E. A. Tudhope. Relevance ranking for one- to three-term queries. In Proc. of RIAO-97, 5th International Conference "Recherche d'Information Assistee par Ordinateur", pages 388400, Montreal, CA, 1997. [6] W. B. Croft, H. R. Turtle, and D. D. Lewis. The use of phrases and structured queries in information retrieval. In A. Bookstein, Y. Chiaramella, G. Salton, and V. V. Raghavan, editors, Proc. ACM-SIGIR International Conference on Research and Development in Information Retrieval, pages 3245, Chicago, 1991. [7] E. F. de Lima and J. O. Pedersen. Phrase recognition and expansion for short, precision-biased queries based on a query log. In Proc. ACM-SIGIRInternational Conference on Research and Development in Information Retrieval, pages 145152, Berkeley, 1999. [8] C. Gutwin, G. Paynter, I. Witten, C. NevillManning, and E. Frank. Improving browsing in digital libraries with keyphrase indexes. Decision Support Systems, 27(1/2):81104, 1998. [9] D. Hawking, N. Craswell, P. Thistlewaite, and D. Harman. Results and challenges in web search evaluation. In Proc. of the Eighth International World-Wide Web Conference, volume 31, pages 13211330, May 1999. [10] D. D. Lewis and W. B. Croft. Term clustering of syntactic phrases. In J.-L. Vidick, editor, Proc. ACM-SIGIRInternational Conference on Research and Development in Information Retrieval, pages 385404, 1990. [11] A. Moffat and J. Zobel. Self-indexing inverted files for fast text retrieval. ACM Transactions on Information Systems, 14(4):349379, 1996. [12] G. W. Paynter, I. H. Witten, S. J. Cunningham, and G. Buchanan. Scalable browsing for large collections: A case study. In Proc. of the 5th ACM International Conference on Digital Libraries, pages 215223, San Antonio, 2000. [13] M. Persin, J. Zobel, and R. Sacks-Davis. Filtered document retrieval with frequency-sorted indexes. Journal of the American Society for Information Science, 47(10):749764, 1996. [14] A. Spink, D. Wolfram, B. J. Jansen, and T. Saracevic. Searching the web: The public and their queries. Journal of the American Society for Information Science, 52(3):226234, 2001. [15] H. Williams, J. Zobel, and P. Anderson. What's next? index structures for efficient phrase querying. In M. Orlowska, editor, Proc. Australasian Database Conference, pages 141152, Auckland, New Zealand, 1999. [16] I. H. Witten, A. Moffat, and T. C. Bell. Managing Gigabytes: Compressing and Indexing Documents and Images. Morgan Kaufmann, San Francisco, California, second edition, 1999. [17] J. Zobel and A. Moffat. Exploring the similarity space. SIGIRForum, 32(1):1834, 1998. 221

common words;evaluation efficiency;stopping;Indexing;nextword index;index representation;phrase query evaluation;query evaluation;phrase query;inverted index

79

Efficient retrieval of similar shapes

We propose an indexing technique for the fast retrieval of objects in 2D images basedon similarity between their boundary shapes. Our technique is robust in the presence of noise andsupports several important notions of similarity including optimal matches irrespective of variations in orientation and/or position. Our method can also handle size-invariant matches using a normalization technique, although optimality is not guaranteedhere. We implementedour method and performed experiments on real (hand-written digits) data. Our experimental results showedthe superiority of our method comparedto search basedon sequential scanning, which is the only obvious competitor. The performance gain of our method increases with any increase in the number or the size of shapes.

Introduction There is an increasing interest in storing andretrieving non-textual objects in databases. For example, this kind of data can be storedin the form of extenders in DB2, DataBlades in In-formix , and cartridges in Oracle. Non-textual objects are frequently in the form of images or shapes. In cases where the key information for description or classification of an object can be found in its boundary, it is natural to store only the boundary anddo retrieval basedon that. Among the areas of applications for boundary shape matching are industrial inspection, object recognition in satellite images, character recognition, classification of chromosomes, andtarget recognition. For example, consider the following query: Query 1 Findall shapes similar to a given shape. A basic question here is how we judge whether two shapes (for example the two shown in Fig. 1) are similar. There is a large body of work in the area of pattern recognition and computer vision on extracting boundary features of a shape and doing shape matching based on those features. The boundary of an object can be described in terms of simple descriptors such as length, diameter, and curvature ([MM86]), chain Fig. 1. Two shape boundaries both representing character `9' codes ([BG80,Bri81]), Fourier descriptors ([PF77,ZR72]) or moments ([BSA91]). Among these features, we use Fourier descriptors as our shape features. Theoretical andexperimen-tal evidence in favor of Fourier descriptors can be found in the literature [PF77,KSP95]. Similar shapes often have differences in size and orientation . For example, consider the two shapes shown in Fig. 1. The Euclidean distance between their Fourier descriptors is 22.88. If we rotate the shape on the right by 30 in the clockwise (cw) direction, the Euclidean distance between their Fourier descriptors drops to zero. A simple approach to remove differences due to shifting, scaling, and rotation is to normalize Fourier descriptors before storing them in a database. However , there are still two problems with normalization. First, normalization is not guaranteedto minimize the distance between two arbitrary shapes. Second, normalization is not always desirable; for example, the shapes `9' and`6' shouldnot be treatedas similar if we are doing character recognition. A solution is to rewrite the query as follows: Query 2 Findall shapes that become similar to a given shape after being rotatedby [-30 , 30 ]. If our shape collection includes, for example, shapes of airplanes , we may write our query insteadas follows: Query 3 Findall shapes similar to a given shape irrespective of rotation. In this paper, we study the issue of efficiently processing these queries. We show how to organize Fourier descriptors in a multidimensional index, and how to efficiently use the index 18 D. Rafiei, A.O. Mendelzon: Efficient retrieval of similar shapes in processing a broadrange of similarity queries. Our goal is to develop an access method that can handle shapes of various sizes andorientations, is much faster than sequential scanning, and does not miss any qualifying data objects in the answers (false positives are acceptable if they can be eliminatedin a post-processing step without much performance degradation). The organization of the rest of the paper is as follows. Section 2 provides some backgroundmaterial on relatedwork, shape representation using Fourier descriptors and shape matching . In Sect. 3, we propose our technique for indexing shapes andprocessing similarity queries. Section 4 presents experimental results. We conclude in Sect. 5. Background 2.1 Related work The following relevant methods for multidimensional indexing andsearch have been proposed: Jagadish [Jag91] proposes a technique for storing and retrieving shape descriptions in a multidimensional index. He maps shapes into their constituent rectangles, keeps a few larger rectangles in a multidimensional index, and uses the area difference between the constituent rectangles of shapes as a measure of similarity. Due to a normalization process, the shape description is invariant under translation and scaling. A problem with this approach is that a shape can be normally coveredby multiple sets of rectangles. This can leadto ambiguity or storing multiple representations of the same shape. Furthermore, it is not possible to do matching in the presence of rotation; for example, two identical shapes may not match if one is rotatedby 45 . Mehrotra andGary [MG93] decompose a shape into several components anduse fixed-sizedsegments of each component as the shape features. Basedon a normalization process, the shape description is made invariant under translation, scaling , androtation. A problem with this approach is that since a shape is broken down into pieces, the overall shape of the boundary is lost. In addition, each shape is described in terms of multiple feature vectors, andthis introduces extra overhead during insertions and retrievals. Berchtoldet al. [BKK97] study the issue of storing polygons so that they can be retrievedbasedon partial similarity matches. They extract almost all possible boundary segments of polygons, transform each segment into a sequence of slope changes, andmap the resulting sequences into their first few Fourier coefficients. Thus, each polygon is representedusing a set of feature points, andthe minimum bounding rectangle of these points for each polygon is storedin a multidimensional index. Due to a normalization, the shape representation is invariant to translation, scaling, androtation, but it is not invariant to the starting point. This problem is handled by storing multiple descriptions of a polygon, each associated to a starting point. Again, representing a polygon in terms of multiple points introduces extra overhead during insertions and retrievals. The QBIC (Query By Image Content) system [FBF + 94] contains a component for approximate shape matching. The system keeps a 20-D feature vector to describe the shape of x x y , y 1 1 y imaginary axis x real axis 0 0 Fig. 2. A boundary and its representation as a complex sequence every object identified in an image. Features, for example, include the area and the circularity, i.e., whether the object is circular or not. To allow fast retrieval, it is suggestedto transform feature vectors using the Karhunen Loeve (KL) transform and keep a few important features (those associatedwith the few largest eigenvalues) in a multidimensional index. However, the choice of proper features andtheir weighting for each application is not an easy task. Some features are abstract quantities which may not easily fit in a distance function computation. In addition, the use of the KL transform makes the multidimensional index rather static. The aforementionedmethods are less general than ours because the notion of similarity is fixedbefore query evaluation; this notion cannot be changedunless a new index structure is created. Our method, instead, provides a set of transformations to express the notion of similarity in a query; yet, the resulting queries are evaluatedusing the same index, without prior knowledge of the specific transformations used. Therefore we have not comparedthe performance of our method with theirs, but with sequential scanning instead. Related work on time series data includes the work of Agrawal et al. [AFS93] on using the discrete Fourier transform for retrieving similar time series andextensions andimprove-ments over this approach [GK95,RM97,RM00]. Similar to our framework, Goldin and Kanellakis [GK95] show that the similarity retrieval will be roughly invariant to simple translations andscales if sequences are normalizedbefore being stored in the index. The authors store in the index both the translation and the scale factors, in addition to normalized sequences , andalso allow those factors to be queriedusing range predicates (see Goldin's Ph.D. thesis [Gol97] for implementation details). A general framework for composing similarity queries is proposed by Jagadish, Mendelzon, and Milo [JMM95]. Our work here can be seen as a special case of this framework over shapes. Our shape matching can also be incorporatedwithin a multimedia query language such as MOQL [L OSO97] where multiple features of images are simultaneously queried. 2.2 Shape representation using Fourier descriptors Given the figure of an object in the complex plane, its boundary can be tracedproducing a 1-D complex function b t of time. For example, a point moving along the boundary shown in Fig. 2 generates the complex function b t = x t + jy t for D. Rafiei, A.O. Mendelzon: Efficient retrieval of similar shapes 19 t = 0, . . . , N - 1 which is periodic with period N. That is, the x-axis of the figure is treatedas the real axis andthe y-axis as the imaginary axis of a sequence of complex numbers. Further information on tracing the boundary of a shape and possible alternatives in representing it can be foundin any image processing textbook such as Gonzalez andWoods [GW92]. It shouldbe notedthat the description is solely basedon the shape of the boundary; objects can still have holes in them, but this is not reflected in the description. Given a boundary function b t , its Fourier transform can be written as B f = 1 N N-1 t=0 b t e -j2tf N (1) where f { (N - 1)/2 , . . . , 0, . . . , (N - 1)/2 } and j = -1 is the imaginary unit. The coefficients B 0 , B 1 , . . ., called Fourier descriptors, describe the shape of the object in the frequency domain. The transformation is loss-less since the energy in the frequency domain is the same as the energy in the spatial domain (due to Parseval's theorem) and also the inverse Fourier transform gives the original boundary function . 2.3 Shape matching using Fourier descriptors Consider two boundary functions b t = x t + jy t and b t = x t +jy t (for t = 0, . . . , N -1).A typical measure of similarity between the two boundaries is the Euclidean distance, which corresponds to mean-square error and which is also directly relatedto the cross-correlation [Raf98]. D 2 (b, b ) = N-1 t=0 |b t - b t | 2 (2) However, the distance computation becomes ambiguous if the two boundaries have different numbers of samples. A solution to avoidthis problem is to findthe Fourier descriptors B and B , respectively, for b and b anduse a fixednumber of lower frequency descriptors (say, 2M +1) to compute the Euclidean distance, i.e., D 2 (B, B ) = M f=-M |B f - B f | 2 . (3) Our proposal The general overview of the proposedmethodis as follows: 1. Obtain the Fourier descriptors of every shape boundary in the database. 2. Compute a fingerprint for every shape, as discussed in Sect. 3.1, andbuilda multidimensional index using the fingerprints. Each fingerprint is storedas a point in the multidimensional index. 3. For basic similarity queries (proximity, nearest neighbours and all-pairs), use the index to retrieve candidate shapes. The qualifying shapes are identified after retrieving their full database records and examining them. 4. For queries that use transformations in their expressions of similarities, if necessary, apply the transformations to the index, as discussed in Sect. 3.4, and retrieve candidate shapes. The full database record of every candidate shape is examinedto findout if it qualifies. We use Fourier descriptors as our shape features. Given a set of shape boundaries, for each boundary b we findits Fourier transform andretain only a fixednumber of lower frequency descriptors. This number, which we denote by 2M + 1, can be chosen, for example to be the average length of a boundary in the spatial domain. If the number of Fourier descriptors happens to be less than 2M + 1, we store zero for higher frequency descriptors. 3.1 Computing a fingerprint To aidin the retrievals that we intendto perform, we apply a few transformations to the descriptors, rather than storing them directly. First, we set B 0 to 0. B 0 is the only descriptor that carries information about the shape location. This setting minimizes the distance function (Eq. 3) with respect to translation. Next, the scale normalization is achieved by dividing every coefficient B f by the amplitude of B 1 , often calledthe fundamental frequency. |B 1 | turns out to be the largest amplitude when the boundary is traced in the counter-clockwise (ccw) direction and the boundary does not cross itself [WW80]. After the normalization, B 0 is 0, so we do not need to store it. Instead, we store the original value of B 0 before the normalization . It shouldbe notedthat the real andthe imaginary parts of the initial value of B 0 represent the shift factors, respectively, along the X and the Y coordinates; the amplitude of the initial value of B 1 represents the scale factor. To totally get ridof B 1 , which already has an amplitude of 1 for all shapes, we do an additional normalization. We shift the starting point such that the phase of B 1 becomes zero. Definition 3.1. Given the Fourier descriptors B -M , . . . , B M of a shape, denote the real part of B 0 by sh x , the imaginary part of B 0 by sh y , the amplitude of B 1 by sc, and the phase of B 1 by p. The shape description is defined as the sequence (sh x , sh y , sc, S -1 , S 2 , S -2 , S 3 , S -3 , . . . , S M , S -M ). (4) where S i = ((B i - (sh x + sh y j))/sc) e -ipj (a complex number) for i = -1, 2, 3, . . .. The Euclidean distance between two shape descriptions, irrespective of variations in location andsize, can be computedas follows: D 2 (S, S ) = M f=-M,f=0,1 |S f - S f | 2 . (5) Such a description is still sensitive to changes in orientation andstarting point of the tracing. We can assume that every data or query shape has a fixed starting point, if we encode its boundary using the same tracing algorithm and perform the same normalization. For example, a tracing algorithm may always start from the top right corner of a shape andtrace it in the ccw direction. In this way, the starting point for two identical shapes will always be the same. Two similar shapes 20 D. Rafiei, A.O. Mendelzon: Efficient retrieval of similar shapes may still have small variations in their starting points, but those variations can be easily resolvedby allowing some variations in starting points. This is discussed in Sect. 3.4.3. There are sophisticatedtechniques to do phase normalization [PF77,WW80]. For example, Wallace et al. [WW80] suggest making the phases of the two coefficients of largest amplitude equal to zero. This is believed to shift the starting point over the axis of symmetry andalso rotate the axis of symmetry such that it coincides with the real axis. However, it shouldbe notedthat none of these techniques are perfect in the sense that a shape can have two or more different phase normalizations, each as goodas the others; or equivalently, two fairly similar shapes may have descriptors which are far from each other. For the purpose of indexing, important features of the description needto be identifiedandplacedin the fingerprint. First, changing the orientation or the starting point of a boundary only affects the phases of descriptors. To insulate the index from such changes, the information about the phases of descriptors is not stored in a fingerprint. Second, as is shown in Fig. 3, the lower frequency descriptors contain information about the general shape, andthe higher frequency descriptors contain information about smaller details. There are strong reasons to believe that for a large class of boundary functions, the lower frequency descriptors contain most of the energy. For example, for continuous piece-wise smooth functions, the amplitude spectrum |S f | decreases at a rate proportional to f -2 [RH74, Page 373]. Thus, we can define a fingerprint of a shape as follows: Definition 3.2. Given a shape description (sh x , sh y , sc, S -1 , S 2 , S -2 , . . . , S M , S -M ), the fingerprint of the shape is defined as (sh x , sh y , sc, |S -1 |, |S 2 |, |S -2 |, . . . , |S k |, |S -k |) where k ( M) is the cut-off frequency. Next we show the completeness of the feature extraction. 3.2 Using fingerprints for indexing The completeness of the indexing methodis basedon the following lemma: Lemma 3.3. The use of a fingerprint, in place of a full shape description for shape matching always returns a superset of the answer set. Proof: For every pair of boundaries S and S of length 2M + 1 andfor every k M, we have M f=-M,f=0,1 |S f - S f | 2 k f=-k,f=0,1 ||S f | - |S f || 2 (6) This is due to the fact that for every term ||S f | - |S f || in the right side of the inequality, there is a term |S f - S f | in the left side and |S f - S f | ||S f | - |S f ||. Thus, storing the fingerprints of shapes in the index does not affect the correctness since the index returns a superset of the answer set. Furthermore, the distance function on the right side of Eq. 6 is invariant to changes in the starting point of the boundary and rotation. However, the index will not be effective if the choice of k results in a large number of false hits or high index dimensionality (the curse of dimensionality). Our experiments in Sect. 4.2 show that the value of k can be chosen as low as 2 which results in storing 5 Fourier amplitudes in the index. There are a large number of multidimensional data structures which can be used for indexing (see the survey by Gaede andGunther [GG98] for details). We use the R*-tree as it is expectedto work well for up to 20 dimensions andthe length of a fingerprint is expectedto be less than 20. 3.3 Basic similarity queries Within this section, we assume that the shapes being com-paredhave the correct sizes, positions, andorientations. Such a match can also be useful, for example before insertions, to prevent storing two replicas of the same image. We consider the three basic similarity queries over a shape database: (a) proximity query 1 ; (b) all-pairs query; and(c) nearest-neighbours query. In a proximity query, we are given a query shape anda threshold , andwe wouldlike to findall database shapes that are within distance of the query shape. To perform a proximity query, both the shape description and its fingerprint are computedas describedin Sect. 3.1, in the same way as each data shape has been. The fingerprint is then used as a search key into the shape index, to retrieve all data shapes that are locatedin its proximity. Note that the index retrieves a superset of the answer set since it only keeps the fingerprints of shape descriptions. The actual result is obtained in an additional step where the Euclidean distance between the full database record of every matching shape andthe query shape is computed. In an all-pairs query, we are given two data sets and a threshold , andwe want to findall pairs of shapes such that one shape is within distance of the other. To perform an all-pairs query, we do a spatial join between the corresponding indices of the two data sets. This is followed by an additional step where the Euclidean distance between the full database records of matching shapes are computed. In a nearest-neighbours query, we are given a query shape, andwe wish to finddata shapes which are the closest to the query shape in distance. To perform a nearest-neighbours query, both the shape description and its fingerprint are computed (as discussedin Sect. 3.1), andthe fingerprint is usedas a search key over the index. Since the index employs the distance between fingerprints for its pruning andthis distance is an underestimate of the real distance between descriptions, a nearest neighbour identified through searching the index may not be the real nearest neighbour. For example, of the two shapes a and b, a couldbe the closest to the query shape based on the distance between full descriptions, but the index will return b if b is the closest basedon the distance between fingerprints . To fix the problem, we pick the nearest neighbour(s) iden-tifiedthrough the index andcompute the distances between full descriptions of the retrievedshapes andthe query shape. If we denote the minimum distance over all retrieved shapes with , the distance from the real nearest neighbours cannot 1 This is often referredto as a range query as well [AFS93,LJF94]. D. Rafiei, A.O. Mendelzon: Efficient retrieval of similar shapes 21 orginal, N=34 4 descriptors are used 6 descriptors are used 8 descriptors are used 10 descriptors are used 12 descriptors are used Fig. 3. Example of reconstructions from Fourier descriptors be greater than ; otherwise the shapes identified through the index are the nearest neighbours. The full algorithm is as follows : Algorithm 1: 1. Using a nearest-neighbours search algorithm (such as [RKV95]), retrieve the nearest neighbour(s) from the index . 2. For every candidate returned in step 1, retrieve its full database recordandcompute its real distance from the query shape. Let NN be the set of all data shapes at the minimum real distances from the query shape; let be this minimum distance. 3. Using as an initial threshold, pose an incremental proximity query to the index (results are returned one at a time andthe threshold can be tightenedduring the process). 4. Get the next data shape within distance of the query shape. If the distance between the data shape and the query shape is less than , then set NN to be the new data shape and to be the new distance; if the distance between the new data shape and the query shape is , then add the new data shape to NN. Repeat this step until there are no more qualifying data shapes. Algorithm 1 is a refinement of the nearest-neighbours algorithm given by Korn et al. [KSF + 96]. The refinement is in the form of tightening the proximity query thresholdin Step 4 as more data shapes are retrieved. There is another incremental refinement of the same algorithm, proposedby Seidl and Kriegel [SK98], which can also be used. 3.4 Queries with transformations A natural way of doing shape matching is to remove certain differences before running a comparison. We can think of this process as applying certain transformations to images before doing a matching. We consider the following four kinds of transformations: 1. Shifting andscaling. 2. Rotation. 3. Change of starting point. 4. Smoothing. In this section, we center our discussion on proximity queries, but the same techniques are applicable to nearest-neighbours andall-pairs queries. Transformations 1 to 3 can be supportedin a multidimensional index by providing a function that computes the distance between a data shape and a query shape; transformations can be applied to shape descriptions inside the function. Transformation 4 can be supported by registering an additional function that checks if an index entry overlaps with a query entry. The transformation can then be appliedto either the index entry or the query entry (or both) before checking for an overlap. Most multidimensional index structures allow users to define such a function. The next four subsections respectively discuss the eval-uations of queries that use individual transformations 1 to 4 in their expressions of similarities. More details on evaluating queries that use a combination of transformations in their expressions of similarities can be foundelsewhere [RM00]. 3.4.1 Match with shifting or scaling In many cases we do not care about the position of a shape within a coordinate system or about its size for matching purposes . To match with shifting or scaling, a fingerprint is com-putedfor the query shape, as describedin Sect. 3.1, andthis fingerprint is usedas a search key for the index. If we are interested in a match invariant under shifting, we simply discard the shift factor of the query point andpermit any value for the shift factor. Similarly, for scaling-invariant matching, we dis-cardthe scale factor of the query point andpermit any value for the scale factor. 22 D. Rafiei, A.O. Mendelzon: Efficient retrieval of similar shapes 3.4.2 Match with rotation We often wish to match shapes irrespective of small variations in orientation. For example, the two shapes shown in Fig. 1 make a perfect match, if one shape is rotatedby 30 . To achieve this, we state in our query the range of the rotation we wish to perform before doing a shape matching. Query 2, for instance, retrieves all database shapes that match a given query shape after one shape is being rotatedby [-30 , 30 ]. Sometimes, we wouldlike to do matching totally invariant to rotation. For example, we may not care about the orientation at all if we are doing airplane recognition. This can be accom-plishedby simply allowing a rotation of [-180 , 180 ] before matching. To perform a match with rotation, a fingerprint is computed for the query shape andis usedas a search key to the index. The search key is used to retrieve all candidates from the index. These candidates include all data points that match the query point irrespective of rotation factor. They also include false positives, i.e., data points that are not in the proximity of the query point for any rotation factor. To discard false positives, we need to retrieve the full database record of every candidate andcheck whether it actually falls in the proximity (say within distance ) of the query shape after being rotatedby some [ 1 , 2 ]. On the other hand, rotating a shape boundary by is equivalent to multiplying every descriptor S f by e j . We can thus rewrite Eq. 5 to make it reflect the rotation. D 2 (S, S ) = M f=-M,f=0,1 |S f - e j .S f | 2 (7) Lemma 3.4. The minimum and the maximum of Eq. 7 take place at = arctan(-X/Y ) + c. where c is an integer, X = f sin f , Y = f cos f and S f .S f = f e j f ( denotes the complex conjugation 2 ). Since we are interestedin the minimum of Eq. 7 when [ 1 , 2 ] and 1 , 2 , the minimum must take place either at an endpoint (i.e., 1 or 2 ) or any point {arctan(-A/B) - , arctan(-A/B) + , arctan(-A/B)} which is inside the region. It is straightforward to compute the distance function for these values andfindout the optimal rotation factor that results in the minimum distance. 3.4.3 Match with changing starting point When we compare two boundaries, we do not care about their starting points. If we use the same tracing algorithm for every boundary, there cannot be large variations in the starting point (though small variations are still possible). However, we may not have much control over the tracing algorithm, andas a result two similar shapes may have different starting points; or even if we use the same tracing algorithm for all boundaries, we may want to remove small variations (if any) before doing a comparison. Shifting the starting point of a boundary by 3 is equivalent to multiplying every descriptor S f by e jf . This operation , similar to rotation, only affects the phases of Fourier 2 The complex conjugate of z = x+yj is defined as z = x-yj. 3 For example, = 2s 0 /N for a boundary of length N means shifting its starting point by s 0 points in ccw direction. descriptors. Thus, we can still use the index to retrieve all candidates. To discard false positives, we need to retrieve the full database record of every candidate and check whether it still qualifies after the starting point is being shiftedby some [ 1 , 2 ]. We can again rewrite Eq. 5 to make it reflect the shift in starting point. D 2 (S, S ) = M f=-M,f=0,1 |S f - e jf .S f | 2 (8) The optimal value for can be obtainedby equating the derivative of the above equation to zero andfinding the roots. This can be done using numerical techniques up to the machine precision [PTVF92]. 3.4.4 Match with smoothing Occasionally, we wish to do matching based on overall shape, irrespective of small variations in details and noise. In such cases, we wouldlike to smooth out sharp edges andsmall variations before doing the comparison. To achieve this, we can apply a moving average transformation to shape boundaries. When an l-point moving average is appliedto a boundary, every point is replacedwith the average of its l surrounding points. On the other hand, applying a moving average to a boundary in the spatial domain corresponds to a vector multiplication in the frequency domain. For example, to apply a 2-point moving average to a boundary with 10 points, we can equivalently multiply its Fourier descriptors by the Fourier transform of vector m 2 = ( 1 2 , 1 2 , 0, 0, 0, 0, 0, 0, 0, 0). This gives us the Fourier descriptors of the smoothed boundary. A distinguishing feature of smoothing, compared to other transformations discussed in this paper, is that its effect on a shape depends on the characteristics of the shape. This is unlike rotation, for instance, where the effect of rotating a data shape by before a comparison is the same as that of rotating the query shape by -. Given a query shape anda desiredmoving average for smoothing, the matching can be performedas follows: 1. Findthe Fourier transform of the desiredmoving average (as demonstrated for 2-point moving average); let us denote this by M. 2. Transforming the query shape: Apply the transformation to the query shape description (sh x , sh y , sc, Q) by replacing Q with Q where Q i = Q i M i for i = -1, -2, 2, . . . , -k, k. 3. Construct a search key by computing the fingerprint of the new shape description. 4. Transforming the index: Apply M to data entries stored in the index before checking for an overlap between a data entry and the search key; this is done inside the function that checks if a data entry from the index overlaps the search key. 5. For every candidate, retrieve its full database record, apply M to it andcheck if the resulting shape falls in the proximity of Q . The transformation can be appliedto the index on the fly as the index is being traversed. The issue of on-the-fly applying single or multiple transformations to an index is studied in the domain of time series data [RM97,RM00]. The same techniques can be appliedto the domain of shapes. D. Rafiei, A.O. Mendelzon: Efficient retrieval of similar shapes 23 a D=0 b D=0.05 c D=0.10 d D=0.15 e D=0.20 f D=0.25 g D=0.30 i D=0.10 h D=0.40 j D=0.20 Fig. 4. Query shapes, shown in the top two rows, andtheir nearest neighbours, shown in the bottom two rows Experiments To determine the effectiveness of our proposed technique, we implementedour methodandran experiments on a dataset of 11,000 real hand-written digits. The data was obtained from the CEDAR CDROM dataset, which was gathered from scannedZIP codes at the Buffalo Post Office [Hul94]. For every digit, the dataset held 1,100 images. Each image was originally in the form of a 16 16 gray-scale image which was convertedinto a binary image (by thresholding) andwas traced to identify a shape boundary. Then, the boundary was encoded using 30 lower Fourier descriptors. For boundaries with length less than 30, zero was padded at the end. For each shape, both its description and its fingerprint are computed, as outlinedin Sect. 3.1, andusedfor the purpose of indexing. As our index, we used Norbert Beckmann's implementation of the R*-tree [BKSS90]. For the nearest-neighbours search, we implementedthe algorithm developedby Roussopoulos et al. [RKV95] as part of Algorithm 1 over R*-tree. We stored 10,000 shapes (1,000 samples of each digit) in the index and usedthe 1,000 remaining samples as queries. We ran each query 10 times andaveragedthe execution times from these runs. All our experiments were conducted on a 168 MHz Ul-trasparc station. We investigatedthe following questions: How effective andpractical is our technique in classifying shapes in a real data domain? How many Fourier coefficients shouldwe store in the index ? Storing larger number of coefficients reduces the number of false positives but increases the index dimensionality , andas a result the search time. How does our technique compare to sequential scanning? 4.1 Shape classification To verify the effectiveness of our proposedtechnique in classifying shapes, we triedto classify all 1,000 query shapes by assigning every query shape to the class of its nearest neighbours . When there was more than one nearest neighbours for a shape, we pickedone randomly. The result was interesting: 96.4% of shapes were classifiedcorrectly. Some of those query shapes are shown, in their gray scale andbinary representation , in the two top rows of Fig. 4 along with their nearest neighbours shown in the two bottom rows of the same figure. Table 1. Various ranges of rotations andtheir effects in correctly classifying the shapes of hand-written digits Rotation factor Fraction of query shapes classifiedcorrectly (%) [0, 0] 96.4% [-10, 10] 96.5% [-20, 20] 96.4% [-30, 30] 96.4% [-40, 40] 96.3% [-50, 50] 96.3% As is shown, query shapes shown in Figs. 4a to 4h are classified correctly with their Euclidean distances from their nearest neighbours varying from 0 to 0.40. The query shape shown in Fig. 4i is not classifiedcorrectly, but its binary representation looks quite similar to that of its nearest neighbour. The query shape shown in Fig. 4j looks different from its nearest neighbour, though their boundaries still look similar. In another experiment, we usedQuery 2 andtriedto identify for each query shape its nearest neighbour irrespective of a rotation factor [-30 , 30 ]. This did not change the overall classification rate, i.e., only 96.4% of shapes were classifiedcorrectly . However, allowing a rotation factor in general did retrieve better matches. Figure 5 shows six query shapes (in the top two rows), their original nearest neighbours (in the middle two rows) and their optimal nearest neighbours (in the bottom two rows) when the rotation factor variedfrom -30 to 30 . As is shown, for example rotating the data shape shown at the bottom of Fig. 5a by 11 in the ccw direction reduces its Euclidean distance from the query shape to 0.30; this is less than the Euclidean distance between the same query shape andits original nearest neighbour. Table 1 summarizes the effect of various rotations in correctly classifying shapes. As is shown, applying a small rotation ( [-10, 10]) to d ata shapes before matches slightly improves the classification rate of hand-written digits; larger rotations, on the other hand, either have no effect or deteriorate the rate of correct classifications . This is because the digit data is generally sensitive to orientations andallowing larger rotations can potentially retrieve more non-identical digits. We later picked1,000 shapes among those storedin the database, applied to each shape a random rotation in the range [-, ] andusedit as a query shape. We only specifiedthe 24 D. Rafiei, A.O. Mendelzon: Efficient retrieval of similar shapes a D=0.34 R=11,D=0.30 b D=0.25 R=-11,D=0.22 c D=0.36 R=9,D=0.35 d D=0.60 R=15,D=0.57 e D=0.45 R=27,D=0.35 f D=0.48 R=-12,D=0.41 Fig. 5. Query shapes (the top two rows), their original nearest neighbours (the middle two rows) and their optimal nearest neighbours (the bottom two rows) varying the rotation factor in [-30 , 30 ] 2 4 6 8 10 0.3 0.4 0.5 0.6 0.7 0.8 Number of Fourier amplitudes Fraction of shapes classified correctly Fig. 6. The fraction of query shapes classifiedcorrectly, varying the number of Fourier amplitudes used for classification rotation interval in our query. As expected, for each query shape, only the shape itself was retrievedfrom the database. 4.2 Varying the cut-off frequency The effectiveness of the index mainly depends on the concentration of the key shape information within a few descriptors of fingerprints. To measure this effectiveness, we ran some experiments varying the number of Fourier descriptors stored in a fingerprint. Figure 6 shows the ratio of query shapes that are classifiedcorrectly (according to the criteria outlinedin Sect. 4.1) to all query shapes varying the number of Fourier amplitudes used for classification. As the number of amplitudes increases up to 6, the ratio of shapes that are classified correctly increases accordingly up to 0.778. This ratio remains the same despite increasing the number of Fourier amplitudes from 6 to 10. Comparedto a full shape description which consists of both the amplitudes and the phases of 30 lower Fourier coefficients, classifying 96.4% of the shapes correctly, a fingerprint does a pretty good job using only 6 amplitudes which make up only 10% of a full shape description and still classifying 0.778% of the shapes correctly. Figure 7a shows the average execution time of Algorithm 1 for 1,000 nearest-neighbours queries, broken into: (1) search time in Step 1 to identify the initial approximate nearest neighbours ; and(2) search time in Step 3 to findthe real nearest neighbours. Figure 7b shows the fraction of index nodes accessed , averaged over 1,000 nearest-neighbours queries, again broken into the fractions accessedin Step 1 andStep 3. As the number of Fourier amplitudes increases, the index selectivity improves, i.e., the index gives fewer false hits. The number of false hits, as is depicted in Fig. 8 for a proximity query, mainly depends on the number of Fourier amplitudes usedin fingerprints andthe output size of the query. Due to the high similarity between different shapes of the same digit, a large fraction of our false hits (for example, 62% when the output size was 11 andthe number of Fourier amplitudes was 6) were other shapes of the same digit depicted by the query shape which were not within the specifieddistance of the query shape. The reduction in false hits reduces the search time since less time is needed to remove those false hits. However, increasing the number of Fourier amplitudes after some point, often calledthe cut-off frequency, either does not reduce the number of false hits or reduces it only slightly. This is because higher frequency amplitudes carry less of the energy than lower frequency ones. On the other hand, the search time increases with the index dimensionality, because the tree becomes deeper. Furthermore, the pruning becomes harder, as is shown in Fig. 7 with the ratio of index nodes that are accessed , because the probability of an arbitrary data bounding rectangle being close to the query point increases with the dimensionality. Given the trade-off between the tree search time and the time spent for removing false hits, it is natural to expect that there is an optimal cut-off frequency. Basedon our experiments , as illustratedin Figs. 6 and7, the optimal cut-off frequency occurs for as few as 6 Fourier amplitudes. 4.3 Comparison to sequential scanning Figure 9 shows the average execution time of our proposed methodcomparedto sequential scanning for 1,000 nearest-neighbours queries. To get its best performance, we used bufferedinput for sequential scanning, in a system with buffer size of 8,192 bytes. For the experiment shown in Fig. 9a, the border length was fixed to 30 while the database size variedfrom 10,000 to 30,000 shapes. Since the size of dataset was limited, we doubled or tripled the size by adding one or two randomly rotated copies of each shape to the database. This doubling did not affect the performance of sequential scanning, which was linear in the input size, but we expected the doubling to deteriorate the performance of our method since high similarity among database shapes would increase the number of false hits. For the experiment shown in Fig. 9b, the number of shapes was fixedat 10,000 while the number of Fourier descriptors used to represent a boundary varied from D. Rafiei, A.O. Mendelzon: Efficient retrieval of similar shapes 25 2 4 6 8 10 0 200 400 600 800 1000 1200 Number of Fourier amplitudes Execution time (msec) : initial NN query : proximity query : Total a 2 4 6 8 10 0 0.2 0.4 0.6 0.8 Number of Fourier amplitudes Fraction of index nodes accessed b : initial NN query : proximity query : Total Fig. 7. Break up of a the execution time and b the fraction of index nodes accessed, for nearest-neighbours queries, varying the number of Fourier amplitudes 2 3 4 5 6 10 20 30 40 50 60 70 Number of Fourier amplitudes # of false hits per every qualifying shape a 0 200 400 600 800 1000 6 8 10 12 14 16 18 20 Output size # of false hits per every qualifying shape b Fig. 8. The average number of false positives for every qualifying shape a varying the number of Fourier amplitudes and fixing the average output size of the query to 11, and b varying the average output size andfixing the number of Fourier amplitudes to 6 1 1.5 2 2.5 3 x 10 4 0 1000 2000 3000 4000 5000 6000 Number of shapes Execution time (msec) : Seq : Index a 20 30 40 50 0 500 1000 1500 2000 2500 3000 Border Length Execution time (msec) : Seq : Index b Fig. 9. a Time per query varying the number of shapes, for nearest-neighbours queries. b Time per query varying the border length, for nearest-neighbours queries 26 D. Rafiei, A.O. Mendelzon: Efficient retrieval of similar shapes 20 to 50. As shown in the figure, increasing either the number of shapes or the border length increases the relative advantage of our method, making it more attractive for large databases. Conclusions We have proposedan indexing technique that can efficiently retrieve images of objects basedon similarity between their boundary shapes. We have used Fourier descriptors as our shape features andhave developedan index organization such that similar shapes can be easily retrievedirrespective of their differences in size, position and orientation. The highlight of our contribution is an index structure that helps find optimal matches between shapes irrespective of various differences between them. Our technique has the following desirable properties : It uses a shape matching mechanism which is well-studied in the area of pattern recognition. It exploits the fact that important features of a large class of shapes are concentratedwithin only a few Fourier descriptors . It can handle shapes of various sizes. It guarantees efficient retrieval of all qualifying shapes. Furthermore, we have presenteda refinement of an earlier nearest-neighbours search algorithm for feature vectors that are truncated, due to the significance of some features over others, before being storedin a R-tree index. Acknowledgement. We thank the Natural Sciences andEngineering Research Council of Canada andthe Institute for Robotics andIntel-ligent Systems for their support of this work. References [AFS93] Agrawal R, Faloutsos C, Swami A (1993) Efficient similarity search in sequence databases. In: Proc. 4th International Conference on Foundations of Data Organizations andAlgorithms (FODO '93), pp 6984, Chicago [BG80] Bribiesca E, Guzman A (1980) How to describe pure form andhow to measure differences in shape using shape numbers . Pattern Recognition 12(2):101112 [BKK97] BerchtoldS, Keim DA, Kriegel HP (1997) Using extended feature objects for partial similarity retrieval. VLDB J 6(4):333348 [BKSS90] Beckmann N, Kriegel HP, Schneider R, Seeger B (1990) The R* tree: an efficient androbust index methodfor points andrectangles. In: Proc. ACM SIGMOD International Conference on Management of Data, pp 322331, Atlantic City [Bri81] Bribiesca E (1981) Arithmetic operations among shapes using shape numbers. Pattern Recognition 13(2):123138 [BSA91] Belkasim SO, Shridhar M, Ahmadi M (1991) Pattern recognition with invariants: a comprehensive study and new results. Pattern Recognition 24:11171138 [FBF + 94] Faloutsos C, Barber R, Flickner M, Niblack W, Petkovic D, Equitz W (1994) Efficient andeffective querying by image content. J Intell Inf Syst 3(3/4):231262 [GG98] Gaede V, Gunther O (1998) Multidimensional access methods. ACM Comput Surv 30(2):170231 [GK95] Goldin DQ, Kanellakis PC (1995) On similarity queries for time-series data: constraint specification and implementation . In: 1st Int. Conference on the Principles and Practice of Constraint Programming, Lecture Notes in Computer Science, vol. 976. Springer, Berlin Heidelberg New York, pp. 137153 [Gol97] Goldin DQ (1997) Constraint query algebras . PhD thesis, Brown University, www.cs.brown.edu/people/dgk/Papers/thesis.ps [GW92] Gonzalez RC, Woods RE (1992) Digital image processing. Addison-Wesley, Reading, Mass., USA [Hul94] Hull JJ (1994) A database for handwritten text recognition research. IEEE Trans Pattern Anal Mache Intell 16(5):550554 [Jag91] Jagadish HV (1991) A retrieval technique for similar shapes. In: Proc. ACM SIGMOD International Conference on Management of Data, pp 208217, Denver, Colo., USA [JMM95] Jagadish HV, Mendelzon AO, Milo T (1995) Similarity-basedqueries . In: Proc. 14th ACM SIGACT-SIGMOD-SIGART Symposium on Principles of Database Systems, pp 3645, San Jose, Calif., USA [KSF + 96] Korn F, Sidiropoulos N, Faloutsos C, Siegel E, Protopa-pas Z (1996) Fast nearest neighbor search in medical image databases. In: Proc. 22nd International Conference on Very Large Data Bases, pp 215226, Mumbai, India [KSP95] Kauppinen H, Seppanen T, Pietikainen M (1995) An experimental comparison of autoregressive andFourier-baseddescriptors in 2D shape classification. IEEE Trans Pattern Anal Mach Intell 17(2):201207 [LJF94] Lin KI, Jagadish HV, Faloutsos C (1994) The TV-tree - an index structure for high-dimensional data. VLDB J 3(4):517542 [L OSO97] Li JZ, Ozsu MT, Szafron D, Oria V (1997) MOQL: a multimedia object query language. In: Proc. 3rd International Workshop on Multimedia Information Systems, pp 1928 [MG93] Mehrotra R, Gary JE (1993) Feature-basedretrieval of similar shapes. In: Proc. 9th International Conference on Data Engineering, pp 108115, Vienna, Austria [MM86] Mokhtarian F, MackworthA (1986) A scale-baseddescrip-tion andrecognition of planar curves andtwo dimensional shapes. IEEE Trans Pattern Anal Mach Intell 8(1):3443 [PF77] Persoon E, Fu KS (1977) Shape discrimination using Fourier descriptors. IEEE Trans Syst Man Cybern 7(2):170179 [PTVF92] Press WH, Teukolsky SA, Vetterling WT, Flannery BP (1992) Numerical recipes in C. Cambridge University, Cambridge, UK [Raf98] Rafiei D (1998) Fourier-transform basedtechniques in efficient retrieval of similar time sequences. PhD thesis, University of Toronto [RH74] RichardCW, Hemami H (1974) Identification of three-dimensional objects using Fourier descriptors of the boundary curve. IEEE Trans Syst Man Cybern 4:371378 [RKV95] Roussopoulos N, Kelley S, Vincent F (1995) Nearest neighbor queries. In: Proc. ACM SIGMOD International Conference on Management of Data, pp 7179, San Jose, Calif., USA [RM97] Rafiei D, Mendelzon AO (1997) Similarity-based queries for time series data. In: Proc. ACM SIGMOD International Conference on Management of Data, pp 1324, Tucson, Ariz., USA D. Rafiei, A.O. Mendelzon: Efficient retrieval of similar shapes 27 [RM00] Rafiei D, Mendelzon AO (2000) Querying time series data based on similarity. IEEE Trans Knowl Data Eng 12(5):675693 [SK98] Seidl T, Kriegel HP (1998) Optimal multi-step nearest neighbour search. In: Proc. ACM SIGMOD International Conference on Management of Data, pp 154165, Seattle, Wash., USA [WW80] Wallace TP, Wintz PA (1980) An efficient three-dimensional aircraft recognition algorithm using normal-izedfourier descriptors. Comput Graph Image Process 13:99126 [ZR72] Zahn CT, Roskies RZ (1972) Fourier descriptors for plane closedcurves. IEEE Trans Comput 21(3):269281

fingerprint;Shape retrieval Similarity retrieval Fourier descriptors;non textual objects;efficiency;database;handwriting recognition;Fourier descriptors;Image databases;search;queries;shape classification;indexing techniques;Similarity queries

8

A Flexible 3D Slicer for Voxelization Using Graphics Hardware

In this paper we present a simple but general 3D slicer for voxelizing polygonal model. Instead of voxelizing a model by projecting and rasterizing triangles with clipping planes, the distance field is used for more accurate and stable voxelization. Distance transform is used with triangles on voxels of each slice. A voxel is marked with opacity only when the shortest distance between it and triangles is judged as intersection. With advanced programmable graphics hardware assistance, surface and solid voxelization are feasible and more efficient than on a CPU.

Introduction Object representation is a broad topic in research. In computer graphics, polygons play a dominant role in 3D graphics because they approximate arbitrary surfaces by meshes. In games and animations , surface representation is the main technique used in rendering and visualization. However, volumetric representation, an alternative method to traditional geometric representation, is well known since the 1980s. It provides a simple and uniform description to measure and model a volumetric objects and establish the research field of volumetric graphics. Voxelization is a process of constructing a volumetric representation of an object. Voxelizing a polygonal object is not only a shift of the representation, it gives an opportunity to manipulate mesh object with volumetric operations such as morphological operation and solid modeling. Many applications, for example, CSG modeling, virtual medicine, haptic rendering, visualization of geometric model, collision detection, 3D spatial analysis, and model fixing, work on volumetric representation or use it as the inter-medium. In this paper, we calculate the accurate distance field on GPU to compute coverage of each voxel in voxelization for polygonal models . Our method works for arbitrary triangulated models without any preprocessing for models, except organizing meshes slab by slab in order to prune the unnecessary computation while voxelizing complex models. By using the power of GPU, Hausdorff distance is guaranteed between each sampled voxel and the polygonal model. Surface voxelization with distance field on a GPU works well and runs faster than on a PC with a CPU. Our method is a reliable and accurate solution for the polygonal model under a given distribution of sampling voxels and a specific opacity criterion. Besides , error tolerance in voxelization is easy to manipulate by adjusting the threshold of opacity criterion for voxelization, which also dominates the smoothness of features and the thickness of surface voxelization. e-mail:wktai@mail.ndhu.edu.tw The rest of paper is organized as follows. Some related works are surveyed in the section 2. In section 3, we present the computation of Hausdorff distance and our framework. The experimental results are illustrated in section 4. Finally, we conclude the proposed approach and point out some future works. Related Works Volume construction approaches are often referred to as scan-conversion or voxelization methods. Researchers mainly focused on modeling aspects such as robustness and accuracy. Wang and Kaufman [Wang and Kaufman 1993] used a method that samples and filters the voxels in 3D space to produce alias-free 3D volume models. They used filters to produce final density from the support of the region that polygons lie inside. Schroeder and Lorensen [Schroeder et al. 1994] created a volumetric model by finding clos-est polygon from distance map and classify the opacity of voxels. Huang et al. [Huang et al. 1998] described separability and min-imality as two desirable features of a discrete surface representation and extended 2D scan line algorithm to perform voxelization. Dachille and Kaufman [Dachille IX and Kaufman 2000] presented an incremental method for voxelizing a triangle with pre-filtering to generate multivalued voxelization. Widjaya et al. [Widjaya et al. 2003] presented the voxelization in common sampling lattices as general 2D lattices including hexagonal lattices and 3D body-center cubic lattices. Ju [Ju 2004] constructed an octree grid for recording intersected edges with the model and expanded nodes to scan-convert a polygon on an octree, and then generate signs from the boundary edges and faces of the octree. In recent years, attention on performance of voxelization raises. More and more studies try to explore the benefits of graphics hardware for more efficient rendering. Chen and Feng [Chen and Fang 1999] presented a slicing-based voxelization algorithm to generate slices of the underlying model in the frame buffer by setting appropriate clipping planes and extracting each slice of the model, which is extended and published in later [Chen and Fang 2000]. Karabassi and Theoharis [Karabassi and Theoharis 1999] projected an object to six faces of its bounding box through standard graphics system for the outermost parts and read back the information from depth buffer. However it works well only on convex objects. Dong et al. [Dong et al. 2004] proposed a real-time voxelization method using GPU acceleration to rasterize and texelize an object into three directional textures and then synthesize textures back to the final volume. 285 Hausdorff Distance Computation and Voxelization In this section, we first discuss the computation of Hausdorff distance between a given triangle and a point. Then we explain how we use GPU to compute the distance field of triangles and modify the rendering pipeline. 3.1 Distance Field Computation For a given 3D point P(x, y, z) and a triangle T (V 0 ,V 1 ,V 2 ), Hausdorff distance is the shortest distance between the point P and any point v on the triangle. A point on triangle T can be parametrically defined by two linearly independent vectors with two weights (s,t) by T (s,t) = B + se 0 + te 1 , (1) where (s,t) D = {(s,t) : s [0, 1],t [0, 1], s + t 1}, and B = V 0 , e 0 = V 1 -V 0 and e 1 = V 2 -V 0 . For any point on triangle T , the distance from T to P is T (s,t) - P , (2) or we can use the squared-distance function instead Q (s,t) = T (s,t) - P 2 , (3) where a point p = ( s , t) exists which makes Q( s , t) minimum. Therefore, the computation of distance can be reduced into a min-imization problem. For an efficient computation, we can expand Q (s,t) as Q (s,t) = as 2 + 2bst + ct 2 + 2ds + 2et + f , (4) where a = e 0 e 0 b = e 0 e 1 c = e 1 e 1 d = e 0 (B - P) e = e 1 (B - P) f = (B - P) (B - P) (5) From analyzing the gradient of Q(s,t), the minimum s and t happens only when Q is zero, where s = be - cd ac - b 2 t = bd - ae ac - b 2 (6) If ( s , t) D, the minimum distance is the distance between p and P ; otherwise, according to the sign of s and t, there are six possible regions that the shortest distance point p may lie on triangle T , as shown in Figure 1. Efficient solutions are well addressed on the book [Schneider and Eberly 2003] by CPU computation with simple calculation and logic classification. However, in GPU, there is no efficient dynamic flow control to determine the shortest point on a triangle. Therefore, instead of directly computing the point of shortest distance on a triangle, we compute the distance from the 3D point to four possible points which may be inside the triangle or on the three boundaries and then the minimum scalar is the shortest distance. These four points are (s 0 ,t 0 ) = ( be -cd ac -b 2 , bd -ae ac -b 2 ) (s 1 ,t 1 ) = (0, e c ) (s 3 ,t 3 ) = ( c +e-b-d a -2b+c , a +d-b-e a -2b+c ) (s 2 ,t 2 ) = (d a , 0), (7) Figure 1: Six regions in s,t coordinate. Space is partitioned by range of parameters s and t for efficient shortest position classification and calculation. Figure 2: Rendering pipeline for generating a distance field of a triangle. A quad is rendered instead of a triangle. Five channels of each vertex of a quad (position, normal, and 4 texture coordinates) is filled with position of quad, position of voxel, and information of a triangle: v 0 , e 0 , e 1 , and normal N respectively. where position (s 0 ,t 0 ) assumes point p is inside the triangle, positions (s 1 ,t 1 ), (s 2 ,t 2 ) and (s 3 ,t 3 ) assume point p is on boundaries of s = 0, t = 0, and s + t = 1. All calculated points are truncated in the range of [0, 1] so that three end vertices of the triangle T are also in consideration and it guarantees these points are on the triangle for distance computation. Therefore, the minimum distance is the shortest distance from the point P to the triangle T . 3.2 Geometry Rendering for Voxelization Voxelization by projection and rasterization faces the difficulty of non-uniform sampling of polygons because polygons with arbitrary orientations are not parallel to projection plane for maximum projection area. Even classifying polygons and projecting them to individual best-fit plane, there still have no guarantee on valid rasterization . However, distance field is omni-directional, i.e., insensitive to the projection plane, and has no assumption on input geometry and therefore no extra preprocessing is required. Our approach is a slice-based approach for distance field generation and voxelization. Figure 2 shows the rendering process for generating the distance field for a triangle. For each triangle T i = {T i (s,t)|v 0 + se 0 +te 1 , s 0,t 0, s +t 1}, a full-filled quad Q i = {q i 0 , q i 1 , q i 2 , q i 3 } is rendered and rasterized to generate the distance field from voxels on a slice to the triangle. Triangle data and voxel positions are associated with the rendering quads. Voxel positions are stored in the channel of vertex normal, the triangle data (base vertex B, vectors e 0 and te 1 , and the normal N) are separately stored in channels of texture coordinates and transmitted to GPU. Voxel positions are linearly interpolated in rasterization of rendering pipeline and pairs of triangle data and voxel positions are sent to pixel processors for Hausdorff distance computation. After distance computation, the shortest distance between a triangle and a 286 voxel is stored in the pixel depth and the pixel color is assigned for identification depending on applications. For example, binary surface voxelization uses color information to identify whether a voxel intersects a geometry such as 0 for empty and 1 for opacity; distance visualization uses color information to display the distance from geometries, etc. The distance field of polygonal objects is constructed incrementally by rendering quads of triangles. Each pixel of depth buffer keeps the shortest distance from its voxel to triangles rendered. Depth buffer updates when a triangle is rendered. Unless distance is recal-culated on different slice or rendered objects deform, quads which have been rendered have no need to be re-rendered again even new geometry are added in the scene. Depth buffer of the viewport is initialized as infinitude. The rendering pseudo code is abstracted as follows: for each triangle t on slice i { Create a quad Q for the triangle t for k = 0 to 3 { // assign a full-filled quad // q is end vertices of quad Q.q[k].position = ScreenBoundary.q[k]; // assign voxel position, and triangle data Q.q[k].normal = Slice[i].q[k]; Q.q[k].tex0 = t.B; Q.q[k].tex1 = t.e0; Q.q[k].tex2 = t.e1; } RenderQuad(Q); } 3.3 Surface Voxelization We use local distance field to reduce work load of GPU because the distance field far away from a triangle is meaningless for surface voxelization. For each triangle, we extend its parallel projected bounding rectangle by a small scalar for effective rasterization especially for triangles perpendicular to the projection plane. Due to coherence and precision in interpolating voxel positions, triangles are rendered with extended bounding rectangles. While a pixel is rasterized by a quad, Hausdorff distance is calculated according to the interpolated voxels, i.e., centers of voxels, and the triangle data. Only if the distance is less than the given threshold, e.g., distance from a uniform voxel center to its corner, the pixel is marked as opacity. Using local distance field could guarantee a small region of Hausdorff distance but greatly improve the performance of surface voxelization. For more efficient voxelization process on GPU, triangles can be culled early by space partitioning. We construct an active triangle list for each slice. Currently we define slabs along Z-axis. According to partitioning planes, triangles are filtered and rearranged slab by slab. Many triangles can be pruned while rendering a slice. It is significantly helpful while voxelizing very complex models. Because distance field is insensitive to projecting directions of triangles , selection of partitioning plane has no influence on effective-ness of voxelization. Experimental Results We implement our fragment program using HLSL on a Pentium 4 3.0 MHz PC with 1G RAM and a nVidia Geforce FX5800 graphics card running on Windows XP with DirectX 9.0c. We use Vertex Shader 1.1 and Pixel Shader 2.0 to implement fragment program in scattering pairs of voxel positions and triangle data and in distance calculation and visualization. Table 1 shows the performance Figure 4: Rendering from the results of voxelization (512 3 ): dragon of 870K faces in 512 3 voxels. of surface voxelization on different models and in different voxel resolutions. Figure 3 and Figure 4 demonstrate quality of voxelization results. In the experiment, opacity threshold is set to the distance from voxel center to its corner. That means if the shortest distance between a voxel and a triangle is less than the threshold, the voxel will be marked as opacity. Note that voxels are normal-ized to cubes in rendering so the scale of output may differ from the original polygonal model. In Table 1, we list average time on surface voxelization per slice, per voxel, and per triangle. In the same resolution, voxelization time is proportional to the complexity of polygonal model. For each voxel, process time is always less than 0.1 ms. Even when voxel resolution increase, GPU still could handle voxelization for complex object in stable throughput which may be increased much more for CPU. Due to speed up by using local distance field and culling for unre-lated geometry, voxelization by distance field can be displayed slice by slice interactively under 128 3 voxel resolution. When voxel resolution is low, voxelization time highly depends on complexity of model. However, when the voxel resolution increases higher, even for a low complexity model, it still need more time to voxelize a model than in lower voxel resolution. On average, resolution of 256 3 could provide a benefit of reliable voxelization both in quality and time cost. Rendering cost is stable for a triangle even when resolution of volume increases while it is linear on a CPU. Voxelization with proposed method is still slower than methods using traditional projection and rasterization by graphic hardware. However, our method is stable, correct and flexible because the opacity of each voxel is determined by thresholding the distance field of objects. Conclusion In this paper, we propose a GPU-based 3D slicer approach for voxelizing polygonal models. We calculate minimum distance between pairs of sampled voxels and triangles of arbitrary models with guarantee of Hausdorff distance. With programmable hardware ver-tex/pixel processors, efficient surface voxelization, solid voxelization , and visualization of the distance field all are feasible on the proposed 3D slicer. However, in current implementation, performance of pixel shader is the bottleneck in overall processing speed. Area of rasterization also has a significant influence on the loading of pixel shader. 287 Model Faces Res. Time(s) Time Slices Time Voxels Time Tri . Res. Time(s) Time Slices Time Voxels Time Tri . Beethoven 5027 128 13.94 0.11 6.65 2.77 256 85.86 0.34 5.12 17.08 Teapot 6320 128 14.31 0.11 6.82 2.26 256 87.62 0.34 5.22 13.86 Cup 7494 128 15.24 0.12 7.27 2.03 256 93.65 0.37 5.58 12.50 Bunny 10000 128 16.24 0.13 7.75 1.62 256 93.98 0.37 5.60 9.40 Bunny 69451 128 43.21 0.34 20.60 0.62 256 231.85 0.91 13.82 3.34 Dragon 871414 128 84.21 0.66 40.15 0.10 256 325.87 1.27 19.42 0.37 Buddha 1087716 128 170.44 1.33 81.27 0.16 256 347.65 1.36 20.72 0.32 Dragon 871414 512 1748.11 3.41 13.02 2.01 * The time unit in Time/Voxels is s Buddha 1087716 512 1825.47 3.57 13.60 1.68 * The time unit in Time/Tri. is ms Table 1: Surface voxelization on different models and in different voxel resolutions. (a) (b) (c) (d) (e) (f) (g) (h) Figure 3: Rendering from the results of voxelization (256 3 ): (a) Beethoven in 256 3 voxels, (b) Teapot in 256 3 voxels, (c) Cup in 256 3 voxels, (d) Bunny of 10000 faces in 256 3 voxels, (e) dragon of 10000 faces in 256 3 voxels, (f) Bunny of 69451 faces in 256 3 voxels, (g) dragon of 870K faces in 256 3 voxels, and (h) Buddha of 1M faces in 256 3 voxels. Therefore, in the near feature, searching a better computational methodology for GPU is one direction to improve performance of distance field computation. In addition, a sophisticated culling for error-free distance computation will be a technique in demand. To improve the quality of voxelization, adaptive dense voxelization and a mechanism for quality measurement and guide on GPU is another interesting topic. References C HEN , H., AND F ANG , S. 1999. Fast voxelization of 3D synthetic objects. ACM Journal of Graphics Tools 3, 4, 3345. C HEN , H., AND F ANG , S. 2000. Hardware accelerated voxelization . Computers and Graphics 24, 3, 433442. D ACHILLE IX, F., AND K AUFMAN , A. E. 2000. Incremental triangle voxelization. In Graphics Interface, 205212. D ONG , Z., C HEN , W., B AO , H., Z HANG , H., AND P ENG , Q. 2004. Real-time voxelization for complex polygonal models. In Proceedings of Pacific Graphics '04 , 4350. H UANG , J., Y AGEL , R., F ILIPPOV , V., AND K URZION , Y. 1998. An accurate method for voxelizing polygon meshes. In Proceedings of IEEE symposium on Volume visualization , 119126. J U , T. 2004. Robust repair of polygonal models. ACM Transactions on Graphics 23 , 3, 888895. K ARABASSI , G. P. E.-A., AND T HEOHARIS , T. 1999. A fast depth-buffer-based voxelization algorithm. ACM Journal of Graphics Tools 4 , 4, 510. S CHNEIDER , P., AND E BERLY , D. H. 2003. Geometry Tools for Computer Graphics . Morgan Kaufmann. S CHROEDER , W. J., L ORENSEN , W. E., AND L INTHICUM , S. 1994. Implicit modeling of swept surfaces and volumes. In Proceedings of IEEE Visualization , 4045. W ANG , S. W., AND K AUFMAN , A. E. 1993. Volume sampled voxelization of geometric primitives. In Proceedings of IEEE Visualization , 7884. W IDJAYA , H., M UELLER , T., AND E NTEZARI ., A. 2003. Voxelization in common sampling lattices. In Proceedings of Pacific Graphics '03 , 497501. 288

Graphics hardware;Hausdorff distance;Voxelization;Distance field;voxelization;local distance field;Object representation;rasterization;Polygonal object;GPU-based 3D slicer approach;GPU;3D slicer;slice-based approach;Rendering;rendering;adaptive dense voxelization;Volumetric representation;pixel shader;opacity;Surface voxelization;polygonal model;Surface representation;Rendering cost;GPU computation;Hausforff distance;object representation;polygonal objects;volumetric representation;triangles;Rendering pipeline;distance transform;volume construction;Modeling;Computational Geometry;geometric representation;hausdorff distance;distance field;Computer Graphics;Polygonal model;3D modelling;infinitude

80

ρ-Queries: Enabling Querying for Semantic Associations on the Semantic Web

This paper presents the notion of Semantic Associations as complex relationships between resource entities. These relationships capture both a connectivity of entities as well as similarity of entities based on a specific notion of similarity called -isomorphism. It formalizes these notions for the RDF data model, by introducing a notion of a Property Sequence as a type. In the context of a graph model such as that for RDF, Semantic Associations amount to specific certain graph signatures. Specifically, they refer to sequences (i.e. directed paths) here called Property Sequences, between entities, networks of Property Sequences (i.e. undirected paths), or subgraphs of ρ-isomorphic Property Sequences. The ability to query about the existence of such relationships is fundamental to tasks in analytical domains such as national security and business intelligence, where tasks often focus on finding complex yet meaningful and obscured relationships between entities. However, support for such queries is lacking in contemporary query systems, including those for RDF. This paper discusses how querying for Semantic Associations might be enabled on the Semantic Web, through the use of an operator ρ. It also discusses two approaches for processing ρ-queries on available persistent RDF stores and memory resident RDF data graphs, thereby building on current RDF query languages.

INTRODUCTION The Semantic Web [13] proposes to explicate the meaning of Web resources by annotating them with metadata that have been described in an ontology. This will enable machines to "understand" the meaning of resources on the Web, thereby unleashing the potential for software agents to perform tasks on behalf of humans. Consequently, significant effort in the Semantic Web research community is devoted to the development of machine processible ontology representation formalisms. Some success has been realized in this area in the form of W3C standards such as the eXtensible Markup Language (XML) [16] which is a standard for data representation and exchange on the Web, and the Resource Description Framework (RDF) [42], along with its companion specification, RDF Schema (RDFS) [17], which together provide a uniform format for the description and exchange of the semantics of web content. Other noteworthy efforts include OWL [25], Topic Maps [53], DAML+OIL [31]. There are also related efforts in both the academic and commercial communities, which are making available tools for semi-automatic [30] and automatic [49][29] semantic (ontology-driven and/or domain-specific) metadata extraction and annotation. With the progress towards realizing the Semantic Web, the development of semantic query capabilities has become a pertinent research problem. Semantic querying techniques will exploit the semantics of web content to provide superior results than present-day techniques which rely mostly on lexical (e.g. search engines) and structural properties (e.g. XQuery [24]) of a document. There are now a number of proposals for querying RDF data including RQL [40], SquishQL [45], TRIPLE [49], RDQL [48]. These languages offer most of the essential features for semantic querying such as the ability to query using ontological concepts, inferencing as part of query answering, and some allow the ability to specify incomplete queries through the use of path expressions. One key advantage of this last feature is that users do not need to have in-depth knowledge of schema and are not required to specify the exact paths that qualify the desired resource entities. However, even with such expressive capabilities, many of these languages do not adequately support a query paradigm that enables the discovery of complex relationships between resources. The pervasive querying paradigm offered by these languages is one in which queries are of the form: "Get all entities that are related to resource A via a relationship R" where R is typically specified as possibly a join condition or path expression, etc. In this approach, a query is a Copyright is held by the author/owner(s). WWW2003, May 20-24, 2003, Budapest, Hungary. ACM 1-58113-680-3/03/0005. 690 specification of which entities (i.e. resources) should be returned in the result. Sometimes the specification describes a relationship that the qualifying entities should have with other entities, e.g. a join expression or a path expression indicating a structural relationship. However, the requirement that such a relationship be specified as part of the query is prohibitive in domains with analytical or investigative tasks such as national/homeland security [11] and business intelligence, where the focus is on trying to uncover obscured relationships or associations between entities and very limited information about the existence and nature of any such relationship is known to the user. In fact, in this scenario the relationship between entities is the subject of the user's query and should being returned as the result of the query as opposed to be specified as part of the query. That is, queries would be of the form "How is Resource A related to Resource B?". For example, a security agent may want to find any relationship between a terrorist act and a terrorist organization or a country known to support such activities. One major challenge in dealing with queries of this nature is that it is often not clear exactly what notion of a relationship is required in the query. For example, in the context of assessing flight security, the fact that two passengers on the same flight are nationals of a country with known terrorist groups and that they have both recently acquired some flight training, may indicate an association due to a similarity. On the other hand, the fact that a passenger placed a phone call to someone in another country that is known to have links to terrorist organizations and activities may indicate another type of association characterized by connectivity. Therefore, various notions of "relatedness" should be supported. This paper intends to make two main contributions. First, we formalize a set of complex relationships for the RDF data model, which we call Semantic Associations. Second, we outline two possible approaches for processing queries about Semantic Associations through the use of an operator (-Queries). One of the two approaches is based on processing -queries on persistent RDF data systems such as RDFSuite [8], while the other is based on processing these queries on a main memory based representation of an RDF model such as JENA [56]. The rest of the paper is organized as follows: Section 2 discusses some background and motivates our work with the help of an example. Section 3 presents the formal framework for Semantic Associations; section 4 discusses implementation strategies for the operator, section 5 reviews some related work, and section 6 concludes the paper. BACKGROUND & MOTIVATION Although there are various knowledge modeling languages that may be used on the Semantic Web such as Topic Maps [55], UML [47], DAML+OIL [31], OWL [25], etc., in this paper we have chosen to formalize Semantic Associations for the RDF data model. It should be clear that we are not suggesting that the notion of Semantic Associations only applies to RDF. On the contrary, the notion is very general and is applicable to any data model that can be represented as a graph. The choice of RDF for formalization does not confer serious problems however. In the first place, some of these other models e.g. DAML+OIL build upon RDF. Secondly, there is work on mappings from other formalisms to RDF [20][41]. Next, we will briefly summarize the RDF data model and then motivate our work with an example. 2.1 RDF RDF [42] is a standard for describing and exchanging semantics of web resources. It provides a simple data model for describing relationships between resources in terms of named properties and their values. The rationale for the model is that by describing what relationships an entity has with other entities, we somehow capture the meaning of the entity. Relationships in RDF, or Properties as they are called, are binary relationships between two resources, or between a resource and a literal value. An RDF Statement, which is a triple of the form (Subject, Property, Object), asserts that a resource, the Subject, has a Property whose value is the Object (which can be either another resource or a literal). This model can be represented as a labeled directed graph, where nodes represent the resources (ovals) or literals (rectangles) and arcs representing properties whose source is the subject and target is the object, and are labeled with the name of the property. For example, in the bottom part of Figure 1, we can see a node &r1 connected by a paints arc to the node &r2, which reflects the fact that &r1 (a painter with first name Pablo, and last name Picasso) painted another resource &r2 (a painting). The meaning of the nodes and arcs is derived from the connection of these nodes and arcs to a vocabulary (the top part of the figure). The vocabulary contains describes types of entities i.e. classes (e.g. Museum) and types of properties (e.g. creates) for the domain. The vocabulary description and is done using the companion specification to RDF called the RDF Schema specification [17]. For example in Figure 1, classes like Painter, Museum and properties such as Paints, are defined. Resources are connected to classes using an rdf:typeof property indicating an instantiation relationship. 2.2 MOTIVATING EXAMPLE Although the focus of our current evaluations involves scenarios in the National Security domain, for brevity and pedagogical reasons, for this paper we have chosen to use a modified version of the example from [40]. We will now illustrate Semantic Associations by way of a simple example shown in Figure 1. The figure shows an RDF model base containing information to be used in the development of a cultural portal, given from two perspectives, reflected in two different schemas (the top part of the figure). The top left section of the picture is a schema that reflects a museum specialist's perspective of the domains using concepts like Museum, Artist, Artifact, etc. The top right section is a schema that reflects a Portal administrator's perspective of the domains using administrative metadata concepts like file-size, mime-type, etc. to describe resources. The lower part of the figure is the model base (or description base in the linguo of [40]), that has descriptions about some Web resources, e.g., museum websites (&r3, &r8), images of artifacts (&r2, &r5, &r7) and for resources that are not directly present on the Web, e.g., people, nodes representing electronic surrogates are created (&r1, &r4, &r6 for the artists Pablo Picasso, Rembrandt, and Rodin August respectively). 691 &r3 &r5 "Reina Sofia Museun" &r7 "oil on canvas" &r2 2000-02-01 "oil on canvas" &r8 "Rodin Museum" "image/jpeg" 2000-6-09 Ext. Resource String Date Integer String title file_siz e last_modified m i m e t y p e Artist Sculptor Artifact Sculpture Museum String String String fname lname creates exhibited sculpts String Painting Painter paints technique material typeOf(instance) subClassOf(isA) subPropertyOf mime-type exhibited technique exhibited title last_modified last_modified title technique exhibited "Rodin" "August" &r6 &r1 fname lname fname lname paints paints creates &r4 "Rembrandt" "Pablo" "Picasso" fname Figure 1: Cultural Portal Information in RDF Typically, a query language allows you to find all entities that are related by a specific relationship. For example, we may ask a query to retrieve all resources related to resource &r1 via a paints relationship, or via a paints.exhibited relationship, and get &r2 as a result for the first query and &r3 as the answer for the second query. However, we are unable to ask queries such as "How are resources &r1 and &r3 related? Such a query should return for example that "&r1 paints &r2 which is exhibited in &r3", indicating a path connecting the two entities. With a query such as this one, the user is trying to determine if there is a relationship between entities, and what the nature of the relationship(s) is(are). It should be possible to ask such a query without any type of specification as to the nature of the relationship, such as using a path expression to give information about the structure of the relationship. For example, the following example RQL query select * from {;Artist}@P{X}.{;Sculpture}@Q{Y}.@R{Z} finds all data paths that traverse the class hierarchies Artist and Sculpture, containing three schema properties, one for each property variable (@variable). However, we notice that the query requires that a property variable be added for every edge in the required path. That is, the user is required to have some idea of at least the structure e.g. length, of the relationship. One approach that some of these systems offer to alleviate this problem is that they provide mechanisms for browsing or querying schemas to allow users to get the information they need. While this may be a reasonable requirement when querying specific domains with a few schemas involved, on the Semantic Web, many schemas may be involved in a query, and requiring a user to browse them all would be a daunting task for the user. In fact, in some cases, such information may not be available to all users (e.g., classified information) even though the data may be used indirectly to answer queries. Furthermore, browsing schemas do not always give the complete picture, especially in the case of RDFS schemas, because, entities may belong to different schemas, creating links between entities that are not obvious from just looking at the schemas. For example in Figure 1, the relationship paints.exhibited.title connecting &r1 to "Reina Soifa Museum", is not apparent by just looking at either schema. So far, we have talked about relationships in terms of a directed path connecting two entities. However, there are some other interesting types of relationships. Let us take for example, resources &r4 and &r6. Both resources could be said to be related because they have both created artifacts (&r5, and &r7) that are exhibited at the same museum (&r8). In this case, having some relationship to the same museum associates both resources. This kind of connectivity is an undirected path between the entities. Another closely related kind of association is class membership. For example, &r1 and &r6 are both Artists, even though of a different kind, and therefore are somewhat associated. Also, &r1 and &r6 could be said to be associated because they both have creations (&r2, and &r7) that are exhibited by a Museum (&r3 and &r8 respectively). In this case, the association is that of a similarity. So, in the first three associations the relationships capture some kind of connectivity between entities, while the last association captures a similarity between entities. Note that the notion of similarity used here is not just a structural similarity, but a semantic similarity of paths (nodes and edges) that the entities are involved in. Nodes are considered similar, if they have a common ancestor class. For example in the relationship involving &r1 and &r6, although one case involves a painting and the other a sculpture, we consider them similar because sculptures and paintings are kinds of Artifacts and sculpting and painting are both kinds of creative activities (the notion of similarity is extended to properties as well). The Semantic Associations shown in this example are fairly simple involving only short paths and are useful only for the purpose of illustration. However, in environments that support information analytics and knowledge discovery involve longer paths, especially undirected paths, which are not easily detectable by users in fast-paced environments. For example at airport security portals, agents may want to quickly determine if a passenger has any kind of link to terrorist organizations or activities. FRAMEWORK The framework described in this section provides a formal basis for Semantic Associations. It builds on the formalization for the RDF data model given in [40], by including a notion of a Property Sequence. A Property Sequence allows us to capture paths in the RDF model and forms the basis for formalizing Semantic Associations as binary relations on Property Sequences. Secondly, we some complex queries called -queries for querying about Semantic Associations. 3.1 Formal Data Model In section 2.1, we describe the RDF data model informally as a labeled directed graph. To recap, the RDF Schema specification [17] provides a special vocabulary for describing classes and properties in a domain. A Property is defined by specifying its domain (the set of classes that it applies to), and its range (either a Literal type e.g. String, Integer, etc, or the classes whose entities it may take as values). Classes are defined in terms of their relationship to other classes using the rdfs:sublassOf property to place them at the appropriate location in a class hierarchy, as well as other user specified properties that may include them in their range or domain thereby linking them to other classes. Properties may also be organized in a hierarchy using the rdfs:subPropertyOf property. 692 The formalization in [40] defines a graph data model along with a type system that connects the RDF Model & Syntax specification with the RDFS schema specification using an interpretation mechanism. It forms the basis for a typed RDF query language called RQL [40]. RQL is fairly expressive and supports a broad range of queries. Its type system T is the set of all possible types that can be constructed from the following types: = C | P | M | U | L | { } | [1: 1 , 2: 2 , ..., n: n ] | (1: 1 + 2: 2 + ... + n: n ) where C indicates a class type, P a property type, M a metaclass type, L a literal type in the set L of literal type names (string, integer, etc.), and U is the type for resource URIs. For the RDF multi-valued types we have {.} as the Bag type, [.] is the Sequence type, and (.) is the Alternative type. The set of values that can be constructed using the resource URIs, literals and class property names is called V. Then, the interpretation of types in T is given naturally by the interpretation function [[ ]], which is a mapping from to the set of values in V. For example, a class C is interpreted as unary relation of type { U }, which is the set of resources (i.e. of type U ) that have an rdf:typeOf property with range C, and includes the interpretations of the subclasses of C. For a property p, [[p]] is given by {[v 1 , v 2 ] | v 1 [[ p.domain ]], v 2 [[ p.range ]] } { [[ p' ]] | p' is a subPropertyOf p} It defines an RDF schema as a 5-tuple RS = (V S , E S , , , H) where: V S is the set of nodes and E S is the set of edges. is an incidence function : E S V S V S , and is a labeling function that maps class and property names to one of the types in T, i.e. : V S E S T. H = (N, &lt;), where N = C P, C and P are the set of class and property names in RS, respectively. H is a well-formed hierarchy, i.e., &lt; is a smallest partial ordering such that: if p 1 , p 2 P and p 1 &lt; p 2 , then p 1 .domain p 2 .domain and p 1 .range p 2 .range. It also formalizes an instance of an RDFS schema called a description base which contains all the asserted instances of classes and properties of an RDF schema. We generalize these definitions to sets of RDF schemas and description bases as basic notion of context for a -query. 3.1.1 Definition 1 The RDFS schema set of RDFS Schemas RSS = {RS i : 1 i n}. Let C = C S1 C S2 ... C S2 where C Si is the set of class names in schema RS i and P = P S1 P S2 ... P Sn , where P Si is the set of property names in RS i then N = C P . [40] defines a description base RD which is an instance of an RDFS schema RS containing all the asserted instances of the classes and properties in RS. We generalize that definition here to the union of instances of the set of schemas in an RDFS schema set. 3.1.2 Definition 2 An instance of an RDF schema set RSS = {RS 1 , RS 2 , .. RS n }, is a description base RDS defined as a 5-tuple = (V DS , E DS , , , ), where V DS = V D1 V D2 ... V Dn and V Di is the set of nodes in the description base of the schema RS i , and E DS is defined similarly. is the incidence function : E DS V DS V DS , is a value function that maps the nodes to the values in V i.e. : V DS V, is a labeling function that maps each node either to one of the container type names (Seq, Bag, Alt) or to a set of class names from the schema set RSS whose interpretations contain the value of the node, and each edge e = [v 1 , v 2 ] to a property name p in RSS, where the interpretation of p contains the pair [ (v 1 ), (v 2 )], i.e., the values of v 1 and v 2 . Formally, : V D E D 2 N {Bag, Seq, Alt} in the following manner: i. For a node n in RDS, (n) = {c | c C and (n) [[c]]} ii. For an edge e from node n 1 to n 2 , (e) = p P and the values of n 1 to n 2 belong in the interpretation of p: [ (n 1 ), (n 2 )] [[p]]. In order capture paths in the RDF graph model, we define a notion of a Property Sequence, represented in the graph as a sequence of edges (i.e. properties). There is a choice to be made in the method for realizing such a notion in a query language such as RQL. One option is to add paths or property sequences as types in a query language making them first class citizens. The second option is to realize them as complex operations such as Cartesian product, on property types. We choose the later approach because attempting to make paths as first class citizens brings up additional issues such as defining path subsumption and so on. We will now define the notion of a Property Sequence. 3.1.3 Definition 3 (Property Sequence) A Property Sequence PS is a finite sequence of properties [P 1 , P 2 , P 3 , ... P n ] where P i is a property defined in an RDF Schema RS j of a schema set RSS. The interpretation of PS is given by: [[PS]] i=1 n [[P i ]] where for ps [[PS]], called an instance of PS, ps[i] [[P i ]] for 1 i n and ps[i][1] = ps[i+1][0]). ps[i][1] refers to the second element of the i th ordered pair and ps[i+1][0] refers to the first element of the i+1 th ordered pair. We define a function NodesOfPS()which returns the set of nodes of a Property Sequence PS, i.e. PS.NodesOfPS()= {C 1 , C 2 , C 3 , ... C k } where C i is a class in either the domain or range of some Property P j in PS, 1 j n. For example in Figure 1, for PS = c reates.exhibited.title, PS .NodesOfPS () = {Artist, Artifact, Museum, Ext. Resource, String}. Let PS = [P 1 , P 2 , P 3 , ... P n ], a description base RDS is said to satisfy or be a model of PS (RDS |= PS) if there exists a sequence of edges e 1 , e 2 , e 3 , ... e n in the description base RDS such that for all i, (e i ) = P i , (e i ) = (v i , v i+1 ) and i=1 n (v i , v i+1 ) = ps for some ps [[PS]]. We define a function PSNodesSequence () on Property Sequence instances that returns its sequence of nodes, i.e. ps.PSNodesSequence()= [v 1 , v 2 , v 3 , ... v n+1 ]. The node v 1 is called the origin of the sequence and v n+1 is called the terminus. Next, we define a set of binary relations on Property Sequences. 693 3.1.4 Definition 4 ( Joined Property Sequences) PS 1 PS 2 is true if: NodesOfPS(PS 1 ) NodesOfPS(PS 2 ) 0. The Property Sequences PS 1 and PS 2 are called joined, and for C (NodesOfPS(PS 1 ) NodesOfPS(PS 2 )), C is called a join node. For example, in Figure 2, the sequences creates.exhibited. and paints.exhibited are joined because they have a join node Museum. &r3 &r5 &r7 "oil on canvas" &r2 "oil on canvas" &r8 Artist Sculptor Artifact Sculpture Museum String String fname lname creates exhibited sculpts String String Painting Painter paints technique material typeOf(instance) subClassOf(isA) subPropertyOf exhibited technique exhibited technique exhibited "Rodin" "August" &r6 &r1 fname lname fname lname paints paints creates &r4 "Rembrandt" "Pablo" "Picasso" fname Figure 2 : Isomorphic Property Sequences 3.1.5 Definition 5 ( -Isomorphic Property Sequences) Two property sequences PS 1 = P 1 , P 2 , P 3 , ... P m , and PS 2 = Q 1 , Q 2 , Q 3 , ... Q m , are called -isomorphic (PS 1 PS 2 ), if for all i, 1 i m: P i = Q i or P i Q i or Q i P i ( means subPropertyOf ) For example in Figure 2, the sequences paints.exhibited and creates.exhibited are isomorphic because paints is considered to be similar to creates, since paints is a subproperty of creates. Note that the example that we use here is somewhat misleading because the example shown for Joined Property Sequences also happens to be -Isomorphic. However, the two notions are quite different because Joined Property Sequences are not required to be similar. 3.1.6 Definition 6 (Length) The length of a Property Sequence is equal to the number of properties in the Property Sequence. In the case of a Joined Property Sequence its length is the sum of all the properties in its constituent Property Sequences, i.e. the length of the undirected path from origin of one Property Sequence to the origin of the other Property Sequence. For example, in Figure 2, the Joined Property Sequences [creates.exhibited, paints.exhibited] has a length of 4. 3.2 Semantic Associations We can now define some binary relations on the domain of entities i.e. resources, based on the different types of Property Sequences. 3.2.1 Definition 7 ( -pathAssociated) -pathAssociated (x, y) is true if there exists a Property Sequence with ps [[PS]] and, either x and y are the origin and terminus of ps respectively, or vice versa, i.e. y is origin and x is terminus. Then ps is said to satisfy -pathAssociated (x, y) written as ps |= -pathAssociated (x, y). 3.2.2 Definition 8 ( -joinAssociated) Let PS 1 and PS 2 be two Property Sequences such that PS 1 PS 2 with a join node C, and there exists ps 1 and ps 2 such that ps 1 [[ PS 1 ]] and ps 2 [[ PS 2 ]] and, n ps1.PSNodesSequence() ps2.PSNodesSequence(), then -joinAssociated (x, y) is true if either of the conditions are satisfied. 1) x is the origin of ps 1 and y is the origin of ps 2 or 2) x is the terminus of ps 1 and y is the terminus of ps 2 . This means that either ps 1 .PSNodesSequence = [ x, b, c ... n, .,., r ] and ps 2 .PSNodesSequence = [ y, , , . . n, , ], or ps 1 .PSNodesSequence = [ a, b, c ... n, .,., r ,x] ] and ps 2 .PSNodesSequence = [ , , , . . n, , y] and n [[ C ]]. We say that (ps 1 , ps 2 ) |= -joinAssociated (x, y). 3.2.3 Definition 9 ( -cpAssociated) This is a special case of Definition 5 that captures an inclusion or sibling relationship (i.e. common parent) between resources. -cpAssociated (x, y) is true if there exists two Property Sequences PS 1 and PS 2 such that PS 1 PS 2 which satisfy joinAssociated (x, y) and, both PS 1 and PS 2 are of the form: rdf.typeOf.(rdfs:subClassOf)*. This relation is used to capture the notion that entities are related if they either belong to the same class or to sibling classes. For example, &r1 and &r6 are related because they are both Artists. We say that (ps 1 , ps 2 ) |= -cpAssociated (x, y). However, in order to reduce the possibility of meaningless associations e.g. both x and y belong to rdfs:Resource, we make further restrictions. We say that -cpAssociated (x, y) is strong if 1) For the join node j of the Joined Property Sequence (inducing the association (i.e. the common parent of x and y), j , where called the ceiling, refers to the most general class in the hierarchy that is to be considered, which is usually user-specified. 2) the length of the Joined Property Sequence inducing the association is minimal. By minimal we mean that it is less than a specific value indicated by context or specified by user. The first restriction ensures that we do go to far up the hierarchy looking for a common parent, while the second ensures that the relationship is not too distant to be meaningful in the user's context. 3.2.4 Definition 10 ( -IsoAssociated) -IsoAssociated (x, y) is true if there exists two property sequences PS 1 and PS 2 such that PS 1 PS 2 , and there exists ps 1 and ps 2 such that ps 1 [[PS 1 ]] and ps 2 [[PS 2 ]] such that, x is the origin of ps 1 and y is the origin of ps 2 . We say that (ps 1 , ps 2 ) |= -IsoAssociated (x, y). 694 We say that x and y are semantically associated if either pathAssociated (x, y), -cpAssociated(x, y), -IsoAssociated(x, y), or -joinAssociated(x, y). 3.3 -Queries A -Query Q is defined as a set of operations that map from a pair of keys (e.g. resource URIs) to the set of Property Sequences PS in the following manner: 1. : U (2) 2 PS 2. : U (2) 2 PS(2) 3. : U (2) 2 PS(2) U (2) = { {x, y} : x, y U and x y }. Similarly, PS (2) is the set of pairs of Property Sequences. In 1., we map from a pair of keys x and y to a set of Property Sequences that induces a pathAssociation of x and y. In 2., we map from (x, y) to a set of binary tuples of Property Sequences that induces either a joinAssociation or a strong -cpAssociation of x and y and in 3., we map from (x, y) to a set of binary tuples of Property Sequences that induces a -isoAssociation. STRATEGIES FOR PROCESSING -QUERIES Our strategy for implementation involves investigating alternative approaches to implementing the -operator, and evaluate their merits and demerits. We consider two major categories. The first category, which we have developed a partial implementation for, involves leveraging existing RDF persistent data storage technologies. Here, a -query processing layer is developed above the RDF data storage layer, which performs some of the computation and, relegates a specific portion of the computation to the data store layer. In the second approach, the implementation involves the use of a memory resident graph representation of the RDF model, along with the use of efficient graph traversal algorithms. We will outline how query processing is done using both approaches. 4.1 Exploiting RDF Data Management Systems In this approach, we leverage existing RDF data storage technologies such as RDFSuite [8] and SESAME [18] and develop a -query processing layer which performs some of the computation and, relegates a specific portion of the computation to the data store layer. Figure 3 gives an illustration of this approach (although, this is somewhat of an oversimplification, it adequate for the purposes of this discussion). Here the processing of a -query is broken down to 4 phases. Phases 2 and 4 occur at the data store layer and phases 1 and 3 occur at the -query processing layer. Phase 1 captures the query, i.e. the resources and context (i.e. schema set). In the second stage, the resources are classified i.e., the classes that entities belong to, within the given context, are identified. This involves a query to the data store layer, which exploits the rdf:typeOf statements to answer the query. Much of the processing is done in the third phase where potential paths involving the entities in the query are discussed by querying a PathGuide (a combination of index structures that stores information about paths that exist between resources classes). There are two kinds of paths that are kept in the PathGuide. The first kind of path is that which is obvious from the schema. The second kind is those paths that exist at the data level but are not evident at the schema level. This is because of the fact that the RDF data model allows multiple classifications of entities. Consequently, every instance of a multiple classification induces a connection between two classes that does not exist at the schema level, and thus is not general to all entities of those classes. Therefore, a query to the PathGuide yields potential property sequences paths between entities, since some of the paths are not general to entire classes but specific to the entities that are multiply classified. For example in Figure 1, the paints.exhibited.title sequence is not a sequence in either the left or right schema, but is present in the description base (i.e. between &r1 and the literal node "Reina Sofia Museum"). The reason for this is &r3`s membership in both the Museum and the Ext.Resource classes, this can be seen as having created an intermediate class node that collapses Museum and the Ext.Resource classes, and consequently links the paints.exhibited sequence to the title property. The fourth stage of query processing involves the validation of the paths found in the PathGuide for the specific entities in the query, by executing queries on the underlying data store. The output of this stage selects only those paths whose queries do not return an empty result. r1 r1 = http://www.xxx.com/yyy r2 r2 = http://www.zzz.net/ A B E C D r1 r1 = http://www.xxx.com/yyy r2 r2 = http://www.zzz.net/ A B E C D r1 r1 = http://www.xxx.com/yyy r2 r2 = http://www.zzz.net/ A B E C D r1 r1 = http://www.xxx.com/yyy r2 r2 = http://www.zzz.net/ 1. Query Entities 2. Classification of Entities 3. Identification of Candidate Paths 4. Pruning of Invalid Paths Figure 3: Illustration of -Query Processing 4.1.1 Issues Two challenges arise from storing all potential paths between classes in the PathGuide indexes. The first is that it causes the size of indexes to be quite large. Second, the potential paths found in the PathGuide in response to a query, could generate a large number of RQL queries that need to be validated at the data store layer, which slows down processing time significantly. However, heuristics could be employed to minimize these problems. For example, to reduce the size of the indices, we could choose to avoid adding every single potential path between classes in the index, but include only those whose support value is at least as large as a user supplied threshold value, where the support value represents the percentage of resources that are involved in 695 multiple classification for any given pair of classes. This means that if very few resources induce a connection between two otherwise unconnected schema classes because of a multiple classification, then we do not include in the indexes, those additional paths created due to the multiple classification, thereby reducing the size of the indices. The rationale for this is that the probability of those paths being involved in the result of a query is low, therefore the additional cost of storing the paths in the indices may not be worth it. A second heuristic is to try to prune the number of paths that need to be validated at the data storage layer. This could be done by assigning weights to Semantic Associations based on the contextual relevance and then validating only those associations with a high relevance weight. Our work in this area is still in progress. An additional problem with processing -queries on existing RDF storage systems is that some of these systems represent each property as a separate relation in a relational database model. Therefore, the computation of a Property Sequence results in a large number of joins which has a negative impact of the speed of query processing. Currently, we do not see any easy solution to this problem. 4.2 Using Graph Algorithms This approach involves the computation of Semantic Associations on a memory-resident graph representation of the RDF model such as that provided by JENA [56], or the memory representation of the schema set as in SESAME [18], to which graph traversals algorithms can be applied. In the case of pathAssociation we can search for paths between entities, and in the case of a -joinAssociation we check if the two entities belong in the same connected component. One issue with this approach is that that trying to find all paths between entities could easily lead to an exponential time algorithm. However, [52] provides promising fast algorithms for solving path problems which may be employed for such computations. In particular, it offers near-linear time algorithms for computing a path expression representing the set of all paths between nodes in a graph. Such a representation may then be processed using contextual information to prune paths that are not very relevant in the given context. In addition, other heuristics may be added. For example, a user may be asked to characterize the desired result set, e.g. shortest paths or longest paths, which will also help to reduce the result set. Similar heuristics to those discussed in the first approach that use context to prune paths based on degree of relevance can also be used here. In that case, the complexity of the problem can be analyzed along the number of semantic paths retrieved Complexity = (n-1) (l=1) (# paths of length l) (probability of keeping path of length l). Another issue is the complexity of graph isomorphism problem which is known to be NP-complete. However, certain classes of graphs have properties making them amenable to efficient manipulation. For example, [12] describes a polynomial time algorithm for detecting isomorphism in rooted directed path graphs, which includes the exact class of graphs that are required for checking -isomorphism. We are currently working on a prototype implementation of this approach. RELATED WORK There is some relationship between our work and that on querying object-oriented and semi-structured data using path expressions [2][3][19][22][23][24][34]. Although, these systems provide powerful and expressive capability, allowing users to query for data without having in-depth schema knowledge, most of them work on the premise that the goal of a query is to find data entities but not complex relationships such as Semantic Associations. Some of these systems [19][22] support paths as first class entities and allow for path variables to be used outside of the FROM clause, i.e. to be returned as a result of a query which suggests that queries for -pathAssociations could be supported. However, they typically assume a simpler data model which is a rooted directed graph without the nuances of RDF such as multiple classification and property hierarchies. Furthermore, the more complex Semantic Associations such as the -joinAssociation and -Isomorphism are not supported, even in systems like [22] which provide some functions that range over path variables, e.g., the difference function which returns the difference in the set of paths that originate from two nodes. With respect to RDF, the current generation of RDF query languages RQL [40], SquishQL [45], RDQL [48], do not support path variables as first class entities and so cannot even be used for querying for path relationships. In the case of the logic-based RDF query languages such as TRIPLE [51], the inference rules required to reason about the full range of the Semantic Associations discussed here, would require functionality beyond FOL. The DISCOVER system [38] provides a keyword proximity search capability over relational databases, which return associations called Joining Sequences. Joining Sequences represent the paths connecting keywords in the query, obtained by traversing foreign key links. However, the semantics associated with these associations is not explicit, but is implicit in the database schema. Thus, the interpretation of the meaning and usefulness of the associations must be done by users. Furthermore, other more complex Semantic Associations such as the -Isomorphism are not captured. There is a common intuition underlying our work and some of the tasks related to data mining, in that they both involve discovering relationships. However, there are significant differences in the goals, methods and results produced by the both kinds of systems. The first difference is articulated in a statement made in [32], where data mining is said to be opportunistic while information access techniques (such as ours) are goal-driven. Traditional data mining [21][26] focuses on discovering patterns and relationships in data that can be used to develop models of the data. In association analysis [7], rules that associate attribute-value pairs are learned from patterns of co-occurrences of attribute values in data, which capture co-occurrence relationships between attributes. On the contrary, we do not try to learn patterns from data rather, we provide specific rules for inferring relationships between entities by looking at property value dependencies, and focus on providing methods for verifying whether these kinds of associations exist between entities. That is, we identify meaningful sequences of binary predicates while data mining association rules involve sets of attribute value pairs. Therefore, we view data mining as a complimentary technology. For example, the association rules learnt from patterns in data can 696 provide knowledge that can be used to guide the search for Semantic Associations or to rank resulting Semantic Associations based on how close the follow the patterns. An initial discussion on Semantic Associations is made in [10]. CONCLUSION & FUTURE WORK Most RDF query systems do not provide adequate querying paradigms to support querying for complex relationships such as Semantic Associations. Support for such querying capabilities is highly desirable in many domains. We have presented a formal framework for these Semantic Associations for the RDF data model, and reviewed some implementation strategies for computing them. There are many open research issues that we plan to focus on in the near future. First, it may be necessary to develop data organization techniques for data that will adequately support the kinds of queries discussed here. Such techniques should eliminate the need for an excessive number of expensive computations such as joins during query processing. Secondly, we plan to develop techniques for dealing with the space complexity problem of the indices used in the PathGuide. For example we may use encoding schemes that compress path information, or heuristics for managing the size of the indices. Another top priority is the development of context-sensitive ranking algorithms that assign higher weights to results that are most relevant in the query context. Finally, we will perform a comparative study of the two implementation strategies discussed in section 4 over a testbed consisting of large amount of automatically extracted metadata generated using the SCORE system [41]. ACKNOWLEDGMENTS Our thanks to Drs. Bob Robinson, John Miller, Krys Kochut, and Budak Arpinar for the illuminating discussions and insightful contributions, and to Boanerges Aleman-Meza on his revision comments. We are also indebted to Dr. Vassilis Christophides whose comments and suggestions were invaluable in preparing the final version of the paper. This work is funded by NSF-ITR-IDM Award # 0219649 titled " Semantic Association Identification and Knowledge Discovery for National Security Applications ." REFERENCES [1] S. Abiteboul, P. Buneman, and D. Suciu. Data on the Web: From Relations to Semistructured Data and XML. Morgan Kaufmann, 1999. [2] S. Abiteboul. Querying Semi-Structured data. In Proc. of ICDT, Jan 1997. http://citeseer.nj.nec.com/abiteboul97querying.html [3] S. Abiteboul, D. Quass, J. McHugh, J. Widom, and J. Wiener. The Lorel Query Language for Semistructured Data. International Journal on Digital Libraries, 1(1):68--88, April 1997. [4] S. Abiteboul, R. Hull, and V. Vianu. Foundations of Databases. Addison-Wesley, 1995. [5] R. Agrawal. Alpha: An Extension of Relational Algebra to Express a Class of Recursive Queries. IEEE Transactions on Software Engineering. 14(7):879-- 885, July 1988. [6] R. Agrawal, A. Borgida, and H.V. Jagadish. Efficient Management of Transitive Relationships in Large Data Bases. In SIGMOD'89, pages 253--262, Portland, Oregon, USA, 1989. [7] R. Agrawal, T. Imielienski and A. Swami. Mining Assocation Rules between Sets of Items in Large Databases. Proc. Conf. On Management of Data. Washington, DC, USA, 207--216. ACM Press, New York, NY USA 1993. [8] S. Alexaki, G. Karvounarakis, V. Christophides, D. Plexousakis, and K. Tolle. The ICS-FORTH RDFSuite: Managing Voluminous RDF Description Bases. In 2nd International Workshop on the Semantic Web, pages 1--13, Hong Kong, 2001. [9] S. Alexaki, G. Karvounarakis, V. Christophides, D. Plexousakis, and K. Tolle. On Storing Voluminous RDF descriptions: The case of Web Portal Catalogs. In 4th International Workshop on the Web and Databases (WebDB), Santa Barbara, CA, 2001. Available at http://139.91.183.30:9090/RDF/publications/webdb2001.pdf [10] K. Anyanwu, A. Sheth, The Operator: Discovering and Ranking Associations on the Semantic Web. SIGMOD Record (Special issue on Amicalola Workshop), December 2002. [11] D. Avant, M. Baum, C. Bertram, M. Fisher, A. Sheth, Y. Warke, "Semantic Technology Applications for Homeland Security," Proc. of the 11 th Intl Conf. on Information and Knowledge Management (CIKM 2002), McLean, VA, November 4-9, 2002, pp. 611--613. [12] L. Babel, I. Ponomarenko, G. Tinhofer. The Isomorphism Problem for Directed Paths and For Rooted Directed Path Graphs. Journal of Algorithms, 21:542--564, 1996. [13] T. Berners-Lee, J. Hendler, and O. Lassila. The Semantic Web. Scientific American, May 2001. [14] B. Berendt, A. Hotho, G. Stumme. Towards Semantic Web Mining. In Proceedings of the International Semantic Web Conference, pp. 264--278, Sardinia, Italy. June 2002. [15] A. Branstadt, V. B. Le, J. P. Spinrad. Graph Classes: A Survey. SIAM 1999. [16] T. Bray, J. Paoli, and C.M. Sperberg-McQueen. Extensible Markup Language (XML) 1.0. W3C Recommendation, February 1998. [17] D. Brickley and R.V. Guha. Resource Description Framework (RDF) Schema Specification 1.0, W3C Candidate Recommendation. 2000. [18] J. Broekstra, A. Kampman, and F. van Harmelen. SESAME: An Architecture for Storing and Querying RDF Data and Schema Information. In D. Fensel, J. Hendler, H. Lieberman, and W. Wahlster, editors, Semantics for the WWW. MIT Press, 2001. 697 [19] P. Buneman, M. Fernandez, D. Suciu. UnQL: A Query Language and Algebra for Semistructured Data Based on Structural Recursion. VLDB Journal, 9(1):76--110, 2000. [20] W. W. Chang, A Discussion of the Relationship Between RDF-Schema and UML. A W3C Note, NOTE-rdf-uml-19980804 . [21] M. Chen, J. Han and P. Yu. Data Mining: An Overview from the Database Perspective. IEEE Trans. On Knowledge and Data Engineering. Vol. 8. No. 6. December 1996. [22] V. Christophides, S. Abiteboul, S. Cluet, and M. Scholl. From Structured Documents to Novel Query Facilities. In Proc. of ACM SIGMOD Conf. on Management of Data, pp. 313--324, Minneapolis, Minnesota, May 1994. [23] V. Christophides, S. Cluet, and G. Moerkotte. Evaluating Queries with Generalized Path Expressions. In Proc. of ACM SIGMOD, pp. 413--422, 1996. [24] D. Chamberlin, D. Florescu, J. Robie, J. Simeon, and M. Stefanescu. XQuery: A Query Language for XML. Working draft, World Wide Web Consortium, June 2001. http://www.w3.org/TR/xquery/ [25] M. Dean, D. Connolly, F. Harmelen, J. Hendler, I. Horrocks, D. McGuinness, P. F. Patel-Schneider, and L. Stein. OWL Web Ontology Language 1.0 Reference, W3C Working Draft 29 July 2002. http://www.w3.org/TR/owl-ref/. [26] U. M. Fayyad, G. Piatetsky-Shapiro, P. Smyth and R. Uthurusany. Advances in Knowledge Discovery and Data Mining.. AAAI/MIT Press 1996. [27] R. Fikes. DAML+OIL query language proposal, August 2001. http://www.daml.org/listarchive/joint-committee/0572 .html. [28] R. H. Guting. GraphDB: Modeling and querying graphs in databases. In Proceedings of the International Conference on Very Large Data Bases, pp. 297--308, 1994. [29] B. Hammond, A. Sheth, and K. Kochut, Semantic Enhancement Engine: A Modular Document Enhancement Platform for Semantic Applications over Heterogeneous Content, in Real World Semantic Web Applications, V. Kashyap and L. Shklar, Eds., IOS Press, December 2002, pp. 29--49. [30] S. Handschuh and S. Staab. Authoring and annotation of web pages in CREAM. In The Eleventh International World Wide Web Conference (WWW2002), Honolulu, Hawaii, USA, 7-11, May, 2002 [31] F. Harmelen, P. F. Patel-Schneider, I. Horrocks, eds. Reference Description of the DAML+OIL (March 2001) ontology markup language. [32] M. Hearst. Distinguishing between Web Data Mining and Information Access. Position statement for Web Data Mining KDD 97. [33] Y. E. Ioannidis, R. Ramakrishnan, L. Winger: Transitive Closure Algorithms Based on Graph Traversal. TODS 18(3): 512--576 (1993). [34] Y. E. Ioannidis, Y. Lashkari. Incomplete Path Expressions and their Disambiguation, In Proc. of the 1994 ACM SIGMOD, International Conference on Management of Data. p.138-149, May 24-27, 1994, Minneapolis, Minnesota, United States. [35] P. Hayes. RDF Model Theory. W3C Working Draft, September 2001. [36] I. Horrocks, S. Tessaris. The Semantics of DQL. http://www.cs.man.ac.uk/~horrocks/Private/DAML/DQL-semantics .pdf [37] I. Horrocks and S. Tessaris. A Conjunctive Query Language for Description Logic Aboxes. In Proc. of AAAI-00, 2000. [38] V. Hristidis and Y. Papakonstanti-nou. DISCOVER: Keyword search in relational databases. In Procs. VLDB, Aug. 2002. [39] ICS-FORTH. The ICS-FORTH RDFSuite web site. Available at http://139.91.183.30:9090/RDF, March 2002. [40] G. Karvounarakis, S. Alexaki, V. Christophides, D. Plexousakis, M. Scholl, RQL: A Declarative Query Language for RDF, WWW2002, May 7-11, 2002, Honolulu, Hawaii, USA. [41] M. S. Lacher and S. Decker. On the Integration of Topic Maps and RDF Data. In Proc. of Semantic Web Working Symposium. Palo Alto. California. August 2001. [42] O. Lassila and R. Swick. Resource Description Framework (RDF) Model and Syntax Specification, W3C Recommendation. 1999. [43] M. Mannino, L. Shapiro, L. Extensions to Query Languages for Graph Traversal Problems. TKDE 2(3): 353--363,1990. [44] A. O. Mendelzon and P. T. Wood. Finding Regular Simple Paths in Graph Databases. SIAM J. Comput., 24(6):1235-1258 , 1995. [45] L. Miller, A. Seaborne, A. Reggiori. Three Implementations of SquishQL, a Simple RDF Query Language. In Proc. of 1 st International Semantic Web Conference. ISWC2002. June 9-12 , 2002, Sardinia, Italy. [46] A. Nanopoulos. Y. Manolopoulos. "Mining Patterns from Graph Traversals", Data and Knowledge Engineering, Vol.37, No.3, pp.243-266, 2001. [47] J. Rumbaugh, I. Jacobson, and G. Booch. The Unified Modeling Language Reference Manual. Addison-Wesley, 1999. [48] A. Seaborne. RDQL: A Data Oriented Query Language for RDF Models. 2001. http://www.hpl.hp.com/semweb/rdql-grammar .html [49] A. Sheth, C. Bertram, D. Avant, B. Hammond, K. Kochut, Y. Warke. Semantic Content Management for Enterprises and the Web, IEEE Internet Computing, July/August 2002, pp. 80--87. [50] A. Sheth, S. Thacker and S. Patel. Complex Relationship and Knowledge Discovery Support in the InfoQuilt System . VLDB Journal. September 25, 2002. [51] M. Sintek and S. Decker. TRIPLE---A Query, Inference, and Transformation Language for the Semantic Web. 698 International Semantic Web Conference (ISWC), Sardinia, June 2002. http://www.dfki.uni-kl.de/frodo/triple/ [52] Tarjan, R. Fast Algorithms for Solving Path Problems. J. ACM Vol. 28, No. 3, July 1891, pp. 594--614. [53] DQL: DAML Query Language. http://www.daml.org/2002/08/dql/ [54] Inkling: RDF query using SquishQL, 2001. http://swordfish.rdfweb.org/rdfquery/. [55] ISO/IEC 13250: 2000 Topic Maps, Jan, 2000. http://www.topicmaps.org/ [56] JENA A Java API for RDF . [57] Whitepaper on National Security and Intelligence, Semagix Inc. 2002. http://www.semagix.com/pdf/national_security.pdf 699

AI;analysis;isomorphism;Complex Data Relationships;RDF;Rooted Directed Path;Semantic Associations;automation;graph traversals;semantic association;Semantic Web Querying;relationship;semantic web;query processing;Property Sequence

81

Energy Management Schemes for Memory-Resident Database Systems

With the tremendous growth of system memories, memory-resident databases are increasingly becoming important in various domains. Newer memories provide a structured way of storing data in multiple chips, with each chip having a bank of memory modules. Current memory-resident databases are yet to take full advantage of the banked storage system, which offers a lot of room for performance and energy optimizations. In this paper, we identify the implications of a banked memory environment in supporting memory-resident databases, and propose hardware (memory-directed) and software (query-directed) schemes to reduce the energy consumption of queries executed on these databases. Our results show that high-level query-directed schemes (hosted in the query optimizer) better utilize the low-power modes in reducing the energy consumption than the respective hardware schemes (hosted in the memory controller), due to their complete knowledge of query access patterns. We extend this further and propose a query restructuring scheme and a multi-query optimization . Queries are restructured and regrouped based on their table access patterns to maximize the likelihood that data accesses are clustered. This helps increase the inter-access idle times of memory modules, which in turn enables a more effective control of their energy behavior. This heuristic is eventually integrated with our hardware optimizations to achieve maximum savings. Our experimental results show that the memory energy reduces by 90% if query restructuring method is applied along with basic energy optimizations over the unoptimized version. The system-wide performance impact of each scheme is also studied simultaneously.

INTRODUCTION Memory-resident databases (also called in-memory databases <A href="81.html#10">[6]) are emerging to be more significant due to the current era of memory-intensive computing. These databases are used in a wide range of systems ranging from real-time trading applications to IP routing. With the growing complexities of embedded systems (like real-time constraints), use of a commercially developed structured 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'04, November 813, 2004, Washington, DC, USA. Copyright 2004 ACM 1-58113-874-1/04/0011 ... $ 5.00. memory database is becoming very critical <A href="81.html#10">[5]. Consequently, device developers are turning to commercial databases, but existing embedded DBMS software has not provided the ideal fit. Embedded databases emerged well over a decade ago to support business systems, with features including complex caching logic and abnormal termination recovery. But on a device, within a set-top box or next-generation fax machine, for example, these abilities are often unnecessary and cause the application to exceed available memory and CPU resources. In addition, current in-memory database support does not consider embedded system specific issues such as energy consumption. Memory technology has grown tremendously over the years, providing larger data storage space at a cheaper cost. Recent memory designs have more structured and partitioned layouts in the form of multiple chips, each having memory banks <A href="81.html#10">[30]. Banked memories are energy efficient by design, as per-access energy consumption decreases with decreasing memory size (and a memory bank is typically much smaller compared to a large monolithic memory). In addition, these memory systems provide low-power operating modes, which can be used for reducing the energy consumption of a bank when it is not being used. An important question regarding the use of these low-power modes is when to transition to one once an idleness is detected. Another important question is whether the application can be modified to take better advantage of these low-power modes. While these questions are slowly being addressed in architecture, compiler, and OS communities, to our knowledge, there has been no prior work that examines the energy and performance behavior of databases under a banked memory architecture. Considering increasingly widespread use of banked memories, such a study can provide us with valuable information regarding the behavior of databases under these memories and potential modifications to DBMSs for energy efficiency. Since such banked systems are also being employed in high-end server systems, banked memory friendly database strategies can also be useful in high-end environments to help reduce energy consumption. Our detailed energy characterization of a banked memory architecture that runs a memory-resident DBMS showed that nearly 59% of overall energy (excluding input/output devices) in a typical query execution is spent in the memory, making this component an important target for optimization (see Figure <A href="81.html#2">1). Moreover, for any system, memory power and energy consumption have become critical design parameters besides cost and performance. Based on these observations, this paper evaluates the potential energy benefits that memory-resident database queries can achieve by making use of banked memory architectures supported with low-power operating modes. Since each memory bank is capable of operating independently, this opens up abundant avenues for energy and performance optimizations. In this paper, we focus on a banked memory architecture and study potential energy benefits when database queries are executed. Specifically, we focus on two important aspects of the problem: Characterizing energy benefits of banked memories using hardware and software techniques: To see whether query execution can make use of available low-power modes, we study both hardware and software techniques. The hardware techniques detect the idleness of memory banks and switch the inactive (idle) banks (during 218 Memory 59% Cache 16% ALU 14% Bus 1% Others 10% Figure 1: Breakup of the energy consumption for various system components. The results are based on the average energy consumption of TPC-H benchmarks <A href="81.html#10">[35] executed on a memory-resident DBMS. query execution) to low-power operating modes. We also present a query-based memory energy optimization strategy, wherein the query plan is augmented by explicit bank turn-off/on instructions that transition memory banks into appropriate operating modes during the course of execution based on the query access pattern. We experimentally evaluate all the proposed schemes and obtain energy consumptions using an energy simulator. Our experiments using TPC-H queries <A href="81.html#10">[35] and a set of queries suitable for handheld devices clearly indicate that both hardware-based and query-directed strategies save significant memory energy. Query restructuring for memory energy savings: We propose a query restructuring scheme and a multi-query optimization strategy to further increase energy benefits coming from using low-power operating modes. The idea behind these schemes is to increase bank inter-access times so that more aggressive low-power modes can be employed and a memory bank can stay in a low-power mode longer once it is transitioned. Our experimental evaluation indicates that this query restructuring strategy does not only reduce energy consumption, but also helps improve overall performance (execution cycles). Apart from providing useful input for database designers, our results can also be used by hardware designers to tune the behavior of low-power modes so that they handle query access patterns better . Similar to the observation that creating a lightweight version of a disk-based database will not serve as a suitable in-memory database, our belief is that taking an in-memory database system and using it on a banked architecture without any modification may not generate the desired results. Therefore, the results presented in this work also shed light on how database design and memory architecture design interact with each other. The remainder of this paper is organized as follows. Section <A href="81.html#2">2 presents related work. Section <A href="81.html#2">3 elaborates on the memory database that we built and also on the memory banking scheme that we employ for our experiments. Section <A href="81.html#3">4 presents in detail the proposed hardware and query-directed energy optimization techniques. The results of our energy evaluation of these schemes are discussed in Section <A href="81.html#5">5. Our experiments also account for the performance overhead incurred in supporting our schemes. Section <A href="81.html#7">6 presents our query restructuring and regrouping scheme, and Section <A href="81.html#8">7 discusses its energy/performance benefits within the context of our banked memory architecture. Finally, Section <A href="81.html#10">8 summarizes the results. RELATED WORK In the past, memory has been redesigned, tuned or optimized to suit emerging fields. Need for customized memory structures and allocation strategies form the foundation for such studies. Copeland et al proposed SafeRAM <A href="81.html#10">[11], a modified DRAM model for safely supporting memory-resident databases alike disk-based systems, and for achieving good performance. In PicoDBMS <A href="81.html#10">[27], Pucheral et al present techniques for scaling down a database to a smart card. This work also investigates some of the constraints involved in mapping a database to an embedded system, especially memory constraints and the need for a structured data layout. Anciaux et al <A href="81.html#10">[3] explicitly model the lower bound of the memory space that is needed for query execution. Their work focuses on light weight devices like personal organizers, sensor networks, and mobile computers. Boncz et al show how memory accesses form a major bottleneck during database accesses <A href="81.html#10">[7]. In their work, they also suggest a few remedies to alleviate the memory bottleneck. An et al analyze the energy behavior of mobile devices when spatial access methods are used for retrieving memory-resident data <A href="81.html#10">[2]. They use a cycle accurate simulator to identify the pros and text Query Optimizer Query Execution Engine Memory Database Queries Data Results Energy & Performance Optimizations (Using Cost Plan) Hardware Optimizations Targeting DBMS Rewrite System Parser System Catalog Figure 2: DBMS architecture. cons of various indexing schemes. In <A href="81.html#10">[1], Alonso et al investigate the possibility of increasing the effective battery life of mobile computers by selecting energy efficient query plans through the optimizer. Although the ultimate goal seems the same, their cost plan and the optimization criterion are entirely different from our scheme. Specifically, their emphasis is on a client-server model optimizing the network throughput and overall energy consumption . Gruenwald et al propose an energy-efficient transaction management system for real-time mobile databases in ad-hoc networks <A href="81.html#10">[16]. They consider an environment of mobile hosts. In <A href="81.html#10">[22], Madden et al propose TinyDB, an acquisitional query processor for sensor networks. They provide SQL-like extensions to sensor networks, and also propose acquisitional techniques that reduce the power consumption of these networks. It should be noted that the queries in such a mobile ad-hoc network or a sensor environment is different from those in a typical DBMS. This has been shown by Imielinksi et al in <A href="81.html#10">[19]. In our model, we base our techniques on a generic banked memory environment and support complex, memory-intensive typical database operations. There are more opportunities for energy optimizations in generic memory databases, which have not yet been studied completely. The approach proposed in this paper is different from prior energy-aware database related studies, as we focus on a banked memory architecture, and use low-power operating modes to save energy. Gassner et al review some of the key query optimization techniques required by industrial-strength commercial query optimizers , using the DB2 family of relational database products as examples <A href="81.html#10">[15]. This paper provides insight into design of query cost plans and optimization using various approaches. In <A href="81.html#10">[23], Manegold studies the performance bottlenecks at the memory hierarchy level and proposes a detailed cost plan for memory-resident databases. Our cost plan and optimizer mimics the PostgreSQL model <A href="81.html#10">[12, 14]. We chose it due to its simple cost models and open source availability. A query restructuring algorithm is proposed by Hellerstein in <A href="81.html#10">[18]. This algorithm uses predicate migration to optimize expensive data retrievals. In <A href="81.html#10">[10], Chaudhuri et al extend this approach to study user-defined predicates and also guarantee an optimal solution for the migration process. Sarawagi et al present a query restructuring algorithm that reduces the access times of data retrieval from tertiary databases <A href="81.html#10">[32]. Monma et al develop the series-parallel algorithm for reordering primitive database operations <A href="81.html#10">[24]. This algorithm optimizes an arbitrarily constrained stream of primitive operations by isolating independent modules. This work forms the basic motivation for our query restructuring algorithm. However, our paper is different from all of the above work in the sense that we reorder queries for reducing energy consumption. Moreover, our database is memory-resident, with the presence of banked memory that gives more freedom for optimizations. SYSTEM ARCHITECTURE For our work, we modified the PostgreSQL DMBS to work with memory-resident data sets as its workload. The block diagram for our setup is shown in Figure <A href="81.html#2">2. The core components are derived from PostgreSQL. The flow of our model is similar to PostgreSQL except that the database is memory resident. A query is parsed for syntax and then sent to the rewrite system. The rewrite system uses the system catalog to generate the query tree, which is then sent to the optimizer. The query optimizer derives the cost of the query in 219 Configuration Registers Self-Monitoring/ Prediction Hardware Memory Controller Bank To/From CPU Module Memory Bus Figure 3: Banked memory architecture. multiple ways using the query tree and issues the best suited plan to the query execution engine. We incorporate our software-based techniques at the optimizer stage of the DBMS. These optimizations are based on the cost that is derived for each of the query plans (the discussion pertaining to the modified cost model is deferred till Section <A href="81.html#3">4). Based on the final query execution plan, the execution engine executes the query by using the database. The database is entirely memory resident and the memory is organized in a banked format (elaborated in the following section). The executor recur-sively iterates the query plan and uses a per-tuple based strategy (pipelined execution, and not bulk processing) to project the output results. The proposed hardware optimizations are at the computer architecture level of the system. Since the base DBMS model is similar to PostgreSQL, we do not elaborate each component in detail ( <A href="81.html#10">[26] provides an elaborate discussion). Instead, we highlight our contributions, and modifications to DBMS (shown in blue in Figure <A href="81.html#2">2) in the following sections. Overall, our strategies require modification to the query optimizer, memory hardware, and system software components. 3.2 Memory Model We use a memory system that contains a memory array organized as banks (rows) and modules (columns), as is shown picto-rially in Figure <A href="81.html#3">3 for a 4 4 memory module array. Such banked systems are already being used in high-end server systems <A href="81.html#10">[30] as well as low-end embedded systems <A href="81.html#10">[31]. The proposed optimizations will, however, apply to most bank-organized memory systems . Accessing a word of data would require activating the corresponding modules of the shown architecture. Such an organization allows one to put the unused banks into a low-power operating mode. To keep the issue tractable, this paper bases the experimental results on a sequential database environment and does not consider a multiprocessing environment (like transaction processing which requires highly complex properties to be satisfied). We assume in our experiments that there is just one module in a bank; hence, in the rest of our discussion, we use the terms "bank" and "module" interchangeably. 3.3 Operating Modes We assume the existence of five operating modes for a memory module: active, standby, nap, power-down, and disabled 1 . Each mode is characterized by its energy consumption and the time that it takes to transition back to the active mode (termed resynchronization time or resynchronization cost). Typically, the lower the energy consumption, the higher the resynchronization time <A href="81.html#10">[30]. Figure <A href="81.html#3">4 shows possible transitions between the various low-power modes (the dynamic energy 2 consumed in a cycle is given for each node) in our model. The resynchronization times in cycles (based on a cycle time of 3.3ns) are shown along the arrows (we assume a negligible cost for transitioning to a lower power mode). The energy and resynchronization values shown in this figure have been obtained from the RDRAM memory data sheet (512MB, 2.5V, 3.3ns cycle time, 8MB modules) <A href="81.html#10">[30]. When a module in standby, nap, or power-down mode is requested to perform a memory transaction, it first goes to the active mode, and 1 Current DRAMs <A href="81.html#10">[30] support up to six energy modes of operation with a few of them supporting only two modes. One may choose to vary the number of modes based on the target memory. 2 We exclusively concentrate on dynamic power consumption that arises due to bit switching, and do not consider the static (leakage) power consumption <A href="81.html#10">[28] in this paper. Full Power (2.063 nJ) Standby (0.743 nJ) Nap (0.035 nJ) Power Down (0.025 nJ) Disabled (0 nJ) 1 16 9000 Figure 4: Available operating modes and their resynchronization costs. then performs the requested transaction. While one could employ all possible transitions given in Figure <A href="81.html#3">4 (and maybe more), our query-directed approach only utilizes the transitions shown by solid arrows. The runtime (hardware-based) approaches, on the other hand, can exploit two additional transitions: from standby to nap, and from nap to power-down. 3.4 System Support for Power Mode Setting Typically, several of the memory modules (that are shown in Figure <A href="81.html#3">3) are controlled by a memory controller which interfaces with the memory bus. For example, the operating mode setting could be done by programming a specific control register in each memory module (as in RDRAM <A href="81.html#10">[30]). Next is the issue of how the memory controller can be told to transition the operating modes of the individual modules. This is explored in two ways in this paper: hardware-directed approach and software-directed (query-directed ) approach. In the hardware-directed approach, there is a Self-Monitoring and Prediction Hardware block (as shown in Figure <A href="81.html#3">3), which monitors all ongoing memory transactions. It contains some prediction hardware (based on the hardware scheme) to estimate the time until the next access to a memory bank and circuitry to ask the memory controller to initiate mode transitions (limited amount of such self-monitored power down is already present in current memory controllers, for example: Intel 82443BX and Intel 820 Chip Sets). In the query-directed approach, the DBMS explicitly requests the memory controller to issue the control signals for a specific module's mode transitions. We assume the availability of a set of configuration registers in the memory controller (see Figure <A href="81.html#3">3) that are mapped into the address space of the CPU (similar to the registers in the memory controller in <A href="81.html#10">[20]). These registers are then made available to the user space (so that the DBMS application can have a control) through operating system calls. Regardless of which strategy is used, the main objective of employing such strategies is to reduce the energy consumption of a query when some memory banks are idle during the query's execution . That is, a typical query only accesses a small set of tables , which corresponds to a small number of banks. The remaining memory banks can be placed into a low-power operating mode to save memory energy. However, it is also important to select the low-power mode to use carefully (when a bank idleness is detected ), as switching to a wrong mode either incurs significant performance penalties (due to large resynchronization costs) or prevents us from obtaining maximum potential energy benefits. Note that energy optimization is our context can be performed from two angles. First, suitable use of low-power operating modes can reduce energy consumption of a given query execution. Second , the query plan can be changed (if it is possible to do so) to further increase energy benefits. In this work, we explore both these angles. POWER MANAGEMENT SCHEMES In a banked architecture, the memory can be managed through either of the following two approaches: (1) a runtime approach wherein the hardware is in full control of operating mode transitions ; and (2) a query-directed scheme wherein explicit bank turn-on/off instructions are inserted in the query execution plan to invoke mode transitions. One also has the option of using both the approaches simultaneously (which we illustrate in later sections). 220 Full Power Standby Nap Power Down idle stanby idle nap idle down resynch stanby resynch nap resynch down Figure 5: Dynamic threshold scheme. 4.1 Hardware-Directed Schemes We explore two hardware-directed approaches that allow the memory system to automatically transition the idle banks to an energy conserving state. The problem then is to detect/predict bank idleness and transition idle banks into appropriate low-power modes. 4.1.1 Static Standby Scheme The first approach is a per-access optimization. Most of the recent DRAMs allow the chips to be put to standby mode immediately after each reference <A href="81.html#10">[30]. After a read/write access, the memory module that gets accessed can be placed into the standby mode in the following cycle. We refer to this scheme as the static standby mode in the rest of our discussion. Note that, while this scheme is not very difficult to implement, it may lead to frequent resynchro-nizations , which can be very harmful as far as execution cycles are concerned. 4.1.2 Dynamic Threshold Scheme Our second hardware-guided approach is based on runtime dynamics of the memory subsystem. The rationale behind this approach is that if a memory module has not been accessed in a while, then it is not likely to be needed in the near future (that is, inter-access times are predicted to be long). A threshold is used to determine the idleness of a module after which it is transitioned to a low-power mode. More specifically, we propose a scheme where each memory module is put into a low-power state with its idle cycles as the threshold for transition. The schematic of our dynamic threshold scheme is depicted in Figure <A href="81.html#4">5. After idle stndby cycles of idleness, the corresponding module is put in the standby mode. Subsequently, if the module is not referenced for another idle nap cycles, it is transitioned to the nap mode. Finally, if the module is not referenced for a further idle down cycles, it is placed into the power-down mode. Whenever the module is referenced, it is brought back into the active mode incurring the corresponding resynchronization costs (based on what low-power mode it was in). It should be noted that even if a single bank experiences a resynchronization cost, the other banks will also incur the corresponding delay (to ensure correct execution). Implementing the dynamic mechanism requires a set of counters (one for each bank) that are decremented at each cycle, and set to a threshold value whenever they expire or the module is accessed. A zero detector for a counter initiates the memory controller to transmit the instructions for mode transition to the memory modules. 4.2 Software-Directed Scheme It is to be noted that a hardware-directed scheme works well independent of the DBMS and the query optimizer used. This is because the idleness predictors are attached to the memory banks and monitor idleness from the perspective of banks. In contrast, a query-directed scheme gives the task of enforcing mode transitions to the query. This is possible because the query optimizer, once it generates the execution plan, has a complete information about the query access patterns (i.e., which tables will be accessed and in what order, etc). Consequently, if the optimizer also knows the table-to-bank mappings, it can have a very good idea about the bank access patterns. Then, using this information, it can proac-tively transition memory banks to different modes. In this section, we elaborate on each step in the particular query-directed approach that we implemented, which includes customized bank allocation, query analysis, and insertion of bank turn-on/off (for explicit power mode control) instructions. 4.2.1 Bank Allocation In the case of software-directed scheme, the table allocation is handled by the DBMS. Specifically, the DBMS allocates the newly-created tables to the banks, and keeps track of the table-to-bank mappings. When a "create table" operation is issued, the DBMS first checks for free space. If there is sufficient free space available in a single bank, the table is allocated from that bank. If a bank is not able to accommodate the entire table, the table is split across multiple banks. Also, while creating a new table, the DBMS tries to reuse the already occupied banks to the highest extent possible; that is, it does not activate a new bank unless it is necessary. Note that the unactivated (unused) banks i.e., the banks that do not hold any data can remain in the disabled mode throughout the execution . However, it also tries not to split tables excessively. In more detail, when it considers an already occupied bank for a new table allocation, the table boundaries are checked first using the available space in that bank. If a bank is more than two-thirds full with the table data, the rest of the bank is padded with empty bits and the new table is created using pages from a new bank. Otherwise, the table is created beginning in the same bank. Irrespective of whether the table is created on a new bank or not, the DBMS creates a new table-to-bank mapping entry after each table creation. In hardware-directed schemes, we avoid these complexities involved in bank allocation as we assume that there is absolutely no software control. Consequently, in the hardware-directed schemes, we use the sequential first touch placement policy. This policy allocates new pages sequentially in a single bank until it gets completely filled, before moving on to the next bank. Also, the table-to -bank mapping is not stored within the DBMS since the mode control mechanism is handled by the hardware. 4.2.2 Estimating Idleness and Selecting the Appropriate Low-Power Mode It should be emphasized that the main objective of our query-directed scheme is to identify bank idleness. As explained above, in order to achieve this, it needs table-to-bank mapping. However , this is not sufficient as it also needs to know when each table will be accessed and how long an access will take (i.e., the query access pattern). To estimate this, we need to estimate the duration of accesses to each table, which means estimating the time taken by the database operations. Fortunately, the current DBMSs already maintain such estimates for query optimization purposes <A href="81.html#10">[12, 15, 29, 33, 34]. More specifically, given a query, the optimizer looks at the query access pattern using the generated query plan. The inter-access times are calculated using the query plan. A query plan elucidates the operations within a query and also the order in which these operations access the various tables in the database. Even in current databases, the query plan generator estimates access costs using query plans <A href="81.html#10">[12]. We use the same access cost estimation methodology. These access costs are measured in terms of page (block) fetches. In our memory-resident database case, a page is basically the block that is brought from memory to the cache. For instance, the cost of sequential scan is defined as follows (taken from <A href="81.html#10">[12]): Cost seq scan = N blocks + CPU N tuples Here, N blocks is the number of data blocks retrieved, N tuples is the number of output tuples, and CPU is the fudge factor that adjusts the system tuple-read speed with the actual memory hierarchy data-retrieval speed. Usually, optimizers use the above cost metric to choose between multiple query plan options before issuing a query. We attach a cost to each page (block) read/write operation to obtain an estimate of the access cost (time) in terms of execution cycles. For instance, the above scan operation is modified as follows: Cost block f etch = T cycles Cost seq scan = N blocks T + CPU N tuples block tuples T In these expressions, T is the delay in cycles to fetch a block from the memory. Thus, our cost plan is projected in terms of access cycles. We extend this to other database operations like JOIN and AGGREGATE based on the cost models defined in <A href="81.html#10">[14, 12]. Given a query, we break down each operation within the plan (including sub-plans) and estimate the access cost (in cycles) for each 221 - &gt; scan A (9000 cycles) - &gt; aggregate (20 cycles) - &gt; scan B (9000 cycles) - &gt; scan A (9000 cycles) - &gt; scan A - &gt; Put A=ON - &gt; aggregate - &gt; Put B=OFF - &gt; scan B - &gt; Put B=ON - &gt; Put A=OFF - &gt; scan A - &gt; Put A=ON (B is already OFF) P2 P1 (i) (ii) Figure 6: Example application of the query-directed scheme. (i) The original execution plan. (b) The augmented execution plan. primitive operation. Our objective in estimating the per-operation time in cycles is to eventually identify the inter-access times of operations in the query (and hence, to put the banks that hold unused tables to low-power modes). There are table accesses associated with each operation, and bank inter-access times depend on the table inter-access times. A query has information of the tables that it accesses. Thus, knowing the inter-access time for each operation leads to the inter-access times for each table as well. A table is mapped to certain banks, and the table-to-bank mapping is available in the query optimizer. Consequently, if the table inter-access time is T , and the resynchronization time is T p (assuming less than T ), then the optimizer can transition the associated modules into a low-power mode (with a unit time energy of E p ) for the initial T - T p period (which would consume a total [T - T p ]E p energy), activate the module to bring it back to the active mode at the end of this period following which the module will resynchronize before it is accessed again (consuming T p E a energy during the transition assuming that E a is the unit time energy for active mode as well as during the transition period). As a result, the total energy consumption with this transitioning is [T - T p ]E p + T p E a without any resynchronization overheads , while the consumption would have been T E a if there had been no transitioning (note that this calculation considers only the idle period). The DBMS optimizer evaluates all possible choices (low-power modes) based on corresponding per cycle energy costs and resynchronization times, and table inter-access time to pick up the best choice. Note that the DBMS can select different low power modes for different idle periods of the same module depending on the duration of each idle period. Specifically, we use the most energy saving low-power mode without increasing the original query execution time (i.e., when the original idleness is over, the module should be up in the active mode ready for the operation). 4.2.3 Inserting Bank-On/Off Instructions The last part of the software-directed scheme is to insert explicit (operating) mode transitioning instructions in the final query execution plan. For this, we introduce place-markers (mapped to system calls) which are interpreted at the low-level (interpreted later by our memory controller, which actually sets the corresponding low-power modes). This is done so that the query execution engine can issue the query without much performance overhead, and with the same transparency. As an example, consider the following. Let tables A and B each have 1000 records, each record being 64 bytes. Consider the query plan depicted in Figure <A href="81.html#5">6(i), taken from PostgreSQL. The query plan reads from bottom to top (P2 follows P1). A scan of table A is done first, followed by a scan of table B. The result of these operations are then used by an aggregate operation. Another (independent ) scan operation on table A follows the aggregate operation. The per step access costs are also shown. From the generated query plan, it is evident that table A is not accessed between point P1 and point P2. Once the results are extracted after the scan at point P1, the banks that hold table A can be put to a low-power mode, and the banks that hold table B can be activated for data extraction. This is illustrated in Figure <A href="81.html#5">6(ii) using place-markers for tables A and B. Banks holding Table A are reactivated at point P2 (banks of Table B remain off). EXPERIMENTAL EVALUATION OF HARDWARE-DIRECTED AND QUERY-DIRECTED SCHEMES In this section, we study the potential energy benefits of our hardware and software-directed schemes. We first explain the experimental setup that we used in our simulations. Then, the set of queries that we used to study our schemes is introduced. After that, we present energy consumption results. While we discuss the energy benefits of using our schemes, we also elaborate the overheads associated with supporting each of our schemes. 5.1 Setup 5.1.1 Simulation Environment As mentioned before, the query-directed schemes are implemented in the query optimizer of the memory database model elaborated in Section <A href="81.html#2">3.1. We interface this DBMS to an enhanced version of the SimpleScalar/Arm simulator <A href="81.html#10">[4] to form a complete database system. The intermediate interface (invoked by DBMS) provides a set of operating system calls (on Linux kernel 2.4.25), which in turn invokes the SimpleScalar simulator. The SimpleScalar simulator models a modern microprocessor with a five-stage pipeline: fetch, decode, issue, write-back, and commit. We implemented our hardware techniques within the framework of the sim-outorder tool from the SimpleScalar suite, extended with the ARM-ISA support <A href="81.html#10">[4]. Specifically, we modeled a processor architecture similar to that of Intel StrongARM SA-1100. The modeled architecture has a 16KB direct-mapped instruction cache and a 8KB direct-mapped data cache (each of 32 byte-length). We also model a 32-entry full associative TLB with a 30-cycle miss latency. The off-chip bus is 32 bit-wide. For estimating the power consumption (and hence, the energy consumption), we use the Wattch simulator from Princeton University <A href="81.html#10">[8]. Our banked memory model is based on <A href="81.html#10">[13,21], as shown in Figure <A href="81.html#3">4. We use values from Figure <A href="81.html#3">4 for modeling the delay (transition cycles) in activation and resynchronization of various power-states . Our simulations account for all performance and energy overheads incurred by our schemes. In particular, the energy numbers we present include the energy spent in maintaining the idleness predictors (in the hardware-directed scheme) and the energy spent in maintaining the table-to-bank mappings (in the query-directed scheme), and in fetching and executing the bank turn-on/off instructions (in the query-directed scheme). The predictors were implemented using decrementing counters (equal to the number of banks) and zero detector based on the discussion in Section <A href="81.html#4">4.1. The predictors are synchronized with the system cycles to maintain consistency of operation, and to minimize the overheads. The query optimizer maintains the table-bank mappings, which is modeled as an array list for instant access. The bank turn-on/off instructions are executed by setting hardware registers, and hence, these instructions are modeled as register operations using the existing instruction set architecture. We present two important statistics in our experimental results. Energy consumption corresponds to the energy consumed in the memory system (including the above mentioned overheads). We also present statistics about the performance overhead (i.e., increase in execution cycles) for each of our schemes. This overhead includes the cycles spent in resynchronization (penalty cycles are modeled based on values in Figure <A href="81.html#3">4) as well as the cycles spent (in the CPU datapath) in fetching and executing the turn-on/off instructions (in the query-directed scheme). 5.1.2 Queries To evaluate our scheme for memory-resident databases, we considered two classes of queries. The first class is a subset of queries from the Transaction Processing Council (TPC-H) benchmark <A href="81.html#10">[35]. TPC-H involves complex queries with a large amount of data accesses. Operations in decision support benchmarks (TPC-D evolved to TPC-H) have good spatial locality with abundant data intensive operations <A href="81.html#10">[9]. This assists us to perform a rigorous test of our schemes. The top part of Table <A href="81.html#6">1 gives details of the TPC-H queries we used and the corresponding database parameters. The selected operations represent a good mix and could be used to build a variety of complicated queries. 222 Table 1: The two classes of queries considered for our experiments. Source Query Description Tables TPC-H Q6 Simple query PART, CUSTOMER, ORDERS, and LINEITEM tables generated using DBGEN with scale 1.0 Q3 Complex query involving JOIN Q4 Complex query involving NEST Q17 Complex query involving JOIN and NEST Queries targeting a simple organizer P1 Simple name and address lookup ADDRESSBOOK populated with 1.3 million entries, 50% subset of FRIENDS and 25% subset of COLLEAGUES P2 Lookup in directory of friends P3 Lookup in directory of colleagues and friends Memory-resident databases run queries that are different from the typical database queries as seen in TPC-H. The second set of queries that we consider are representative of applications that execute on handheld devices. The typical operations that are performed on an organizer were imitated on our setup (we name the queries P1, P2, P3). The first query involves a simple address lookup using a `NAME' as input. The SQL for query P1 is shown on the left section of Table <A href="81.html#6">2. Recent organizers <A href="81.html#10">[17, 25] provide an ordered view of the underlying addressbook database. For instance, organizers provide the creation of folders. A "friends" folder can be a collection of personnel with a tag set as "friend" in the addressbook . We defined folder as a restrained/customized view of the same database (address book). Intuitively, query P2 strives to do a lookup of friends living in a particular city. The "friends" view and hence the query P2 is defined on the right section of Table <A href="81.html#6">2. Query P3 combines views (folders). For this we defined a new folder called "colleagues". P3 aims to find friends and/or colleagues whose names start with an `a', living in a particular `CITY'. Since P3 is very similar to P2 with some extra fields, we do not present the SQL for P3. The intermediate tables and results during query execution are also stored in the memory. 5.1.3 Default Parameters For our experiments, we populate our database using the DBGEN software from TPC-H benchmark suite with a scale factor of 1.0. Our organizer database is populated with 1.3 million records. For dynamic threshold scheme, we use 10, 100 and 10,000 cycles as idle stndby , idle nap , and idle down , respectively. For all schemes, the banks are in power-down mode before their first access. On/Off instructions are inserted based on the inter-access times of table. We use the same cycles as in idle stndby , idle nap , and idle down for inserting instructions. As an example, consider the inter-access (T) of a table as 25 cycles, which lies between 10 (idle stndby ) and 100 (idle nap ) cycles. We insert an On/Off instruction at the beginning of T to put a table to standby mode for 24 cycles, taking into consideration the resynchronization period of 1 cycle as well. Similar technique is applied for inter-access times that fall in between other power modes. A single page transfer time is needed for access cost calculation in software-directed scheme. We derive this by executing the TPC-H queries on the SimpleScalar simulator (with the SA-1100 model) and by studying the cycle times for transferring a data block from memory to the cache. For all experiments, the default configuration is the 512MB RDRAM memory with 8MB banks. In the following section, we study the energy implications of our hardware and software schemes using this setup. We then present the performance overheads. 5.2 Query Energy Evaluation Figure <A href="81.html#6">7 shows the normalized memory energy consumption for our hardware-directed schemes. While presenting our results, we normalize all values with respect to the base case, which is the version with no query optimizations. "Static Standby" in Figure <A href="81.html#6">7 indicates the static standby scheme. We see that, by simply putting the modules to standby mode after each access, this scheme is able to achieve an average 55% reduction in memory energy consumption of TPC-H queries when compared to the unoptimized case. The energy improvements are less pronounced in the case of handheld Table 2: SQL for organizer queries Query P1 Query P2 SELECT CREATE VIEW a name, friends AS P2: a address, SELECT SELECT a city, a name, a address, a office phone, a address, a home phone, a home phone, a city, a mobile phone a mobile phone, a home phone, FROM a email, a mobile phone friends a web, FROM WHERE a specialnotes addressbook a city = `[CITY]' FROM WHERE GROUP BY addressbook a tag = `[FRIEND]' a name; WHERE GROUP BY a name = '[NAME]'; a name; queries (37% reduction on the average). This is mainly because of the different number of tables manipulated by these two types of queries. In the TPC-H case, multiple tables are scattered across various banks and hence, there is a potential of placing more memory banks into low-power modes. In the case of handheld queries, there is just one table scattered across multiple banks, which makes putting modules to a low-power mode more difficult as modules are tightly connected, as far as query access patterns are concerned. We also observe from Figure <A href="81.html#6">7 that the dynamic threshold scheme further extends these improvements through its ability to put a bank into any of the possible low-power modes. On an average, there is a 60% (43%) energy improvement in TPC-H (handheld) queries. Figure <A href="81.html#6">7 also shows the normalized energy behavior of our query-directed scheme (denoted On/Off Instr). It is evident that this scheme outperforms the best hardware-directed scheme (by an average of 10%) in saving the memory energy consumption. This is because of two main reasons. First, when a bank idleness is estimated, the query-directed scheme has a very good idea about its length (duration). Therefore, it has a potential of choosing the most appropriate low-power mode for a given idleness. Second, based on its idleness estimate, it can also preactivate the bank. This 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Q6 Q3 Q4 Q17 P1 P2 P3 No r m aliz ed E n er g y Static Standby Dynamic Threshold On/Of f Instr Figure 7: Energy consumption of hardware and software-directed modes. The values shown are normalized to the version with no energy optimizations. 223 eliminates the time and energy that would otherwise have spent in resynchronization. Consequently, the average memory energy consumption of the query-directed scheme is just 32% of the unoptimized version for TPC-H queries, and 44% in case of organizer (handheld) queries [i.e., an additional 8% (13%) improvement over the hardware schemes for TPC-H (handheld) queries]. 5.3 Performance Overhead Analysis Our techniques are very effective in reducing the memory energy consumption. As mentioned earlier, transitions from the low-power modes to the active mode come with an overhead of resynchronization (in terms of both performance and energy). The energy values reported in previous section take into consideration the extra energy needed to activate the modules as well. In this part, we quantify the basic performance overheads that are faced in supporting our schemes. Figure <A href="81.html#7">8 shows the performance overheads for both the hardware and software-directed schemes. The static standby scheme has the maximum overhead, which is expected. This is especially the case when queries generate frequent memory accesses. The memory is brought down to the standby mode after each access, and is resyn-chronized in another access that follows immediately. As a result, the performance worsens as bad as 28% for the static standby case. On the other hand, for the dynamic threshold scheme, the performance overhead is slightly better since the banks are not blindly put to a low-power mode after each access. This verifies our prediction that when a module goes to low-power mode, it would either remain for a while in that mode or may even be transitioned into a lower power mode. The query-directed scheme has the least overhead (&lt;2%). The main reason for this is the ability of pre-activating a bank before it is actually accessed. Therefore, considering both performance and energy results, one may conclude that the query-directed scheme is better than the hardware-directed schemes. However, it is also to be noted that the query-directed scheme requires access to the query optimizer. In comparison, the hardware-based schemes can work with any query optimizer. Therefore, they might be better candidates when it is not possi-ble/profitable to modify the query plan. QUERY RESTRUCTURING The approaches presented above mainly try to optimize energy consumption without modifying the queries themselves (except maybe for the query-directed scheme where we insert turn on/off instructions in the query plan). In this section, we go one step further, and demonstrate that even larger energy savings are possible if one has the flexibility of reorganizing query operations. We show how this can be achieved in the context of both individual queries and multiple queries (optimized simultaneously). Our main objective in restructuring queries is to increase memory bank inter-access times. Note that when bank inter-access time is increased, we can either remain in a given low-power operating mode longer, thereby feeling the potential impact of resynchronization less (i.e., amortizing the cost of resynchronization); or we can switch to a more energy saving mode (as we now have a longer idleness), which means more energy savings. We present different query restructuring strategies for achieving this. When considering a single query, the bank inter-access times can be increased by re-ordering query operations. On the other hand, the primary goal of the heuristic that targets at multiple-queries is to cluster the usage of tables from multiple queries together, so that the overall table accesses are localized. That is, assuming that we have multiple queries to optimize, our objective is to interleave these query executions in such a way that the reuse of individual tables (or of table portions) is maximized. In other words, when a table is accessed, we want to execute all other query operations (potentially coming from different queries) to that table (one after another), before we move to the next table. This also tends to cluster accesses to the same bank, and tends to increase the bank inter-access times (which is very important from an energy perspective as explained above). In the following, we first study intra-query restructuring and then inter-query restructuring. After these two steps, bank turn-on/off instructions are inserted at the relevant points, depending on the bank access patterns. 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Q6 Q3 Q4 Q17 P1 P2 P3 Norm a l i z ed P e r f orm a n c e Static Standby Dynamic Threshold On/Off Instr Figure 8: The performance overhead involved in supporting our schemes. There is an average overhead of 15%, 8%, and 2% for standby, dynamic and on/off schemes, respectively, over the unoptimized version. Step 1 (intra-query optimization): A query is first examined to see if there are any potential reuse regions. If there are any reusable regions, their accesses are grouped together. We achieve this by examining the query execution plan. The query plan is studied to see if there are any advantages in rearranging the operations (primitives) in a query based on its table usage. Operations that require the same (set of) table(s) are then grouped together (i.e., they are scheduled to be executed one after another). The detailed procedure is shown in Figure <A href="81.html#8">9. Each operation in the query plan is first scanned and placed into a table group based on the table(s) that it accesses. Then, the operations are rearranged in the query plan (taking into account the dependencies between them) based on their corresponding table groups. For this, we look at the query plan tree. The path from each leaf node to the root, called stream, is investigated. The ultimate goal is to schedule operations (nodes in the plan tree) based on their table groups. We try to schedule operations within one table group (which is currently active) before scheduling the operations from another table group (which is not active) in an attempt to increase the bank inter-access times. That is, a stream is traversed from bottom to top, and each node within the stream is put to the schedule queue (as they are encountered) based on its table group. It should be emphasized that we preserve the original semantics of the operations (constraints) in the algorithm. This procedure is repeated for each stream in the tree, and until all streams have the most energy-efficient schedule based on their table accesses. At the end of this step, an energy-aware schedule queue gets generated for the considered query (saved in schedule list). Step 2 (inter-query optimization): Tables are examined to optimize multiple queries simultaneously. For each table that is accessed, all accesses arising from multiple queries to the particular table are grouped together. In this step, the schedule list from multiple queries are grouped together. Each list is scanned to identify nodes that access a given table. The nodes that access the same table are then scheduled to execute together (without disturbing the dependency constraints). In fact, the nodes from multiple queries are just grouped (combined ) not reordered. Thus, in this step, the constraint flow for each schedule list (taken care of in Step 1) is automatically maintained . Additional conditional flow checks could be reinforced at this stage if desired. Figure <A href="81.html#8">10 shows the regrouping procedure. f inal schedule list stores the final consolidated schedule of operations from all the queries. Step 3 (energy optimizations): Include energy optimizations by inserting On/Off instructions into the final schedule list. In this step, the access costs are calculated for each operation in the f inal schedule list as shown in Section <A href="81.html#4">4.2.2. Each operation is attached with an access cost, and the turn-on/off instructions are inserted based on the table inter-access times. The methodology used for adding these instructions to the f inal schedule list is the same as in Section <A href="81.html#4">4.2.2, and the on/off markers are placed as elaborated in Section <A href="81.html#5">4.2.3. As an example, consider two queries Q1 and Q2. Their original 224 table_group is a table-to-operations mapping list. schedule_list stores the final schedule of operations. /* identify the group to which an * operation belongs */ operation_rearragement (){ for (each operation in query i) { identify the table(s) in i; for (each table j in i) { add operation to table_group[j]; } } schedule_operations(); } /* schedule operations */ schedule_operations() { schedule_list = empty; do { for (each stream in query plan tree) { start from leaf node; for (each node in stream) { identify its constraint nodes that follow; /* the rest are independent nodes */ group(constraint nodes); group(independent nodes); check for new violations; add new constraints if necessary; save the schedule_list; move up a node in the stream; } move to the next stream; } } until no more changes } /* group nodes */ group(node_list) { if(node_list is constraint node list) { for (each node in node_list) { lookup table_group of node; add node to schedule_list based on table_group; /* preserve the dependency order */ preserve flow of node_list in schedule_list; } } else { /* set of independent nodes */ add node to schedule_list based on table_group; /* no need to preserve constraint flow */ regroup to put all table_group nodes together; } Figure 9: Reorganizing operations within a query to optimize for energy (Step 1). The query tree is investigated from the bottom to top for grouping operations based on their table accesses . group_multiple_queries { for (each schedule_list) { do { pick an unscheduled node i in schedule_list; /* i.e. pick a node without a &quot;complete&quot; tag */ for (other schedule_lists) { if (node j has same table_group as node i) { schedule node j after node i; mark node j as &quot;complete&quot;; /* with respect to multi-query schedule */ } } } until all node in schedule_list is &quot;complete&quot; } } Towards the end of the procedure, final_schedule_list stores the entire list of &quot;complete&quot; schedule. Figure 10: Grouping schedule list from multiple queries (Step 2). Operations from multiple queries are grouped based on their table accesses using their corresponding schedule lists. query plan is shown in Figure <A href="81.html#9">11(i). Q1 is revised as the table accesses are optimizable. Figure <A href="81.html#9">11(ii) shows the result after applying Step 1. Step 2 results in the output depicted in Figure <A href="81.html#9">11(iii). Finally , in Step 3, we insert on/off instructions in appropriate places (see Figure <A href="81.html#9">11(iv)). EXPERIMENTAL EVALUATION OF QUERY RESTRUCTURING In this section, we evaluate our query restructuring approach by extending our database and queries discussed in Section <A href="81.html#5">5.1.2. As before, our focus is on memory energy consumption. We also study the impact of the technique on the overall performance. Towards the end, other alternative options are also elaborated. 7.1 Multi-Query Setup Since simultaneous processing of multiple queries is needed to validate our approach, we considered a combination of queries, which we term as scenarios in the rest of this paper. Among the queries considered in Section <A href="81.html#5">5.1.2, there can be multiple combinations of queries that arrive sequentially, and that (which) are optimizable using our technique. The various combination (scenarios) of organizer queries and their naming schemes are shown in Table <A href="81.html#8">3. For instance, P12 indicates that P1 is sequentially processed along with P2. The combination scenarios for TPC-H queries are shown in Table <A href="81.html#8">4. The combinations shown in these tables are the prominent ones and the behavior of other combinations are very similar to these, hence, they are not included in this paper. Table 3: Scenarios for organizer queries. Type Legend Combination Two query combinations P11 P1 + P1 P12 P1 + P2 P23 P2 + P3 Three query combination P123 P1 + P2 + P3 Table 4: Scenarios for TPC-H queries. Type Legend Combination Two query combinations S11 Q6 + Q6 S12 Q6 + Q3 S13 Q6 + Q4 S14 Q6 + Q17 S23 Q3 + Q4 S24 Q3 + Q17 S34 Q4 + Q17 Three query combinations S222 Q3 + Q3 + Q3 S123 Q6 + Q3 + Q4 Four query combinations S1111 Q6 + Q6 + Q6 + Q6 S1234 Q6 + Q3 + Q4 + Q17 7.2 Query Energy Evaluation In this section, we evaluate the query energy of the various scenarios we presented in the previous section. We first study the improvements obtained from our query restructuring heuristic, and further extend our study to combine query restructuring with various hardware and software-directed schemes (of Section <A href="81.html#3">4) meant to improve the energy consumption. Figure <A href="81.html#9">12 shows the sole contribution of query restructuring scheme in improving the energy consumption. The energy reduces by an average 55% from the unoptimized version when our query restructuring scheme is used. By just grouping similar accesses (to ensure data reuse), query restructuring can achieve significant reduction in the energy consumption of multiple queries. In order to identify the benefits coming solely from Step 1 (intra-query optimization) in our query restructuring scheme, we also combined Step 1 and Step 3, and compared it with our query-directed scheme (studied in Section <A href="81.html#4">4.2 -- which is simply Step 3 of our query restructuring). Figure <A href="81.html#9">13 shows the results. There is up to 19% improvement in energy when operations are shuffled with a query based on their table usage. When the query restructuring scheme is combined with hardware-directed schemes, there is further improvement in energy savings (Figure <A href="81.html#9">14). The static standby scheme works only for small queries that have a uniform access pattern, but when complex queries are encountered, the dynamic runtime scheme outperforms the static standby scheme due to its good prediction of 225 - &gt; function(A) (9000 cycles) - &gt; hash join - &gt; scan B (9000 cycles) - &gt; scan A (9000 cycles) Q1 - &gt; aggregate (20 cycles) - &gt; scan B (9000 cycles) - &gt; scan A (9000 cycles) Q2 - &gt; hash join - &gt; scan B (9000 cycles) - &gt; function(A) (9000 cycles) - &gt; scan A (9000 cycles) Q1 - &gt; aggregate (20 cycles) - &gt; scan B (9000 cycles) - &gt; scan A (9000 cycles) Q2 (i) Original Queries (ii) After applying Step 1 - &gt; aggregate (from Q2) - &gt; hash join (from Q1) - &gt; scan B (from Q2) - &gt; scan B (from Q1) - &gt; function(A) (from Q1) - &gt; scan A (from Q2) - &gt; scan A (from Q1) Q1 + Q2 (iii) After applying Step 2 - &gt; aggregate (from Q2) - &gt; Put B=OFF (A is already OFF) - &gt; hash join (from Q1) - &gt; scan B (from Q2) - &gt; scan B (from Q1) - &gt; Put B=ON - &gt; Put A=OFF - &gt; function(A) (from Q1) - &gt; scan A (from Q2) - &gt; scan A (from Q1) - &gt; Put A=ON (B is already OFF) (iv) After applying Step 3 Q1 + Q2 Figure 11: Example of query restructuring and regrouping based on energy behavior. the application behavior. This can be seen in Figure <A href="81.html#9">14, where the dynamic threshold scheme performs better in the TPC-H scenarios than for the handheld query scenarios. The savings obtained by putting a module into multiple low-power modes for longer periods are more than the savings obtained by periodically putting a module to just standby mode. The software-directed schemes perform similar to dynamic the runtime threshold strategy when combined with the query restructuring algorithm. In Figure <A href="81.html#9">14, the insertion of explicit turn-on/off instructions improves the energy by an average 78%, when compared to the unoptimized version. This result is comparable to the improvements obtained using the dynamic threshold scheme. In fact, the dynamic threshold scheme performs slightly better for some TPC-H queries (e.g., S12, S13, and S14). This situation occurs due to the following factor. When multiple queries are combined using query restructuring, it becomes difficult to predict the inter-access times since each query has a varying access pattern, and combining random access patterns complicates the job of the predictor (and requires a more sophisticated predictor). The runtime schemes work at the hardware instruction level without any knowledge of the DBMS application. But, this illustrates how a simple software technique implemented at the query optimizer (by just analyzing the high-level query structure) is able to achieve improvements as good as an equivalent but expensive hardware technique . As mentioned earlier in the paper, when queries are restructured and grouped, the memory access pattern changes. The bank turn-on/off instructions can be inserted only in prominent "hot" and "cold" access regions, respectively. There are a few modules, which is beyond the control of software. For instance, we insert turn-on/off instructions based on tables. A given table could be scattered across many modules. Our predictor estimates the inter-access time for which the entire table needs to be put to low-power mode. However, even during a table access, there are regions (modules) that are hardly used. Dynamic runtime scheme is extremely good in handling this situation by its ability to put individual modules to a low-power state based on just that module's access. This implies that the combination of hardware and software schemes form the best strategy when query restructuring is deployed. Figure <A href="81.html#9">14 also shows the case when both dynamic runtime scheme and the turn-on/off instructions are used in tandem after 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 P11 P12 P23 P123 S11 S12 S13 S14 S23 S24 S34 S222 S123 S1111 S1234 N o r m al i z ed E ner gy Figure 12: Contribution of query restructuring towards energy improvements. The energy values shown are normalized to the version with no optimizations. query restructuring. The benefits obtained from such a hardware-software interaction is prominent. There is an average 90% reduction in the memory energy consumption across the applications . In some cases, there is up to 95% improvement in the energy consumption. These results clearly show that query restructuring combined with the use of low-power operating modes can lead to significant energy savings. 7.3 Performance Overhead Analysis Query restructuring combined with the use of low-power modes has an impact on the performance. In Figure <A href="81.html#10">15, we present the normalized system-wide performance of our query restructuring scheme. It is evident that the performance improves by an average of 48% when multiple queries are restructured and grouped. The improvement in performance is mainly due to the improved locality utilization in the memory hierarchy. That is, the data brought to the cache by one query is reused by other queries (as a result of restructuring). We do not present here detailed cache behavior statistics due to lack of space. Figure <A href="81.html#10">16 shows the normalized performance for the combination schemes as well. When static standby scheme is used with query restructuring, the performance improvements obtained from query restructuring gets negated by the resynchronization overhead from the standby mode for each access. Thus, the performance 0% 2% 4% 6% 8% 10% 12% 14% 16% 18% 20% Q6 Q3 Q4 Q17 P1 P2 P3 E n erg y I m prov e m e n t s Figure 13: Benefits obtained by restructuring operations within a query (contribution of Step 1). 0 0.1 0.2 0.3 0.4 0.5 0.6 P11 P12 P23 P123 S11 S12 S13 S14 S23 S24 S34 S222 S123 S1111 S1234 Norm a l i z ed E n erg y Restructuring + Static Standby Restructuring + Dynamic Threshold Restructuring + On/Off Instr Restructuring + On/Off Instr + Dynamic Threshold Figure 14: Energy consumption reduces significantly when low-power modes are utilized along with query restructuring scheme. Values shown are normalized to the unoptimized version . Best energy savings comes from a hybrid hardware-software scheme. 226 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 P11 P12 P23 P123 S12 S13 S14 S23 S24 S34 S222 S123 S1111 S1234 Norm a l i z ed P e r f orm a nc e Figure 15: Performance improvement obtained from basic query restructuring over the unoptimized version. 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 P11 P12 P23 P123 S12 S13 S14 S23 S24 S34 S222 S123 S1111 S1234 N o r m al i z e d P e r f o r m anc e Restructuring + Static Standby Restructuring + Dynamic Threshold Restructuring + On/Off Instr Restructuring + On/Off Instr + Dynamic Threshold Figure 16: Overall performance after applying energy optimizations along with query restructuring. Values shown are normalized to the unoptimized version. worsens in some cases by even 65% for complicated queries. However , overall, there is still a 10% performance improvement for all applications on the average. The turn-on/off instructions have the least performance overhead, and hence, preserve the performance improvements obtained from query restructuring. From Figure <A href="81.html#10">16, this combination shows a 47% improvement in performance (negating the improvements obtained from basic query restructuring by a mere 1%). Dynamic runtime threshold on the other hand negates the performance improvements from query restructuring by average 6%. Combining turn-on/off instructions with dynamic runtime threshold shows an average performance improvement of 45% for applications, which implies a 3% overhead addition from the low-power schemes towards query restructuring. Thus, it is clear that query restructuring with both turn-on/off instructions and runtime threshold forms the best alternative from both energy consumption and performance perspectives. CONCLUDING REMARKS This paper is an attempt to study the potential of employing low-power operating modes to save memory energy during query execution. We propose hardware-directed and software-directed (query-directed) schemes that periodically transition the memory to low-power modes in order to reduce the energy consumption of memory-resident databases. Our experimental evaluations using two sets of queries clearly demonstrate that query-directed schemes perform better than hardware-directed schemes since the query optimizer knows the query access pattern prior to query execution, and can make use of this information in selecting the most suitable mode to use when idleness is detected. This scheme brings about 68% reduction in energy consumption. In addition, the query-directed scheme can also preactivate memory banks before they are actually needed to reduce potential performance penalty. Our query restructuring scheme based on memory bank accesses provides another scope for optimization. One can re-order operations within a query to increase bank inter-access times. It is also possible to go beyond this, and consider the access patterns of multiple queries at the same time. Multiple queries are optimized based on their table accesses, i.e., all accesses to a table are clustered as much as possible. This scheme is able to put memory banks to low-power operating modes for longer periods of time due to fewer table activations. There is up to 90% improvement in energy and 45% improvement in performance when queries are restructured and regrouped based on their table accesses. Overall, we can conclude that a suitable combination of query restructuring and low-power mode management can bring large energy benefits without hurting performance. REFERENCES [1] R. Alonso and S. Ganguly. Query optimization for energy efficiency in mobile environments. In Proc. of the Fifth Workshop on Foundations of Models and Languages for Data and Objects, 1993. [2] N. An, S. Gurumurthi, A. Sivasubramaniam, N. Vijaykrishnan, M. Kandemir, and M.J. Irwin. Energy-performance trade-offs for spatial access methods on memory-resident data. The VLDB Journal, 11(3):179197, 2002. [3] N. Anciaux, L. Bouganim, and P. Pucheral. On finding a memory lower bound for query evaluation in lightweight devices. Technical report, PRiSM Laboratoire de recherche en informatique, 2003. [4] T. M. Austin. The simplescalar/arm toolset. SimpleScalar LLC. http://www.simplescalar.com/. [5] Birdstep Technology. Database Management In Real-time and Embedded Systems - Technical White Paper, 2003. http://www.birdstep.com/collaterals/. [6] Bloor Research Ltd. Main Memory Databases, November 1999. [7] P.A. Boncz, S. Manegold, and M.L. Kersten. Database architecture optimized for the new bottleneck: Memory access. In The VLDB Journal, pages 5465, 1999. [8] D. Brooks, V. Tiwari, and M. Martonosi. Wattch: a framework for architectural-level power analysis and optimizations. In Proc. International Symposium on Computer Architecture, 2000. [9] Q. Cao, P. Trancoso, J.-L Larriba-Pey, J. Torrellas, R. Knighten, and Y. Won. Detailed characterization of a quad pentium pro server running tpc-d. In Proc. of the International Conference on Computer Design, 1999. [10] S. Chaudhuri and K. Shim. Optimization of queries with user-defined predicates. ACM Transactions on Database Systems, 24(2):177228, 1999. [11] G.P. Copeland, T. Keller, R. Krishnamurthy, and M. Smith. The case for safe ram. In Proc. of the Fifteenth International Conference on Very Large Data Bases, pages 327335, 1989. [12] Database Management System, The PostgreSQL Global Development Group. PostgreSQL 7.2, 2001. http://www.postgresql.org/. [13] V. Delaluz, M. Kandemir, N. Vijaykrishnan, A. Sivasubramaniam, and M.J. Irwin. Dram energy management using software and hardware directed power mode control. In Proc. of the International Symposium on High-Performance Computer Architecture, 2001. [14] Z. Fong. The design and implementation of the postgres query optimizer. Technical report, University of California, Berkeley. http://s2k-ftp.cs.berkeley.edu:8000/postgres/papers/. [15] P. Gassner, G.M. Lohman, K.B. Schiefer, and Y. Wang. Query optimization in the ibm db2 family. Data Engineering Bulletin, 16(4):418, 1993. [16] Le Gruenwald and S.M. Banik. Energy-efficient transaction management for real-time mobile databases in ad-hoc network environments. In Proc. of the Second International Conference on Mobile Data Management, 2001. [17] Handspring. Handspring Organizers, 2004. http://www.handspring.com/products/. [18] J.M. Hellerstein. Optimization techniques for queries with expensive methods. ACM Transactions on Database Systems, 23(2):113157, 1998. [19] T. Imielinski, S. Viswanathan, and B.R. Badrinath. Energy efficient indexing on air. In Proc. of ACM SIGMOD Conference, 1994. [20] Intel Corporation. Intel 440BX AGPset: 82443BX Host Bridge/Controller Data Sheet, April 1998. [21] A.R. Lebeck, X. Fan, H. Zeng, and C.S. Ellis. Power aware page allocation. In Proc. of the International Conference on Architectural Support for Programming Languages and Operating Systems, 2000. [22] S. Madden, M.J. Franklin, J.M. Hellerstein, and W. Hong. The design of an acquisitional query processor for sensor networks. In Proc. of the ACM SIGMOD International Conference on Management of Data, pages 491502. ACM Press, 2003. [23] S. Manegold. Understanding, Modeling, and Improving Main-Memory Database Performance. Ph.d. thesis, Universiteit van Amsterdam, Amsterdam, The Netherlands, December 2002. [24] C.L. Monma and J.B. Sidney. Sequencing with series-parallel precedence constraints. Mathematics of Operations Research, 4:215224, 1979. [25] Palm Inc. Palm Handhelds, 2004. http://www.palm.com/products/. [26] The PostgreSQL Global Development Group. PostgreSQL 7.2 Developers Guide, 2002. http://www.postgresql.org/docs/. [27] P. Pucheral, L. Bouganim, P. Valduriez, and C. Bobineau. Picodbms: Scaling down database techniques for the smartcard. The VLDB Journal, 12(1):120132, 2001. [28] J.M. Rabaey, A. Chandrakasan, and B. Nikolic. Digital Integrated Circuits. Prentice Hall, second edition, 2002. [29] R. Ramakrishnan and J. Gehrke. Database Management Systems. McGraw-Hill publishers, third edition, 2002. [30] Rambus Inc. Rambus RDRAM 512MB Datasheet, 2003. [31] Samsung Microelectronics. Mobile 512MB DRAM Chip Series. http://www.samsung.com/Products/Semiconductor/. [32] S. Sarawagi and M. Stonebraker. Reordering query execution in tertiary memory databases. In The VLDB Journal, pages 156167, 1996. [33] A. Silberschatz, H.F. Korth, and S. Sudarshan. Database System Concepts. McGraw-Hill, fourth edition, 2001. [34] Sleepycat Software. Berkeley DB V4.2, 2004. http://www.sleepycat.com/docs/index.html. [35] Transaction Processing Performance Council. TPC-H Benchmark Revision 2.0.0, 2003. 227

hardware energy scheme;query-directed energy management;power consumption;memory-resident databases;database;energy;low-power modes;query-directed scheme;banked memory;multi-query optimization;DRAM;energy optimization;query restructuring

82

Enforcing Security and Safety Models with an Information Flow Analysis Tool

Existing security models require that information of a given security level be prevented from "leaking" into lower-security information. High-security applications must be demonstrably free of such leaks, but such demonstration may require substantial manual analysis. Other authors have argued that the natural way to enforce these models automatically is with information-flow analysis, but have not shown this to be practicable for general purpose programming languages in current use. Modern safety-critical systems can contain software components with differing safety integrity levels, potentially operating in the same address space. This case poses problems similar to systems with differing security levels; failure to show separation of data may require the entire system to be validated at the higher integrity level. In this paper we show how the information flow model enforced by the SPARK Examiner provides support for enforcing these security and safety models. We describe an extension to the SPARK variable annotations which allows the specification of a security or safety level for each state variable, and an extension to the SPARK analysis which automatically enforces a given information flow policy on a SPARK program.

INTRODUCTION Software is often used to develop security-critical applications . Some of these applications are required to manage information of different classification levels, ensuring that each user may only access the data for which he or she has adequate authorisation. This requirement has two components; the developer must produce a design and implementation which supports this model and its constraints, and the implementation must support verification that the constraints are never violated. It is not enough for the implementation to be correct; for it to be secure, it must be seen to be correct. In this paper we examine the problem of developing and verifying a security-critical application containing this security structure (often termed multi-level security). We will analyse recent work on information flow and security and show how the information flow analysis of the SPARK Examiner tool is appropriate for solving this problem. We will show how it is also important for the efficient analysis of safety-critical systems. We will then describe modifications to the current definition of the SPARK Ada language[2] and the Examiner which permit complete validation of Ada programs against defined security models such as the Bell-LaPadula model[3]. We intend that the additional flow analysis features described here will appear in a future commercial Examiner release. Although we anticipate possible minor changes to the syntax and analysis performed, we expect the released version to implement the analysis described here in all substantial aspects. EXISTING WORK In this section we identify typical standards which will inform the development and verification of a security-critical or safety-critical application. We then analyse a survey paper on information flow and security, and compare its conclusions against the information flow model of the SPARK annotated subset of Ada95. 2.1 Standards The Common Criteria for IT Security [5] specify the development and verification activities that are suitable for applications at differing levels of security. These Evaluation Assurance Levels (EALs) range from EAL-1 (lowest) to EAL-7 (highest). As the EAL number rises, the required activities become more rigorous; at the higher levels, formal specification and analysis of the software becomes required. 39 For safety-related applications there are similar concepts for the safety criticality of software; RTCA DO-178B[10] for civil avionics software defines criticality levels E (lowest ) through to A (most critical). As one would expect, the required development and verification activities become increasingly more onerous (and expensive) with increasing criticality level. As a result, if an avionics system were to contain a "core" of critical functionality at Level A and a larger body of utility code at Level D then either the entire system would have to be developed and verified at Level A or a rigorous argument would have to be applied that the Level D code could not in any way affect the integrity of the Level A code. Another notation for safety criticality is the Safety Integrity Level (SIL), described in IEC 61508[8]. SIL-1 is the lowest level of safety integrity, and SIL-4 the highest; DO-178B level A approximates to SIL-3/SIL-4 when comparing the development and verification activities required. In this paper we assume that we are attempting to validate systems at EAL-5 or greater (for security) and RTCA Level B or greater (for safety). We are therefore required to provide a rigorous and comprehensive justification for any statements which we make about the separation of data. Therefore we now look at how such statements may be expressed . 2.2 Information Flow "Information flow" as it applies to conventional imperative computer programs has a range of definitions, and is often confused with "data flow"; we shall take the definition as expressed in Barnes[2] p.13: data flow analysis is concerned with the direction of data flow; whereas information flow analysis also considers the coupling between variables. An example of the difference between data flow and information flow comes from the following (Ada) code: while (X &lt; 4) loop A := A + 2 * B; X := X * 2; end loop; Here, data flow analysis would state that data flows from X to X and from A and B to A. Information flow analysis would note additionally that the final value of A is affected by the initial value of X, and hence that there is information flow from X to A. Sabelfeld and Myers[11] recently surveyed the use of information flow analysis in software. They viewed the fundamental problem of maintaining security as one of tracking the flow of information in computing systems. They examined the various ways that secure information could leak into less secure information, overtly and covertly. They identify in particular the concept of implicit information flows, an example of which is given in the while-loop example above. Their survey characterised language-based information flow as requiring: 1. semantics-based security, so that rigorous argument could be made about a variation of a high-security value not affecting a low-security value; and 2. a security type system, so that a program or expression can be assigned security values (types) and the type changes of the program or expression can be characterised . They concluded that "standard" security practices do not and cannot enforce the end-to-end confidentiality required by common security models. They characterised the existing work in security and information flow analysis, but notably did not address the work of Bergeretti and Carre[4]. 2.3 Security models The Bell-LaPadula (BLP) model of computer security[3] enforces two properties: 1. no process may read data from a higher security level; and 2. no process may write data to a lower security level. Such multi-level security has a number of problems; Anderson [1] provides a list of them including: "blind write-up": the inability to inform low-security data whether a write to high-security data has happened correctly; "downgrading": moving information from a high security level to a lower level is sometimes desirable; and "TCB bloat": a large subset of the operating system may end up in the Trusted Computing Base (TCB). The Dolev-Yao security model[6] makes secret information indivisible; it cannot be leaked in part but only in total. 2.4 Information Flow in Ada Bergeretti and Carre wrote a seminal paper[4] describing a practical implementation of information flow analysis in the SPADE Pascal language (although the principles were applicable to most conventional imperative programming languages). This was managed by the composition of matrices representing information flow dependencies between variable imports and exports. Notably, conditional and infinite loops were permissible and analysable within the framework of such a language; these were managed by computing the transitive closure of the information flow matrix corresponding to one execution of the loop. This information flow model was implemented in the SPARK annotated Ada subset[2]. The subset requires each subprogram to be annotated with the required information flow ("derives" annotations) if information flow analysis is required . The SPARK subset is enforced by the SPARK Examiner tool which checks the required information flow of each subprogram against the actual information flow. The annotations (and other SPARK rules such as the ban on circular dependency) are necessary to make this information flow analysis tractable. The result of this is that it is possible, in a fully-analysed SPARK program, to be certain that a given exported variable is independent of a given imported variable. This is a key step towards supporting multi-level security in SPARK Ada, and makes it easier to write demonstrably secure and correct code, but is not a complete solution. In Section 3 we will describe how SPARK Examiner analysis may be extended better to implement these checks. We now examine case studies of the use of SPARK, and the utility of information flow in real applications. 40 2.5 Security and Safety Applications An example of a high-security application which was partly developed in SPARK is the MULTOS CA[7]. This was developed and verified at a high level of integrity, approximating to the Common Criteria EAL-5. The delivered system had a defect rate of 0.04 errors per thousand lines of code (KLOC). This showed that SPARK was a practical language for implementing high-security applications. Part of the analysis required during development and verification of the CA was to show that secret data could not leak out directly or indirectly in unclassified outputs. An example of a safety-critical application with mixed integrity levels where information flow analysis was helpful was the SHOLIS information system described by King et al.[9]. This application mixed SIL-4 and non-critical code, and was written in the SPARK subset of Ada83. Static verification was used, including information flow analysis and partial program proof, to verify significant properties of SHOLIS. There was a successful argument based partly on information flow analysis that the high-SIL code was not compromised by the low-SIL code; however, this argument had to be made manually, based on the validated information flow of each subprogram, as the Examiner did not provide such tracing facilities at the program level. IMPROVING SPARK ANALYSIS The outstanding need in SPARK is to be able to mark state variables in packages with a (numerical) level of security and / or safety. This has been implemented by allowing an optional aggregate after a package ("own") variable declaration . 3.1 Marking security levels Suppose that a package Classify was defined as follows to create various secrecy levels: package Classify is -- Security levels UNCLASSIFIED : constant := 0; RESTRICTED : constant := 1; CONFIDENTIAL : constant := 2; SECRET : constant := 3; TOPSECRET : constant := 4; end Classify; Then we define a package KeyStore to store and manage a symmetric encryption key SymmetricKey, designed to mutate after each encryption according to a rotation parameter RotorValue . The key is clearly a high-security data item, with the rotor value requiring less security. KeyStore marks its state variables with the field Integrity 1 thus: --# inherit Classify; package KeyStore --# own SymmetricKey(Integrity =&gt; Classify.SECRET); --# RotorValue(Integrity =&gt; Classify.RESTRICTED); is procedure Rotate; --# global in RotorValue; --# in out SymmetricKey; 1 Integrity was chosen to make sense for both security and safety applications --# derives SymmetricKey from --# SymmetricKey, RotorValue; procedure Encrypt(C : in MessageBlock; E : out MessageBlock); --# global in SymmetricKey; --# derives E from C, SymmetricKey; ... end KeyStore; Any security analysis must show, in the case of this code, that SymmetricKey data cannot leak into RotorValue or data; i.e. there must be no subprogram (or main program) whose information annotation shows SymmetricKey as an import to RotorValue. Note that in this case package Classify will need to be inherited by all relevant package, but will never be withed and so never compiled. 3.2 Implementation We now define how the integrity levels are marked in the SPARK language, and how they are enforced by the SPARK Examiner. The extra information flow checking is invoked using the /policy=X command line switch to the Examiner. Current valid policy values are security and safety. 3.2.1 Variable declaration The above Integrity property on package own variables is a static property; at any point in static analysis of the code, the actual and required integrity level of an own variable is known. Own variables without Integrity levels are taken to have a default integrity level of Natural'Last (i.e. very highly classified) if imported under a security policy, and Natural'First (i.e. unclassified) if exported under a security policy. This gives the most paranoid checking so that data may not leak from an input to an output through an intermediate variable with unspecified integrity. If the analysis policy is safety then the default integrity values for unspecified own variables are reversed. 3.2.2 Integrity checks Information flow is declared explicitly in derives annotations for subprograms; an example from our cryptographic KeyStore is the Rotate subprogram which changes the key based on the value of RotorValue: procedure Rotate; --# global in RotorValue; --# in out SymmetricKey; --# derives SymmetricKey from --# SymmetricKey, RotorValue; Using policy security the Examiner will then check that the integrity level of the export SymmetricKey (SECRET) is no less than the integrity levels of any import (SECRET and RESTRICTED ). With policy safety the checks would be that the integrity of the export is no more than the integrity levels of any import, i.e. that high-safety data exports cannot be contaminated by low-safety data imports. Information flow is also checked at each procedure call by substituting in actual parameters (which may be own variables ) for formal parameters and rechecking the procedure's known information flow for integrity flows. As examples we 41 give the procedures which get and set the rotor and key values: procedure GetKey(This_Key : out Key); --# global in SymmetricKey; --# derives This_Key from SymmetricKey; procedure GetRotor(This_Rotor : out Rotor); --# global in RotorValue; --# derives This_Rotor from RotorValue; procedure SetKey(New_Key : in Key); --# global out SymmetricKey; --# derives SymmetricKey from New_Key; procedure SetRotor(New_Rotor : in Rotor); --# global out RotorValue; --# derives RotorValue from New_Rotor; Because RotorValue is RESTRICTED and SymmetricKey is SECRET , the analysis requires a check at each invocation of these subprograms that the actual parameters mapped to formal parameters do not violate integrity flow checks. 3.2.3 Analysis techniques not adopted One analysis technique which we considered (but rejected) was to track the integrity flows within each subprogram body and raise warnings at each individual statement where a violation occurs. We now explain how this would have worked and why we rejected it. In the following code, the programmer generates a key and rotor, encrypts a message with it, then tries to create a new rotor based on the old key. -- Key and rotor generation R1 := KeyStore.MakeRotor(34,56,22,55); KeyStore.SetRotor(R1); K1 := KeyStore.MakeKey(66,11,2,4); KeyStore.SetKey(K1); -- Encrypt a message KeyStore.Encrypt(C =&gt; Clear, E =&gt; Encrypted); -- Get a copy of the (changed) key and break -it down into data KeyStore.GetKey(K1); KeyStore.DecomposeKey(K1,I1,I2,I3,I4); -- Build a new rotor R1 := KeyStore.MakeRotor(I1,I2,I3,I4); KeyStore.SetRotor(R1); The statement-by-statement information flow would proceed as shown in Table 1, where C denotes Clear, RV denotes RotorValue, SK denotes SymmetricKey, and E denotes Encrypted . The final step causes RotorValue to exceed its assigned integrity level, and would generate a static integrity flow error. The problem with this technique arises from the need to track the integrity levels of local variables. It quickly becomes clear that for practical analysis reasons each local variable involved needs to be assigned an integrity level. This is possible, and is done by declaring them as own variables with integrity levels in a package embedded in the subprogram in question, but is cumbersome. It also requires substantial rework of any existing code which we may want to retro-analyse. R1 K1 SK RV I1 C E Instruction 0 0 1 0 0 1 R1 := 0 0 1 1 SetRotor 0 0 1 1 K1 := 0 0 3 1 1 SetKey 0 0 3 1 1 3 Encrypt 0 3 3 1 1 3 GetKey 0 3 3 1 3 1 3 Decompose 3 3 3 1 3 1 3 R1 := 3 3 3 3 3 1 3 SetRotor Table 1: Information flow for example 3.2.4 Example of checking We placed the code described above into procedure Operate in a package Crypto and annotated the declaration with the correct derives annotation thus: --# inherit KeyStore,Classify,BitString; package Crypto --# own Clear, --# Encrypted(Integrity =&gt; Classify.SECRET); is procedure Operate; --# global out KeyStore.RotorValue,Encrypted; --# in out KeyStore.SymmetricKey; --# in Clear; --# derives --# KeyStore.SymmetricKey, --# KeyStore.RotorValue --# from --# KeyStore.SymmetricKey --# & --# Encrypted --# from --# Clear, --# KeyStore.SymmetricKey --# ; end Crypto; Analysis using /policy=security gives the following static semantic error: Unit name: Crypto Unit type: package specification Unit has been analysed, any errors are listed below 1 error(s) or warning(s) Line 30 --# KeyStore.SymmetricKey ^1 *** ( 1) Semantic Error :175: Information flow from KeyStore.SymmetricKey to KeyStore.RotorValue violates the selected information flow policy.. which correctly identifies the potential leak. 3.3 Case study SHOLIS is the Royal Navy's Ship Helicopter Operating Limits Information System [9] designed to assist landing of 42 helicopters on Royal Navy Type 23 frigates. Failure of this system could result in the death of helicopter pilots and passengers , loss of a helicopter and damage to the ship. This is intolerable for normal operation, hence SIL-4 reliability is required to give sufficient confidence that such an accident will not happen during the in-service lifetime of the system. SHOLIS is an on-demand system rather than a continuously operating system, and so has a required probability of failure to function on demand of approximately 10 -4 ; a more precise probability would be specified in the system safety case. However, the bulk of the SHOLIS code does not relate to critical system functionality. The code specific to SIL-4 must be analysed at a deep level; the rest of the code can be analysed less deeply as long as it can be shown not to affect the SIL-4 data adversely. 3.3.1 Original analysis The original SHOLIS code is Ada 83 and consists of 75 source files and shadow files amenable to SPARK analysis; 26.5KLoC of non-blank non-comment non-annotation code. It is a substantial program and therefore a suitable test to see if integrity level checking scales. Without any own variables annotated, a full SPARK analysis by an Examiner with policy=safety generated no integrity errors, as we would expect. 3.3.2 Identifying critical outputs To enable easy marking of variable safety criticality we added a single package Safety: package Safety is NSC : constant := 0; SC : constant := 1; end Safety; We took as an example the safety-critical output Alarm2Z in a package RMR which represents an alarm signal on the front panel. This was annotated: --# own Alarm2Z : BasicTypes.OkFailT --# (Integrity =&gt; Safety.SC); A re-analysis of package RMR raised no integrity errors. The other package that used Alarm2Z was EVENT which was the main event handler. A SPARK of this package raised integrity flow errors in subprogram Sync, where Alarm2Z depended on a large range of other inputs, which had not yet been marked as safety-critical. This was what we would expect so far and confirmed that basic safety integrity analysis was working. 3.3.3 Extending analysis The next phase of work started in the SENSOR package which was near the middle of the package hierarchy. We set the package variables representing current speed, heading, roll, pitch and wind velocity state to be safety-critical and then ran a trial SPARK analysis. This indicated many own variables in this and other packages which caused integrity flow errors since they had no explicit integrity level. For each of these packages in turn, we: 1. marked all of the package own variables as non-critical (Safety.NSC); 2. re-analysed the package specification; 3. analysed the package body to ensure that there were no integrity flow errors at subprogram call points; and 4. if necessary, transformed NSC variables to SC status and re-ran. Eventually we converged on a stable SC/NSC partition of the variables. 3.3.4 Declassification The DisplayBuffers state variable in I/O package sio5 was a point where safety-critical and non-safety critical data merged. It was necessary during the actual project to produce a manual justification that the buffer would never be filled with non-safety critical data when there was safety-critical data to be added and displayed to a user. We therefore set its integrity to NSC and ignored all integrity errors relating to flows from SIO5.DisplayBuffers. 3.3.5 Results There were 1282 integrity flow errors, but every single one of these referred to a flow from SIO5.DisplayBuffers to Fault.Faults, as expected. Therefore only one manual argument is needed to validate the separation of SC and NSC data at this level. Of the 233 package specification variables which were given integrity levels, 110 were NSC and 123 were SC. 3.3.6 Lessons learned SHOLIS was developed using a now out-of-date version of the SPARK Examiner which did not support proof work involving abstract state; as a consequence there were many more public own variables than you would expect in a well-designed modern SPARK program. This made the conversion work slower; at the top level, as noted above, it in-creased the time required beyond what was available for the study. The "TCB bloat" problem noted in Section 2.3 did not seem to be a problem. While working up the calling hierarchy there was a small amount of returning to lower levels to make NSC variables into SC, but this did not spread out of control. There is a clear need for a declassification mechanism, as discussed in more detail in Section 4.3. Being able to suppress the integrity flow errors from DisplayBuffers would have made the transformation process easier. 3.4 Possible tactical extensions Given the preceding work, it is relatively simple to extend the own variable annotation to allow other fields in the aggregate. Within the security domain, there are considerations which mean that security cannot be considered on a linear scale. An example is a set of documents on an international military aviation development where markings might include NATO RESTRICTED, UK RESTRICTED and ICELAND RESTRICTED. They are all at the same level of security, but apply to different nationalities in different ways. A UK citizen could receive the first two, a German could receive the first one only, and a Russian could not receive any. The nationality information could be represented by an additional field, which might be an array of booleans mapping each country code to an Allowed/Forbidden code. 43 3.5 Security policies So far we have considered enforcing the Bell-LaPadula security policy. However, there are other policies which may be desirable for enforcement. One example is a policy where information at security level N may only flow into other information at security level N; there is no concept of ordering on these security levels, they may simply not be mixed. There is further work to be done on investigating whether other information flow policies are desirable and useful for security-critical or safety-critical code. In principle they should not be complicated to enforce. The /policy=X command line switch provides a hook to specify different policies in future. ISSUES FOR FUTURE RESEARCH In this section we discuss the limitations of the existing work and examine how the analysis techniques may be extended in the future. 4.1 Difficulties with analysis The concept of "label creep" as identified by Sabelfeld and Myers refers to the tendency of data to increase in security level as program execution progresses; assigning a SECRET value to one field of a large CONFIDENTIAL record will make the entire record SECRET . It remains to be seen how SPARK security programs should be designed in order to minimise label creep. Concurrency is more complex because there is the possibility that security information may leak from the observable scheduling order of the processes. This is addressed to some extent because SPARK analysis of Ravenscar programs does information flow across tasks. 4.2 Wider analysis One extension suggestion that has come from a software development project within the company is the idea of subprogram integrity level. The motivation is similar to that for the SHOLIS analysis; that only part of a given program is safety-critical, and that verification activities (proof, unit testing levels, coverage analysis etc.) may be better focused on the safety-critical parts. Subprograms are a more useful unit of division for these activities than package state. The algorithm for identifying a subprogram's actual integrity level is to examine its exports. If it only exports own variables then the subprogram integrity level is the maximum of all exported own variable integrity levels. If some exported variables are formal parameters then each invocation of the subprogram must be examined for own variables that may be actual parameters, and the maximum integrity level taken over all possible exported own variables. Functions are taken to be equivalent to a procedure with a single exported parameter. There are two clear choices for implementing this strategy: 1. whole-program analysis, determining subprogram integrity level once all invocations are known; or 2. partial-program analysis, annotating each critical subprogram declaration with its integrity level and checking at declaration and each invocation that the maximum integrity level is not violated. The first choice is minimal-impact, but does not admit analysis until the whole program is written which is likely to prove troublesome; the integrity level of many subprograms will not be known until very late in the development process, by which time testing should already be ramped up. The second choice is higher impact; it requires an extra annotation to be added to the SPARK language and checked by the Examiner, and requires developers to add the annotation to each potentially critical subprogram as it is specified . However, the benefits of the partial program analysis are likely to outweigh these drawbacks. 4.3 Subverting the analysis Declassification is occasionally necessary in security programs ; this is when the assigned security level of information is deliberately reduced. An example would be a security filter which took SECRET information, stripped out sensitive fields and output information which was no more than CONFIDENTIAL level. This can be done in SPARK by hiding the body of a declassifying subprogram from the Examiner, but this is clearly not an optimal solution. A better solution should be found. 4.4 Considerations for certification In very high security applications it may be necessary to certify the object code as well as the source code. It remains to be seen whether and how the information known from the SPARK source code analysis can be carried over to inform an object code analysis. There are other ways by which secure information can be observed, such as covert or timing channels. A full implementation of multi-level security information analysis should be followed by an analysis of how much information could be leaked this way. CONCLUSIONS In this paper we have described recent work on applying information flow analysis techniques to enforcing multi-level security in a single software application. We have shown how the requirements listed by Sabelfeld and Myers[11] are partially satisfied by the information flow analysis possible with SPARK Ada and the SPARK Examiner. We have further shown that the existing SPARK language and analysis may be extended to enforce the Bell-LaPadula security model with relatively little change. SPARK Ada has already proven itself in high-security and safety-critical application development. It now appears to be an effective choice of language to partition data of differing criticality, and provide a low-cost but robust argument of safety or security for an application. SPARK already provides the semantics-based security required by Sabelfeld and Myers; the extensions to own variable annotations now provide the complementary security type system. For end-to-end confidentiality in a secure system, we believe that SPARK Ada's extended information flow analysis provides a hard-to-refute justification of data security. 5.1 Acknowledgements The authors are grateful to Peter Amey, Neil White and Will Ward from Praxis Critical Systems for their feedback on this paper and the prototype integrity checking facilities of the Examiner. 44 REFERENCES [1] R. J. Anderson. Security engineering: a guide to building dependable distributed systems. Wiley Computer Publishing, 2001. ISBN 0-471-38922-6. [2] J. Barnes. High Integrity Software: The SPARK Approach to Safety And Security. Addison Wesley, April 2003. [3] D. E. Bell and L. LaPadula. Secure computer systems. Technical Report ESR-TR-73-278, Mitre Corporation, November 1973. v. I and II. [4] J.-F. Bergeretti and B. A. Carre. Information-flow and data-flow analysis of while-programs. ACM Transactions on Programming Languages and Systems, 7(1):3761, January 1985. [5] Common Criteria. Common Criteria for Information Technology Security Evaluation, August 1999. [6] D. Dolev and A. Yao. On the security of public-key protocols. IEEE Transactions on Information Theory, 2(29):198208, August 1983. [7] A. Hall and R. Chapman. Correctness by construction: Developing a commercial secure system. IEEE Software, pages 1825, Jan/Feb 2002. [8] International Electrotechnical Commission. IEC Standard 61508, Functional Safety of Electrical / Electronic / Programmable Electronic Safety-Related Systems, March 2000. [9] S. King, J. Hammond, R. Chapman, and A. Pryor. Is proof more cost effective than testing? IEEE Transactions on Software Engineering, 26(8):675686, August 2000. [10] RTCA / EUROCAE. RTCA DO-178B / EUROCAE ED-12: Software Considerations in Airborne Systems and Equipment Certification, December 1992. [11] A. Sabelfeld and A. C. Myers. Language-based information-flow security. IEEE Journal on Selected Areas in Communications, 21(1), January 2003. 45

security level;SPARK Ada;integrity;Information flow;Dolev-Yao;subprogram;SPARK;information flow;safety;security;Bell-LaPadula

83

Entropy and Self-Organization in Multi-Agent Systems

Emergent self-organization in multi-agent systems appears to contradict the second law of thermodynamics. This paradox has been explained in terms of a coupling between the macro level that hosts self-organization (and an apparent reduction in entropy), and the micro level (where random processes greatly increase entropy). Metaphorically, the micro level serves as an entropy "sink," permitting overall system entropy to increase while sequestering this increase from the interactions where self-organization is desired. We make this metaphor precise by constructing a simple example of pheromone-based coordination, defining a way to measure the Shannon entropy at the macro (agent) and micro (pheromone) levels, and exhibiting an entropy-based view of the coordination.

INTRODUCTION Researchers who construct multi-agent systems must cope with the world's natural tendency to disorder. Many applications require a set of agents that are individually autonomous (in the sense that each agent determines its actions based on its own state and the state of the environment, without explicit external command), but corporately structured. We want individual local decisions to yield coherent global behavior. Self-organization in natural systems (e.g., human culture, insect colonies) is an existence proof that individual autonomy is not incompatible with global order. However, widespread human experience warns us that building systems that exhibit both individual autonomy and global order is not trivial. Not only agent researchers, but humans in general, seek to impose structure and organization on the world around us. It is a universal experience that the structure we desire can be achieved only through hard work, and that it tends to fall apart if not tended. This experience is sometimes summarized informally as "Murphy's Law," the observation that anything that can go wrong, will go wrong and at the worst possible moment. At the root of the ubiquity of disorganizing tendencies is the Second Law of Thermodynamics, that "energy spontaneously tends to flow only from being concentrated in one place to becoming diffused and spread out." [9] Adding energy to a system can overcome the Second Law's "spontaneous tendency" and lead to increasing structure. However, the way in which energy is added is critical. Gasoline in the engines of construction equipment can construct a building out of raw steel and concrete, while the same gasoline in a bomb can reduce a building to a mass of raw steel and concrete. Agents are not immune to Murphy. The natural tendency of a group of autonomous processes is to disorder, not to organization. Adding information to a collection of agents can lead to increased organization, but only if it is added in the right way. We will be successful in engineering agent-based systems just to the degree that we understand the interplay between disorder and order. The fundamental claim of this paper is that the relation between self-organization in multi-agent systems and thermodynamic concepts such as the second law is not just a loose metaphor, but can provide quantitative, analytical guidelines for designing and operating agent systems. We explain the link between these concepts, and demonstrate by way of a simple example how they can be applied in measuring the behavior of multi-agent systems. Our inspiration is a model for self-organization proposed by Kugler and Turvey [7], which suggests that the key to reducing disorder in a multi-agent system is coupling that system to another in which disorder increases. Section 2 reviews this model and relates it to the problem of agent coordination. Section 3 describes a test scenario that we have devised, inspired by self-organization in pheromone systems, and outlines a method for measuring entropy in this scenario. Section 4 reports our experimental results. Section 5 summarizes our conclusions. AN ENTROPY MODEL FOR SELF-ORGANIZATION In the context of biomechanical systems, Kugler and Turvey [7] suggest that self-organization can be reconciled with second-law tendencies if a system includes multiple coupled levels of dynamic activity. Purposeful, self-organizing behavior occurs at the macro level. By itself, such behavior would be contrary to the second law. However, the system includes a micro level whose dynamics generate increasing disorder. Thus the system as a whole is increasingly disordered over time. Crucially, the behavior of elements at the macro level is coupled to the micro level dynamics. To understand this model, we begin with an example, then abstract out the underlying dynamics, and finally comment on the legitimacy of identifying processes at this level with principles from thermodynamics. 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. AGENTS'01, May 28-June 1, 2001, Montral, Quebec, Canada. Copyright 2001 ACM 1-58113-326-X/01/0005...$5.00. 124 2.1 An Example: Pheromones The parade example of such a system is the self-organization of an insect colony (such as the construction of minimal spanning tree networks among nests and food sources by ants, or the erection of multi-storied structures with regularly spaced pillars and floors by tropical termites), through pheromone-based coordination [1, 11]. Pheromones are scent markers that insects use in two ways. First, they deposit pheromones in the environment to record their state. For example, a foraging ant just emerging from the nest in search of food might deposit nest pheromone, while an ant that has found food and is carrying it will deposit food pheromone. ([15] documents use of multiple pheromones by insects.) Second, they orient their movements to the gradient of the pheromone field. In the example of foraging ants, those seeking food climb the gradient of the food pheromone, while those carrying food climb the gradient of the nest pheromone. The most realistic models of the ants' pheromone-climbing behavior incorporates a stochastic element in their movement. That is, they do not follow the gradient deterministically, but use its strength to weight a roulette wheel from which they determine their movement. The environment in which pheromones are deposited plays a critical role in such a system. It is not passive, but active, and performs three information-processing functions with the pheromones. 1. It aggregates deposits of the same flavor of pheromone from different ants, thus providing a form of data fusion across multiple agents at different times, based on their traversal of a common location. 2. It evaporates pheromones over time, thus forgetting obsolete information. This dynamic is usefully viewed as a novel approach to truth maintenance. Conventional knowledge bases remember every assertion unless there is cause to retract it, and execute truth maintenance processes to detect and resolve the conflicts that result when inconsistent assertions coexist. Insect systems forget every assertion unless it is regularly reinforced. 3. Evaporation provides a third function, that of disseminating information from the location at which it was deposited to nearby locations. An ant does not have to stumble across the exact location at which pheromone was deposited in order to access the information it represents, but can sense the direction from its current location to the pheromone deposit in the form of the gradient of evaporated pheromone molecules. 2.2 The Model in Detail In the Kugler-Turvey model, ants and their movements constitute the macro level of the system, while pheromone molecules constitute the micro level. The purposeful movement of ants, constructing minimal paths among their nests and food sources, achieve a reduction in disorder at the macro level, made possible because the agents at this level are coupled to the micro level, where the evaporation of pheromone molecules under Brownian motion results in an overwhelming growth in disorder. As a result, the disorder of the overall system increases, in keeping with the Second Law, in spite of the emergence of useful order at the macro level. Figure 1 illustrates the interplay among these processes, and how this model of agent coordination differs from more classical views. Classically, agents are considered to perceive one another directly, reason about this perception, and then take rational action. The Kugler-Turvey model views coordination as being mediated by an environment that agents change by their actions (e.g., depositing pheromones), a process known as "stigmergy" [4]. Processes in the environment generate structures that the agents perceive, thus permitting ordered behavior at the agent level. At the same time, these processes increase disorder at the micro level, so that the system as a whole becomes less ordered over time. Research in synthetic pheromones [2, 12, 13] draws directly on this model of coordination, but the model is of far broader applicability. In a multi-commodity market, individual agents follow economic fields generated by myriad individual transactions, and self-organization in the demand and supply of a particular commodity is supported by an environment that distributes resources based on the other transactions in the system. The movement of currency in such a system provides similar functions to those of pheromones in insect systems. More broadly, we hypothesize that a coupling of ordered and disordered systems is ubiquitous in robust self-organizing systems, and that the lack of such a coupling correlates with architectures that do not meet their designers' expectations for emergent cohesiveness. 2.3 A Caveat At this point, readers with a background in physics and chemistry may be uneasy. These disciplines formulated the Second Law within a strict context of processes that result in energy changes. The fundamental physical measures associated with the second law are temperature T, heat Q, and (thermodynamic) entropy S, related by the definition Equation 1 T dQ dS = Statistical mechanics identifies this macroscopic measure with the number of microscopically defined states accessible to the system by the relation Equation 2 S k ln where k is Boltzmann's constant, 1.4E-16 erg/deg. M icro New to n ian ; F o rce F ield ; E n tro p y F lo w (E n tro p y ) F lo w (E n tro p y ) M acro No n -New to n ian F lo w F ield "Neg en tro p y" Pe rcep tio n Percep tion P h ero m o n e Rational Action (Entropy ) Ra tio na l A cti on (Ent ropy ) P h ero m o n e Dynam ics P ercep tio n Ratio n al A ctio n Ag en t 1 Ag en t 2 T rad itio n al A g en t Dynam ics Key M icro New to n ian ; F o rce F ield ; E n tro p y F lo w (E n tro p y ) F lo w (E n tro p y ) M acro No n -New to n ian F lo w F ield "Neg en tro p y" Pe rcep tio n Percep tion P h ero m o n e Rational Action (Entropy ) Ra tio na l A cti on (Ent ropy ) P h ero m o n e Dynam ics P ercep tio n Ratio n al A ctio n Ag en t 1 Ag en t 2 T rad itio n al A g en t Dynam ics Key Figure 1. Comparison of Conventional and Pheromone-Based Models of Coordination 125 Thus defined, thermodynamic entropy has strong formal similarities [10] to information entropy [14] Equation 3 = i i i p p S log where i ranges over the possible states of the system and p i is the probability of finding the system in state i. These formal similarities have led to a widespread appropriation of the notion of "entropy" as a measure of macro-level disorder, and of the Second Law as describing a tendency of systems to become more chaotic. Our approach participates to this appropriation. It has been objected [8] that such appropriation completely ignores the role of energy intrinsic to both thermodynamic definitions (via T and dQ in the macro definition and k in the micro definition). Such an objection implicitly assumes that energy is logically prior to the definition, and that unless information processes are defined in terms of energy changes, it is illegitimate to identify their changes in entropy with those of thermodynamics. An alternative approach to the question would argue that in fact the prior concept is not ergs but bits, the universe is nothing but a very large cellular automaton with very small cells [3, 6], and physics and chemistry can in principle be redefined in terms of information-theoretic concepts. Our approach is sympathetic with this view. While we are not prepared at this point to define the precise correspondence between ergs and bits, we believe that physical models are an under-exploited resource for understanding computational systems in general and multi-agent systems in particular. The fact that the thermodynamic and information approaches work in different fundamental units (ergs vs. bits) is not a reason to separate them, but a pole star to guide research that may ultimately bring them together. EXPERIMENTAL SETUP We experiment with these concepts using a simple model of pheromone-based behavior. In this section we describe the experiment and how one measures entropy over it. 3.1 The Coordination Problem Consider two agents, one fixed and one mobile, who desire to be together. Neither knows the location of the other. The mobile agent, or walker, could travel to the destination of the stationary one, if it only knew where to go. The stationary agent, or target, deposits pheromone molecules at its location. As the pheromone molecules diffuse through the environment, they create a gradient that the walker can follow. Initially, the walker is at (30,30) and the target is at (50,50) in a 100x100 field. Every time step, the target deposits one molecule at (50,50). Both the walker and the pheromone molecules move by computing an angle [0,2] relative to their current heading and taking a step of constant length (1 for the walker, 2 for the pheromone molecule) in the resulting direction. Thus both molecules and walkers can be located at any real-valued coordinates in the field. Molecules move every cycle of the simulation and the walker every five cycles, so altogether the molecules move ten times as fast as the walker. Molecules fall off of the field when they reach the edge, while the walker bounces off the edges. Molecules choose the heading for their next step from a uniform random distribution, and so execute an unbiased random walk. The walker computes its heading from two inputs. 1. It generates a gradient vector Gr from its current location to each molecule within a specified radius , with magnitude Equation 4 &lt; = i r i r g G 2 r where r i is the distance between the walker and the ith molecule and g is a "gravitational constant" (currently 1). 2. It generates a random vector Rr with random heading and length equal to a temperature parameter T. The vector sum R G r r + , normalized to the walker's step length (1 in these experiments), defines the walker's next step. Including Rr in the computation permits us to explore the effectiveness of different degrees of stochasticity in the walker's movement, following the example of natural pheromone systems. The state of the walker defines the macro state of the system, while the states of the molecules define the micro state. This model can easily be enhanced in a number of directions, including adding multiple walkers and multiple targets, and permitting walkers and targets to deposit pheromone molecules of various flavors. The simple configuration is sufficient to demonstrate our techniques and their potential for understanding how the walker finds the target. 3.2 Measuring Entropy Computing the Shannon or Information Entropy defined in Equation 3 requires that we measure 1. the set of states accessible to the system and 2. the probability of finding the system in each of those states. 3.2.1 Measuring the Number of System States In most computational systems, the discreteness of digital computation makes counting system states straightforward (though the number of possible states is extremely high). We have purposely defined the movement of our walker and molecules in continuous space to highlight the challenge of counting discrete system states in an application embedded in the physical world (such as a robotic application). Our approach is to superimpose a grid on the field, and define a state on the basis of the populations of the cells of the grid. We can define state, and thus entropy, in terms either of location or direction. Location-based state is based on a single snapshot of the system, while direction-based state is based on how the system has changed between successive snapshots. Each approach has an associated gridding technique. For location-based entropy, we divide the field with a grid. Figure 2 shows a 2x2 grid with four cells, one spanning each quarter of the field. The state of this system is a four-element vector reporting the number of molecules in each cell (in the example, reading row-wise from upper left, &lt;1,1,3,2&gt;. The number of possible states in an nxn grid with m particles is n 2m . The parameters in location-based gridding are the number of divisions in each direction, their orientation, and the origin of the grid. 126 Rectangular grids are easiest to manage computationally, but one could also tile the plane with hexagons. For direction-based entropy, we center a star on the previous location of each particle and record the sector of the star into which the particle is found at the current step. Figure 3 shows a four-rayed star with a two particles. The state of the system is a vector with one element for each particle in some canonical order. Counting sectors clockwise from the upper left, the state of this example is &lt;2,3&gt;. The number of possible states with an n-pointed star and m particles is mn. The parameters in direction-based gridding are the number of rays in the star and the rotation of the star about its center. In both techniques, the analysis depends critically on the resolution of the grid (the parameter n) and its origin and orientation (for location) or rotation (for direction). To understand the dependency on n, consider two extremes. If n is very large, the chance of two distributions of particles on the field having the same state is vanishingly small. For N distributions, each will be a distinct state, each state will have equal probability 1/N, and the entropy will be log(N). This state of affairs is clearly not informative. At the other extreme, n = 1, all distributions represent the same state, which therefore occurs with probability 1, yielding entropy 0, again not informative. We choose the gridding resolution empirically by observing the length scales active in the system as it operates. To understand the dependency on origin/orientation or rotation, consider two particles in the same cell. After they move, will they still be in the same cell (keeping entropy the same) or in different cells (increasing entropy)? Exactly the same movements of the two particles could yield either result, depending on how the grid is registered with the field. We follow Gutowitz's technique [5] of measuring the entropy with several different origins and taking the minimum, thus minimizing entropy contributions resulting from the discrete nature of the grid. 3.2.2 Measuring the Probabilities In principle, one could compute the probability of different system states analytically. This approach would be arduous even for our simple system, and completely impractical for a more complex system. We take a Monte Carlo approach instead. We run the system repeatedly. At each step in time, we estimate the probability of each observed state by counting the number of replications in which that state was observed. The results reported here are based on 30 replications. Shannon entropy has a maximum value of log(N) for N different states, achieved when each state is equally probable. To eliminate this dependence on N, we normalize the entropies we report by dividing by log(N) (in our case, log(30)), incidentally making the choice of base of logarithms irrelevant. EXPERIMENTAL RESULTS We report the behavior of entropy first in the micro system, then in the unguided and guided macro system, and finally in the complete system. 4.1 Entropy in the Micro System Figure 4 shows locational entropy in the micro system (the pheromone molecules), computed from a 5x5 grid. Entropy increases with time until it saturates at 1. The more molecules enter the system and the more they disperse throughout the field, the higher the entropy grows. Increasing the grid resolution has no effect on the shape of this increase, but reduces the time to saturation, because the molecules must spread out from a single (0,0) (100,0) (0,100) (100,100) (0,0) (100,0) (0,100) (100,100) Figure 2. Location-based gridding. (0,0) (100,0) (0,100) (100,100) 1 2 (0,0) (100,0) (0,100) (100,100) 1 2 Figure 3. Direction-based gridding. 0 50 100 150 200 250 Time 0.2 0.4 0.6 0.8 1 o r c i My p o r t n E Figure 4. Micro Entropy x Time (5x5 Grid) 127 location and the finer the grid, the sooner they can generate a large number of different states. Directional entropy also increases with time to saturation. This result (not plotted) can be derived analytically. The molecule population increases linearly with time until molecules start reaching the edge. Then the growth slows, and eventually reaches 0. Let M be the population of the field at equilibrium, and consider all M molecules being located at (50,50) through the entire run. Initially, all are stationary, and each time step one additional molecule is activated. Then the total number of possible system states for a 4-star is 4M, but the number actually sampled during the period of linear population growth is 4t, since the stationary molecules do not generate any additional states. Thus the entropy during the linear phase is log(4t)/log(4M). As the growth becomes sublinear, the entropy asymptotically approaches 1, as with locational entropy. 4.2 Entropy in the Unguided Macro System Figure 5 shows the path of a walker uncoupled to the micro system (when the target is emitting no pheromone molecules). With no coupling to the micro field, the walker is just a single molecule executing a random walk. Figure 6 shows that locational entropy for this walker increases over time, reflecting the increased number of cells accessible to the walker as its random walk takes it farther from its base. The grid size (15 divisions in each direction) is chosen on the basis of observations of the guided walker, discussed in the next section. The directional entropy (not plotted) is constant at 1, since the walker chooses randomly at each step from all available directions. 4.3 Entropy in Guided Macro System Now we provide the walker with a micro field by emitting pheromone molecules from the target. Figure 7 shows the path followed by a typical walker with radius = 20 and T = 0. This path has three distinct parts. Initially, the walker wanders randomly around its origin at (30,30), until the wavefront of molecules diffusing from (50,50) encounters its radius. In this region, the walker has no guidance, because no molecules are visible. Once the walker begins to sense molecules, it moves rather directly from the vicinity of (30,30) to (50,50), following the pheromone gradient. When it arrives at (50,50), it again receives no guidance from the molecules, because they are distributed equally in all directions. So it again meanders. The clouds of wandering near the start and finish have diameters in the range of 5 to 10, suggesting a natural grid between 20x20 and 10x10. We report here experiments with a 15x15 grid. Because of their initial random walk around their origin, walkers in different runs will be at different locations when they start to move, and will follow slightly different paths to the target (Figure 8). 30 35 40 45 50 55 x 30 35 40 45 50 55 y Figure 5. Unguided Walker Path. Axes are location in the (100x100) field. 0 50 100 150 200 250 Time 0.2 0.4 0.6 0.8 1 o r c a My p o r t n E Figure 6. Unguided Walker Locational Entropy (15x15 Grid) 30 35 40 45 50 55 x 30 35 40 45 50 55 y Figure 7. Guided Walker Path ( = 20, T = 0) 30 35 40 45 50 55 x 30 35 40 45 50 55 y Figure 8. Ensemble of Guided Walkers ( = 20, T = 0) 128 The dots in Figure 9 and Figure 10 show the directional and locational entropies across this ensemble of guided walkers as a function of time. The solid line in each case plots the normalized median distance from the walkers to the target (actual maximum 28), while the dashed line plots the normalized median number of molecules visible to the walkers (actual maximum 151). The lines show how changes in entropy and reduction in distance to the target are correlated with the number of molecules that the walker senses at any given moment. At the beginning and end of the run, when the walkers are wandering without guidance, directional entropy is 1, corresponding to a random walk. During the middle portion of the run, when the walker is receiving useful guidance from the micro level, the entropy drops dramatically. As the temperature parameter T is increased in the range 50 to 100, the bottom of the entropy well rises, but the overall shape remains the same (plot not shown). The locational entropy presents a different story. The minimization method for avoiding discreteness artifacts has the effect of selecting at each time step the offset that best centers the cells on the walkers. At the beginning of the run and again at the end, most walkers are close together, and fall within the same cell (because we chose a cell size comparable to these clouds). Walkers leave the starting cloud at different times, since those closer to the target sense the pheromones sooner, and follow different paths, depending on where they were when the pheromone reached them. Thus they spread out during this movement phase, and cluster together again once they reach the target. The effect of raising T to 100 on locational entropy is that the right end of the curve rises until the curve assumes a similar shape (plot not shown) to Figure 6. Comparison of Figure 6 and Figure 10 shows that though the directed portion of the walker's movement has higher entropy than the undirected portions, coupling the walker to the micro level does reduce the walker's overall entropy. Even at its maximum, the entropy of the guided walker is much lower than that of the random one, demonstrating the basic dynamics of the Kugler-Turvey model. The different behavior of locational and directional entropy is instructive. Which is more orderly: a randomly moving walker, or one guided by pheromones? The expected location of a random walker is stationary (though with a non-zero variance), while that of a guided walker is non-stationary. In terms of location, the random walker is thus more regular, and the location entropy reflects this. However, the movement of the guided walker is more orderly than that of the random walker, and this difference is reflected in the directional entropy. This difference highlights the importance of paying attention to dynamical aspects of agent behavior. Our intuition that the guided walker is more orderly than the random one is really an intuition about the movement of this walker, not its location. 4.4 Entropy in the Overall System Central to the Kugler-Turvey model is the assertion that entropy increase at the micro level is sufficient to ensure entropy increase in the overall system even in the presence of self-organization and concomitant entropy reduction at the micro level. Our experiment illustrates this dynamic. As illustrated in Figure 4, by time 60, normalized entropy in the micro system has reached the maximum level of 1, indicating that each of the 30 replications of the experiment results in a distinct state. If each replication is already distinct on the basis of the locations of the pheromone molecules alone, adding additional state elements (such as the location of the walker) cannot cause two replications to become the same. Thus by time 60 the normalized entropy of the entire system must also be at a maximum. In particular, decreases in macro entropy, such as the decrease in locational entropy from time 80 on seen in Figure 10, do not reduce the entropy of the overall system. One may ask whether the reduction in macro (walker) entropy is causally related to the increase in micro entropy, or just coincidental. After all, a static gradient of pheromone molecules would guide the walker to the target just as effectively, but would be identical in every run, and so exhibit zero entropy. This argument neglects whatever process generates the static gradient in the first place. An intelligent observer could produce the gradient, but then the behavior of the system would hardly be "self-organizing." In our scenario, the gradient emerges as a natural consequence of a completely random process, the random walk of the pheromone molecules emerging from the target. The gradient can then reduce the entropy of a walker at the macro level, but the price paid for this entropy reduction is the increase in entropy generated by the random process that produces and maintains the gradient. 0 10 20 30 40 50 Time 0.2 0.4 0.6 0.8 1 y p o r t n E e c n a t s i Dd n as e l u c e l o M Figure 9. Guided walker: dots = directional entropy (4 star), solid line = median distance to target (max 28), dashed line = median visible molecules (max 151). 0 50 100 150 200 250 Time 0.2 0.4 0.6 0.8 1 o r c a My p o r t n E e c n a t s i Dd n as e l u c e l o M Figure 10. Guided walker: dots = locational entropy (15x15 grid), solid line = median distance to target (max 28), dashed line = median visible molecules (max 151). 129 One may also ask whether our hypothesis requires a quantitative relation between entropy loss at the macro level and entropy gain at the micro level. A strict entropy balance is not required; the micro level might generate more entropy than the macro level loses. In operational terms, the system may have a greater capacity for coordination than a particular instantiation exploits. What is required is that the entropy increase at the micro level be sufficient to cover the decrease at the macro level, and this we have shown. SUMMARY To be effective, multi-agent systems must yield coordinated behavior from individually autonomous actions. Concepts from thermodynamics (in particular, the Second Law and entropy) have been invoked metaphorically to explain the conditions under which coordination can emerge. Our work makes this metaphor more concrete and derives several important insights from it. This metaphor can be made quantitative, through simple state partitioning methods and Monte Carlo simulation. These methods show how coordination can arise through coupling the macro level (in which we desire agent self-organization with a concomitant decrease in entropy) to an entropy-increasing process at a micro level (e.g., pheromone evaporation). Our demonstration focuses on synthetic pheromones for the sake of expositional simplicity, but we believe that the same approach would be fruitful for understanding self-organization with other mechanisms of agent coordination, such as market systems. This confirmation of the Kugler-Turvey model encourages us as agent designers to think explicitly in terms of macro and micro levels, with agent behaviors at the macro level coupled in both directions (causally and perceptually) to entropy-increasing processes at the micro level. Some form of pheromone or currency is a convenient mechanism for creating such an entropy-increasing process. Researchers must distinguish between static and dynamic order in a multi-agent system. We have exhibited a system that appears intuitively to be self-organizing, and shown that the measure of order underlying this intuition is dynamic rather than static. ACKNOWLEDGMENTS This work is supported in part by the DARPA JFACC program under contract F30602-99-C-0202 to ERIM CEC. The views and conclusions in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Defense Advanced Research Projects Agency or the US Government. REFERENCES [1] E. Bonabeau, M. Dorigo, and G. Theraulaz. Swarm Intelligence: From Natural to Artificial Systems. New York, Oxford University Press, 1999. [2] S. Brueckner. Return from the Ant: Synthetic Ecosystems for Manufacturing Control. Thesis at Humboldt University Berlin, Department of Computer Science, 2000. [3] E. Fredkin. Finite Nature. In Proceedings of The XXVIIth Recontre de Moriond, 1992. [4] P.-P. Grass. La Reconstruction du nid et les Coordinations Inter-Individuelles chez Bellicositermes Natalensis et Cubitermes sp. La thorie de la Stigmergie: Essai d'interprtation du Comportement des Termites Constructeurs. Insectes Sociaux, 6:41-80, 1959. [5] H. A. Gutowitz. Complexity-Seeking Ants. In Proceedings of Third European Conference on Artifical Life, 1993. [6] B. Hayes. Computational Creationism. American Scientist, 87(5):392-396, 1999. [7] P. N. Kugler and M. T. Turvey. Information, Natural Law, and the Self-Assembly of Rhythmic Movement. Lawrence Erlbaum, 1987. [8] F. L. Lambert. Shuffled Cards, Messy Desks, and Disorderly Dorm Rooms - Examples of Entropy Increase? Nonsense! Journal of Chemical Education, 76:1385, 1999. [9] F. L. Lambert. The Second Law of Thermodynamics. 2000. Web Page, http://www.secondlaw.com/. [10] J. Lukkarinen. Re: continuing on Entropy. 2000. Email Archive, http://necsi.org:8100/Lists/complex-science/Message/2236 .html. [11] V. D. Parunak. 'Go to the Ant': Engineering Principles from Natural Agent Systems. Annals of Operations Research, 75:69-101, 1997. [12] V. D. Parunak and S. Brueckner. Ant-Like Missionaries and Cannibals: Synthetic Pheromones for Distributed Motion Control. In Proceedings of Fourth International Conference on Autonomous Agents (Agents 2000), pages 467-474, 2000. [13] Peeters, P. Valckenaers, J. Wyns, and S. Brueckner. Manufacturing Control Algorithm and Architecture. In Proceedings of Second International Workshop on Intelligent Manufacturing Systems, pages 877-888, K.U. Leuven, 1999. [14] E. Shannon and W. Weaver. The Mathematical Theory of Communication. Urbana, IL, University of Illinois, 1949. [15] ithsonian Institution. Encyclopedia Smithsonian: Pheromones in Insects. 1999. Web Page, http://www.si.edu/resource/faq/nmnh/buginfo/pheromones.ht m. 130

thermodynamic;Pheromones;entropy;Entropy;coordination;autonomy;pheromones;multi-agent system;self-organization;Self-Organization

84

Entropy-based Sensor Selection Heuristic for Target Localization

We propose an entropy-based sensor selection heuristic for localization. Given 1) a prior probability distribution of the target location, and 2) the locations and the sensing models of a set of candidate sensors for selection, the heuristic selects an informative sensor such that the fusion of the selected sensor observation with the prior target location distribution would yield on average the greatest or nearly the greatest reduction in the entropy of the target location distribution. The heuristic greedily selects one sensor in each step without retrieving any actual sensor observations. The heuristic is also computationally much simpler than the mutual-information-based approaches. The effectiveness of the heuristic is evaluated using localization simulations in which Gaussian sensing models are assumed for simplicity. The heuristic is more effective when the optimal candidate sensor is more informative.

INTRODUCTION The recent convergence of micro-electro-mechanical systems (MEMS) technology, wireless communication and networking technology, and low-cost low-power miniature digital hardware design technology has made the concept of wireless sensor networks viable and a new frontier of research [2, 1]. The limited on-board energy storage and the limited wireless channel capacity are the major constraints of wireless sensor networks. In order to save precious resources, a sensing task should not involve more sensors than necessary . From the information-theoretic point of view, sensors are tasked to observe the target in order to increase the information (or to reduce the uncertainty) about the target state. The information gain attributable to one sensor may be very different from that attributable to another when sensors have different observation perspectives and sensing uncertainties . Selective use of informative sensors reduces the number of sensors needed to obtain information about the target state and therefore prolongs the system lifetime. In the scenario of localization or tracking using wireless sensor networks, the belief state of the target location can be gradually improved by repeatedly selecting the most informative unused sensor until the required accuracy (or uncertainty) level of the target state is achieved. There have been several investigations into information-theoretic approaches to sensor fusion and management. The idea of using information theory in sensor management was first proposed in [8]. Sensor selection based on expected information gain was introduced for decentralized sensing systems in [12]. The mutual information between the predicted sensor observation and the current target location distribution was proposed to evaluate the expected information gain about the target location attributable to a sensor in [11, 6]. On the other hand, without using information theory, Yao et. al. [16] found that the overall localization accuracy depends on not only the accuracy of individual sensors but also the sensor locations relative to the target location during the development of localization algorithms. We propose a novel entropy-based heuristic for sensor selection based on our experiences with target localization. It is computationally more efficient than mutual-information-based methods proposed in [11, 6]. 36 We use the following notations throughout this paper: 1. S is the set of candidate sensors for selection, i S is the sensor index; 2. x is the realization of the random vector that denotes the target location; 3. x t is the actual target location; 4. ^ x is the maximum likelihood estimate of the target location ; 5. x i is the deterministic location of sensor i; 6. z i is the realization of the random variable that denotes the observation of sensor i about the target location; 7. z ti is the actual observation of sensor i about the target location; 8. z v i is the realization of the random variable that denotes the view of sensor i about the target location. The rest of this paper is organized as follows. Section 2 describes the heuristic in detail. Section 3 evaluates the heuristic using simulations. Section 4 discusses the discrepancy between the heuristic and the mutual information based approaches. Section 5 outlines future work. Section 6 concludes the paper. Section 7 acknowledges the sponsors. SENSOR SELECTION HEURISTIC This Sect. formulates the sensor selection problem in localization , presents the details of the entropy-based sensor selection heuristic, and discusses the relation between the entropy difference proposed in this paper and mutual information used in previous work about sensor selection. 2.1 Sensor Selection Problem in Localization There are several information measures. In this paper, we use Shannon entropy [14] to quantify the information gain (or uncertainty reduction) about the target location due to sensor observation. We adopt the greedy sensor selection strategy used in mutual-information-based approaches [11, 6]. The greedy strategy gradually reduces the uncertainty of the target location distribution by repeatedly selecting the currently unused sensor with maximal expected information gain. The observation of the selected sensor is incorporated into the target location distribution using sequential Bayesian filtering [3, 7]. The greedy sensor selection and the sequential information fusion continue until the uncertainty of the target location distribution is less than or equal to the required level. The core problem of the greedy sensor selection approach is how to efficiently evaluate the expected information gain attributable to each candidate sensor without actually retrieving sensor data. The sensor selection problem is formulated as follows. Given 1. the prior target location distribution: p(x), 2. the locations of candidate sensors for selection: x i , i S, 3. the sensing models of candidate sensors for selection: p(z i |x), i S, the objective is to find the sensor ^i whose observation z ^i minimizes the expected conditional entropy of the posterior target location distribution, ^i = arg min iS H(x|z i ) . (1) Equivalently, the observation of sensor ^i maximizes the expected target location entropy reduction, ^i = arg max iS (H(x) - H(x|z i )) . (2) H(x) - H(x|z i ) is one expression of I(x; z i ), the mutual information between the target location x and the predicted sensor observation z i , I(x; z i ) = p(x, z i ) log p(x, z i ) p(x)p(z i ) dxdz i , (3) where p(x, z i ) = p(z i |x)p(x) and p(z i ) = p(x, z i )dx. Thus, the observation of sensor ^i maximizes the mutual information I(x; z i ), ^i = arg max iS I(x; z i ) . (4) Sensor selection based on (4) is the maximal mutual information criterion proposed in [11, 6]. The target location x could be of up to three dimensions. The sensor observation z i (e.g. the direction to a target in a three-dimensional space ) could be of up to two dimensions. Therefore I(x; z i ) is a complex integral in the joint state space (x, z i ) of up to five dimensions. The complexity of computing I(x; z i ) could be more than that low-end sensor nodes are capable of. If the observation z i is related to the target location x only through the sufficient statistics z(x), then I(x; z i ) = I(z(x); z i ) . (5) If z(x) has fewer dimensions than x, then I(z(x); z i ) is less complex to compute than I(x; z i ). In the above special scenario , I(z(x); z i ) has been proposed to replace I(x; z i ) to reduce the complexity of computing mutual information in [11]. In this paper, we propose an alternative entropy-based sensor selection heuristic. In general, the entropy-based sensor selection heuristic is computationally much simpler than the mutual information based approaches. However, the observation of the sensor selected by the heuristic would still yield on average the greatest or nearly the greatest entropy reduction of the target location distribution. 2.2 Entropy-based Sensor Selection Heuristic During the development of wireless sensor networks for localization, we have observed that the localization uncertainty reduction attributable to a sensor is greatly effected by the difference of two quantities, namely, the entropy of the distribution of that sensor's view about the target location , and the entropy of that sensor's sensing model for the actual target location. A sensor's view about the target location is the geometric projection of the target location onto that sensor's observation perspective. For example, a direction-of-arrival (DOA) sensor's view of the target location is the direction from the sensor to the target. The view of sensor i about the target location is denoted as z v i ,which is a function of the target location x and the sensor location x i , z v i = f(x, x i ) . (6) z v i usually has less dimensions than x. The probability distribution of the view of sensor i about the target location, p(z v i ), is the projection of the target location distribution p(x) onto the observation perspective of sensor i, p(z v i )dz v i = z v i f(x,x i )z v i +dz v i p(x)dx . (7) Alternatively, p(z v i ) can be regarded as the `noise free' prediction of the sensor observation distribution p(z i ) based on the target location distribution p(x). 37 0 0.2 0.4 0.6 0.8 1 1.2 x 10 -4 East North 38 o 36 o 0 o 100 200 300 400 50 100 150 200 250 300 350 400 Figure 1: A DOA sensor's view about the target location. Thestatespaceof thetarge t location is gridded in 1 1 cells. Theimagedepicts theprobability distribution of thetarget location. Theactual target location is (200, 200), denoted by marker +. From the perspective of the DOA sensor denoted by the square, only the direction to the target is observable. The view of the DOA sensor about the target is in the interval [36 o , 38 o ] if and only if the target is inside the sector delimited by 36 o lineand 38 o line. In practice, the state space of the target location and the sensor view can be discretized by griding for numerical analysis . The discrete representation of p(z v i ) can be computed as follows. 1. Let X be the grid set of the target location x; 2. Let Z be the grid set of the sensor view z v i ; 3. For each grid point z v i Z, initialize p(z v i ) to zero; 4. For each grid point x X , determine the corresponding grid point z v i Z using equation (6), and update its probability as p(z v i ) = p(z v i ) + p(x); 5. Normalize p(z v i ) to make the total probability of the sensor view be 1. The numerical computation of p(z v i ) for a DOA sensor is illustrated in Fig. 1 and Fig. 2. The entropy of the probability distribution of the view of sensor i, H v i , is H v i = p (z v i ) log p(z v i )dz v i . (8) Given the discrete representation of p(z v i ) with a grid size of z v i , H v i can be numerically computed as H v i = p (z v i ) log p(z v i )z v i . (9) The sensing model of sensor i for the actual target location x t is p(z i |x t ), which describes the probability distribution of the observation of sensor i given that the target is at x t . The sensing model incorporates observation uncertainty from all sources, including the noise corruption to the signal, the signal modeling error of the sensor estimation algorithm, and the inaccuracy of the sensor hardware. For a single-modal target location distribution p(x), we can use the maximum 10 20 30 40 50 60 70 80 0 0.02 0.04 0.06 0.08 0.1 0.12 DOA (degree) Probability Figure2: Thediscreteprobability distribution of a DOA se nsor's vie w. Thestatespaceof theDOA sensor view is gridded in 2 o intervals. The target location distribution and theDOA sensor location are illustrated in Fig. 1. Marker X denotes the probability of the DOA view interval [36 o , 38 o ], which is the summation of theprobability of all target locations inside the sector delimited by 36 o lineand 38 o line in Fig. 1. Please note that the sensor view distribution does not depends on the sensing uncertainty characteristics at all. likelihood estimate ^ x of the target location to approximate the actual target location x t . Thus the entropy of the sensing model of sensor i for the actual target location x t is approximated as H si = p (z i |^x) log p(z i |^x)dz i . (10) For a multi-modal target location distribution p(x) with M peaks ^ x (m) , where m = 1, . . . , M , the entropy of the sensing model of sensor i for the actual target location x t can be approximated as a weighted average of the entropy of the sensing model for all modes, H si = M m=1 p(^ x (m) ) p(z i |^x (m) ) log p(z i |^x (m) )dz i . (11) Given a target location distribution p(x), the target location with maximum likelihood or local maximum likelihood can be found using standard search algorithms. We have repeatedly observed that the incorporation of the observation of sensor i with larger entropy difference H v i - H si yields on average larger reduction in the uncertainty of the posterior target location distribution p(x|z i ). Therefore, given a prior target location distribution and the location and the sensing uncertainty model of a set of candidate sensors for selection, the entropy difference H v i - H si can sort candidate sensors into nearly the same order as mutual information I(x; z i ) does. Specifically, the sensor with the maximal entropy difference H v i - H si also has the maximum or nearly the maximal mutual information I(x; z i ). Hence we propose to use the entropy difference H v i - H si as an alternative to mutual information I(x; z i ) for selecting 38 the most informative sensor. The entropy-based heuristic is to compute H v i - H si for every candidate sensor i S and then to select sensor ^i such that ^i = arg max iS (H v i - H si ) . (12) In Sect. 3, the validity of the heuristic is evaluated using simulations and the complexity of the heuristic is analyzed for two-dimensional localization. The entropy-based sensor selection heuristic works nearly as well as the mutual-information -based approaches. In addition, the heuristic is computationally much simpler than mutual information. 2.3 Relation of Entropy Difference and Mutual Information A brief analysis of the relation between entropy difference H v i - H si and mutual information I(x; z i ) helps to reveal fundamental properties of our sensor selection heuristic. Mutual information I(x; z i ) has another expression, namely, H(z i ) - H(z i |x). The entropy difference H v i - H si is closely related to H(z i ) - H(z i |x). H(z i ) is the entropy of the predicted sensor observation distribution p(z i ), H(z i ) = p (z i ) log p(z i )dz i . (13) The predicted sensor observation distribution p(z i ) becomes the sensor's view distribution p(z v i ) when the sensing model p(z i |x) is deterministic without uncertainty. The uncertainty in the sensing model p(z i |x) makes H(z i ) larger than the sensor's view entropy H v i defined in (8). H v i closely approximates H(z i ) when the entropy of the sensing model p(z i |x) is small relative to H v i . H(z i |x) is actually the expected entropy of the sensing model p(x) averaged for all possible target locations, H(z i |x) = p (x, z i ) log p(z i |x)dxdz i = p(x){-p (z i |x) log p(z i |x)dz i }dx . (14) When p(x) is a single-modal distribution, H si is defined in (10), which is the entropy of the sensing model for the most likely target location estimate ^ x. When p(x) is a multi-modal distribution, H si is defined in (11), which is the average entropy of the sensing model for all target locations with local maximal likelihood. When the entropy of the sensing model, p(z i |x) log p(z i |x)dz i , changes gradually with x, H si can reasonably approximate H(z i |x). The entropy difference H v i - H si reasonably approximates the mutual information H(z i ) - H(z i |x) when H si is small relative to H v i and the entropy of the sensing model changes gradually with x. However, selection of the most informative sensor does not require an exact evaluation of sensor information utility. Instead, an order of sensors in terms of information utility is needed. H v i - H si could sort sensors into approximately the same order as mutual information does. Therefore, a sensor with the maximal entropy difference H v i - H si also has the maximal or nearly the maximal mutual information. The correlation between the entropy difference H v i - H si and mutual information I(x; z i ) is analyzed using simulations in Sect. 3. Section 4 discusses the discrepancy between the heuristic and the mutual information based approaches. HEURISTIC EVALUATION This Sect. presents the evaluation of the entropy-based sensor selection heuristic using simulations. The computational complexity of the heuristic is also analyzed. The Gaussian noise model has been widely assumed for sensor observations in many localization and tracking algorithms, e.g. the Kalman filter [9]. Successes of these algorithms indicate that the Gaussian sensing model is a reasonable first-order-approximation of the reality. As a starting point, we assume Gaussian sensing models in the evaluative simulations for simplicity. The simple Gaussian sensing models assumed here are not accurate especially when sensors are very close to the target. To avoid the problem of over-simplified sensing models in the simulations, we only analyze sensors with some middle distance range to the target. The heuristic will be evaluated further under more realistic sensing models in the future. Four scenarios of sensor selection for localization have been studied. Three of them involve DOA sensors, range sensors, or time-difference-of-arrival (TDOA) sensors respectively. One of them involves all of the above sensors mixed together. In every sensor selection scenario, both the entropy difference H v i - H si and mutual information I(x; z i ) are evaluated and compared for all candidate sensors. In all sensor selection scenarios, the entropy difference H v i - H si can sort all candidate sensors into nearly the same order as mutual information I(x; z i ) does. Therefore, the sensor with the maximal entropy difference H v i - H si selected by the heuristic always has the maximum or nearly the maximal mutual information I(x; z i ). The larger the entropy difference H v i - H si and mutual information I(x; z i ) are, the more consistent their sensor selection decisions are. 3.1 Selection of DOA Sensors Consider now entropy-based sensor selection when all candidate sensors are DOA sensors, as depicted in Fig. 3. The prior probability distribution p(x) of the target location x is non-zero in a limited area. We assume the unbiased Gaussian sensing models for DOA sensors in some middle distance range to the target. Specifically, given a target location such that 10 x - x i 600, the probability distribution of DOA observation z i is assumed to be p(z i |x) = 1 2 e -(z i -z v i ) 2 /(2 2 ) , (15) where z v i = f(x, x i ) is the direction from sensor i to the target location x. For many DOA estimation algorithms like the approximate maximum likelihood (AML) algorithm [4], DOA estimation usually becomes much more uncertain when the candidate sensor is either very near or very far from the target. In this scenario, we exclude sensors that are either outside the study area or within a distance of 10 to the area of non-zero p(x). The entropy difference H v i - H si and mutual information I(x; z i ) of DOA sensors are evaluated and compared in five cases. In each case, Gaussian sensing models of the same standard deviation are assumed for all 100 candidate sensors . However, the standard deviation varies with the case. As shown in fig. 4, mutual information I(x; z i ) vs the entropy difference H v i - H si is plotted for all candidate sensors in all cases. Mutual information I(x; z i ) increases nearly monotonically with the entropy difference H v i - H si . The larger the entropy difference H v i -H si and mutual information I(x; z i ) are, the more correlated they are. Therefore, 39 0 0.2 0.4 0.6 0.8 1 1.2 x 10 -4 100 200 300 400 50 100 150 200 250 300 350 400 Figure 3: Scenario of sensor selection for localization using DOA sensors exclusively. The image depicts theprior probability distribution p(x) of thetarget location x. p(x) is zero outside the solid rectangle. The actual target location is (200, 200), denoted by marker +. The squares denote candidate DOA sensors for selection. 100 DOA sensors are uniformly randomly placed outside the dotted rectangle. The gap between the solid rectangle and the dotted rect-angleis 10. the entropy difference H v i - H si sorts DOA sensors in nearly the same order as mutual information I(x; z i ) does, especially when the entropy difference H v i - H si is large. The candidate DOA sensor selected by the proposed heuristic has the maximal entropy difference H v i - H si , and also has the maximal mutual information I(x; z i ). 3.2 Selection of Range Sensors and TDOA Sensors This Subsect. evaluates the entropy-based sensor selection heuristic for range sensors and TDOA sensors respectively. Fig. 5 shows the sensor selection scenario in which all candidate sensors can only measure the range to the target . The prior probability distribution p(x) of the target location x is non-zero in a limited area. We assume the unbiased Gaussian sensing models p(z i |x) for range sensors used in [13]. When the actual range is small relative to the standard deviation of the Gaussian sensing model, p(z i |x) is significantly greater than zero even for negative values of range observation z i . Because a range of negative value has no physical meaning, the above Gaussian sensing model is not valid for short ranges. To avoid the above difficulty of the Gaussian sensing model, we only consider candidate sensors in some middle distance range to the target. Specifically , in this range sensor selection scenario, we exclude sensors that are either outside the study area or within a distance of 32 to the area of non-zero p(x). Fig. 6 shows the sensor selection scenario in which only TDOA sensors are used. The prior probability distribution p(x) of the target location x is non-zero in a limited area. As in [15], the signal arrival time difference observed by every TDOA sensor is relative to a common reference sensor. We -2 -1 0 1 2 3 4 0 0.5 1 1.5 2 2.5 3 3.5 4 Entropy difference (bit) Mutual information (bit) = 32 = 16 = 8 = 4 = 2 Figure4: Mutual information I(x; z i ) vs entropy difference H v i - H si of DOA sensors. Each symbol denotes (H v i - H si , I(x; z i )) pair evaluated for one candidatese nsor. Theprior targe t location distribution and the candidate sensor placements are shown in Fig. 3. Five cases with different standard deviation of Gaussian sensing models are studied. In each case, all candidate sensors are assumed to have the same value. also assume the unbiased Gaussian sensing models p(z i |x) for TDOA sensors. In order to be comparable with scenarios of DOA sensors and range sensors, we only consider TDOA sensors in middle range distance to the target. Specifically, we exclude TDOA sensors that are either outside the study area or within a distance of 10 to the area of non-zero p(x). Following the same approach to the heuristic evaluation for DOA sensors, the entropy difference H v i -H si and mutual information I(x; z i ) of every candidate sensor are evaluated and compared for range sensor selection scenario in Fig. 5 and for TDOA sensor selection scenario in Fig. 6 respectively . Mutual information I(x; z i ) vs the entropy difference H v i - H si is plotted in Fig. 7 for all range sensors and in Fig. 8 for all TDOA sensors. In both scenarios, mutual information I(x; z i ) increases nearly monotonically with the entropy difference H v i - H si . The larger the entropy difference H v i -H si and mutual information I(x; z i ) are, the more correlated they are. Using the proposed heuristic, both the selected range sensor and the selected TDOA sensor have the maximal entropy difference H v i - H si , and also have nearly the maximal mutual information I(x; z i ). 3.3 Selection of Mixed Sensors In order to evaluate the entropy-based sensor selection heuristic across different sensing modalities, this Subsect. is devoted to the sensor selection scenario in which candidate sensors are a mixture of DOA sensors, range sensors and TDOA sensors. Fig. 9 shows the sensor selection scenario for mixed candidate sensors. Each candidate sensor is randomly assigned one of three sensing modalities, namely, DOA, range, and TDOA. Gaussian sensing models are assumed for all candidate sensors with middle range distance to the target. Each 40 0 0.5 1 1.5 2 2.5 3 3.5 x 10 -4 100 200 300 400 50 100 150 200 250 300 350 400 Figure 5: Scenario of sensor selection for localization using rangese nsors. Theimagede picts theprior probability distribution p(x) of thetarget location x. p(x) is zero outside the solid rectangle. The actual target location is (200, 200), denoted by marker +. The circles denote candidate range sensors for selection . 100 range sensors are uniformly randomly placed outside the dotted rectangle. The gap between the solid rectangle and the dotted rectangle is 32. candidate sensor is also randomly assigned one of five values of the standard deviation of the sensing model, namely, 2, 4, 8, 16, and 32. 100 candidate sensors are uniformly randomly placed in the vicinity of the prior target location estimation. In order to avoid the difficulties of Gaussian sensing models for DOA sensors and range sensors close to the target, we exclude sensors either outside the study area or within a distance of 32 to the non-zero area of the prior target location distribution p(x). The entropy difference H v i - H si and mutual information I(x; z i ) of every candidate sensor are evaluated and plotted in Fig. 10. The correlation between H v i - H si and I(x; z i ) of mixed sensors is very similar to the correlation between H v i - H si and I(x; z i ) of sensors with single modality . Across various sensing modalities, mutual information I(x; z i ) increases nearly monotonically with the entropy difference H v i - H si . Therefore, across various sensing modalities , the candidate sensor with the maximal entropy difference H v i - H si , selected by the proposed heuristic, has the maximal mutual information I(x; z i ). 3.4 Computational Complexity Computational complexity analysis is an important part of the evaluation of the heuristic. We will analyze the complexity of the heuristic and compare it to the complexity of the mutual-information-based approaches. For two-dimensional localization, the target location x is two-dimensional. The sensor's view z v i of the target location x is one-dimensional. The sensor observation z i is one-dimensional . We assume that all random variables are gridded for numerical computation. Specifically, the area with non-trivial p(x) is gridded into n n. The interval with 0 0.5 1 1.5 2 2.5 3 3.5 x 10 -4 100 200 300 400 50 100 150 200 250 300 350 400 Figure 6: Scenario of sensor selection for localization using TDOA sensors. The image depicts the prior probability distribution p(x) of thetarget location x. p(x) is zero outside the solid rectangle. The actual target location is (200, 200), denoted by marker +. The triangles denote candidate TDOA sensors for selection. Every TDOA observation is relative to a common reference sensor denoted by marker . 100 TDOA sensors are uniformly randomly placed outside the dotted rectangle. The gap between the solid rectangle and the dotted rectangle is 10. non-trivial p(z i ) or p(z v i ) is also gridded into n. We assume there are K candidate sensors for selection. K is usually a small number. The proposed heuristic evaluates the entropy difference H v i - H si of all sensors and then selects the one with the maximal H v i - H si . As shown in (7), p(z v i ) can be computed from p(x) with cost O(n 2 ). As shown in (8), H v i can be computed from p(z v i ) with cost O(n). As shown in (10) and (11), H si can be computed from p(z i |x) with cost O(n). The cost to compute H v i - H si for one candidate sensor is O(n 2 ). Therefore, the total cost for the heuristic to select one out of K candidate sensors is O(n 2 ). The mutual-information-based approaches evaluate the mutual information I(x; z i ) of all sensors and then select the one with the maximal I(x; z i ). As shown in (3), I(x; z i ) can be directly computed from p(x) and p(z i |x) with cost of O(n 3 ). Therefore, the total cost to select one out of K candidate sensors is O(n 3 ). As we mentioned early in Subsect. 2.1, the computational cost of mutual information I(x; z i ) could be reduced in some special scenarios. In general, however , the heuristic is computationally much simpler than the mutual-information-based approaches. DISCREPANCY BETWEEN HEURISTIC AND MUTUAL INFORMATION As shown in Sect. 3, when the mutual information I(x; z i ) is close to 0 bit, the entropy difference H v i - H si might not sort candidate sensors into exactly the same order as the mutual information does. Such discrepancy is caused by the dispersion of the correlation between the entropy difference 41 -2 -1 0 1 2 3 4 5 0 1 2 3 4 5 Entropy difference (bit) Mutual information (bit) = 32 = 16 = 8 = 4 = 2 Figure7: Mutual information I(x; z i ) vs entropy difference H v i - H si of range senors. Each symbol denotes (H v i - H si , I(x; z i )) pair evaluated for one candidate sensor. The prior target location distribution and the candidate sensor placements are shown in Fig. 5. Five cases with different standard deviation of Gaussian sensing models are studied. In each case, all candidate sensors are assumed to have the same value. H v i - H si and the mutual information I(x; z i ) when the mutual information is small. In this Sect., we examine such correlation dispersion and evaluate its impact on the discrepancy of sensor selection decisions of the entropy-based heuristic and the mutual information based approaches. 4.1 Dispersion In this Subsect., we describe the dispersion of the correlation between the entropy difference H v i - H si and the mutual information I(x; z i ) when the mutual information is small. We also examine possible sources for such correlation dispersion. Close examination on the convex part of the mutual information vs. entropy difference curve in Fig. 7 and Fig. 8 reveals that the correlation between the mutual information I(x; z i ) and the entropy difference H v i - H si is not strictly monotonic. Instead, there is obvious dispersion of the correlation . The convex part corresponds to the situation in which candidate sensors are not very informative because the mutual information between the target location distribution and the sensor observation is close to 0 bit. In another words, when candidate sensors are not very informative, the entropy difference H v i -H si might not sort candidate sensors into the same order as the mutual information I(x; z i ) does. Given a set of candidate sensors whose observation could only reduce a little amount of uncertainty of the target location distribution, the sensor selected on the basis of the maximum entropy difference H v i - H si might not have the maximum mutual information I(x; z i ). Thus, there might be discrepancy between the sensor selection decision of the entropy-based heuristic and that of the mutual information based approaches if no candidate sensor is very informative. -4 -2 0 2 4 6 0 1 2 3 4 5 6 Entropy difference (bit) Mutual information (bit) = 32 = 16 = 8 = 4 = 2 Figure8: Mutual information I(x; z i ) vs entropy difference H v i - H si of TDOA senors. Each symbol denotes (H v i - H si , I(x; z i )) pair evaluated for one candidatese nsor. Theprior targe t location distribution and the candidate sensor placements are shown in Fig. 6. Five cases with different standard deviation of Gaussian sensing models are studied. In each case, all candidate sensors are assumed to have the same value. There might be multiple causes of such correlation dispersion between the entropy difference H v i -H si and the mutual information I(x; z i ). As pointed out in Subsect. 2.3, the entropy difference H v i -H si can be viewed as an approximation of the mutual information I(x; z i ). Thus, the order of sensors sorted by the entropy difference H v i -H si is intrinsically an approximation of that by the mutual information I(x; z i ). In practice, the discretization of the state space of the target location random variable and the sensor view random variable might also introduce inaccuracy into the evaluation of H v i . Besides, as shown in (10) and (11), the maximum likelihood estimate of the target location is used to approximate the actual target location when evaluating the entropy of the sensing model for the actual target location. 4.2 Impact In this Subsect., we examine the impact of the dispersion of the correlation between the entropy difference H v i - H si and the mutual information I(x; z i ) when the mutual information is small. The analysis shows that such correlation dispersion causes very little degradation to the quality of sensor selection decision of the entropy-based heuristic. As shown by the convex part of the mutual information vs. entropy difference curve in Fig. 7 and Fig. 8, there is dispersion of the correlation between the entropy difference H v i -H si and the mutual information I(x; z i ) when candidate sensors are not very informative. We model such dispersion using a uniform distribution bounded by a parallelogram illustrated in Fig. 11. A candidate sensor could assume any position (H v i - H si , I(x; z i )) within the parallelogram with uniform probability. As shown in Fig. 11, the geometry of the parallelogram is defined by parameters a, b and c. a 42 0 0.2 0.4 0.6 0.8 1 1.2 x 10 -4 100 200 300 400 50 100 150 200 250 300 350 400 Figure 9: Scenario of sensor selection for localization using sensors with various modalities. The imagedepicts theprior probability distribution p(x) of thetarge t location x. p(x) is zero outside the solid rectangle. The actual target location is (200, 200), denoted by marker +. The squares, the circles, and the triangles denote DOA sensors, range sensors and TDOA sensors respectively. Every TDOA observation is relative to a common reference sensor denoted by marker . Each sensor is randomly chosen to bea DOA se nsor, a rangese nsor, or a TDOA sensor. Each sensor is also randomly assigned one of five values of the standard deviation of Gaussian sensing models, namely, 2, 4, 8, 16, and 32. The sizeof a symbol indicates themagnitudeof . 100 sensors of various sensing modalities and values areuniformly randomly place d outsidethedotte d rectangle. The gap between the solid rectangle and the dotted rectangle is 32. is the variation scope of entropy difference H v i - H si among the set of candidate sensors. c indicates the variation scope of the mutual information I(x; z i ) among the set of candidate sensors. b describes the magnitude of dispersion of the correlation between the entropy difference H v i - H si and the mutual information I(x; z i ). Although the bounded uniform distribution is not accurate, it captures the major features of the correlation dispersion revealed by simulations in Sect. 3. We choose this dispersion model for simplicity. As the first order approximation, the simple dispersion model does help to reveal some major characteristics of the impact of the correlation dispersion on the heuristics-based sensor selection . A typical dispersion scenario is illustrated in Fig. 11. The mutual information I(x; z i ) of candidate sensors varies from 0 bit to 1 bit. Correspondingly, the entropy difference H v i - H si of candidate sensors changes from -2 bit to 0 bit. For any value of the entropy difference H v i -H si , the disperse of the mutual information I(x; z i ) is 0.1 bit. Given the above scenario, we run 10, 000 simulations. In each simulation, 8 candidate sensors randomly assume their (H v i -H si , I(x; z i )) pairs within the specified dispersion range. In each simulation , we identify both the sensor with the maximum entropy -2 0 2 4 6 0 1 2 3 4 5 6 Entropy difference (bit) Mutual information (bit) TDOA sensor DOA sensor range sensor Figure10: Mutual information I(x; z i ) vs entropy difference H v i - H si of mixed senors. Each symbol denotes (H v i - H si , I(x; z i )) pair evaluated for one candidate sensor. The prior target location distribution and the candidate sensor placements are shown in Fig. 9. difference H v i - H si and the sensor with the maximum mutual information I(x; z i ). With 87.8% chance, the sensor selected by the entropy-based heuristic also has the maximum mutual information. Even when the heuristic fails to select the sensor of the maximum mutual information, the mutual information of the selected sensor is on average only about 0.026 bit less than the maximum mutual information. Overall, the mutual information of the sensor selected by the entropy-based heuristic is about 0.026(1-87.8%) = 0.0032 bit less than the maximum mutual information. Therefore, most of the time, the correlation dispersion does not cause discrepancy of the sensor selection decisions between the entropy-based heuristic and the mutual information based approaches. Over all, the entropy-based heuristic introduces very little degradation to the quality of the sensor select decision even when candidate sensors are not very informative. We have analyzed the impact of the correlation dispersion for different configurations of a, b, c, and the number of candidate sensors. In table 1 , a = 2 bit, b = 0.1 bit and c = 1 bit are fixed. We only change the number of candidate sensors. The chance for the heuristic to successfully select the sensor with the maximum mutual information decreases as the number of candidate sensors increases. When the heuristic fails to select the sensor with the maximum mutual information, the degradation of sensor selection decision based on the heuristic compared to that based on the mutual information does not change with the number of candidate sensors. Thus, the overall degradation of sensor selection decision based on the heuristic compared to that based on mutual information also increases as the number of candidate sensors increases. In table 2 , a = 2 bit and c = 1 bit are fixed and the number of candidate sensors are fixed to be 8. We only change the dispersion width b. The chance for the heuristic to successfully select the sensor with the maximum mutual 43 -2 -1.5 -1 -0.5 0 0 0.5 1 1.5 2 Entropy difference (bit) Mutual information (bit) b a c Figure 11: Discrepancy between the entropy-based sensor selection heuristic and the mutual information based approaches when candidate sensors are not very informative. The dispersion of the correlation between the entropy difference H v i - H si and themutual information I(x; z i ) is modeled by a uniform distribution bounded by a parallelogram. The geometry of the parallelogram is defined by parameters a, b and c. Candidate sensors are denoted by marker whosecoordinates are(H v i - H si , I(x; z i )). The entropy-based heuristic selects the rightmost sensor, which has the maximum entropy difference H v i - H si and is enclosed by a square marker. The mutual information based approaches selects the top sensor, which has the maximum mutual information I(x; z i ) and is enclosed by a diamond-shaped marker. The above two selected sensors might not be the same. In the scenario of this figure, a = 2 bits, b = 0.1 bit, c = 1 bit, and 8 candidatesensors areavailable for selection. information decreases as the dispersion width b increases. When the heuristic fails to select the sensor with the maximum mutual information, the degradation of sensor selection decision based on the heuristic compared to that based on the mutual information increases as the dispersion width b increases. Thus, the overall degradation of sensor selection decision based on the heuristic compared to that based on mutual information also increases as the dispersion width b increases. In table 3 , a = 2 bit and b = 0.1 bit are fixed and the number of candidate sensors are fixed to be 8. We only change the mutual information variation scope c. The chance for the heuristic to successfully select the sensor with the maximum mutual information increases as the mutual information variation scope c increases. When the heuristic Table 1: Impact Change with Number of Sensors Number of Candidate Sensors 4 8 16 Chance of Success (%) 93.6 87.8 78.2 Degradation per Failure (bit) 0.026 0.026 0.026 Overall Degradation (bit) 0.0016 0.0032 0.0058 Table2: Impact Changewith Dispersion Width Dispersion Width b (bit) 0.05 0.1 0.2 Chance of Success (%) 93.6 87.8 78.1 Degradation per Failure (bit) 0.013 0.026 0.054 Overall Degradation (bit) 0.0008 0.0032 0.012 Table3: Impact Changewith Mutual Info. Scope Mutual Info. Scope c (bit) 0.5 1 2 Chance of Success (%) 78.2 87.8 93.6 Degradation per Failure (bit) 0.027 0.026 0.025 Overall Degradation (bit) 0.0058 0.0032 0.0016 fails to select the sensor with the maximum mutual information , the degradation of sensor selection decision based on the heuristic compared to that based on the mutual information does not change much with the mutual information variation scope c. Thus, the overall degradation of sensor selection decision based on the heuristic compared to that based on mutual information decreases as the mutual information variation scope c increases. In table 4 , b = 0.1 bit is fixed and the number of candidate sensors are fixed to be 8. We proportionally change the entropy difference variation scope a and the mutual information variation scope c so that c/a = 1/2 is fixed. The chance for the heuristic to successfully select the sensor with the maximum mutual information increases as the entropy difference variation scope a and the mutual information variation scope c proportionally increase. When the heuristic fails to select the sensor with the maximum mutual information , the degradation of sensor selection decision based on the heuristic compared to that based on the mutual information does not change. Thus, the overall degradation of sensor selection decision based on the heuristic compared to that based on mutual information decreases as the entropy difference variation scope a and the mutual information variation scope c proportionally increase. FUTURE WORK When the sensors is selected for tracking a temporally continuous source, the prior target location distribution at time t + 1 can be obtained from the posterior target location distribution at time t by using the target dynamic model as described in [11]. However, when the sensor selection heuristic is applied to locate a temporally discontinuous source such as a bird call, it is not straightforward to obtain the prior target location distribution used in the sequential Bayesian fusion. One possible solution to the above problem could be as follows. First, all sensors buffer the signal once an event Table4: Impact Changewith Entropy Diff. Scopec and Mutual Info. Scope a in Proportion Entropy Diff. Scope a (bit) 1 2 4 Mutual Info. Scope c (bit) 0.5 1 2 Chance of Success (%) 78.2 87.8 93.6 Degradation per Failure (bit) 0.026 0.026 0.026 Overall Degradation (bit) 0.0058 0.0032 0.0016 44 such as a bird call is detected. Then, all triggered sensors elect a leader that received the strongest signal intensity using a protocol similar to that described in [10]. Finally, the leader can pick a few sensors to generate an initial prior target location distribution assuming a certain sensing model. With the initial prior target location distribution, we can apply the sensor selection heuristic to incrementally reduce the uncertainty of the target location distribution. We plan to implement and test the above mechanism in the future. 5.2 Discretization of State Space There is a trade offof computational efficiency and numerical accuracy in the discretization of the state space of random variables such as the target location and the sensor view. The bigger the grid size is, the fewer grids are involved in the computation. However, a bigger grid size also introduces more inaccuracy into the evaluation of the entropy difference heuristic. In the future, we must study more details about the trade offin order to choose a proper grid size. 5.3 Sensing Uncertainty Model We have assumed Gaussian sensing models in the simulations as the first step to evaluate the heuristic. Inaccuracy of sensing models diminishes the effectiveness of any sensor selection criterion. We plan to construct a more realistic sensing model for the AML-based DOA estimation. We have implemented AML algorithm for real-time DOA estimation on a wireless sensor network testbed [5]. We will first analyze the sensing uncertainty characteristic of the AML algorithm, and then experimentally validate and refine it using the testbed. We will also evaluate the effectiveness of the entropy-based sensor selection heuristic using realistic sensing models and implement the heuristic on the real-time wireless sensor network testbed for localization. CONCLUSION We have proposed an entropy-based sensor selection heuristic for localization. The effectiveness of the heuristic has been evaluated using simulations in which Gaussian sensing models are assumed for simplicity. Simulations have shown that the heuristic selects the sensor with nearly the maximal mutual information between the target location and the sensor observation. Given the prior target location distribution, the sensor locations, and the sensing models, on average, the sensor selected by the heuristic would yield nearly the greatest reduction in the entropy of the posterior target location distribution. The heuristic is more effective when the optimal candidate sensor is more informative. As mutual-information -based sensor selection approaches [11, 6] do, the heuristic greedily selects one sensor in each step without retrieving any actual sensor observations. In addition, in general , our heuristic is computationally much simpler than the mutual-information-based approaches. ACKNOWLEDGMENTS This material is based upon work partially supported by the National Science Foundation (NSF) under Cooperative Agreement #CCR-0121778, and DARPA SensIT program under contract AFRL/IFG 315 330-1865 and AROD-MURI PSU 50126. REFERENCES [1] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. Wireless sensor networks: a survey. Computer Networks, 38(4):393442, March 2002. [2] G. Asada, M. Dong, T. Lin, F. Newberg, G. Pottie, W. Kaiser, and H. Marcy. Wireless integrated network sensors: low power systems on a chip. In Proc. the European Solid State Circuits Conference, The Hague, Netherlands, 1998. [3] J. M. Bernardo and A. F. M. Smith. Bayesian theory. Wiley, New York, 1996. [4] J. Chen, R. Hudson, and K. Yao. Maximum-likelihood source localization and unknown sensor location estimation for wideband signals in the near-field. IEEE T. Signal Proces., 50(8):18431854, August 2002. [5] J. Chen, L. Yip, J. Elson, H. Wang, D. Maniezzo, R. Hudson, K. Yao, and D. Estrin. Coherent acoustic array processing and localization on wireless sensor networks. Proc. the IEEE, 91(8):11541162, August 2003. [6] E. Ertin, J. Fisher, and L. Potter. Maximum mutual information principle for dynamic sensor query problems. In Proc. IPSN'03, Palo Alto, CA, April 2003. [7] S. Haykin. Adpative filter theory. Prentice Hall, New Jersey, USA, 1996. [8] K. Hintz and E. McVey. A measure of the information gain attributable to cueing. IEEE T. Syst. Man Cyb., 21(2):434442, 1991. [9] R. E. Kalman. A new approach to linear filtering and prediction problems. Trans. of the ASMEJournal of Basic Engineering, 82(Series D):3545, 1960. [10] J. Liu, J. Liu, J. Reich, P. Cheung, and F. Zhao. Distributed group management for track initiation and maintenance in target localization applications. In Proc. International Workshop on Informaiton Processing in Sensor Networks (IPSN), Palo Alto, CA, April 2003. [11] J. Liu, J. Reich, and F. Zhao. Collaborative in-network processing for target tracking. EURASIP JASP: Special Issues on Sensor Networks, 2003(4):378391, March 2003. [12] J. Manyika and H. Durrant-Whyte. Data fusion and sensor management: a decentralized information-theoretic approach. Ellis Horwood, New York, 1994. [13] A. Savvides, W. Garber, S. Adlakha, R. Moses, and M. B. Srivastava. On the error characteristics of multihop node localization in ad-hoc sensor networks. In Proc. IPSN'03, Palo Alto, CA, USA, April 2003. [14] C. E. Shannon. A mathematical theory of communication. Bell Systems Technical Journal, 27(6):379423 and 623656, 1948. [15] T. Tung, K. Yao, C. Reed, R. Hudson, D. Chen, and J. Chen. Source localization and time delay estimation using constrained least squares and best path smoothing. In Proc. SPIE'99, volume 3807, pages 220223, July 1999. [16] K. Yao, R. Hudson, C. Reed, D. Chen, and F. Lorenzelli. Blind beamforming source localization on a sensor array system. In AWAIRS project presentation at UCLA, USA, December 1997. 45

Shannon entropy;entropy;target localization;localization;target tracking;wireless sensor networks;mutual information;information-directed resource management;sensor selection;heuristic;information fusion

85

Estimating the Global PageRank of Web Communities

Localized search engines are small-scale systems that index a particular community on the web. They offer several benefits over their large-scale counterparts in that they are relatively inexpensive to build, and can provide more precise and complete search capability over their relevant domains. One disadvantage such systems have over large-scale search engines is the lack of global PageRank values. Such information is needed to assess the value of pages in the localized search domain within the context of the web as a whole. In this paper, we present well-motivated algorithms to estimate the global PageRank values of a local domain. The algorithms are all highly scalable in that, given a local domain of size n, they use O(n) resources that include computation time, bandwidth, and storage. We test our methods across a variety of localized domains, including site-specific domains and topic-specific domains. We demonstrate that by crawling as few as n or 2n additional pages, our methods can give excellent global PageRank estimates.

INTRODUCTION Localized search engines are small-scale search engines that index only a single community of the web. Such communities can be site-specific domains, such as pages within 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 prot or commercial advantage and that copies bear this notice and the full citation on the rst page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specic permission and/or a fee. KDD'06, August 2023, 2006, Philadelphia, Pennsylvania, USA. Copyright 2006 ACM 1-59593-339-5/06/0008 ... $ 5.00. the cs.utexas.edu domain, or topic-related communities-for example, political websites. Compared to the web graph crawled and indexed by large-scale search engines, the size of such local communities is typically orders of magnitude smaller. Consequently, the computational resources needed to build such a search engine are also similarly lighter. By restricting themselves to smaller, more manageable sections of the web, localized search engines can also provide more precise and complete search capabilities over their respective domains. One drawback of localized indexes is the lack of global information needed to compute link-based rankings. The PageRank algorithm [3], has proven to be an effective such measure. In general, the PageRank of a given page is dependent on pages throughout the entire web graph. In the context of a localized search engine, if the PageRanks are computed using only the local subgraph, then we would expect the resulting PageRanks to reflect the perceived popularity within the local community and not of the web as a whole. For example, consider a localized search engine that indexes political pages with conservative views. A person wishing to research the opinions on global warming within the conservative political community may encounter numerous such opinions across various websites. If only local PageRank values are available, then the search results will reflect only strongly held beliefs within the community. However, if global PageRanks are also available, then the results can ad-ditionally reflect outsiders' views of the conservative community (those documents that liberals most often access within the conservative community). Thus, for many localized search engines, incorporating global PageRanks can improve the quality of search results. However, the number of pages a local search engine indexes is typically orders of magnitude smaller than the number of pages indexed by their large-scale counterparts. Localized search engines do not have the bandwidth, storage capacity, or computational power to crawl, download, and compute the global PageRanks of the entire web. In this work, we present a method of approximating the global PageRanks of a local domain while only using resources of the same order as those needed to compute the PageRanks of the local subgraph. Our proposed method looks for a supergraph of our local subgraph such that the local PageRanks within this supergraph are close to the true global PageRanks. We construct this supergraph by iteratively crawling global pages on the current web frontier--i.e., global pages with inlinks from pages that have already been crawled. In order to provide 116 Research Track Paper a good approximation to the global PageRanks, care must be taken when choosing which pages to crawl next; in this paper, we present a well-motivated page selection algorithm that also performs well empirically. This algorithm is derived from a well-defined problem objective and has a running time linear in the number of local nodes. We experiment across several types of local subgraphs, including four topic related communities and several site-specific domains. To evaluate performance, we measure the difference between the current global PageRank estimate and the global PageRank, as a function of the number of pages crawled. We compare our algorithm against several heuristics and also against a baseline algorithm that chooses pages at random, and we show that our method outperforms these other methods. Finally, we empirically demonstrate that, given a local domain of size n, we can provide good approximations to the global PageRank values by crawling at most n or 2n additional pages. The paper is organized as follows. Section 2 gives an overview of localized search engines and outlines their advantages over global search. Section 3 provides background on the PageRank algorithm. Section 4 formally defines our problem, and section 5 presents our page selection criteria and derives our algorithms. Section 6 provides experimental results, section 7 gives an overview of related work, and, finally, conclusions are given in section 8. LOCALIZED SEARCH ENGINES Localized search engines index a single community of the web, typically either a site-specific community, or a topic-specific community. Localized search engines enjoy three major advantages over their large-scale counterparts: they are relatively inexpensive to build, they can offer more precise search capability over their local domain, and they can provide a more complete index. The resources needed to build a global search engine are enormous. A 2003 study by Lyman et al. [13] found that the `surface web' (publicly available static sites) consists of 8.9 billion pages, and that the average size of these pages is approximately 18.7 kilobytes. To download a crawl of this size, approximately 167 terabytes of space is needed. For a researcher who wishes to build a search engine with access to a couple of workstations or a small server, storage of this magnitude is simply not available. However, building a localized search engine over a web community of a hundred thousand pages would only require a few gigabytes of storage . The computational burden required to support search queries over a database this size is more manageable as well. We note that, for topic-specific search engines, the relevant community can be efficiently identified and downloaded by using a focused crawler [21, 4]. For site-specific domains, the local domain is readily available on their own web server. This obviates the need for crawling or spidering, and a complete and up-to-date index of the domain can thus be guaranteed. This is in contrast to their large-scale counterparts, which suffer from several shortcomings. First, crawling dynamically generated pages--pages in the `hidden web'--has been the subject of research [20] and is a non-trivial task for an external crawler. Second, site-specific domains can enable the robots exclusion policy. This prohibits external search engines' crawlers from downloading content from the domain, and an external search engine must instead rely on outside links and anchor text to index these restricted pages. By restricting itself to only a specific domain of the internet , a localized search engine can provide more precise search results. Consider the canonical ambiguous search query, `jaguar', which can refer to either the car manufacturer or the animal. A scientist trying to research the habitat and evolutionary history of a jaguar may have better success using a finely tuned zoology-specific search engine than querying Google with multiple keyword searches and wading through irrelevant results. A method to learn better ranking functions for retrieval was recently proposed by Radlinski and Joachims [19] and has been applied to various local domains, including Cornell University's website [8]. PAGERANK OVERVIEW The PageRank algorithm defines the importance of web pages by analyzing the underlying hyperlink structure of a web graph. The algorithm works by building a Markov chain from the link structure of the web graph and computing its stationary distribution. One way to compute the stationary distribution of a Markov chain is to find the limiting distribution of a random walk over the chain. Thus, the PageRank algorithm uses what is sometimes referred to as the `random surfer' model. In each step of the random walk, the `surfer' either follows an outlink from the current page (i.e. the current node in the chain), or jumps to a random page on the web. We now precisely define the PageRank problem. Let U be an m m adjacency matrix for a given web graph such that U ji = 1 if page i links to page j and U ji = 0 otherwise. We define the PageRank matrix P U to be: P U = U D -1 U + (1 - )ve T , (1) where D U is the (unique) diagonal matrix such that U D -1 U is column stochastic, is a given scalar such that 0 1, e is the vector of all ones, and v is a non-negative, L 1 normalized vector, sometimes called the `random surfer' vector . Note that the matrix D -1 U is well-defined only if each column of U has at least one non-zero entry--i.e., each page in the webgraph has at least one outlink. In the presence of such `dangling nodes' that have no outlinks, one commonly used solution, proposed by Brin et al. [3], is to replace each zero column of U by a non-negative, L 1 -normalized vector. The PageRank vector r is the dominant eigenvector of the PageRank matrix, r = P U r. We will assume, without loss of generality, that r has an L 1 -norm of one. Computationally, r can be computed using the power method. This method first chooses a random starting vector r (0) , and iteratively multiplies the current vector by the PageRank matrix P U ; see Algorithm 1. In general, each iteration of the power method can take O(m 2 ) operations when P U is a dense matrix . However, in practice, the number of links in a web graph will be of the order of the number of pages. By exploiting the sparsity of the PageRank matrix, the work per iteration can be reduced to O(km), where k is the average number of links per web page. It has also been shown that the total number of iterations needed for convergence is proportional to and does not depend on the size of the web graph [11, 7]. Finally, the total space needed is also O(km), mainly to store the matrix U . 117 Research Track Paper Algorithm 1: A linear time (per iteration) algorithm for computing PageRank. ComputePR (U ) Input: U : Adjacency matrix. Output: r: PageRank vector. Choose (randomly) an initial non-negative vector r (0) such that r (0) 1 = 1. i 0 repeat i i + 1 U D -1 U r (i-1) { is the random surfing probability }r (i) + (1 - )v {v is the random surfer vector.} until r (i) - r (i-1) &lt; { is the convergence threshold.} r r (i) PROBLEM DEFINITION Given a local domain L, let G be an N N adjacency matrix for the entire connected component of the web that contains L, such that G ji = 1 if page i links to page j and G ji = 0 otherwise. Without loss of generality, we will partition G as: G = L G out L out G within , (2) where L is the n n local subgraph corresponding to links inside the local domain, L out is the subgraph that corresponds to links from the local domain pointing out to the global domain, G out is the subgraph containing links from the global domain into the local domain, and G within contains links within the global domain. We assume that when building a localized search engine, only pages inside the local domain are crawled, and the links between these pages are represented by the subgraph L. The links in L out are also known, as these point from crawled pages in the local domain to uncrawled pages in the global domain. As defined in equation (1), P G is the PageRank matrix formed from the global graph G, and we define the global PageRank vector of this graph to be g. Let the n-length vector p be the L 1 -normalized vector corresponding to the global PageRank of the pages in the local domain L: p = E L g E L g 1 , where E L = [ I | 0 ] is the restriction matrix that selects the components from g corresponding to nodes in L. Let p denote the PageRank vector constructed from the local domain subgraph L. In practice, the observed local PageRank p and the global PageRank p will be quite different. One would expect that as the size of local matrix L approaches the size of global matrix G, the global PageRank and the observed local PageRank will become more similar. Thus, one approach to estimating the global PageRank is to crawl the entire global domain, compute its PageRank, and extract the PageRanks of the local domain. Typically, however, n N , i.e., the number of global pages is much larger than the number of local pages. Therefore , crawling all global pages will quickly exhaust all local resources (computational, storage, and bandwidth) available to create the local search engine. We instead seek a supergraph ^ F of our local subgraph L with size O(n). Our goal Algorithm 2: The FindGlobalPR algorithm. FindGlobalPR (L, L out , T , k) Input: L: zero-one adjacency matrix for the local domain , L out : zero-one outlink matrix from L to global subgraph as in (2), T : number of iterations, k: number of pages to crawl per iteration. Output: ^ p: an improved estimate of the global PageRank of L. F L F out L out f ComputePR(F ) for (i = 1 to T ) {Determine which pages to crawl next} pages SelectNodes(F , F out , f , k) Crawl pages, augment F and modify F out {Update PageRanks for new local domain} f ComputePR(F ) end {Extract PageRanks of original local domain & normalize} ^ p E L f E L f 1 is to find such a supergraph ^ F with PageRank ^ f , so that ^ f when restricted to L is close to p . Formally, we seek to minimize GlobalDif f ( ^ f ) = E L ^ f E L ^ f 1 - p 1 . (3) We choose the L 1 norm for measuring the error as it does not place excessive weight on outliers (as the L 2 norm does, for example), and also because it is the most commonly used distance measure in the literature for comparing PageRank vectors, as well as for detecting convergence of the algorithm [3]. We propose a greedy framework, given in Algorithm 2, for constructing ^ F . Initially, F is set to the local subgraph L, and the PageRank f of this graph is computed. The algorithm then proceeds as follows. First, the SelectNodes algorithm (which we discuss in the next section) is called and it returns a set of k nodes to crawl next from the set of nodes in the current crawl frontier, F out . These selected nodes are then crawled to expand the local subgraph, F , and the PageRanks of this expanded graph are then recomputed. These steps are repeated for each of T iterations. Finally, the PageRank vector ^ p, which is restricted to pages within the original local domain, is returned. Given our computation , bandwidth, and memory restrictions, we will assume that the algorithm will crawl at most O(n) pages. Since the PageRanks are computed in each iteration of the algorithm, which is an O(n) operation, we will also assume that the number of iterations T is a constant. Of course, the main challenge here is in selecting which set of k nodes to crawl next. In the next section, we formally define the problem and give efficient algorithms. NODE SELECTION In this section, we present node selection algorithms that operate within the greedy framework presented in the previous section. We first give a well-defined criteria for the page selection problem and provide experimental evidence that this criteria can effectively identify pages that optimize our problem objective (3). We then present our main al-118 Research Track Paper gorithmic contribution of the paper, a method with linear running time that is derived from this page selection criteria . Finally, we give an intuitive analysis of our algorithm in terms of `leaks' and `flows'. We show that if only the `flow' is considered, then the resulting method is very similar to a widely used page selection heuristic [6]. 5.1 Formulation For a given page j in the global domain, we define the expanded local graph F j : F j = F s u T j 0 , (4) where u j is the zero-one vector containing the outlinks from F into page j, and s contains the inlinks from page j into the local domain. Note that we do not allow self-links in this framework. In practice, self-links are often removed, as they only serve to inflate a given page's PageRank. Observe that the inlinks into F from node j are not known until after node j is crawled. Therefore, we estimate this inlink vector as the expectation over inlink counts among the set of already crawled pages, s = F T e F T e 1 . (5) In practice, for any given page, this estimate may not reflect the true inlinks from that page. Furthermore, this expectation is sampled from the set of links within the crawled domain, whereas a better estimate would also use links from the global domain. However, the latter distribution is not known to a localized search engine, and we contend that the above estimate will, on average, be a better estimate than the uniform distribution, for example. Let the PageRank of F be f . We express the PageRank f + j of the expanded local graph F j as f + j = (1 - x j )f j x j , (6) where x j is the PageRank of the candidate global node j, and f j is the L 1 -normalized PageRank vector restricted to the pages in F . Since directly optimizing our problem goal requires knowing the global PageRank p , we instead propose to crawl those nodes that will have the greatest influence on the PageRanks of pages in the original local domain L: influence(j) = kL |f j [k] - f[k]| (7) = E L (f j - f ) 1 . Experimentally, the influence score is a very good predictor of our problem objective (3). For each candidate global node j, figure 1(a) shows the objective function value Global Diff(f j ) as a function of the influence of page j. The local domain used here is a crawl of conservative political pages (we will provide more details about this dataset in section 6); we observed similar results in other domains. The correlation is quite strong, implying that the influence criteria can effectively identify pages that improve the global PageRank estimate. As a baseline, figure 1(b) compares our objective with an alternative criteria, outlink count. The outlink count is defined as the number of outlinks from the local domain to page j. The correlation here is much weaker. .00001 .0001 .001 .01 0.26 0.262 0.264 0.266 Influence Objective 1 10 100 1000 0.266 0.264 0.262 0.26 Outlink Count Objective (a) (b) Figure 1: (a) The correlation between our influence page selection criteria (7) and the actual objective function (3) value is quite strong. (b) This is in contrast to other criteria, such as outlink count, which exhibit a much weaker correlation. 5.2 Computation As described, for each candidate global page j, the influence score (7) must be computed. If f j is computed exactly for each global page j, then the PageRank algorithm would need to be run for each of the O(n) such global pages j we consider, resulting in an O(n 2 ) computational cost for the node selection method. Thus, computing the exact value of f j will lead to a quadratic algorithm, and we must instead turn to methods of approximating this vector. The algorithm we present works by performing one power method iteration used by the PageRank algorithm (Algorithm 1). The convergence rate for the PageRank algorithm has been shown to equal the random surfer probability [7, 11]. Given a starting vector x (0) , if k PageRank iterations are performed, the current PageRank solution x (k) satisfies: x (k) - x 1 = O( k x (0) - x 1 ), (8) where x is the desired PageRank vector. Therefore, if only one iteration is performed, choosing a good starting vector is necessary to achieve an accurate approximation. We partition the PageRank matrix P F j , corresponding to the subgraph F j as: P F j = ~ F ~ s ~ u T j w , (9) where ~ F = F (D F + diag(u j )) -1 + (1 - ) e + 1 e T , ~ s = s + (1 - ) e + 1 , ~ u j = (D F + diag(u j )) -1 u j + (1 - ) e + 1 , w = 1 - + 1 , and diag(u j ) is the diagonal matrix with the (i, i) th entry equal to one if the i th element of u j equals one, and is zero otherwise. We have assumed here that the random surfer vector is the uniform vector, and that L has no `dangling links'. These assumptions are not necessary and serve only to simplify discussion and analysis. A simple approach for estimating f j is the following. First, estimate the PageRank f + j of F j by computing one PageRank iteration over the matrix P F j , using the starting vector = f 0 . Then, estimate f j by removing the last 119 Research Track Paper component from our estimate of f + j (i.e., the component corresponding to the added node j), and renormalizing. The problem with this approach is in the starting vector. Recall from (6) that x j is the PageRank of the added node j. The difference between the actual PageRank f + j of P F j and the starting vector is - f + j 1 = x j + f - (1 - x j )f j 1 x j + | f 1 - (1 - x j ) f j 1 | = x j + |x j | = 2x j . Thus, by (8), after one PageRank iteration, we expect our estimate of f + j to still have an error of about 2x j . In particular , for candidate nodes j with relatively high PageRank x j , this method will yield more inaccurate results. We will next present a method that eliminates this bias and runs in O(n) time. 5.2.1 Stochastic Complementation Since f + j , as given in (6) is the PageRank of the matrix P F j , we have: f j (1 - x j ) x j = ~ F ~ s ~ u T j w f j (1 - x j ) x j = ~ F f j (1 - x j ) + ~ sx j ~ u T j f j (1 - x j ) + wx j . Solving the above system for f j can be shown to yield f j = ( ~ F + (1 - w) -1 ~ s ~ u T j )f j . (10) The matrix S = ~ F +(1-w) -1 ~ s ~ u T j is known as the stochastic complement of the column stochastic matrix P F j with respect to the sub matrix ~ F . The theory of stochastic complementation is well studied, and it can be shown the stochastic complement of an irreducible matrix (such as the PageRank matrix) is unique. Furthermore, the stochastic complement is also irreducible and therefore has a unique stationary distribution as well. For an extensive study, see [15]. It can be easily shown that the sub-dominant eigenvalue of S is at most +1 , where is the size of F . For sufficiently large , this value will be very close to . This is important, as other properties of the PageRank algorithm, notably the algorithm's sensitivity, are dependent on this value [11]. In this method, we estimate the length vector f j by computing one PageRank iteration over the stochastic complement S, starting at the vector f : f j Sf. (11) This is in contrast to the simple method outlined in the previous section, which first iterates over the ( + 1) ( + 1) matrix P F j to estimate f + j , and then removes the last component from the estimate and renormalizes to approximate f j . The problem with the latter method is in the choice of the ( + 1) length starting vector, . Consequently, the PageRank estimate given by the simple method differs from the true PageRank by at least 2x j , where x j is the PageRank of page j. By using the stochastic complement, we can establish a tight lower bound of zero for this difference. To see this, consider the case in which a node k is added to F to form the augmented local subgraph F k , and that the PageRank of this new graph is (1 - x k )f x k . Specifi-cally , the addition of page k does not change the PageRanks of the pages in F , and thus f k = f . By construction of the stochastic complement, f k = Sf k , so the approximation given in equation (11) will yield the exact solution. Next, we present the computational details needed to efficiently compute the quantity f j -f 1 over all known global pages j. We begin by expanding the difference f j -f , where the vector f j is estimated as in (11), f j - f Sf - f = F (D F + diag(u j )) -1 f + (1 - ) e + 1 e T f +(1 - w) -1 ( ~ u T j f )~ s - f. (12) Note that the matrix (D F +diag(u j )) -1 is diagonal. Letting o[k] be the outlink count for page k in F , we can express the k th diagonal element as: (D F + diag(u j )) -1 [k, k] = 1 o[k]+1 if u j [k] = 1 1 o[k] if u j [k] = 0 Noting that (o[k] + 1) -1 = o[k] -1 - (o[k](o[k] + 1)) -1 and rewriting this in matrix form yields (D F +diag(u j )) -1 = D -1 F -D -1 F (D F +diag(u j )) -1 diag(u j ). (13) We use the same identity to express e + 1 = e - e ( + 1) . (14) Recall that, by definition, we have P F = F D -1 F +(1 -) e . Substituting (13) and (14) in (12) yields f j - f (P F f - f) -F D -1 F (D F + diag(u j )) -1 diag(u j )f -(1 - ) e ( + 1) + (1 - w) -1 ( ~ u T j f )~ s = x + y + ( ~ u T j f )z, (15) noting that by definition, f = P F f , and defining the vectors x, y, and z to be x = -F D -1 F (D F + diag(u j )) -1 diag(u j )f (16) y = -(1 - ) e ( + 1) (17) z = (1 - w) -1 ~ s. (18) The first term x is a sparse vector, and takes non-zero values only for local pages k that are siblings of the global page j. We define (i, j) F if and only if F [j, i] = 1 (equiva-lently , page i links to page j) and express the value of the component x[k ] as: x[k ] = k :(k,k )F ,u j [k]=1 f [k] o[k](o[k] + 1) , (19) where o[k], as before, is the number of outlinks from page k in the local domain. Note that the last two terms, y and z are not dependent on the current global node j. Given the function h j (f ) = y + ( ~ u T j f )z 1 , the quantity f j - f 1 120 Research Track Paper can be expressed as f j - f 1 = k x[k] + y[k] + ( ~ u T j f )z[k] = k:x[k]=0 y[k] + ( ~ u T j f )z[k] + k:x[k]=0 x[k] + y[k] + ( ~ u T j f )z[k] = h j (f ) k :x[k]=0 y[k] + ( ~ u T j f )z[k] + k:x[k]=0 x[k] + y[k] + ( ~ u T j f )z[k] .(20) If we can compute the function h j in linear time, then we can compute each value of f j - f 1 using an additional amount of time that is proportional to the number of non-zero components in x. These optimizations are carried out in Algorithm 3. Note that (20) computes the difference between all components of f and f j , whereas our node selection criteria, given in (7), is restricted to the components corresponding to nodes in the original local domain L. Let us examine Algorithm 3 in more detail. First, the algorithm computes the outlink counts for each page in the local domain. The algorithm then computes the quantity ~ u T j f for each known global page j. This inner product can be written as (1 - ) 1 + 1 + k:(k,j)F out f [k] o[k] + 1 , where the second term sums over the set of local pages that link to page j. Since the total number of edges in F out was assumed to have size O( ) (recall that is the number of pages in F ), the running time of this step is also O( ). The algorithm then computes the vectors y and z, as given in (17) and (18), respectively. The L 1 NormDiff method is called on the components of these vectors which correspond to the pages in L, and it estimates the value of E L (y + ( ~ u T j f )z) 1 for each page j. The estimation works as follows. First, the values of ~ u T j f are discretized uniformly into c values {a 1 , ..., a c }. The quantity E L (y + a i z) 1 is then computed for each discretized value of a i and stored in a table. To evaluate E L (y + az) 1 for some a [a 1 , a c ], the closest discretized value a i is determined, and the corresponding entry in the table is used. The total running time for this method is linear in and the discretization parameter c (which we take to be a constant). We note that if exact values are desired, we have also developed an algorithm that runs in O( log ) time that is not described here. In the main loop, we compute the vector x, as defined in equation (16). The nested loops iterate over the set of pages in F that are siblings of page j. Typically, the size of this set is bounded by a constant. Finally, for each page j, the scores vector is updated over the set of non-zero components k of the vector x with k L. This set has size equal to the number of local siblings of page j, and is a subset of the total number of siblings of page j. Thus, each iteration of the main loop takes constant time, and the total running time of the main loop is O( ). Since we have assumed that the size of F will not grow larger than O(n), the total running time for the algorithm is O(n). Algorithm 3: Node Selection via Stochastic Complementation. SC-Select (F , F out , f , k) Input: F : zero-one adjacency matrix of size corresponding to the current local subgraph, F out : zero-one outlink matrix from F to global subgraph, f : PageRank of F , k: number of pages to return Output: pages: set of k pages to crawl next {Compute outlink sums for local subgraph} foreach (page j F ) o[j] k:(j,k)F F [j, k] end {Compute scalar ~u T j f for each global node j } foreach (page j F out ) g[j] (1 - ) 1 +1 foreach (page k : (k, j) F out ) g[j] g[j] + f[k] o[k]+1 end end {Compute vectors y and z as in (17) and (18) } y -(1 - ) e ( +1) z (1 - w) -1 ~ s {Approximate y + g[j] z 1 for all values g[j] } norm diffs L 1 NormDiffs (g, E L y, E L z) foreach (page j F out ) {Compute sparse vector x as in (19)} x 0 foreach (page k : (k, j) F out ) foreach (page k : (k, k ) F )) x[k ] x[k ] f [k] o[k](o[k]+1) end end x x scores[j] norm diffs[j] foreach (k : x[k] &gt; 0 and page k L) scores[j] scores[j] - |y[k] + g[j] z[k]| + |x[k]+y[k]+g[j]z[k])| end end Return k pages with highest scores 5.2.2 PageRank Flows We now present an intuitive analysis of the stochastic complementation method by decomposing the change in PageRank in terms of `leaks' and `flows'. This analysis is motivated by the decomposition given in (15). PageRank `flow' is the increase in the local PageRanks originating from global page j. The flows are represented by the non-negative vector ( ~ u T j f )z (equations (15) and (18)). The scalar ~ u T j f can be thought of as the total amount of PageRank flow that page j has available to distribute. The vector z dictates how the flow is allocated to the local domain; the flow that local page k receives is proportional to (within a constant factor due to the random surfer vector) the expected number of its inlinks. The PageRank `leaks' represent the decrease in PageRank resulting from the addition of page j. The leakage can be quantified in terms of the non-positive vectors x and y (equations (16) and (17)). For vector x, we can see from equation (19) that the amount of PageRank leaked by a local page is proportional to the weighted sum of the Page-121 Research Track Paper Ranks of its siblings. Thus, pages that have siblings with higher PageRanks (and low outlink counts) will experience more leakage. The leakage caused by y is an artifact of the random surfer vector. We will next show that if only the `flow' term, ( ~ u T j f )z, is considered, then the resulting method is very similar to a heuristic proposed by Cho et al. [6] that has been widely used for the "Crawling Through URL Ordering" problem. This heuristic is computationally cheaper, but as we will see later, not as effective as the Stochastic Complementation method. Our node selection strategy chooses global nodes that have the largest influence (equation (7)). If this influence is approximated using only `flows', the optimal node j is: j = argmax j E L ~ u T j f z 1 = argmax j ~ u T j f E L z 1 = argmax j ~ u T j f = argmax j (D F + diag(u j )) -1 u j + (1 - ) e + 1 , f = argmax j f T (D F + diag(u j )) -1 u j . The resulting page selection score can be expressed as a sum of the PageRanks of each local page k that links to j, where each PageRank value is normalized by o[k]+1. Interestingly, the normalization that arises in our method differs from the heuristic given in [6], which normalizes by o[k]. The algorithm PF-Select, which is omitted due to lack of space, first computes the quantity f T (D F +diag(u j )) -1 u j for each global page j, and then returns the pages with the k largest scores. To see that the running time for this algorithm is O(n), note that the computation involved in this method is a subset of that needed for the SC-Select method (Algorithm 3), which was shown to have a running time of O(n). EXPERIMENTS In this section, we provide experimental evidence to verify the effectiveness of our algorithms. We first outline our experimental methodology and then provide results across a variety of local domains. 6.1 Methodology Given the limited resources available at an academic institution , crawling a section of the web that is of the same magnitude as that indexed by Google or Yahoo! is clearly infeasible. Thus, for a given local domain, we approximate the global graph by crawling a local neighborhood around the domain that is several orders of magnitude larger than the local subgraph. Even though such a graph is still orders of magnitude smaller than the `true' global graph, we contend that, even if there exist some highly influential pages that are very far away from our local domain, it is unrealis-tic for any local node selection algorithm to find them. Such pages also tend to be highly unrelated to pages within the local domain. When explaining our node selection strategies in section 5, we made the simplifying assumption that our local graph contained no dangling nodes. This assumption was only made to ease our analysis. Our implementation efficiently handles dangling links by replacing each zero column of our adjacency matrix with the uniform vector. We evaluate the algorithm using the two node selection strategies given in Section 5.2, and also against the following baseline methods: Random: Nodes are chosen uniformly at random among the known global nodes. OutlinkCount: Global nodes with the highest number of outlinks from the local domain are chosen. At each iteration of the FindGlobalPR algorithm, we evaluate performance by computing the difference between the current PageRank estimate of the local domain, E L f E L f 1 , and the global PageRank of the local domain E L g E L g 1 . All PageRank calculations were performed using the uniform random surfer vector. Across all experiments, we set the random surfer parameter , to be .85, and used a convergence threshold of 10 -6 . We evaluate the difference between the local and global PageRank vectors using three different metrics : the L 1 and L norms, and Kendall's tau. The L 1 norm measures the sum of the absolute value of the differences between the two vectors, and the L norm measures the absolute value of the largest difference. Kendall's tau metric is a popular rank correlation measure used to compare PageRanks [2, 11]. This metric can be computed by counting the number of pairs of pairs that agree in ranking, and subtracting from that the number of pairs of pairs that disagree in ranking. The final value is then normalized by the total number of n 2 such pairs, resulting in a [ -1, 1] range, where a negative score signifies anti-correlation among rankings, and values near one correspond to strong rank correlation. 6.2 Results Our experiments are based on two large web crawls and were downloaded using the web crawler that is part of the Nutch open source search engine project [18]. All crawls were restricted to only `http' pages, and to limit the number of dynamically generated pages that we crawl, we ig-nored all pages with urls containing any of the characters `?', `*', `@', or `='. The first crawl, which we will refer to as the `edu' dataset, was seeded by homepages of the top 100 graduate computer science departments in the USA, as rated by the US News and World Report [16], and also by the home pages of their respective institutions. A crawl of depth 5 was performed, restricted to pages within the `.edu' domain, resulting in a graph with approximately 4.7 million pages and 22.9 million links. The second crawl was seeded by the set of pages under the `politics' hierarchy in the dmoz open directory project[17]. We crawled all pages up to four links away, which yielded a graph with 4.4 million pages and 17.3 million links. Within the `edu' crawl, we identified the five site-specific domains corresponding to the websites of the top five graduate computer science departments, as ranked by the US News and World Report. This yielded local domains of various sizes, from 10,626 (UIUC) to 59,895 (Berkeley). For each of these site-specific domains with size n, we performed 50 iterations of the FindGlobalPR algorithm to crawl a total of 2n additional nodes. Figure 2(a) gives the (L 1 ) difference from the PageRank estimate at each iteration to the global PageRank, for the Berkeley local domain. The performance of this dataset was representative of the typical performance across the five computer science site-specific local domains. Initially, the L 1 difference between the global and local PageRanks ranged from .0469 (Stanford ) to .149 (MIT). For the first several iterations, the 122 Research Track Paper 0 5 10 15 20 25 30 35 40 45 50 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 Number of Iterations Global and Local PageRank Difference (L1) Stochastic Complement PageRank Flow Outlink Count Random 0 10 20 30 40 50 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 Number of Iterations Global and Local PageRank Difference (L1) Stochastic Complement PageRank Flow Outlink Count Random 0 5 10 15 20 25 0.16 0.18 0.2 0.22 0.24 0.26 0.28 0.3 0.32 0.34 Number of Iterations Global and Local PageRank Difference (L1) Stochastic Complement PageRank Flow Outlink Count Random (a) www.cs.berkeley.edu (b) www.enterstageright.com (c) Politics Figure 2: L 1 difference between the estimated and true global PageRanks for (a) Berkeley's computer science website, (b) the site-specific domain, www.enterstageright.com, and (c) the `politics' topic-specific domain. The stochastic complement method outperforms all other methods across various domains. three link-based methods all outperform the random selection heuristic. After these initial iterations, the random heuristic tended to be more competitive with (or even outperform , as in the Berkeley local domain) the outlink count and PageRank flow heuristics. In all tests, the stochastic complementation method either outperformed, or was competitive with, the other methods. Table 1 gives the average difference between the final estimated global PageRanks and the true global PageRanks for various distance measures. Algorithm L 1 L Kendall Stoch. Comp. .0384 .00154 .9257 PR Flow .0470 .00272 .8946 Outlink .0419 .00196 .9053 Random .0407 .00204 .9086 Table 1: Average final performance of various node selection strategies for the five site-specific computer science local domains. Note that Kendall's Tau measures similarity, while the other metrics are dissimilarity measures. Stochastic Complementation clearly outperforms the other methods in all metrics. Within the `politics' dataset, we also performed two site-specific tests for the largest websites in the crawl: www.adam-smith .org, the website for the London based Adam Smith Institute, and www.enterstageright.com, an online conservative journal. As with the `edu' local domains, we ran our algorithm for 50 iterations, crawling a total of 2n nodes. Figure 2 (b) plots the results for the www.enterstageright.com domain. In contrast to the `edu' local domains, the Random and OutlinkCount methods were not competitive with either the SC-Select or the PF-Select methods. Among all datasets and all node selection methods, the stochastic complementation method was most impressive in this dataset, realizing a final estimate that differed only .0279 from the global PageRank, a ten-fold improvement over the initial local PageRank difference of .299. For the Adam Smith local domain, the initial difference between the local and global PageRanks was .148, and the final estimates given by the SC-Select , PF-Select, OutlinkCount, and Random methods were .0208, .0193, .0222, and .0356, respectively. Within the `politics' dataset, we constructed four topic-specific local domains. The first domain consisted of all pages in the dmoz politics category, and also all pages within each of these sites up to two links away. This yielded a local domain of 90,811 pages, and the results are given in figure 2 (c). Because of the larger size of the topic-specific domains, we ran our algorithm for only 25 iterations to crawl a total of n nodes. We also created topic-specific domains from three political sub-topics: liberalism, conservatism, and socialism. The pages in these domains were identified by their corresponding dmoz categories. For each sub-topic, we set the local domain to be all pages within three links from the corresponding dmoz category pages. Table 2 summarizes the performance of these three topic-specific domains, and also the larger political domain. To quantify a global page j's effect on the global PageRank values of pages in the local domain, we define page j's impact to be its PageRank value, g[j], normalized by the fraction of its outlinks pointing to the local domain: impact(j) = o L [j] o[j] g[j], where, o L [j] is the number of outlinks from page j to pages in the local domain L, and o[j] is the total number of j's outlinks. In terms of the random surfer model, the impact of page j is the probability that the random surfer (1) is currently at global page j in her random walk and (2) takes an outlink to a local page, given that she has already decided not to jump to a random page. For the politics local domain, we found that many of the pages with high impact were in fact political pages that should have been included in the dmoz politics topic, but were not. For example, the two most influential global pages were the political search engine www.askhenry.com, and the home page of the online political magazine, www.policy-review .com. Among non-political pages, the home page of the journal "Education Next" was most influential. The journal is freely available online and contains articles regarding various aspect of K-12 education in America. To provide some anecdotal evidence for the effectiveness of our page selection methods, we note that the SC-Select method chose 11 pages within the www.educationnext.org domain, the PF-Select method discovered 7 such pages, while the OutlinkCount and Random methods found only 6 pages each. For the conservative political local domain, the socialist website www.ornery.org had a very high impact score. This 123 Research Track Paper All Politics: Algorithm L 1 L 2 Kendall Stoch. Comp. .1253 .000700 .8671 PR Flow .1446 .000710 .8518 Outlink .1470 .00225 .8642 Random .2055 .00203 .8271 Conservativism: Algorithm L 1 L 2 Kendall Stoch. Comp. .0496 .000990 .9158 PR Flow .0554 .000939 .9028 Outlink .0602 .00527 .9144 Random .1197 .00102 .8843 Liberalism: Algorithm L 1 L 2 Kendall Stoch. Comp. .0622 .001360 .8848 PR Flow .0799 .001378 .8669 Outlink .0763 .001379 .8844 Random .1127 .001899 .8372 Socialism: Algorithm L 1 L Kendall Stoch. Comp. .04318 .00439 .9604 PR Flow .0450 .004251 .9559 Outlink .04282 .00344 .9591 Random .0631 .005123 .9350 Table 2: Final performance among node selection strategies for the four political topic-specific crawls. Note that Kendall's Tau measures similarity, while the other metrics are dissimilarity measures. was largely due to a link from the front page of this site to an article regarding global warming published by the National Center for Public Policy Research, a conservative research group in Washington, DC. Not surprisingly, the global PageRank of this article (which happens to be on the home page of the NCCPR, www.nationalresearch.com), was approximately .002, whereas the local PageRank of this page was only .00158. The SC-Select method yielded a global PageRank estimate of approximately .00182, the PF-Select method estimated a value of .00167, and the Random and OutlinkCount methods yielded values of .01522 and .00171, respectively. RELATED WORK The node selection framework we have proposed is similar to the url ordering for crawling problem proposed by Cho et al. in [6]. Whereas our framework seeks to minimize the difference between the global and local PageRank, the objective used in [6] is to crawl the most highly (globally) ranked pages first. They propose several node selection algorithms, including the outlink count heuristic, as well as a variant of our PF-Select algorithm which they refer to as the `Page-Rank ordering metric'. They found this method to be most effective in optimizing their objective, as did a recent survey of these methods by Baeza-Yates et al. [1]. Boldi et al. also experiment within a similar crawling framework in [2], but quantify their results by comparing Kendall's rank correlation between the PageRanks of the current set of crawled pages and those of the entire global graph. They found that node selection strategies that crawled pages with the highest global PageRank first actually performed worse (with respect to Kendall's Tau correlation between the local and global PageRanks) than basic depth first or breadth first strategies. However, their experiments differ from our work in that our node selection algorithms do not use (or have access to) global PageRank values. Many algorithmic improvements for computing exact PageRank values have been proposed [9, 10, 14]. If such algorithms are used to compute the global PageRanks of our local domain, they would all require O(N ) computation, storage, and bandwidth, where N is the size of the global domain. This is in contrast to our method, which approximates the global PageRank and scales linearly with the size of the local domain. Wang and Dewitt [22] propose a system where the set of web servers that comprise the global domain communicate with each other to compute their respective global PageRanks . For a given web server hosting n pages, the computational , bandwidth, and storage requirements are also linear in n. One drawback of this system is that the number of distinct web servers that comprise the global domain can be very large. For example, our `edu' dataset contains websites from over 3,200 different universities; coordinating such a system among a large number of sites can be very difficult. Gan, Chen, and Suel propose a method for estimating the PageRank of a single page [5] which uses only constant bandwidth , computation, and space. Their approach relies on the availability of a remote connectivity server that can supply the set of inlinks to a given page, an assumption not used in our framework. They experimentally show that a reasonable estimate of the node's PageRank can be obtained by visiting at most a few hundred nodes. Using their algorithm for our problem would require that either the entire global domain first be downloaded or a connectivity server be used, both of which would lead to very large web graphs. CONCLUSIONS AND FUTURE WORK The internet is growing exponentially, and in order to navigate such a large repository as the web, global search engines have established themselves as a necessity. Along with the ubiquity of these large-scale search engines comes an increase in search users' expectations. By providing complete and isolated coverage of a particular web domain, localized search engines are an effective outlet to quickly locate content that could otherwise be difficult to find. In this work, we contend that the use of global PageRank in a localized search engine can improve performance. To estimate the global PageRank, we have proposed an iterative node selection framework where we select which pages from the global frontier to crawl next. Our primary contribution is our stochastic complementation page selection algorithm. This method crawls nodes that will most significantly impact the local domain and has running time linear in the number of nodes in the local domain. Experimentally , we validate these methods across a diverse set of local domains, including seven site-specific domains and four topic-specific domains. We conclude that by crawling an additional n or 2n pages, our methods find an estimate of the global PageRanks that is up to ten times better than just using the local PageRanks. Furthermore, we demonstrate that our algorithm consistently outperforms other existing heuristics. 124 Research Track Paper Often times, topic-specific domains are discovered using a focused web crawler which considers a page's content in conjunction with link anchor text to decide which pages to crawl next [4]. Although such crawlers have proven to be quite effective in discovering topic-related content, many irrelevant pages are also crawled in the process. Typically, these pages are deleted and not indexed by the localized search engine. These pages can of course provide valuable information regarding the global PageRank of the local domain . One way to integrate these pages into our framework is to start the FindGlobalPR algorithm with the current subgraph F equal to the set of pages that were crawled by the focused crawler. The global PageRank estimation framework, along with the node selection algorithms presented, all require O(n) computation per iteration and bandwidth proportional to the number of pages crawled, T k. If the number of iterations T is relatively small compared to the number of pages crawled per iteration, k, then the bottleneck of the algorithm will be the crawling phase. However, as the number of iterations increases (relative to k), the bottleneck will reside in the node selection computation. In this case, our algorithms would benefit from constant factor optimizations. Recall that the FindGlobalPR algorithm (Algorithm 2) requires that the PageRanks of the current expanded local domain be recomputed in each iteration. Recent work by Langville and Meyer [12] gives an algorithm to quickly recompute PageRanks of a given webgraph if a small number of nodes are added. This algorithm was shown to give speedup of five to ten times on some datasets. We plan to investigate this and other such optimizations as future work. In this paper, we have objectively evaluated our methods by measuring how close our global PageRank estimates are to the actual global PageRanks. To determine the benefit of using global PageRanks in a localized search engine, we suggest a user study in which users are asked to rate the quality of search results for various search queries. For some queries, only the local PageRanks are used in ranking , and for the remaining queries, local PageRanks and the approximate global PageRanks, as computed by our algorithms , are used. The results of such a study can then be analyzed to determine the added benefit of using the global PageRanks computed by our methods, over just using the local PageRanks. Acknowledgements. This research was supported by NSF grant CCF-0431257, NSF Career Award ACI-0093404, and a grant from Sabre, Inc. REFERENCES [1] R. Baeza-Yates, M. Marin, C. Castillo, and A. Rodriguez. Crawling a country: better strategies than breadth-first for web page ordering. World-Wide Web Conference, 2005. [2] P. Boldi, M. Santini, and S. Vigna. Do your worst to make the best: paradoxical effects in pagerank incremental computations. Workshop on Web Graphs, 3243:168180, 2004. [3] S. Brin and L. Page. The anatomy of a large-scale hypertextual web search engine. Computer Networks and ISDN Systems, 33(17):107117, 1998. [4] S. Chakrabarti, M. van den Berg, and B. Dom. Focused crawling: a new approach to topic-specific web resource discovery. World-Wide Web Conference, 1999. [5] Y. Chen, Q. Gan, and T. Suel. Local methods for estimating pagerank values. Conference on Information and Knowledge Management, 2004. [6] J. Cho, H. Garcia-Molina, and L. Page. Efficient crawling through url ordering. World-Wide Web Conference, 1998. [7] T. H. Haveliwala and S. D. Kamvar. The second eigenvalue of the Google matrix. Technical report, Stanford University, 2003. [8] T. Joachims, F. Radlinski, L. Granka, A. Cheng, C. Tillekeratne, and A. Patel. Learning retrieval functions from implicit feedback. http://www.cs.cornell.edu/People/tj/career. [9] S. D. Kamvar, T. H. Haveliwala, C. D. Manning, and G. H. Golub. Exploiting the block structure of the web for computing pagerank. World-Wide Web Conference, 2003. [10] S. D. Kamvar, T. H. Haveliwala, C. D. Manning, and G. H. Golub. Extrapolation methods for accelerating pagerank computation. World-Wide Web Conference, 2003. [11] A. N. Langville and C. D. Meyer. Deeper inside pagerank. Internet Mathematics, 2004. [12] A. N. Langville and C. D. Meyer. Updating the stationary vector of an irreducible markov chain with an eye on Google's pagerank. SIAM Journal on Matrix Analysis, 2005. [13] P. Lyman, H. R. Varian, K. Swearingen, P. Charles, N. Good, L. L. Jordan, and J. Pal. How much information 2003? School of Information Management and System, University of California at Berkely, 2003. [14] F. McSherry. A uniform approach to accelerated pagerank computation. World-Wide Web Conference, 2005. [15] C. D. Meyer. Stochastic complementation, uncoupling markov chains, and the theory of nearly reducible systems. SIAM Review, 31:240272, 1989. [16] US News and World Report. http://www.usnews.com. [17] Dmoz open directory project. http://www.dmoz.org. [18] Nutch open source search engine. http://www.nutch.org. [19] F. Radlinski and T. Joachims. Query chains: learning to rank from implicit feedback. ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2005. [20] S. Raghavan and H. Garcia-Molina. Crawling the hidden web. In Proceedings of the Twenty-seventh International Conference on Very Large Databases, 2001. [21] T. Tin Tang, D. Hawking, N. Craswell, and K. Griffiths. Focused crawling for both topical relevance and quality of medical information. Conference on Information and Knowledge Management, 2005. [22] Y. Wang and D. J. DeWitt. Computing pagerank in a distributed internet search system. Proceedings of the 30th VLDB Conference, 2004. 125 Research Track Paper

node selection;Experimentation;global PageRank;Algorithms;crawling;site specific domain;localized search engines

86

Evaluating Similarity Measures: A Large-Scale Study in the Orkut Social Network

Online information services have grown too large for users to navigate without the help of automated tools such as collaborative filtering, which makes recommendations to users based on their collective past behavior. While many similarity measures have been proposed and individually evaluated, they have not been evaluated relative to each other in a large real-world environment. We present an extensive empirical comparison of six distinct measures of similarity for recommending online communities to members of the Orkut social network. We determine the usefulness of the different recommendations by actually measuring users' propensity to visit and join recommended communities. We also examine how the ordering of recommendations influenced user selection, as well as interesting social issues that arise in recommending communities within a real social network.

INTRODUCTION The amount of information available online grows far faster than an individual's ability to assimilate it. For example, consider "communities" (user-created discussion groups) within Orkut, a social-networking website (http://www.orkut.com) 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'05, August 2124, 2005, Chicago, Illinois, USA. Copyright 2005 ACM 1-59593-135-X/05/0008 ... $ 5.00. affiliated with Google. The original mechanisms for users to find communities were labor-intensive, including searching for keywords in community titles and descriptions or browsing other users' memberships. Four months after its January 2004 debut, Orkut had over 50,000 communities, providing the necessity and opportunity for data-mining for automated recommendations. There are now (May 2005) over 1,500,000 communities. While there are many forms of recommender systems [3], we chose a collaborative filtering approach [13] based on overlapping membership of pairs of communities. We did not make use of semantic information, such as the description of or messages in a community (although this may be an area of future work). Our recommendations were on a per-community , rather than a per-user basis; that is, all members of a given community would see the same recommendations when visiting that community's page. We chose this approach out of the belief, which was confirmed, that community memberships were rich enough to make very useful recommendations without having to perform more compu-tationally intensive operations, such as clustering of users or communities or computing nearest neighbor relations among users. Indeed, Sarwar et al. have found such item-based algorithms to be both more efficient and successful than user-based algorithms [13]. By measuring user acceptance of recommendations , we were able to evaluate the absolute and relative utility of six different similarity measures on a large volume of data. MEASURES OF SIMILARITY The input data came from the membership relation M = {(u, c) | u U , c C} , where C is the set of communities with at least 20 members and U the set of users belonging to at least one such community. When we began our experiment in May 2004, |C| = 19, 792, |U | = 181, 160, and |M| = 2, 144, 435. Table 1 summarizes the distribution. All of our measures of community similarity involve the overlap between two communities, i.e., the number of com-Table 1: Distribution of community memberships min max median Users per community 20 9077 50 230.5 Communities per user 1 4173 6 28.0 678 Research Track Poster mon users. If a base community b and a (potentially) related community r are considered as sets of users, the overlap is |B R|, where we use capital letters to represent the set containing a community's members. Note that overlap cannot be the sole factor in relatedness, as the size of communities varies greatly. If we only considered overlap, practically every community would be considered related to the "Linux" community, which was the most popular, with 9,077 members . The similarity measures in the next section normalize the overlap in different ways. 2.1 Similarity Measure Functions Each similarity measure we consider is presented as a (possibly asymmetric) function of b and r indicating how appropriate the related community r is as a recommendation for the base community b. We do not use the result of the function as an absolute measure of similarity, only to rank recommendations for a given base community. 2.1.1 L1-Norm If we consider the base and related communities to be vectors b and r , where the i th element of a vector is 1 if user i is a member and 0 if not, we can measure the overlap as the product of their L1-norms: L1( b , r ) = b r b 1 r 1 This quantity can also be expressed in set notation, where we use a capital letter to represent the set containing a community's members: L1(B, R) = |B R| |B| |R| Note that this evaluates to the overlap between the two groups divided by the product of their sizes. When the base community is held constant (as when we determine the base community's recommendations), this evaluates to the overlap divided by the size of the related community, favoring small communities. Kitts et al. [9] reported this to be a successful measure of similarity in their recommender system. 2.1.2 L2-Norm Similarly, we can measure the overlap with the product of the L2-norms ("cosine distance" [3, 6, 12]) of b and r : L2( b , r ) = b r b 2 r 2 In set notation: L2(B, R) = |B R| |B| |R| Note that the square-root in the denominator causes L2 to penalize large communities less severely than L1. Observe that the L2-norm presented here is equivalent to the widely used cosine coefficient applied to binary data. Moreover, while Pearson correlation has been used previ-ously in recommender systems where ranking data is available , we did not use this measure here since it is generally considered inappropriate for binary data. 2.1.3 Pointwise Mutual-Information: positive correlations (MI1) Information theory motivates other measures of correlation , such as "mutual information" [2]. We chose pointwise mutual information where we only count "positive" correlations (membership in both B and R). Such a formulation essentially focuses on how membership in one group is pre-dictive of membership in another (without considering how base non-membership in a group effects membership in another group), yielding: M I1(b, r) = P(r, b) lg P(r, b) P(r) P(b) 2.1.4 Pointwise Mutual-Information: positive and negative correlations (MI2) Similarly, we can compute the pointwise mutual information with both positive and negative correlations (e.g., membership in both B and R, or non-membership in both groups). Again, we don't compute the full expected mutual information, since we believe cross-correlations (e.g., how membership in B affects non-membership in R) tend to be distortive with the recommendation task since such cross-correlations are plentiful but not very informative. This yields: M I2(b, r) = P(r, b) lg P(r, b) P(r) P(b) + P( r, b) lg P( r, b) P( r) P( b) 2.1.5 Salton (IDF) Salton proposed a measure of similarity based on inverse document frequency scaling (tf-idf) [12]: IDF (b, r) = P (r|b) (- lg P(r)) IDF (B, R) = |B R| |B| (- lg |R| |U | ) 2.1.6 Log-Odds We first considered the standard log-odds function, which measures the relative likelihood that presence or absence in a base community predicts membership in a related community : LogOdds0(b, r) = lg P(r|b) P(r| b) Empirically, we found this generated the exact same rankings as using the L1-Norm, which makes sense because: 1. Logarithm is monotonic and, while affecting scores, does not affect rankings. 2. Constant factors, such as |B|, do not affect rankings. 3. For |B| |U |, P(r| b) P(r) We formulated a different log-odds metric, which measures whether membership in the base community is likelier to predict membership or absence in the related community: LogOdds(b, r) = lg P(r|b) P( r|b) 679 Research Track Poster Table 2: Average size of top-ranked community for each measure measure Average size rank 1 rank 2 rank 3 L1 332 482 571 L2 460 618 694 MI1 903 931 998 MI2 966 1003 1077 IDF 923 985 1074 LogOdds 357 513 598 Table 3: Agreement in top-ranked results between measures. For example, MI1 and IDF rank the same related community first for 98% of base communities . Correlations greater than 85% are in bold. L1 .70 L2 .41 .60 MI1 .39 .57 .96 MI2 .41 .59 .98 .97 IDF .88 .79 .46 .44 .46 LogOdds 2.2 Discussion For a given measure, we refer to the related community yielding the highest value to be the top-ranked related community relative to a base community. The average size of top-ranked communities for each measure, which varies greatly, is shown in Table 2. Table 3 shows how often two functions yield the same top-ranking result. Table 4 shows the top recommendations for the "I love wine" community. Note that MI1, MI2, and IDF favor very large communities, while L1 and LogOdds favor small communities. Note that in addition to the obvious correlations between the two mutual-information functions (96%), there is a very strong correlation between IDF and the mutual-information functions (97-98%). Manipulation of the formulas for MI1 and IDF shows: M I1(b, r) = P(r, b) lg P(r, b) P(r) P(b) = P(r|b)P(b) lg P(r|b) - P(r|b)P(b) lg P(r) = P(r|b)P(b) lg P(r|b) -P(r|b) [1 - P( b)] lg P(r) = P(r|b)[P(b) lg P(r|b) + P(r|b)P( b) lg P(r)] -P(r|b) lg P(r) Substituting IDF (b, r) = -P (r|b) lg P(r), we get: M I1(b, r) = P(r|b) P(b) lg P(r|b) + P( b) lg P(r) +IDF (b, r) Since for virtually all communities b, P (b) P ( b), we can approximate: M I1(b, r) IDF (b, r) + P(r|b) P( b) lg P(r) Thus, MI1 yields a ranking that can be thought of as starting with the ranking of IDF and perturbing the score of each element in the ranking by P(r|b) P( b) lg P(r), which generally is not great enough to change the relative ranking of the top scores, leading to MI1 and IDF often giving the same ranking to top-scoring communities. (Note that this perturbation quantity is given only to explain the high correlation between MI1 and IDF. Statistically, it is meaningless, since b and b cannot simultaneously hold.) EXPERIMENT DESIGN We designed an experiment to determine the relative value of the recommendations produced by each similarity measure . This involved interleaving different pairs of recommendations and tracking user clicks. Specifically, we measured the efficacy of different similarity measures using pair-wise binomial sign tests on click-through data rather than using traditional supervised learning measures such as preci-sion/recall or accuracy since there is no "true" labeled data for this task (i.e., we do not know what are the correct communities that should be recommended to a user). Rather, we focused on the task of determining which of the similarity measures performs best on a relative performance scale with regard to acceptance by users. 3.1 Combination When a user viewed a community page, we hashed the combined user and community identifiers to one of 30 values , specifying an ordered pair of similarity measures to compare. Let S and T be the ordered lists of recommendations for the two measures, where S = (s 1 , s 2 , . . . , s |S| ) and T = (t 1 , t 2 , . . . , t |T | ) and |S| = |T |. The recommendations of each measure are combined by Joachims' "Combined Rank-ing" algorithm [7], restated in Figure 1. The resulting list is guaranteed to contain the top k S and k T recommendations for each measure, where k T k S k T + 1 [7, Theorem 1]. 3.2 Measurements Whenever a user visited a community, two measures were chosen and their recommendations interleaved, as discussed above. This was done in a deterministic manner so that a given user always saw the same recommendations for a given community. To minimize feedback effects, we did not regenerate recommendations after the experiment began. A user who views a base community (e.g., "I love wine") is either a member (denoted by "M") or non-member (denoted by "n"). (We capitalize "M" but not "n" to make them eas-ier to visually distinguish.) In either case, recommendations are shown. When a user clicks on a recommendation, there are three possibilities: (1) the user is already a member of the recommended community ("M"), (2) the user joins the recommended community ("j"), or (3) the user visits but does not join the recommended community ("n"). The combination of base and related community memberships can be combined in six different ways. For example "Mj" denotes a click where a member of the base community clicks on a recommendation to another community to which she does not belong and joins that community. Traditionally, analyses of recommender systems focus on "Mj", also known informally as "if you like this, you'll like that" or formally as "similarity" or "conversion". "Mn" recommendations are considered distracters, having negative utility, since they waste a user's time with an item not of interest. Before running the experiment, we decided that the measures should be judged on their "Mj" performance. Other interpretations are possible: "Mn" links could be considered to have positive utility for any of the following 680 Research Track Poster Table 4: Top recommendations for each measure for the "I love wine" community, with each recommended community's overlap with the base community and size. The size of "I love wine" is 2400. L1 L2 MI1 MI2 IDF LogOdds 1 Ice Wine (Eiswein) (33/51) Red Wine (208/690) Japanese Food/Sushi Lovers (370/3206) Japanese Food/Sushi Lovers (370/3206) Japanese Food/Sushi Lovers (370/3206) Japanese Food/Sushi Lovers (370/3206) 2 California Pinot Noir (26/41) Cheeses of the World (200/675) Red Wine (208/690) Red Wine (208/690) Photography (319/4679) Photography (319/4679) 3 Winery Visitor Worldwide (44/74) I love red wine! (170/510) Cheeses of the World (200/675) Cheeses of the World (200/675) Red Wine (208/690) Linux (299/9077) Figure 1: Joachims' "Combine Rankings" algorithm [7] Input: ordered recommendation lists S = (s 1 , s 2 , . . . , s |S| ) and T = (t 1 , t 2 , . . . , t |T | ) where |S| = |T | Call: combine (S, T, 0, 0, ) Output: combined ordered recommendation list D combine(S, T, k s , k t , D){ if (k s &lt; |S| k t &lt; |T |) if (k s = k t ) { if (S[k s + 1] / D){D := D + S[k s + 1]; } combine(S, T, k s + 1, k t , D); } else { if (T [k t + 1] / D){D := D + T [k t + 1]; } combine(S, T, k s , k t + 1, D); } } } Table 5: Clicks on recommendations, by membership status in the base and recommended communities, as counts and as percentages of total clicks. The last column shows the conversion rate, defined as the percentage of non-members clicking on a related community who then joined it ( j n+j ). membership in base community membership in recommended community M (member) n (non-member) j (join) total conversion rate M (member): number of clicks 36353 184214 212982 433549 54% percent of total clicks 4% 20% 24% 48% n (non-member): number of clicks 8771 381241 77905 467917 17% percent of total clicks 1% 42% 9% 52% total: number of clicks 45124 565455 290887 901466 34% percent of total clicks 5% 63% 32% 100% 681 Research Track Poster reasons: 1. As the user found the link sufficiently interesting to click on, it was of more utility than a link not eliciting a click. 2. The user is genuinely interested in the related community but does not want to proclaim her interest, as membership information is public and some communities focus on taboo or embarrassing topics. For example , a recommendation given for the popular "Choco-late" community is "PMS". Note that this effect is specific to social networks and not, for example, Usenet groups, where the user's list of communities is not revealed to other users. Similarly, it is unclear how to value clicks from a base community that the user does not belong to. Does an "nj" click indicate failure, since the base community was not joined by the user, but the recommended community was, indicating a degree of dissimilarity? Or is it of positive utility, since it helped a user find a community of interest? For these reasons, we tracked all clicks, recording the user's membership status in the base and recommended communities for later analysis. (We did not track whether users returned to communities in the future because of the logging overhead that would be required.) 3.3 User Interface On community pages, our recommendations were provided in a table, each cell of which contained a recommended community's name, optional picture, and link (Figure 2). Recommendations were shown by decreasing rank from left to right, top to bottom, in up to 4 rows of 3. For aesthetic reasons, we only showed entire rows; thus, no recommendations were displayed if there were fewer than 3. We also provided a control that allowed users to send us comments on the recommendations. RESULTS We analyzed all accesses between July 1, 2004, to July 18, 2004, of users who joined Orkut during that period. The system served 4,106,050 community pages with recommendations , which provides a lower bound on the number of views. (Unfortunately, we could not determine the total number of views due to browser caching.) There were 901,466 clicks on recommendations, 48% by members of the base community, 52% by non-members (Table 5). Clicks to related communities to which the user already belonged were rare, accounting for only 5% of clicks. The most common case was for a non-member of a base community to click through and not join a related community (42%). We defined conversion rate (also called precision) as the percentage of non-members who clicked through to a community who then joined it. The conversion rate was three times as high (54%) when the member belonged to the base community (from which the recommendation came) than not (17%). 4.1 Relative performance of different measures We compared each measure pairwise against every other measure by analyzing clicks of their merged recommendations . If the click was on a recommendation ranked higher by measure L2 than measure L1, for example, we considered it a "win" for L2 and a loss for L1. If both measures ranked it equally, the result was considered to be a tie. Table 6 shows the outcomes of all clicks, with conversions by members ("Mj") and non-members of the base community ("nj") broken out separately. We say that a measure dominates another if, in their pairwise comparison, the former has more "wins". For example, L2 dominates L1. This definition, combined with the data in Table 6, yielded a total order (to our surprise) among the measures: L2, MI1, MI2, IDF, L1, LogOdds. The same total order occurred if only "nj" clicks were considered. The order was different if all clicks were considered: L2, L1, MI1, MI2, IDF, LogOdds. 4.2 Conversion rates There was great variance in conversion rate by recommended community. We examined the 93 recommended communities that were clicked through to more than 1000 times. Unsurprisingly, the ten with the lowest conversion rate all were about sex (e.g., Amateur Porn). Note that members of the base community were far more willing than non-members to join, perhaps because they had already shown their willingness to join a sex-related community. At the other extreme, none of the ten with the highest conversion rate were sexual (e.g., Flamenco). Table 7 provides selected data by each membership combination. Unsurprisingly , for all 93 base communities, members were more likely than non-members to join the recommended community. 4.3 User comments Users were also able to submit feedback on related communities . Most of the feedback was from users who wanted recommendations added or removed. Some complained about inappropriate recommendations of sexual or political communities, especially if they found the displayed image offensive. A few objected to our generating related community recommendations at all, instead of allowing community creators to specify them. In one case, poor recommendations destroyed a community: The creator of a feminist sexuality community disbanded it both because of the prurient recommendations that appeared on her page and the disruptive new members who joined as a result of recommendations from such communities. We agreed with her that the recommendations were problematic and offered to remove them. While anecdotal, this example illustrates how a recommendation can have unanticipated consequences that cannot be captured in simple statistical measures. (An informal discussion of users' behavior when we allowed them to choose related communities can be found elsewhere [14].) POSITIONAL EFFECTS During the above experiment, we became curious how the relative placement of recommendations affected users' selections and performed a second experiment. 5.1 Design After determining that L2 was the best measure of similarity , we recomputed the recommendations and studied the effect of position on click-through. While in our original experiment we displayed up to 12 recommendations in decreasing rank, for this experiment we displayed up to 9 recommendations in random order, again ensuring that each 682 Research Track Poster Table 6: The relative performance of each measure in pairwise combination on clicks leading to joins, divided by base community membership status, and on all clicks. Except where numbers appear in italics, the superioriority of one measure over another was statistically significant (p &lt; .01) using a binomial sign test [10]. measures M j n j all clicks win equal loss win equal loss win equal loss L2 MI1 6899 2977 4993 2600 1073 1853 30664 12277 20332 L2 MI2 6940 2743 5008 2636 1078 1872 31134 11260 19832 L2 IDF 6929 2697 5064 2610 1064 1865 30710 11271 20107 L2 L1 7039 2539 4834 2547 941 1983 28506 13081 23998 L2 LogOdds 8186 1638 4442 2852 564 1655 34954 6664 18631 MI1 MI2 3339 9372 1855 1223 3401 683 14812 37632 7529 MI1 IDF 3431 8854 1891 1139 3288 629 14671 37049 7758 MI1 LogOdds 7099 3546 3341 2514 1213 1193 29837 13869 13921 MI1 L1 6915 1005 6059 2547 407 2338 27786 4308 29418 MI2 IDF 1564 11575 1031 533 4266 359 6003 47885 4490 MI2 LogOdds 6920 3959 3177 2484 1418 598 2881 15308 13188 MI2 L1 6830 950 6419 2383 362 2333 26865 3872 29864 IDF L1 6799 1006 6304 2467 392 2352 27042 4069 29755 IDF LogOdds 6691 3804 3096 2452 1378 1085 28224 15013 13330 L1 LogOdds 6730 518 5975 2521 108 2059 31903 2097 24431 Table 7: Conversion rates by status of membership in base community, for communities to which more than 1000 clicks on recommendations occurred. member of base community non-member of base community Related community MM Mj Mj conversion rate nM nn nM conversion rate 10 communities with highest conversion rates 583 2273 6984 75% 198 3454 2017 37% 10 communities with lowest conversion rates 326 1984 826 29% 68 26287 472 1.8% all 93 communities 13524 54415 52614 46% 3488 127819 19007 17% user always saw the same ordering of recommendations for a given community. By randomizing the position of recommendations , we sought to measure ordering primacy effects in the recommendations as opposed to their ranked quality. 5.2 Results We measured all 1,279,226 clicks on related community recommendations from September 22, 2004, through October 21, 2004. Table 8 shows the relative likelihood of clicks on each position. When there was only a single row, the middle recommendation was clicked most, followed by the leftmost, then rightmost recommendations, although the differences were not statistically significant. When there were two or three rows, the differences were very significant (p &lt; .001), with preferences for higher rows. P-values were computed using a Chi-Squared test comparing the observed click-through rates with a uniform distribution over all positions [10]. CONCLUSION AND FUTURE PLANS Orkut's large number of community memberships and users allowed us to evaluate the relative performance of six different measures of similarity in a large-scale real-world study. We are not aware of any comparable published large-scale experiments. We were surprised that a total order emerged among the similarity measures and that L2 vector normalization showed the best empirical results despite other measures, such as log-odds and pointwise mutual information , which we found more intuitive. For future work, we would like to see how recommendations handpicked by community owners compare. Just as we can estimate communities' similarity through common users, we can estimate users' similarity through common community memberships: i.e., user A might be similar to user B because they belong to n of the same communities . It will be interesting to see whether L2 also proves superior in such a domain. We could also take advantage of Orkut's being a social network [8], i.e., containing information on social connections between pairs of users. In addition to considering common community memberships, we could consider distance between users in the "friendship graph". Users close to each other (e.g., friends or friends-of -friends) might be judged more likely to be similar than distant strangers, although some users might prefer the latter type of link, since it would introduce them to someone they would be unlikely to meet otherwise, perhaps from a different country or culture. Similarly, friendship graph information can be taken into account when making community recommendations, which would require that recommendations be computed on a per-user (or per-clique), rather than per-community, basis. In such a setting, we could make community recommendations based on weighted community overlap vectors where weights are determined based on the graph distances of other community members to a given user. This is a fertile area for future work and yet another example of how the interaction 683 Research Track Poster Figure 2: Displays of recommendations for three different communities Table 8: The relative likelihood of clicks on link by position when there are (a) one, (b) two, or (c) three rows of three recommendations. (a) n=28108, p=.12 (b) n=24459, p&lt;.001 (c) n=1226659, p&lt;.001 1.00 1.01 .98 1.04 1.05 1.08 1.11 1.06 1.04 .97 .94 .92 1.01 .97 .99 1.01 .94 .87 of data mining and social networks is becoming an exciting new research area [4] [11]. ACKNOWLEDGMENTS This work was performed while Ellen Spertus was on sabbatical from Mills College and a visiting scientist at Google. We are grateful to Patrick Barry, Alex Drobychev, Jen Fitz-patrick , Julie Jalalpour, Dave Jeske, Katherine Lew, Marissa Mayer, Tom Nielsen, Seva Petrov, Greg Reshko, Catherine Rondeau, Adam Sawyer, Eric Sachs, and Lauren Simpson for their help on the Orkut project and to Corey Anderson, Alex Drobychev, Oren Etzioni, Alan Eustace, Keith Golden, Carrie Grimes, Pedram Keyani, John Lamping, Tom Nielsen, Peter Norvig, Kerry Rodden, Gavin Tachibana, and Yonatan Zunger for their help on this research or its exposition. REFERENCES [1] Breese, J.; Heckerman, D.; Kadie, C. Empirical Analysis of Predictive Algorithms for Collaborative Filtering. In Proceedings of the Fourteenth Conference on Uncertainty in Artificial Intelligence (Madison, Wisconsin, 1998). Morgan Kaufmann. [2] Cover, T.M., and Thomas, J.A. Elements of Information Theory. Wiley, New York, 1991. [3] Deshpande, M., and Karypis, G. Item-Based Top-N Recommendation Algorithms. ACM Transactions on Information Systems 22(1) (January 2004), 143-177. [4] Domingos, P. Prospects and Challenges for Multi-Relational Data Mining. ACM SIGKDD Exploration Newsletter 5(1) (July 2003). [5] Dumais, S.; Joachims, T.; Bharat, K.; Weigend, A. SIGIR 2003 Workshop Report: Implicit Measures of User Interests and Preferences. SIGIR Forum 37(2) (Fall 2003). [6] Harman, D. Ranking Algorithms. In W. B. Frakes and R. Baeza-Yates (ed.), Information Retrieval: Data Structures & Algorithms (chapter 14). Upper Saddle River, NJ, USA: Prentice Hall, 1992. [7] Joachims, T. Evaluating Retrieval Performance Using Clickthrough Data. In Proceedings of the SIGIR Workshop on Mathematical/Formal Methods in Information Retrieval (2002). ACM Press, New York, NY. [8] Kautz, H.; Selman, Bart; Shah, M. Referral Web: Combining Social Networks and Collaborative Filtering. Communications of the ACM 45(8) (March 1997). [9] Kitts, B.; Freed, D.; Vrieze, M. Cross-Sell: A Fast Promotion-Tunable Customer-Item Recommendation Method based on Conditionally Independent Probabilities. In Proceedings of the Sixth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (Boston, 2000). ACM Press, New York, NY, 437-446. [10] Lehmann, E.L. Testing Statistical Hypotheses (second edition). Springer-Verlag, 1986. [11] Raghavan, P. Social Networks and the Web (Invited Talk). In Advances in Web Intelligence: Proceedings of the Second International Atlantic Web Intelligence Conference, May 2004. Springer-Verlag, Heidelberg. [12] Salton, G. Automatic Text Processing: The Transformation, Analysis, and Retrieval of Information by Computer. Addison Wesley, Reading, MA, 1989. [13] Sarwar, B.; Karypis, G.; Konstan, J.; Reidl, J. Item-Based Collaborative Filtering Recommendation Algorithms. In Proceedings of the Tenth International Conference on the World Wide Web (WWW10) (Hong Kong, 2001). ACM Press, New York, NY, 285-295. [14] Spertus, Ellen. Too Much Information. Orkut Media Selections, January 19, 2005. Available online at "http://media.orkut.com/articles/0078.html". 684 Research Track Poster

collaborative filtering;online communities;community;recommender system;social network;social networks;similarity measure;Data mining

87

Evaluation and Evolution of a Browse and Search Interface: Relation Browser++

We present in this paper the design and an evaluation of a novel interface called the Relation Browser++ (RB++) for searching and browsing large information collections. RB++ provides visualized category overviews of an information space and allows dynamic filtering and exploration of the result set by tightly coupling the browsing and searching functions. A user study was conducted to compare the effectiveness, efficiency and user satisfaction of completing various types of searching and browsing using the RB++ interface and a traditional form-fillin interface for a video library. An exploration set of tasks was also included to examine the effectiveness of and user satisfaction with the RB++ when applied to a large federal statistics website. The comparison study strongly supported that RB++ was more effective, efficient, and satisfying for completing data exploration tasks. Based on the results, efforts to automatically populate the underlying database using machine learning techniques are underway. Preliminary implementations for two large-scale federal statistical websites have been installed on government servers for internal evaluation.

INTRODUCTION The size and breadth of large government websites and digital libraries makes it difficult for people to quickly grasp what content is and is not available. Dynamic overviews and previews of collections can help people decide if it is worthwhile to look further [8]. As they do look further, it is helpful for their searching and browsing to quickly see partitions of the collection and how many items are available in different partitions. We believe that government website users will be well-served by highly interactive user interfaces that support alternative views of the collections, partitions, and results sets. This paper describes a user interface that aims to provide agile control for browse and search, reports results from a user study comparing this interface to a typical WWW search interface, and describes the ongoing evolution of the interface, including efforts to automate discovery of topical categories and assignment of webpages to those categories. Faceted category structure is one way to help people understand the composition of an information collection. A faceted approach provides different ways to slice and dice the information space, which allows people to look at the information space from different perspectives. Allowing people to explore the relationships among different facets may further deepen their understanding and support new insights. The relation browser (RB) is an interface which provides an overview of the collection by displaying different categories and enables people to explore the relationships among these categories [13]. The different facet values also serve as selectable objects that may be employed as query widgets for a search so that the entire space can quickly be partitioned with simple mouse moves and with consequent immediate display of the resulting partition in the results panel. Figure 1 shows the mock-up interface of an early version of the relation browser in the domain of U.S. federal statistics websites. The web pages in the site were sliced into four different facets: by topic, data type, region, and date. The numbers beside the bars indicate the number of websites associated with the attributes. By mousing over any of the topics, the distribution of the specific topics in other facets are visualized as graphic bars. The underlying data for this instance of the interface was manually extracted from a small set of 200 webpages contained in more than 70 federal statistical agency websites. 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. Conference'04, Month 12, 2004, City, State, Country. Copyright 2004 ACM 1-58113-000-0/00/0004...$5.00. Copyright held by the author 179 Figure 1. Relation Browser (RB) This early version of the RB has been redesigned based on user studies and experience applying the interface to more than a dozen different database instances [13]. The new version is called RB++, which improves the RB significantly in several ways (see Figure 2) [23,24]. First, RB++ displays multiple facets (categories) visually and on the same screen rather than only two facets with tab options to others. The multiple facets provide an overview of the information space. The facet values are visually represented by graphic bars with different lengths, which indicate the number of items associated with them. Second, RB++ allows more flexibility to explore relationships. One of the features of RB++ is that you can restrict the information items (partition the information space) by mousing over any bars and other bars are proportionally highlighted to show the conditional distribution across all the facets. Note that the previous RB was limited to visualizing pairwise relationships with one main facet. Third, the RB++ added a dynamic filtering function for the result set (see Figure 3). Once the search results are displayed in the table, further filtering can be done by typing in keywords (string patterns) in the boxes located immediately above the result fields. The filtering is dynamic, which means that with each character typed in or removed from the boxes, RB++ matches the string patterns in the boxes with the corresponding field of the results. Only the matched results are then displayed immediately in the results panel and the matched string in the results is highlighted. This dynamic feature gives users instant and constant feedback about the filtered results and how many items they will get with different keywords, which allows users to try out different filtering keywords very easily and efficiently. Fourth, the RB++ provides an overview of the results set and tightly couples the overview and results set panels. The overview panel is dynamically updated to give users a contextualized overview of the updated result set. These new features give users more power to understand and explore the information collection and give them a flexible and rapid way to find the information they want. A linguistic model of BNF grammar to model the user interaction with the interface is provided in section 2.3 to help reveal the dynamic nature of the RB++. In the paper, we argue that the RB++ interface will bring users added values beyond simple searching and browsing by in fact combining these search strategies seamlessly. In the next section, the methodology of a user study is described. The results of the user study are then presented and discussed. Limitations of the interface and current efforts to deal with data classification are then described. Figure 2. Initial display of RB++ with visualized category overview on the top Figure 3. RB++ with dynamic filtering of the results (note the changes in the overview and updated results) METHODOLOGY The purpose of the user study was two-fold: first, we wanted to compare the effectiveness, efficiency and user satisfaction associated with completing certain tasks using RB++ against that obtained by the traditional form-fillin search interface (baseline interface). Second, we wanted to explore if the RB++ interface would lead to new interaction patterns with the interface and if so, to determine what these new interaction patterns might be. 180 Seventeen undergraduate and graduate students were recruited from the UNC-Chapel Hill campus for this study. They came from various schools and departments. There were 10 females and 7 males with an age range from 19 to 44, (15 were in their 20s, and all were familiar with www browsers. The participants were given $15 for their participation. The data from the first two participants was used as a pilot test; based on which the experimental protocol and instruments were revised. The data from the other 15 participants was used for the data analysis. The study proceeded in two phases, a within subjects comparison across two different interfaces for the same database, and an exploratory investigation with a single government statistical website instance of RB++. 2.1 Phase One: RB++ to Baseline Comparison for a Film Database The first phase was a comparison study in which participants used both the RB++ interface and the baseline interface. The order of using these interfaces was counter balanced. The domain of the information items in both interfaces was the video collection in the UNC-CH library (http://www.lib.unc.edu/house/mrc/index.html?page=filmograph y) that contains about 10000 films. The library online video search interface (Filmfinder) was used as our baseline interface. FilmFinder is a fairly typical www form-fillin search interface (see Figure 4 and Figure 5), where users can specify queries within fields such as title, release year, director, description, genre, origin, and format. Figure 4. Filmfinder with Form-fillin Interface Figure 5. Results Page of Filmfinder All participants were run individually in sessions ranging from 60-90 minutes and all sessions were video taped. The protocol for the first phase was as follows: First, a demographic pre-test questionnaire was completed. Second, the participant was trained for the first interface assigned in their condition. The training consisted of: an introduction to the features of the interface, a demo of each type of task with the interface, and participant practice using the interface until s/he was comfortable with it. Third, the participant used the interface to complete 10 search tasks. Tasks were assigned to participants one by one by handing them pieces of paper for each task. A timer was used to count time used to complete each task except for task 10 (see description of task 10 below). After each task, a short satisfaction questionnaire was completed by the participant. Fourth, a usability questionnaire was filled out after the participant finished using the first interface. Next, the participant was trained for the second interface and the same procedures were used to complete 10 more search tasks. Finally, an open-ended questionnaire about perceived differences and preferences for the two interfaces was completed. The tasks were classified into three different types: 1. Simple look up task. Tasks 1 to 3 in each task set were of this type. For example, "Check if the movie titled "The Matrix" is in the library movie collection." 2. Data exploration and analysis tasks. Tasks 4 to 9 in each task set were of this type. This kind of task requires users to understand and make sense of the information collection, which could be a starting point for them to further their searching or browsing. Two examples of this type are: "In which decade did "Steven Spielberg" direct the most movies?"; and "How many movie titles does the library hold that were released in the year 2000?" 3. Task 10 was a free exploration task, which asked participants to find five favorite videos without any time constraints. The tasks assigned for the two interfaces were different but comparable. For example, the comparable tasks for two interfaces simply substituted different video titles or directors. 2.2 Phase Two: Explore RB++ for EIA Website The second phase was an exploratory study of the RB++ applied to roughly 10,000 pages in the Energy Information Administration (EIA) website. Based on intensive manual 181 inspection of the EIA website, four facets were identified with associated facet values: fuel type (with the facet values: alternatives, coal, electricity, natural gas, nuclear, petroleum, and renewable); geography (state level, regional level, national level, and international level); sector (commercial, electric utility, industrial, and residential); and process (delivery, import/export, price/cost, production, resources/reserves, and usage). All the facets were displayed on the overview panel (see Figure 6). The results panel displayed the title, page size, and description of the web pages. Figure 6. RB++ interface applied to EIA website The protocol of the second phase was as follows: First, the RB++ EIA application was introduced to the participant. Second, the participant practiced using the interface until s/he was comfortable with it. Third, the participant used the interface to complete four tasks 1 . The process was recorded and a short satisfaction questionnaire was filled out after finishing each task. Fourth, an open-ended questionnaire was completed after finishing all the tasks. Lastly, the participant was briefly interviewed. Data collected included both quantitative and qualitative data from the two phases of the study. Data collected for the first phase included performance data (time spent finishing tasks), error rates of tasks, ratings on the satisfaction questionnaire after finishing each task, ratings on the usability questionnaire after finishing each interface, and comments on the open questionnaire about perceived differences and preferences for the two interfaces. Data collected for the second phase included ratings on the satisfaction questionnaire after finishing each task, comments on the post-session questionnaire and the verbal comments made in the interview. 1 Tasks for the second phase study: 1. I want to learn the current status of Chinese nuclear energy. 2. Find the most recent weekly data on petroleum prices in the USA. 3. Find the statistical data on coal production across different states in the year 2001. 4. What kinds of information can I and can I not find from the website? 2.3 Modeling User Interaction To help us form hypotheses and analyze and make sense of the experimental data, we employed a linguistic model, called BNF grammar, to model the user's interaction with the interface. BNF grammar was originally used by Reisner to describe the dialog grammar of an interactive graphics system [15], where the user's interaction with a system was seen as an action language and BNF grammar was used to formally describe the language. The BNF grammar consists of a set of rules which define higher level user behaviors in terms of lower level ones. Each rule can be composed of terminals, non-terminals, and a set of symbols. Terminals usually represent the lowest level of user behavior, such as pressing a key or clicking a mouse button and can not be further defined. Non-terminals represent a high level abstraction and can be defined in terms of other non-terminals and terminals. Terminals are written with upper case letters and non-terminals are written with lower case letters. The "::=" symbol is read as " is defined as". The "+", "|" and "-" symbols are used at the right hand side of rules to connect, respectively, sequence of user behavior, set of options, and concurrent user behaviors. With the BNF grammar, we can describe the user's interaction with the RB++ as follows: A1 information seeking ::= explore collection(A3) | (formulate query(A2) + CLICK SEARCH BUTTON + navigate results(A5)) A2 formulate query ::= (explore collection(A3) + form query(A4)) | form query(A4) A3 explore collection ::= (CLICK VISUAL BAR-OBSERVE VISUAL BAR + explore collection(A3)) | (MOUSE OVER VISUAL BAR-OBSERVE VISUAL BAR + explore collection(A3)) A4 form query ::= (CLICK VISUAL BAR + form query(A4)) | (TYPE IN KEYWORD + form query(A4)) A5 navigate results ::= (browse results(A6) + navigate results(A5)) | (CLICK RESTART BUTTON + information seeking(A1)) A6 browse results ::= (show results(A7)-OBSERVE RESULTS + browse results(A6)) | (CLICK RESULT ITEM + browse results(A6)) | (CLICK SORTING BUTTON + browse results(A6))| (explore results(A8) + browse results(A6)) A7 show results ::= CLICK SIDEBAR A8 explore results ::= (observe system state(A9) + explore results(A8)) | (filter results(A10) + explore results(A8)) A9 observe system state ::= (OBSERVE VISUAL BAR + observe system state(A9)) | (OBSERVE NUMBER + observe system state (A9)) A10 filter results ::= CLICK VISUAL BAR | MOUSE OVER VISUAL BAR | TYPE IN KEYWORD The interaction with baseline interface can be described as: B1 information seeking ::= formulate query(B2) + CLICK SEARCH BUTTON + navigate results(B4) B2 formulate query ::= (TYPE IN KEYWORD + formulate query(B2)) | (select item(B3) + formulate query(B2)) 182 B3 select item ::= CLICK PULL DOWN MENU + CLICK ITEM B4 navigate results ::= (browse results(B5) + navigate results(B4)) | (CLICK NEW SEARCH LINK + information seeking(B1)) B5 browse results ::= (show results(B6)-OBSERVE RESULTS + browse results(B5)) | (show results(B6)-COUNT RESULTS + browse results(B5)) | (CLICK ITEM + browse results(B5)) | (CLICK SORTING LINK + browse results(B5)) B6 show results ::= CLICK SIDEBAR | (CLICK SIDEBAR + CLICK NEXT PAGE LINK) The number of rules and options within rules reflects the interactive nature and number of alternative choices provided by these two interfaces. Note that we used the terminals such as CLICK SEARCH BUTTON and CLICK VISUAL BAR which strictly speaking are not the lowest level of user behaviors, however, using higher level abstraction as terminals is suitable for interactive display-based systems [4] and ensures later data analysis. Many rules are defined recursively and consist of several options, which essentially reflect the interactivity of the graphical user interface (GUI). For example, a fairly interactive user behavior in RB++, "browse results (A6)", consists of either `OBSERVE RESULTS', `CLICK RESULT ITEM', `CLICK SORTING BUTTON', explore results, or any combination of the above. From the BNF definition, we can see that RB++ is a more interactive interface than the baseline because it involves more rules and recursive definitions. However, it is not necessarily a complicated interface, since the rules for the RB++ interface are largely composed of a set of options instead of a sequence of user behaviors, which means that many rules are not executed for some types of tasks. Based on the BNF grammars, we hypothesize that for the simple search tasks, the RB++ interface will not necessarily be significantly different from the baseline interface, but for complicated searching and browsing tasks, that require more interaction or collection exploration, the RB++ will be significantly more effective, efficient, and satisfying than the baseline. For simple look up type tasks, both interfaces involve the sequence of user actions: formulate query, CLICK SEARCH BUTTON, and navigate results (see rule A1 and B1). Navigation of results is simple for this type of task in that it only involves the judgment of zero or non-zero results, which is trivial in both interfaces. Formulation of the query in this case involves typing in keywords and/or selecting the items from the interfaces (see rule A2, A4 and B2, B3). Even though item selection in the baseline interface involves two clicks (see rule B3) which means a slightly longer time to execute than in RB++, which only needs one click on the visual bar for item selection (see rule A4), we expected no significant difference. For type 2 tasks that involve data exploration and analysis, interaction with the visual bars of the RB++ interface provides an effective and efficient interaction style. Two typical sequences of user behaviors to complete type 2 tasks are: explore the collection by clicking (or mousing over) and observing the visual bars (see rule A1 and A3), or formulate a query and then explore the results by observing the visual bars (see rule A1, A5, A6, A8 and A9). With the traditional interface to finish type 2 tasks, users have to formulate a query and then literally scan and count all the results (see rule B1, B4, B5, and B6), which is time consuming. We also hypothesized that users would exhibit rich interaction during their navigation of the results with RB++ (see rule A5 to A10). Actions of typing in keywords and clicking visual bars to filter results (rule A10) would be used frequently and interchangeably by the users to finish complex search tasks, especially when large numbers of results are returned. RESULTS Table 1 lists the average time (in seconds) across all the participants to finish tasks 1 to 9 using the two different interfaces. Notice that we allowed the participants to stop the task if they felt that the task was too hard or too time-consuming to finish. It turned out that there were five participants who stopped task 5 and eight participants stopped task 6 before completion when they used the FilmFinder. Performance data of these participants were discarded for the unfinished tasks. Table 1. Performance data (in seconds) Task 1 (.879) 2 (.522) 3 (.026) 4 (.000) 5 (.000) RB++ 14.4 16.1 17.0 18.9 15.7 FilmFinder 14.7 14.4 29.7 40.7 204.0 Task 6 (.000) 7 (.000) 8 (.000) 9 (.000) 10 RB++ 12.7 13.5 27.1 20.6 N/A FilmFinder 328.0 87.2 101.3 112.8 N/A Paired sample t tests on the performance data were computed and the p values are shown in the parenthesis for each task. We can see that except for the first two tasks (which were type 1 tasks), the performance differences between the two interfaces were all statistically significant at the .05 level. Clearly, RB++ supported superior performance for type 2 tasks. We also counted error rates for tasks 1 to 9, which are listed in Table 2. The error rate was calculated as the number of participants who gave the wrong answer to the task divided by the total number of participants. We can see that except for the 8th task, no participants got wrong answers for any of the tasks using the RB++ interface. The error rates of the baseline interface were much higher than that of the RB++ interface, especially for tasks 5, 6, and 7. Notice that we did not consider those participants who gave up the task 5 or 6 using the Filmfinder, so the actual denominators used for calculating the error rates for these tasks were smaller than the total number of participants. Table 2. Error rates Task 1 2 3 4 5 RB++ 0/15 0/15 0/15 0/15 0/15 FilmFinder 0/15 0/15 0/15 2/15 5/10 Task 6 7 8 9 10 183 RB++ 0/15 0/15 1/15 0/15 N/A FilmFinder 4/7 13/15 2/15 5/15 N/A We also did paired sample t tests on the three satisfaction questions 2 which were completed after each task. Each response was given on a 5 point scale from strongly agree (5) to strongly disagree (1). For the first three tasks (simple lookups), there were no statistically significant differences between the two interfaces on any of the 3 questions. On the exploratory tasks (4-9), statistically significant differences favoring the RB++ were found on all three of the satisfaction questions. We also compared the results for the seven overall usability questions on each interface asked after participants had done the tasks with each interface. Their responses were also given on a five point scale from strongly agree to strongly disagree. In each of the seven ratings, statistically significant differences were found favoring the RB+ interface. Clearly, satisfaction with the RB++ was greater than that with the Filmfinder. There were also open-ended questions that the participants answered after finishing both interfaces. All of the participants considered the RB++ interface to be easier to use, especially for the complex searches. They commented on the easy use of the visual display with the multiple categories, which made it easy to combine the search criteria and narrow down the data, and they also thought it was good to be able to manipulate the search results in multiple ways. Thirteen out of 15 participants indicated that the RB++ interface gave them more confidence to complete the tasks. It was easy to go back and forth and to verify the results and the informative overview panel gave the participants more confidence to finish tasks. There was one participant who thought that both interfaces gave equal confidence and there was one participant who thought that the Filmfinder interface gave more confidence since he was more familiar with the Filmfinder and he felt somewhat confused by the dynamic feature of the RB++, but he acknowledged the usefulness of the dynamic feature in narrowing the results in the results panel. When asked which interface better helped them gain an understanding of the library movie collection, the RB++ interface was chosen by all the participants. Again the visual display of the multiple categories and the cross reference of these categories was considered to be useful features for them to understand the whole collection. In addition, 10 out of the 15 participants indicated that they were more likely to use the RB++ interface if both were available. Three participants chose both interfaces, depending on the type of tasks, and two participants chose the FilmFinder because of its familiarity and aesthetic appeal. For the question on the best thing about the RB++, participants pointed out the visual display of the multiple categories, its cross reference ability, the dynamic matching ability of the searching boxes and the one screen display of the results as opposed to the multiple page display of results in Filmfinder. As the worst thing about the RB++, participants indicated that it was not as 2 It is easy to use the interface I feel satisfied with the results I got I feel confident with the results I got aesthetically appealing as the Filmfinder and not quite as intuitive to use as the Filmfinder. Two participants specifically mentioned that the constant changing and updating of the interface made it a bit confusing. 3.2 Phase Two Results During the second phase we also asked participants to fill out the satisfaction questionnaires after finishing each task and these ratings were predictably high (all means above 3.5) More importantly, participants were also required to answer a set of open-ended questions after finishing the second phase. For the first question: "What is your overall impression of this interface for finding the statistical data?" the overall impression was positive. Participants used phrases such as "fairly easy to use", "very helpful in finding the information", "good for quick searching". There were also a couple of negative comments such as: "interface still came up with many results after filtering", "title of the results are not descriptive enough". Only one participant said that he did not like the interface, because of the poor categorization of information items under some categories which made him frustrated. When answering the second question: "Was it helpful to understanding what is available at EIA?" all the participants thought the interface was helpful in that regard, which was largely attributed to the visual display of the categories, which gave them a sense of what the website covered. One participant wished that there were more categories displayed. The questionnaire also asked if the search boxes were helpful in completing the tasks. Participants gave high praise to this feature with comments such as "it's great to be taken directly to the page but not to have your results lost", "I like the way it narrows the focus and sort of guides a person to the info sought", "I didn't have to be concerned with performing a complex search that may return a null set-the results reflected my search string instantly". Two participants also commented that the feature was somewhat limited in use since relevant information may not appear in the title, or description. DISCUSSION The results strongly support that the RB++ interface was more effective and efficient in completing type 2 tasks than the baseline interface and that users felt more confident and satisfied with the RB++ in completing type 2 tasks. The higher effectiveness, efficiency and satisfaction gained in the RB++ resulted mainly from two aspects: the visual display of the statistical summary of the information items and the dynamic keyword searching capability in the results panel. The visualization bars helped the users understand relative proportions of items at a glance and use the posting numbers directly, which is much faster than literally counting. If we look at the BNF grammar, completion of type 2 tasks in RB++ only required participants to explore the collection (see first option of rule A1) without submitting queries to the database and then observing and counting returned results, which are necessary steps for the baseline interface to complete the same tasks (see rule B1). The dynamic search boxes allow users do further filtering based on certain criteria and give users feedback on the filtered results instantly and continuously, which not only encourages the users to use this function, but also improves their efficiency. Another 184 interface feature: displaying all the results on one screen might also help improve the efficiency and satisfaction, as several users mentioned. Several components were tightly coupled in the interface with displayed search results. The search boxes are tightly coupled with the results, which means that any input in the search boxes will invoke instant filtering on the results. The visual bars are tightly coupled with the results and as such they support two functions. One is that any operations on the visual bars such as mouse over and selection, invoke the instant filtering of the results. The other is that any update of the results also updates the summary statistics in the visualization on the bars. Coupling provides users more ways to interact with the system and make the interaction more natural and smooth (see rule A8, A9, A10), which suggests a different interaction style for finding information than traditional search interfaces which tend to require discrete, well-defined turn-taking between the user and system. Traditionally, when users get to the results page, all they can do is browse the results. If they want to refine the results, they have to go back to the search interface, type in the refined keywords, click the search, and browse the new results, which not only interrupts the normal results browsing interaction, but also loses the current result set. RB++ encourages users to get an initial manageable result set and then refine it using one interface window without the need to go back and forth. Instead of displaying a set of static results, RB++ offers an effective and efficient means for users to understand the results by displaying summary statistics bars which give both visual and numeric data (see rule A9), and to explore results by providing ways to dynamically and continuously filter (see rule A10). The result set can be as large as displaying the whole collection, or as small as only one item, which depends on the initial query on the collection. In the second phase study, most of the participants completed their search tasks without doing a second query on the initial interface. The study showed that participants could utilize the initial interface to get an initial result set by selecting relevant categories and then narrow down results and find relevant web pages by exploring the results set. Typing in keywords (or string patterns) in search boxes was found to be the most frequently used means to explore and filter result set. These features were highly appreciated by the participants as seen from their comments. RELATED WORK Many information access interfaces try to provide a starting point for users by presenting overviews of the collection [9]. Overviews can help users understand the whole collection and select or eliminate sources from consideration. Overviews can direct users to subcollections quickly, where they can explore the details. Usually two types of overviews are employed: category overview and graphic overview. The category approach of Yahoo is a good example for the category overview. The HiBrowse interface for viewing category labels hierarchically based on the facets is another example [14]. A more recent information access interface using the category overview by presenting faceted metadata is the Flamenco interface [21]. The last two interfaces not only present the category labels to the users but also inform the users of the number of documents under each category. However, these interfaces do not allow users to employ simple mouse moves to quickly explore the relationship between different categories (or facets). The Flamenco interface could do this as part of its browsing and searching efforts, but it requires many commitments from users such as clicking the category and waiting. The previous version of the relation browser [13] presented various categories and allowed users to explore the relations by mouse over operation, but the interface only allowed the users to mouse over the main category. The graphic overview is another type of overview, which usually employs various information visualization techniques. Lin [12] used the Kohenen map to visually present a topical overview of the collection. Each block on the map represents a subcollection with similar topics which are labeled by one or two salient words extracted from the subcollection. The adjacency of blocks indicates the topic similarity between subcollections. Wise, et al. [19] developed a three dimensional interface to visually present various topics. Zhang, et al. [25] exacted the key concepts from a collection and visually presented the concepts in a spring-embedded graph. Similar concepts were clustered together and usually represented as subtopics. The graphical overview is visually appealing, but the usability of this kind of interface has yet to be explored. 3-D interfaces are more problematic than 2-D interface in terms of ease of use and learnability. It seems that textual labels of category structure are more understandable than graphical representation. Some research has been conducted on how to present the retrieved results in context. Hearst [10] used clustering techniques to cluster retrieval results on the fly and presented different clusters with labeled words to the users to help them understand of the results. Chen, and Dumais [3] employed classification techniques to categorize retrieved results based on the existing category structure and displayed them in hierarchical categories. Zamir and Etzioni [22] developed an interface that used on-the-fly clustering of metasearch results. These interfaces cluster or categorize the retrieval results on the fly, so scaling is problematic. The RB++ categorizes the collection offline and uses a uniform category structure to present overviews of the collection and the retrieval results. Consequently, RB++ can be scaled up easily. However, because RB++ depends on the metadata to reside on the client side to achieve its dynamics, it also suffers a different kind of scalability limitation. To date, we have had good success and response with data sets with tens of thousands of records and a dozen or so facets, however data sets with millions of records and scores of facets are problematic. 3 . There also has been some work on fast location of specific information items. Sorting is a prevalent means to help users locate a specific item. However, users still need to visually go through a list of items. The Alphaslider [1] is a visual component to help users quickly locate a known string of items, but it's not very easy to use, especially for novice users. Besides, The Alphaslider can only locate the information items based on the first letter alphabetically. RB++ provides an easy and flexible way to locate the information items by typing in string patterns and the patterns can be matched anywhere in the information items. A similar technique is actually used in some applications such as the address box of Internet Explorer 3 Various RB++ examples are available at http://idl.ils.unc.edu/rave/examples.html 185 browser, but the patterns are limited to matching from the beginning of the query string. Dynamic query was a new type of interface [16] that inspired the original relation browser work. The interface visually displays the information items and provides the visual controlling components to explore the information items by tightly coupling search and visual display of results. RB++ uses this design concept, but instead of providing a visual interface, RB++ employs a more understandable (especially for topical overview) category structure for the information items. Moreover, the search box is a very effective and efficient component for the non categorized attributes of the items, while the visual controlling components such as sliders or check boxes can only be used for controlling categorical attributes of the items. Query preview [18], attribute explorer [17] and other interfaces [11], and [20] provide similar ways to explore the relationships between different facets of the classification. These interfaces worked for structured information such as that found in databases. Of course, all these types of interfaces depend on good underlying categorization of data. Our long-term goal is to make the interface work for unstructured textual information. The search boxes provided are a first step in this direction, although they are currently limited to search within the fields specified in the results display. LIMITATIONS OF RB++ AND ONGOING WORK One constraint in RB++ is the limited number of categories that can be displayed, which is affected by two factors. One factor is screen real estate. We can partially alleviate the issue of screen real estate by utilizing a Zoomable User Interface (ZUI) to display the categories. We have experimented with integrating the Jazz toolkit [2] into the interface and this provides aproximately a ten-fold increase in the number of facet values that can be supported within each facet, although at the expense of some of the mouseover dynamics since the mouse must now be used for zooming as well as normal hovering. Another factor is size of the memory to hold the client-side distribution counts data, the number of which increases exponentially with increased number of displayed categories. One way to solve the issue is to only calculate part of the distribution counts data, which hopefully are most frequently used during the user's interaction with the interface. Other approaches, such as employing novel data structures were also suggested by Doan, et al. [5]. However, all these solutions have to sacrifice the interactivity of the interface that depends on client-side metadata to support rapid mouse activities. For example, preloading partial distribution counts data for large numbers of categories make some distributional data and visualization unavailable when users try to re-partition the information space by mouse moves. At present, the best development path seems to be hierarchical partitioning of very large information spaces with multiple RB++ instances that require a new download for each of the cascading subsets. Another constraint of RB++ is the limited matching function of the search boxes. The interface currently matches input string patterns to the corresponding result fields on the lexical level. Matching in this level is sufficient in many cases such as matching with fields with numbers or short textual strings such as titles, but for the fields with more semantic bearing strings such as descriptions of web pages, a more sophisticated match function based on semantics might be needed--perhaps a kind of full-text engine in each text field, although the close coupling with the facet panel will then be in doubt. Currently, the interface provides a uniform category structure for both the entire collection and the retrieved results set. This is good for its consistency. However, for the retrieved result set, a more fine-grained category structure might be better for users to understand it and conduct string searches. Overall, the RB++ represents an example of a highly interactive user interface that offers improved performance and satisfaction for users of large websites and digital libraries. It can find application as the entry point for a large website or as a way to work with large results sets returned from search engines as long as the data is structured in advance. WORK ON AUTOMATIC CLUSTERING AND CLASSIFICATION Over the past few years we have created more than two dozen instances of data sets using RB and RB++, demonstrating its applicability as an interface to many different types of data. If the data resides in a database, it is possible to map the scheme to the underlying RB++ scheme (see [24] for details on the system architecture) and simply import the data automatically. For many WWW applications, this is not possible so we have been developing ways to automate facet discovery and webpage assignment to those facets. The basic approach is to crawl the website(s), create term-document representations and then use machine learning techniques to cluster the webpages and extract candidate labels, use human judgment to select the best labels, and then classify the webpages into those categories using the statistical model produced in the process. See Efron et al [6,7] for details of the techniques we have applied to date. To illustrate the current state of development, consider the EIA example used in the study reported here. The categories displayed on the EIA RB++ instance were originally created manually, which certainly did not scale well to many other large information collections such as various other government statistical web sites. Figure 7 is a screen shot of the RB++ instance for the Bureau of Labor Statistics web site that uses topic facets and webpage assignment that were automatically determined. About 13,000 HTML web pages were crawled from web site and a soft clustering method was then applied to those pages. For each webpage, the statistical model yielded a probability of belonging to every cluster. Thus, every page `belongs' to every cluster at some level of probability. The first topic column contains all the pages with highest probability of belonging to those facet values. The second column contains all the pages with the second highest probability values for those facet values. The third and fourth columns are the months and years of page update extracted from web pages themselves-facets that are know to be problematic but used for illustration since our primary emphasis was on topic discovery. The display of two topical columns reflects the underlying characteristic of soft clustering where items can appear in multiple clusters. However, our discussions with BLS staff and demonstrations to other potential users show that this two-column display is confusing. Therefore, only the first topical column was kept in a later version of the RB++ BLS instance (see figure 8). An 186 additional facet: geographical coverage, which is an important facet of government statistical web sites, was added in this version. The assignments were made using a rule based classification method to classify the web pages into categories of various geographical coverages. Figure 7. First version of RB instance for BLS web site Figure 8. Latest version of RB instance for BLS web site. In addition to the BLS instance, we have used these techniques to create RB++ instances for the FedStats website and are working to create instances for other federal statistical agencies. At present, the FedStats and BLS instances are installed on FedStats servers and are being tested by federal statistical agency personnel. We also hope to address the important facet of time coverage of data itself in future work. Overall, the RB++ user interface has evolved to a state where it can be easily applied to many kinds of well-structured data. The user testing reported here demonstrates the efficacy of the interface for search and browse tasks that would be very difficult to execute with SQL syntax or form fillin interfaces. Our current efforts are to develop techniques for automatically populating the RB++ database with unstructured data from the WWW. To date, these efforts have led to promising prototypes for several federal statistical websites. ACKNOWLEDGMENTS This work was supported by NSF Grant EIA 0131824. The authors wish to thank Paul Solomon and other anonymous reviewers for their valuable comments on the paper and Tim Shearer for designing the underlying database structure of the interface. We also thank Jonathan Elsas and Miles Efron for developing the clustering software and techniques. REFERENCES [1] Ahlberg, C., and Shneiderman, B. The alphaslider: a compact and rapid selector. In Proceedings of the SIGCHI conference on human factors in computing systems. Boston, Massachusetts. 1994. [2] Bederson, B., Meyer, J., and Good, L. Jazz:An Extensible zoomable user interface graphics toolkit in java. In ACM UIST2000, 171-180. [3] Chen, H, and Dumais, S. Bringing order to the web: Automatically categorizing searching results. In Proceedings of the SIGCHI conference on human factors in computing systems. The Hague, Amsterdam. 2000 [4] Dix, A., Finlay, J. Abowd, G. Beale, R, and Finley, J. Human-Computer Interaction (2 nd Ed.). Prentice Hall, Hillsdale, NJ, 1998 [5] Doan, K., Plaisant, C., Shneiderman, B., and Bruns, T. Interface and Data Architecture for Query Preview in Networked Information Systems. ACM Transactions on Information Systems, July 1999, Vol. 17, No. 3, 320-341. [6] Efron, M., Marchionini, G. and Zhang, J. Implications of the recursive representation problem for automatic concept identification in on-line governmental information. ASIST SIG-CR Workshop, Long Beach, CA, 2003. [7] Efron, M., Elsas, J., Marchionini, G., and Zhang J.. Machine learning for information architecture in a large governmental website. Joint Conference on Digital Libraries 2004 (Tuscon, AZ, June 7-11, 2004) [8] Greene, S., Marchionini, G., Plaisant, C., & Shneiderman, B. (2000). Previews and overviews in digital libraries: Designing surrogates to support visual information seeking. Journal of the American Society for Information Science, 51(4), 380-393. [9] Hearst, M. User interfaces and visualization. In Modern information retrieval. Ed. by Baeza-Yates, R., and Ribeiro-Neto, B. Chapter 10, ACM Press, New York, NY, 1999 257-324. [10] Hearst, M. and Pedersen, P. Reexamining the cluster hypothesis: Scatter/Gather on retrieval results, Proceedings of 19th Annual International ACM/SIGIR Conference, Zurich, 1996 [11] Lanning, T., Wittenburg, K., Heinrichs, M., Fyock, C., and Li, G. Multidimensional information visualization through sliding rods. AVI'02 Palermo, Italy. [12] Lin, X. Map displays for information retrieval. Journal of the American society for information science. 1997 48(1), 40-54. 187 [13] Marchionini, G., and Brunk, B. Toward a general relation browser: A GUI for information architects. Journal of Digital Information. Article No. 179, 2003-04-09 2003 4(1). http://jodi.ecs.soton.ac.uk/Articles/v04/i01/Marchionini/ [14] Pollitt A. S., Ellis G. P., and Smith M. P. HIBROWSE for Bibliographic Databases Journal of Information Science, 1994 20 (6), 413-426. [15] Reisner, P., Formal Grammar and human factor design of an interactive graphics system. IEEE Trans. on Software Engineering, 7(2), 229-240, 1981 [16] Shneiderman, B., Dynamic queries for visual information seeking, IEEE Software 11, 6 (1994), 70-77. [17] Spence, R, and Tweedie, L. The attribute explore: information synthesis via exploration. Interacting with Computers. 1998 11, 137-146. [18] Tanin, E., Lotem, A., Haddadin, I., Shneiderman, B., Plaisant, C., and Slaughter, L. Facilitating data exploration with query previews: a study of user performance and preference. Behaviour & information technology. 2000 19(6). 393-403. [19] Wise, J., Thomas, J., Pennock, K., Lantrip, D., Pottier, M. and Schur, A. Visualizing the non-visual: spatial analysis and interaction with information from text documents. In Proc. of the Information visualization Symposium 95, pages 51-58. IEEE Computer Society Press, 1995

searching;efficiency;user satisfaction;user study;information search;Information Storage and Retrieval;interaction patterns with interface;Interface design;visualization;Relation Browser++;browsing;Browse and Search Interface;RB++;interactive system;Faceted category structure;information browsing;effectiveness;search;dynamic query;facets;browse;Human Factor;visual display;Modeling User Interaction;satisfaction;User interface;category overview;user interface

88

Event Threading within News Topics

With the overwhelming volume of online news available today, there is an increasing need for automatic techniques to analyze and present news to the user in a meaningful and efficient manner. Previous research focused only on organizing news stories by their topics into a flat hierarchy. We believe viewing a news topic as a flat collection of stories is too restrictive and inefficient for a user to understand the topic quickly. In this work, we attempt to capture the rich structure of events and their dependencies in a news topic through our event models. We call the process of recognizing events and their dependencies event threading. We believe our perspective of modeling the structure of a topic is more effective in capturing its semantics than a flat list of on-topic stories. We formally define the novel problem, suggest evaluation metrics and present a few techniques for solving the problem. Besides the standard word based features, our approaches take into account novel features such as temporal locality of stories for event recognition and time-ordering for capturing dependencies. Our experiments on a manually labeled data sets show that our models effec-tively identify the events and capture dependencies among them.

INTRODUCTION News forms a major portion of information disseminated in the world everyday. Common people and news analysts alike are very interested in keeping abreast of new things that happen in the news, but it is becoming very difficult to cope with the huge volumes 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'04, November 813, 2004, Washington,DC,USA. Copyright 2004 ACM 1-58113-874-1/04/0011 ... $ 5.00. of information that arrives each day. Hence there is an increasing need for automatic techniques to organize news stories in a way that helps users interpret and analyze them quickly. This problem is addressed by a research program called Topic Detection and Tracking (TDT) [3] that runs an open annual competition on standardized tasks of news organization. One of the shortcomings of current TDT evaluation is its view of news topics as flat collection of stories. For example, the detection task of TDT is to arrange a collection of news stories into clusters of topics. However, a topic in news is more than a mere collection of stories: it is characterized by a definite structure of inter-related events. This is indeed recognized by TDT which defines a topic as `a set of news stories that are strongly related by some seminal real-world event' where an event is defined as `something that happens at a specific time and location' [3]. For example, when a bomb explodes in a building, that is the seminal event that triggers the topic. Other events in the topic may include the rescue attempts, the search for perpetrators, arrests and trials and so on. We see that there is a pattern of dependencies between pairs of events in the topic. In the above example, the event of rescue attempts is `influenced' by the event of bombing and so is the event of search for perpetrators. In this work we investigate methods for modeling the structure of a topic in terms of its events. By structure, we mean not only identifying the events that make up a topic, but also establishing dependencies--generally causal--among them. We call the process of recognizing events and identifying dependencies among them event threading, an analogy to email threading that shows connections between related email messages. We refer to the resulting interconnected structure of events as the event model of the topic. Although this paper focuses on threading events within an existing news topic, we expect that such event based dependency structure more accurately reflects the structure of news than strictly bounded topics do. From a user's perspective, we believe that our view of a news topic as a set of interconnected events helps him/her get a quick overview of the topic and also allows him/her navigate through the topic faster. The rest of the paper is organized as follows. In section 2, we discuss related work. In section 3, we define the problem and use an example to illustrate threading of events within a news topic. In section 4, we describe how we built the corpus for our problem. Section 5 presents our evaluation techniques while section 6 describes the techniques we use for modeling event structure. In section 7 we present our experiments and results. Section 8 concludes the paper with a few observations on our results and comments on future work. 446 RELATED WORK The process of threading events together is related to threading of electronic mail only by name for the most part. Email usually incorporates a strong structure of referenced messages and consistently formatted subject headings--though information retrieval techniques are useful when the structure breaks down [7]. Email threading captures reference dependencies between messages and does not attempt to reflect any underlying real-world structure of the matter under discussion. Another area of research that looks at the structure within a topic is hierarchical text classification of topics [9, 6]. The hierarchy within a topic does impose a structure on the topic, but we do not know of an effort to explore the extent to which that structure reflects the underlying event relationships. Barzilay and Lee [5] proposed a content structure modeling technique where topics within text are learnt using unsupervised methods, and a linear order of these topics is modeled using hidden Markov models. Our work differs from theirs in that we do not constrain the dependency to be linear. Also their algorithms are tuned to work on specific genres of topics such as earthquakes, accidents, etc., while we expect our algorithms to generalize over any topic. In TDT, researchers have traditionally considered topics as flat-clusters [1]. However, in TDT-2003, a hierarchical structure of topic detection has been proposed and [2] made useful attempts to adopt the new structure. However this structure still did not ex-plicitly model any dependencies between events. In a work closest to ours, Makkonen [8] suggested modeling news topics in terms of its evolving events. However, the paper stopped short of proposing any models to the problem. Other related work that dealt with analysis within a news topic includes temporal summarization of news topics [4]. PROBLEM DEFINITION AND NOTATION In this work, we have adhered to the definition of event and topic as defined in TDT. We present some definitions (in italics) and our interpretations (regular-faced) below for clarity. 1. Story: A story is a news article delivering some information to users. In TDT, a story is assumed to refer to only a single topic. In this work, we also assume that each story discusses a single event. In other words, a story is the smallest atomic unit in the hierarchy (topic event story). Clearly, both the assumptions are not necessarily true in reality, but we accept them for simplicity in modeling. 2. Event: An event is something that happens at some specific time and place [10]. In our work, we represent an event by a set of stories that discuss it. Following the assumption of atomicity of a story, this means that any set of distinct events can be represented by a set of non-overlapping clusters of news stories. 3. Topic: A set of news stories strongly connected by a seminal event. We expand on this definition and interpret a topic as a series of related events. Thus a topic can be represented by clusters of stories each representing an event and a set of (directed or undirected) edges between pairs of these clusters representing the dependencies between these events. We will describe this representation of a topic in more detail in the next section. 4. Topic detection and tracking (TDT) :Topic detection detects clusters of stories that discuss the same topic; Topic tracking detects stories that discuss a previously known topic [3]. Thus TDT concerns itself mainly with clustering stories into topics that discuss them. 5. Event threading: Event threading detects events within in a topic, and also captures the dependencies among the events. Thus the main difference between event threading and TDT is that we focus our modeling effort on microscopic events rather than larger topics. Additionally event threading models the relatedness or dependencies between pairs of events in a topic while TDT models topics as unrelated clusters of stories. We first define our problem and representation of our model formally and then illustrate with the help of an example. We are given a set of news stories on a given topic and their time of publication. We define a set of events with the following constraints: (1) (2) (3) While the first constraint says that each event is an element in the power set of S, the second constraint ensures that each story can belong to at most one event. The last constraint tells us that every story belongs to one of the events in . In fact this allows us to define a mapping function from stories to events as follows: iff (4) Further, we also define a set of directed edges which denote dependencies between events. It is important to explain what we mean by this directional dependency: While the existence of an edge itself represents relatedness of two events, the direction could imply causality or temporal-ordering. By causal dependency we mean that the occurrence of event B is related to and is a consequence of the occurrence of event A. By temporal ordering , we mean that event B happened after event A and is related to A but is not necessarily a consequence of A. For example, consider the following two events: `plane crash' (event A) and `subse-quent investigations' (event B) in a topic on a plane crash incident. Clearly, the investigations are a result of the crash. Hence an arrow from A to B falls under the category of causal dependency. Now consider the pair of events `Pope arrives in Cuba'(event A) and `Pope meets Castro'(event B) in a topic that discusses Pope's visit to Cuba. Now events A and B are closely related through their association with the Pope and Cuba but event B is not necessarily a consequence of the occurrence of event A. An arrow in such scenario captures what we call time ordering. In this work, we do not make an attempt to distinguish between these two kinds of dependencies and our models treats them as identical. A simpler (and hence less controversial) choice would be to ignore direction in the dependencies altogether and consider only undirected edges. This choice definitely makes sense as a first step but we chose the former since we believe directional edges make more sense to the user as they provide a more illustrative flow-chart perspective to the topic. To make the idea of event threading more concrete, consider the example of TDT3 topic 30005, titled `Osama bin Laden's Indict-ment' (in the 1998 news). This topic has 23 stories which form 5 events. An event model of this topic can be represented as in figure 1. Each box in the figure indicates an event in the topic of Osama's indictment. The occurrence of event 2, namely `Trial and Indictment of Osama' is dependent on the event of `evidence gathered by CIA', i.e., event 1. Similarly, event 2 influences the occurrences of events 3, 4 and 5, namely `Threats from Militants', `Reactions 447 from Muslim World' and `announcement of reward'. Thus all the dependencies in the example are causal. Extending our notation further, we call an event A a parent of B and B the child of A, if . We define an event model to be a tuple of the set of events and set of dependencies . Trial and (5) (3) (4) CIA announces reward Muslim world Reactions from Islamic militants Threats from (2) (1) Osama Indictment of CIA gathered by Evidence Figure 1: An event model of TDT topic `Osama bin Laden's indictment'. Event threading is strongly related to topic detection and tracking , but also different from it significantly. It goes beyond topics, and models the relationships between events. Thus, event threading can be considered as a further extension of topic detection and tracking and is more challenging due to at least the following difficulties . 1. The number of events is unknown. 2. The granularity of events is hard to define. 3. The dependencies among events are hard to model. 4. Since it is a brand new research area, no standard evaluation metrics and benchmark data is available. In the next few sections, we will describe our attempts to tackle these problems. LABELED DATA We picked 28 topics from the TDT2 corpus and 25 topics from the TDT3 corpus. The criterion we used for selecting a topic is that it should contain at least 15 on-topic stories from CNN headline news. If the topic contained more than 30 CNN stories, we picked only the first 30 stories to keep the topic short enough for annota-tors . The reason for choosing only CNN as the source is that the stories from this source tend to be short and precise and do not tend to digress or drift too far away from the central theme. We believe modeling such stories would be a useful first step before dealing with more complex data sets. We hired an annotator to create truth data. Annotation includes defining the event membership for each story and also the dependencies . We supervised the annotator on a set of three topics that we did our own annotations on and then asked her to annotate the 28 topics from TDT2 and 25 topics from TDT3. In identifying events in a topic, the annotator was asked to broadly follow the TDT definition of an event, i.e., `something that happens at a specific time and location'. The annotator was encouraged to merge two events A and B into a single event C if any of the stories discusses both A and B. This is to satisfy our assumption that each story corresponds to a unique event. The annotator was also encouraged to avoid singleton events, events that contain a single news story, if possible. We realized from our own experience that people differ in their perception of an event especially when the number of stories in that event is small. As part of the guidelines, we instructed the annotator to assign titles to all the events in each topic. We believe that this would help make her understanding of the events more concrete. We however, do not use or model these titles in our algorithms. In defining dependencies between events, we imposed no restrictions on the graph structure. Each event could have single, multiple or no parents. Further, the graph could have cycles or orphan-nodes . The annotator was however instructed to assign a dependency from event A to event B if and only if the occurrence of B is `either causally influenced by A or is closely related to A and follows A in time'. From the annotated topics, we created a training set of 26 topics and a test set of 27 topics by merging the 28 topics from TDT2 and 25 from TDT3 and splitting them randomly. Table 1 shows that the training and test sets have fairly similar statistics. Feature Training set Test set Num. topics 26 27 Avg. Num. Stories/Topic 28.69 26.74 Avg. Doc. Len. 64.60 64.04 Avg. Num. Stories/Event 5.65 6.22 Avg. Num. Events/Topic 5.07 4.29 Avg. Num. Dependencies/Topic 3.07 2.92 Avg. Num. Dependencies/Event 0.61 0.68 Avg. Num. Days/Topic 30.65 34.48 Table 1: Statistics of annotated data EVALUATION A system can generate some event model using certain algorithms, which is usually different from the truth model (we assume the annotator did not make any mistake ). Comparing a system event model with the true model requires comparing the entire event models including their dependency structure. And different event granularities may bring huge discrepancy between and . This is certainly non-trivial as even testing whether two graphs are isomorphic has no known polynomial time solution. Hence instead of comparing the actual structure we examine a pair of stories at a time and verify if the system and true labels agree on their event-memberships and dependencies . Specifically, we compare two kinds of story pairs: Cluster pairs ( ) : These are the complete set of un-ordered pairs of stories and that fall within the same event given a model . Formally, (5) where is the function in that maps stories to events as defined in equation 4. Dependency pairs ( ) : These are the set of all ordered pairs of stories such that there is a dependency from the event of to the event of in the model . (6) Note the story pair is ordered here, so is not equivalent to . In our evaluation, a correct pair with wrong 448 (B-&gt;D) Cluster pairs (A,C) Dependency pairs (A-&gt;B) (C-&gt;B) (B-&gt;D) D,E D,E (D,E) (D,E) (A-&gt;C) (A-&gt;E) (B-&gt;C) (B-&gt;E) (B-&gt;E) Cluster precision: 1/2 Cluster Recall: 1/2 Dependency Recall: 2/6 Dependency Precision: 2/4 (A-&gt;D) True event model System event model A,B C A,C B Cluster pairs (A,B) Dependency pairs Figure 2: Evaluation measures direction will be considered a mistake. As we mentioned earlier in section 3, ignoring the direction may make the problem simpler, but we will lose the expressiveness of our representation . Given these two sets of story pairs corresponding to the true event model and the system event model , we define recall and precision for each category as follows. Cluster Precision (CP) : It is the probability that two randomly selected stories and are in the same true-event given that they are in the same system event. (7) where is the story-event mapping function corresponding to the model . Cluster Recall(CR) : It is the probability that two randomly selected stories and are in the same system-event given that they are in the same true event. (8) Dependency Precision(DP) : It is the probability that there is a dependency between the events of two randomly selected stories and in the true model given that they have a dependency in the system model . Note that the direction of dependency is important in comparison. (9) Dependency Recall(DR) : It is the probability that there is a dependency between the events of two randomly selected stories and in the system model given that they have a dependency in the true model . Again, the direction of dependency is taken into consideration. (10) The measures are illustrated by an example in figure 2. We also combine these measures using the well known F1-measure commonly used in text classification and other research areas as shown below. (11) where and are the cluster and dependency F1-measures respectively and is the Joint F1-measure ( ) that we use to measure the overall performance. TECHNIQUES The task of event modeling can be split into two parts: clustering the stories into unique events in the topic and constructing dependencies among them. In the following subsections, we describe techniques we developed for each of these sub-tasks. 6.1 Clustering Each topic is composed of multiple events, so stories must be clustered into events before we can model the dependencies among them. For simplicity, all stories in the same topic are assumed to be available at one time, rather than coming in a text stream. This task is similar to traditional clustering but features other than word distributions may also be critical in our application. In many text clustering systems, the similarity between two stories is the inner product of their tf-idf vectors, hence we use it as one of our features. Stories in the same event tend to follow temporal locality, so the time stamp of each story can be a useful feature. Additionally, named-entities such as person and location names are another obvious feature when forming events. Stories in the same event tend to be related to the same person(s) and locations(s). In this subsection, we present an agglomerative clustering algorithm that combines all these features. In our experiments, however , we study the effect of each feature on the performance sepa-rately using modified versions of this algorithm. 6.1.1 Agglomerative clustering with time decay (ACDT) We initialize our events to singleton events (clusters), i.e., each cluster contains exactly one story. So the similarity between two events, to start with, is exactly the similarity between the corresponding stories. The similarity Ѵ between two stories and is given by the following formula: Ѵ ״ (12) Here , , are the weights on different features. In this work, we determined them empirically, but in the future, one can consider more sophisticated learning techniques to determine them. ״ is the cosine similarity of the term vectors. is 1 if there is some location that appears in both stories, otherwise it is 0. is similarly defined for person name. We use time decay when calculating similarity of story pairs, i.e., the larger time difference between two stories, the smaller their similarities. The time period of each topic differs a lot, from a few days to a few months. So we normalize the time difference using the whole duration of that topic. The time decay adjusted similarity 449 Ѵ is given by Ѵ Ѵ (13) where and are the time stamps for story 1 and 2 respectively. T is the time difference between the earliest and the latest story in the given topic. is the time decay factor. In each iteration, we find the most similar event pair and merge them. We have three different ways to compute the similarity between two events and : Average link: In this case the similarity is the average of the similarities of all pairs of stories between and as shown below: Ѵ Ѵ (14) Complete link: The similarity between two events is given by the smallest of the pair-wise similarities. Ѵ Ѵ (15) Single link: Here the similarity is given by the best similarity between all pairs of stories. Ѵ Ѵ (16) This process continues until the maximum similarity falls below the threshold or the number of clusters is smaller than a given number . 6.2 Dependency modeling Capturing dependencies is an extremely hard problem because it may require a `deeper understanding' of the events in question. A human annotator decides on dependencies not just based on the information in the events but also based on his/her vast repertoire of domain-knowledge and general understanding of how things operate in the world. For example, in Figure 1 a human knows `Trial and indictment of Osama' is influenced by `Evidence gathered by CIA' because he/she understands the process of law in general. We believe a robust model should incorporate such domain knowledge in capturing dependencies, but in this work, as a first step, we will rely on surface-features such as time-ordering of news stories and word distributions to model them. Our experiments in later sections demonstrate that such features are indeed useful in capturing dependencies to a large extent. In this subsection, we describe the models we considered for capturing dependencies. In the rest of the discussion in this subsection, we assume that we are already given the mapping and we focus only on modeling the edges . First we define a couple of features that the following models will employ. First we define a 1-1 time-ordering function that sorts stories in ascending order by their time of publication. Now, the event-time-ordering function is defined as follows. ش ش (17) In other words, time-orders events based on the time-ordering of their respective first stories. We will also use average cosine similarity between two events as a feature and it is defined as follows. Ѵ ״ (18) 6.2.1 Complete-Link model In this model, we assume that there are dependencies between all pairs of events. The direction of dependency is determined by the time-ordering of the first stories in the respective events. Formally, the system edges are defined as follows. (19) where is the event-time-ordering function. In other words, the dependency edge is directed from event to event , if the first story in event is earlier than the first story in event . We point out that this is not to be confused with the complete-link algorithm in clustering. Although we use the same names, it will be clear from the context which one we refer to. 6.2.2 Simple Thresholding This model is an extension of the complete link model with an additional constraint that there is a dependency between any two events and only if the average cosine similarity between event and event is greater than a threshold . Formally, Ѵ (20) 6.2.3 Nearest Parent Model In this model, we assume that each event can have at most one parent. We define the set of dependencies as follows. Ѵ (21) Thus, for each event , the nearest parent model considers only the event preceding it as defined by as a potential candidate. The candidate is assigned as the parent only if the average similarity exceeds a pre-defined threshold . 6.2.4 Best Similarity Model This model also assumes that each event can have at most one parent. An event is assigned a parent if and only if is the most similar earlier event to and the similarity exceeds a threshold . Mathematically, this can be expressed as: Ѵ Ѵ (22) 6.2.5 Maximum Spanning Tree model In this model, we first build a maximum spanning tree (MST) using a greedy algorithm on the following fully connected weighted, undirected graph whose vertices are the events and whose edges are defined as follows: Ѵ (23) Let be the set of edges in the maximum spanning tree of . Now our directed dependency edges are defined as follows. Ѵ (24) 450 Thus in this model, we assign dependencies between the most similar events in the topic. EXPERIMENTS Our experiments consists of three parts. First we modeled only the event clustering part (defining the mapping function ) using clustering algorithms described in section 6.1. Then we modeled only the dependencies by providing to the system the true clusters and running only the dependency algorithms of section 6.2. Finally, we experimented with combinations of clustering and dependency algorithms to produce the complete event model. This way of experimentation allows us to compare the performance of our algorithms in isolation and in association with other components. The following subsections present the three parts of our experimentation . 7.1 Clustering We have tried several variations of the algorithm to study the effects of various features on the clustering performance. All the parameters are learned by tuning on the training set. We also tested the algorithms on the test set with parameters fixed at their optimal values learned from training. We used agglomerative clus-Model best T CP CR CF P-value cos+1-lnk 0.15 0.41 0.56 0.43 cos+all -lnk 0.00 0.40 0.62 0.45 cos+Loc+avg -lnk 0.07 0.37 0.74 0.45 cos+Per+avg -lnk 0.07 0.39 0.70 0.46 cos+TD+avg -lnk 0.04 0.45 0.70 0.53 2.9e-4* cos+N(T)+avg-lnk 0 .41 0.62 0.48 7.5e-2 cos+N(T)+T+avg-lnk 0.03 0.42 0.62 0.49 2.4e-2* cos+TD+N(T)+avg-lnk 0 .44 0.66 0.52 7.0e-3* cos+TD+N(T)+T+avg-lnk 0.03 0.47 0.64 0.53 1.1e-3* Baseline(cos+avg-lnk) 0.05 0.39 0.67 0.46 Table 2: Comparison of agglomerative clustering algorithms (training set) tering based on only cosine similarity as our clustering baseline. The results on the training and test sets are in Table 2 and 3 respectively . We use the Cluster F1-measure (CF) averaged over all topics as our evaluation criterion. Model CP CR CF P-value cos+1-lnk 0.43 0.49 0.39 cos+all -lnk 0.43 0.62 0.47 cos+Loc+avg -lnk 0.37 0.73 0.45 cos+Per+avg -lnk 0.44 0.62 0.45 cos+TD+avg -lnk 0.48 0.70 0.54 0.014* cos+N(T)+avg-lnk 0.41 0.71 0.51 0.31 cos+N(T)+T+avg-lnk 0.43 0.69* 0.52 0.14 cos+TD+N(T)+avg-lnk 0.43 0.76 0.54 0.025* cos+TD+N(T)+T+avg-lnk 0.47 0.69 0.54 0.0095* Baseline(cos+avg-lnk) 0.44 0.67 0.50 Table 3: Comparison of agglomerative clustering algorithms (test set) P-value marked with a means that it is a statistically significant improvement over the baseline (95% confidence level, one tailed T-test). The methods shown in table 2 and 3 are: Baseline: tf-idf vector weight, cosine similarity, average link in clustering. In equation 12, , . And in equation 13. This F-value is the maximum obtained by tuning the threshold. cos+1-lnk: Single link comparison (see equation 16) is used where similarity of two clusters is the maximum of all story pairs, other configurations are the same as the baseline run. cos+all-lnk: Complete link algorithm of equation 15 is used. Similar to single link but it takes the minimum similarity of all story pairs. cos+Loc+avg-lnk: Location names are used when calculating similarity. in equation 12. All algorithms starting from this one use average link (equation 14), since single link and complete link do not show any improvement of performance. cos+Per+avg-lnk: in equation 12, i.e., we put some weight on person names in the similarity. cos+TD+avg-lnk: Time Decay coefficient in equation 13, which means the similarity between two stories will be decayed to if they are at different ends of the topic. cos+N(T)+avg-lnk: Use the number of true events to control the agglomerative clustering algorithm. When the number of clusters is fewer than that of truth events, stop merging clusters. cos+N(T)+T+avg-lnk: similar to N(T) but also stop agglomeration if the maximal similarity is below the threshold . cos+TD:+N(T)+avg-lnk: similar to N(T) but the similarities are decayed, in equation 13. cos+TD+N(T)+T+avg-lnk: similar to TD+N(Truth) but calculation halts when the maximal similarity is smaller than the threshold . Our experiments demonstrate that single link and complete link similarities perform worse than average link, which is reasonable since average link is less sensitive to one or two story pairs. We had expected locations and person names to improve the result, but it is not the case. Analysis of topics shows that many on-topic stories share the same locations or persons irrespective of the event they belong to, so these features may be more useful in identifying topics rather than events. Time decay is successful because events are temporally localized, i.e., stories discussing the same event tend to be adjacent to each other in terms of time. Also we noticed that providing the number of true events improves the performance since it guides the clustering algorithm to get correct granularity. However, for most applications, it is not available. We used it only as a "cheat" experiment for comparison with other algorithms. On the whole, time decay proved to the most powerful feature besides cosine similarity on both training and test sets. 7.2 Dependencies In this subsection, our goal is to model only dependencies. We use the true mapping function and by implication the true events . We build our dependency structure using all the five models described in section 6.2. We first train our models on the 26 training topics. Training involves learning the best threshold for each of the models. We then test the performances of all the trained models on the 27 test topics. We evaluate our performance 451 using the average values of Dependency Precision (DP), Dependency Recall (DR) and Dependency F-measure (DF). We consider the complete-link model to be our baseline since for each event, it trivially considers all earlier events to be parents. Table 4 lists the results on the training set. We see that while all the algorithms except MST outperform the baseline complete-link algorithm , the nearest Parent algorithm is statistically significant from the baseline in terms of its DF-value using a one-tailed paired T-test at 95% confidence level. Model best DP DR DF P-value Nearest Parent 0.025 0.55 0.62 0.56 0.04* Best Similarity 0.02 0.51 0.62 0.53 0.24 MST 0.0 0.46 0.58 0.48 Simple Thresh. 0.045 0.45 0.76 0.52 0.14 Complete-link 0 .36 0.93 0.48 Table 4: Results on the training set: Best is the optimal value of the threshold . * indicates the corresponding model is statistically significant compared to the baseline using a one-tailed, paired T-test at 95% confidence level. In table 5 we present the comparison of the models on the test set. Here, we do not use any tuning but set the threshold to the corresponding optimal values learned from the training set. The results throw some surprises: The nearest parent model, which was significantly better than the baseline on training set, turns out to be worse than the baseline on the test set. However all the other models are better than the baseline including the best similarity which is statistically significant. Notice that all the models that perform better than the baseline in terms of DF, actually sacrifice their recall performance compared to the baseline, but improve on their precision substantially thereby improving their performance on the DF-measure. We notice that both simple-thresholding and best similarity are better than the baseline on both training and test sets although the improvement is not significant. On the whole, we observe that the surface-level features we used capture the dependencies to a reasonable level achieving a best value of 0.72 DF on the test set. Although there is a lot of room for improvement, we believe this is a good first step. Model DP DR DF P-value Nearest Parent 0.61 0.60 0.60 Best Similarity 0.71 0.74 0.72 0.04* MST 0.70 0.68 0.69 0.22 Simple Thresh. 0.57 0.75 0.64 0.24 Baseline (Complete-link) 0.50 0.94 0.63 Table 5: Results on the test set 7.3 Combining Clustering and Dependencies Now that we have studied the clustering and dependency algorithms in isolation, we combine the best performing algorithms and build the entire event model. Since none of the dependency algorithms has been shown to be consistently and significantly better than the others, we use all of them in our experimentation. From the clustering techniques, we choose the best performing Cos+TD. As a baseline, we use a combination of the baselines in each components , i.e., cos for clustering and complete-link for dependencies. Note that we need to retrain all the algorithms on the training set because our objective function to optimize is now JF, the joint F-measure. For each algorithm, we need to optimize both the clustering threshold and the dependency threshold. We did this empirically on the training set and the optimal values are listed in table 6. The results on the training set, also presented in table 6, indicate that cos+TD+Simple-Thresholding is significantly better than the baseline in terms of the joint F-value JF, using a one-tailed paired T-test at 95% confidence level. On the whole, we notice that while the clustering performance is comparable to the experiments in section 7.1, the overall performance is undermined by the low dependency performance. Unlike our experiments in section 7.2 where we had provided the true clusters to the system, in this case, the system has to deal with deterioration in the cluster quality. Hence the performance of the dependency algorithms has suffered substantially thereby lowering the overall performance. The results on the test set present a very similar story as shown in table 7. We also notice a fair amount of consistency in the performance of the combination algorithms. cos+TD+Simple-Thresholding outperforms the baseline significantly. The test set results also point to the fact that the clustering component remains a bottleneck in achieving an overall good performance. DISCUSSION AND CONCLUSIONS In this paper, we have presented a new perspective of modeling news topics. Contrary to the TDT view of topics as flat collection of news stories, we view a news topic as a relational structure of events interconnected by dependencies. In this paper, we also proposed a few approaches for both clustering stories into events and constructing dependencies among them. We developed a time-decay based clustering approach that takes advantage of temporal-localization of news stories on the same event and showed that it performs significantly better than the baseline approach based on cosine similarity. Our experiments also show that we can do fairly well on dependencies using only surface-features such as cosine-similarity and time-stamps of news stories as long as true events are provided to the system. However, the performance deteriorates rapidly if the system has to discover the events by itself. Despite that discouraging result, we have shown that our combined algorithms perform significantly better than the baselines. Our results indicate modeling dependencies can be a very hard problem especially when the clustering performance is below ideal level. Errors in clustering have a magnifying effect on errors in dependencies as we have seen in our experiments. Hence, we should focus not only on improving dependencies but also on clustering at the same time. As part of our future work, we plan to investigate further into the data and discover new features that influence clustering as well as dependencies. And for modeling dependencies, a probabilistic framework should be a better choice since there is no definite answer of yes/no for the causal relations among some events. We also hope to devise an iterative algorithm which can improve clustering and dependency performance alternately as suggested by one of the reviewers. We also hope to expand our labeled corpus further to include more diverse news sources and larger and more complex event structures. Acknowledgments We would like to thank the three anonymous reviewers for their valuable comments. This work was supported in part by the Center 452 Model Cluster T Dep. T CP CR CF DP DR DF JF P-value cos+TD+Nearest-Parent 0.055 0.02 0.51 0.53 0.49 0.21 0.19 0.19 0.27 cos+TD+Best -Similarity 0.04 0.02 0.45 0.70 0.53 0.21 0.33 0.23 0.32 cos+TD+MST 0.04 0.00 0.45 0.70 0.53 0.22 0.35 0.25 0.33 cos+TD+Simple -Thresholding 0.065 0.02 0.56 0.47 0.48 0.23 0.61 0.32 0.38 0.0004* Baseline (cos+Complete-link) 0.10 0 .58 0.31 0.38 0.20 0.67 0.30 0.33 Table 6: Combined results on the training set Model CP CR CF DP DR DF JF P-value cos+TD+Nearest Parent 0.57 0.50 0.50 0.27 0.19 0.21 0.30 cos+TD+Best Similarity 0.48 0.70 0.54 0.31 0.27 0.26 0.35 cos+TD+MST 0.48 0.70 0.54 0.31 0.30 0.28 0.37 cos+TD+Simple Thresholding 0.60 0.39 0.44 0.32 0.66 0.42 0.43 0.0081* Baseline (cos+Complete-link) 0.66 0.27 0.36 0.30 0.72 0.43 0.39 Table 7: Combined results on the test set 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 authors' and do not necessarily reflect those of the sponsor. REFERENCES [1] J. Allan, J. Carbonell, G. Doddington, J. Yamron, and Y. Yang. Topic detection and tracking pilot study: Final report. In Proceedings of the DARPA Broadcast News Transcription and Understanding Workshop, pages 194218, 1998. [2] J. Allan, A. Feng, and A. Bolivar. Flexible intrinsic evaluation of hierarchical clustering for tdt. volume In the Proc. of the ACM Twelfth International Conference on Information and Knowledge Management, pages 263270, Nov 2003. [3] James Allan, editor. Topic Detection and Tracking:Event based Information Organization. Kluwer Academic Publishers, 2000. [4] James Allan, Rahul Gupta, and Vikas Khandelwal. Temporal summaries of new topics. In Proceedings of the 24th annual international ACM SIGIR conference on Research and development in information retrieval, pages 1018. ACM Press, 2001. [5] Regina Barzilay and Lillian Lee. Catching the drift: Probabilistic content models, with applications to generation and summarization. In Proceedings of Human Language Technology Conference and North American Chapter of the Association for Computational Linguistics(HLT-NAACL), pages 113120, 2004. [6] D. Lawrie and W. B. Croft. Discovering and comparing topic hierarchies. In Proceedings of RIAO 2000 Conference, pages 314330, 1999. [7] David D. Lewis and Kimberly A. Knowles. Threading electronic mail: a preliminary study. Inf. Process. Manage., 33(2):209217, 1997. [8] Juha Makkonen. Investigations on event evolution in tdt. In Proceedings of HLT-NAACL 2003 Student Workshop, pages 4348, 2004. [9] Aixin Sun and Ee-Peng Lim. Hierarchical text classification and evaluation. In Proceedings of the 2001 IEEE International Conference on Data Mining, pages 521528. IEEE Computer Society, 2001. [10] Yiming Yang, Jaime Carbonell, Ralf Brown, Thomas Pierce, Brian T. Archibald, and Xin Liu. Learning approaches for detecting and tracking news events. In IEEE Intelligent Systems Special Issue on Applications of Intelligent Information Retrieval, volume 14 (4), pages 3243, 1999. 453

Complete-Link model;Event;Intelligent Information Retrieval;Event Threading;threading;meaningful and efficient analysis and presentation of news;Information browsing and organization;Nearest Parent Model;information searching;Dependency modeling;Agglomerative clustering with time decay;dependency;News topic modeling;Topic detection and tracking;clustering;temporal localization of news stories