{ "paper_id": "I13-1014", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T07:14:33.871484Z" }, "title": "Detecting Spammers in Community Question Answering", "authors": [ { "first": "Zhuoye", "middle": [], "last": "Ding", "suffix": "", "affiliation": { "laboratory": "", "institution": "Fudan University", "location": { "addrLine": "12110240006,zhouyaqian,qz" } }, "email": "" }, { "first": "Yeyun", "middle": [], "last": "Gong", "suffix": "", "affiliation": { "laboratory": "", "institution": "Fudan University", "location": { "addrLine": "12110240006,zhouyaqian,qz" } }, "email": "" }, { "first": "Yaqian", "middle": [], "last": "Zhou", "suffix": "", "affiliation": { "laboratory": "", "institution": "Fudan University", "location": { "addrLine": "12110240006,zhouyaqian,qz" } }, "email": "" }, { "first": "Qi", "middle": [], "last": "Zhang", "suffix": "", "affiliation": { "laboratory": "", "institution": "Fudan University", "location": { "addrLine": "12110240006,zhouyaqian,qz" } }, "email": "" }, { "first": "Xuanjing", "middle": [], "last": "Huang", "suffix": "", "affiliation": { "laboratory": "", "institution": "Fudan University", "location": { "addrLine": "12110240006,zhouyaqian,qz" } }, "email": "xjhuang@fudan.edu.cn" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "As the popularity of Community Question Answering(CQA) increases, spamming activities also picked up in numbers and variety. On CQA sites, spammers often pretend to ask questions, and select answers which were published by their partners or themselves as the best answers. These fake best answers cannot be easily detected by neither existing methods nor common users. In this paper, we address the issue of detecting spammers on CQA sites. We formulate the task as an optimization problem. Social information is incorporated by adding graph regularization constraints to the text-based predictor. To evaluate the proposed approach, we crawled a data set from a CQA portal. Experimental results demonstrate that the proposed method can achieve better performance than some state-of-the-art methods.", "pdf_parse": { "paper_id": "I13-1014", "_pdf_hash": "", "abstract": [ { "text": "As the popularity of Community Question Answering(CQA) increases, spamming activities also picked up in numbers and variety. On CQA sites, spammers often pretend to ask questions, and select answers which were published by their partners or themselves as the best answers. These fake best answers cannot be easily detected by neither existing methods nor common users. In this paper, we address the issue of detecting spammers on CQA sites. We formulate the task as an optimization problem. Social information is incorporated by adding graph regularization constraints to the text-based predictor. To evaluate the proposed approach, we crawled a data set from a CQA portal. Experimental results demonstrate that the proposed method can achieve better performance than some state-of-the-art methods.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Due to the massive growth of Web 2.0 technologies, user-generated content has become a primary source of various types of content. Community Question Answering (CQA) services have also attracted continuously growing interest. They allow users to submit questions and answer questions asked by other users. A huge number of users contributed enormous questions and answers on popular CQA sites such as Yahoo! Answers 1 , Baidu Zhidao 2 , Facebook Questions 3 , and so on. According to a statistic from Yahoo, Yahoo! Answers receives more than 0.82 million questions and answers per day 4 .", "cite_spans": [ { "start": 585, "end": 586, "text": "4", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "On CQA sites, users are primary contributors of content. The volunteer-driven mechanism brings many positive effects, including the rapid growth in size, great user experience, immediate response, and so on. However, the open access and reliance on users have also made these systems becoming targets of spammers. They post advertisements or other irrelevant answers aiming at spreading advertise or achieving other goals. Some spammers directly publish content to answer questions asked by common users. Additionally, another kind of spammers (we refer them as \"best answer spammers\") create multiple user accounts, and use some accounts to ask a question, the others to provide answers which are selected as the best answers by themselves. They deliberately organize themselves in order to deceive readers. This kind of spammers are even more hazardous, since they are neither easily ignored nor identifiable by a human reader. Google Confucius CQA system also reported that best answer spammers may generate amounts of fake best answers, which could have a non-trivial impact on the quality of machine learning model (Si et al., 2010) .", "cite_spans": [ { "start": 1120, "end": 1137, "text": "(Si et al., 2010)", "ref_id": "BIBREF20" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "With the increasing requirements, spammer detection has received considerable attentions, including e-mails(L. Gomes et al., 2007; C.Wu et al., 2005) , web spammer (Cheng et al., 2011) , review spammer (Lim et al., 2010; N.Jindal and B.Liu, 2008; ott et al., 2011) , social media spammer (Zhu et al., 2012; Bosma et al., 2012; Wang, 2010) . However, little work has been done about spammers on CQA sites. Filling this need is a challenging task. The existing approaches of spam detection can be roughly into two directions. The first direction usually relied on costly human-labeled training data for building spam classifiers based on textual features (Y. Liu et al., 2008; Y.Xie et al., 2008; Ntoulas et al., 2006; Gyongyi and Molina, 2004) . However, since fake best answers are well designed and lack of easily identifiable textual patterns, text-based methods cannot achieve satisfactory performance. Another direction relied solely on hyperlink graph in the web (Z. Gyongyi et al., 2004; Krishnan and Raj, 2006; Benczur et al., 2005) . Although making good use of link information, link-based methods neglect the contentbased information. Moreover, unlike the web, there is no explicit link structure on CQA sites. So two intuitive research questions are: (1) Is there any useful link-based structure for spammer detection in CQA? (2) If so, can the two techniques, i.e., content-based model and link-based model, be integrated together to complement each other for CQA spammer detection?", "cite_spans": [ { "start": 111, "end": 130, "text": "Gomes et al., 2007;", "ref_id": null }, { "start": 131, "end": 149, "text": "C.Wu et al., 2005)", "ref_id": "BIBREF6" }, { "start": 164, "end": 184, "text": "(Cheng et al., 2011)", "ref_id": "BIBREF4" }, { "start": 202, "end": 220, "text": "(Lim et al., 2010;", "ref_id": "BIBREF14" }, { "start": 221, "end": 246, "text": "N.Jindal and B.Liu, 2008;", "ref_id": "BIBREF16" }, { "start": 247, "end": 264, "text": "ott et al., 2011)", "ref_id": "BIBREF18" }, { "start": 288, "end": 306, "text": "(Zhu et al., 2012;", "ref_id": "BIBREF26" }, { "start": 307, "end": 326, "text": "Bosma et al., 2012;", "ref_id": "BIBREF2" }, { "start": 327, "end": 338, "text": "Wang, 2010)", "ref_id": "BIBREF21" }, { "start": 657, "end": 674, "text": "Liu et al., 2008;", "ref_id": null }, { "start": 675, "end": 694, "text": "Y.Xie et al., 2008;", "ref_id": "BIBREF23" }, { "start": 695, "end": 716, "text": "Ntoulas et al., 2006;", "ref_id": "BIBREF17" }, { "start": 717, "end": 742, "text": "Gyongyi and Molina, 2004)", "ref_id": "BIBREF9" }, { "start": 972, "end": 993, "text": "Gyongyi et al., 2004;", "ref_id": "BIBREF9" }, { "start": 994, "end": 1017, "text": "Krishnan and Raj, 2006;", "ref_id": "BIBREF11" }, { "start": 1018, "end": 1039, "text": "Benczur et al., 2005)", "ref_id": "BIBREF0" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "To address the problems, in this paper, we first investigate the link-based structure in CQA. Then we formulate the task as an optimization problem in the graph with an efficient solution. We learn a content-based predictor as an objective function. The link-based information is incorporated into textual predictor by the way of graph regularization. Finally, to evaluate the proposed approach, we crawled a large data set from a commercial CQA site. Experimental results demonstrate that our proposed method can improve the accuracy of spammer detection.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The major contributions of this work can be summarized as follows: (1) To the best of our knowledge, our work is the first study on spammer detection on CQA sites; (2) Our proposed optimization model can integrate the advantages of both content-based model and link-based model for CQA spammer detection. (3) Experimental results demonstrate that our method can improve accuracy of spammer detection.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The remaining of the paper is organized as follows: In section 2, we review a number of the state-of-the-art approaches in related areas. Section 3 analyzes the social network of CQA sites. Section 4 presents the proposed method. Experimental results in test collections and analysis are shown in section 5. Section 6 concludes this paper.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Most of current studies on spam detection can be roughly divided into two categories: contentbased model and link-based model.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "2" }, { "text": "Content-based method targets at extracting ev-idences from textual descriptions of the content, treating the text corpus as a set of objects with associated attributes, and applying some classification methods to detect spam(P. Heymann et al., 2007; C.Castillo et al., 2007; Y.Liu et al., 2008; Y.Xie et al., 2008) . Fetterly proposed quite a few statistical properties of web pages that could be used to detect content spam(D. Fetterly et al., 2004) . Benevenuto went a step further by addressing the issue of detecting video spammers and promoters and applied the state-of-the-arts supervised classification algorithm to detect spammers and promoters (Benevenuto et al., 2009) . Lee proposed and evaluated a honeypot-based approach for uncovering social spammers in online social systems (Lee et al., 2010) . Wang proposed to improve spam classification on a microblogging platform (Wang, 2010) . An alternative web spam detection technique relies on link analysis algorithms, since a hyperlink often reflects some degree of similarity among pages (Gyngyi and Garcia-Molina, 2005; Gyongyi et al., 2006; Zhou et al., 2008) . Corresponding algorithms include TrustRank(Z. Gyongyi et al., 2004) and AntiTrustRank (Krishnan and Raj, 2006) , which used a seed set of Web pages with labels of trustiness or badness and propagate these labels through the link graph. Moreover, Benczur developed an algorithm called SpamRank which penalized suspicious pages when computing PageRank (Benczur et al., 2005) .", "cite_spans": [ { "start": 228, "end": 249, "text": "Heymann et al., 2007;", "ref_id": null }, { "start": 250, "end": 274, "text": "C.Castillo et al., 2007;", "ref_id": "BIBREF3" }, { "start": 275, "end": 294, "text": "Y.Liu et al., 2008;", "ref_id": "BIBREF22" }, { "start": 295, "end": 314, "text": "Y.Xie et al., 2008)", "ref_id": "BIBREF23" }, { "start": 428, "end": 450, "text": "Fetterly et al., 2004)", "ref_id": null }, { "start": 653, "end": 678, "text": "(Benevenuto et al., 2009)", "ref_id": "BIBREF1" }, { "start": 790, "end": 808, "text": "(Lee et al., 2010)", "ref_id": "BIBREF12" }, { "start": 884, "end": 896, "text": "(Wang, 2010)", "ref_id": "BIBREF21" }, { "start": 1050, "end": 1082, "text": "(Gyngyi and Garcia-Molina, 2005;", "ref_id": "BIBREF8" }, { "start": 1083, "end": 1104, "text": "Gyongyi et al., 2006;", "ref_id": "BIBREF10" }, { "start": 1105, "end": 1123, "text": "Zhou et al., 2008)", "ref_id": "BIBREF25" }, { "start": 1172, "end": 1193, "text": "Gyongyi et al., 2004)", "ref_id": "BIBREF9" }, { "start": 1212, "end": 1236, "text": "(Krishnan and Raj, 2006)", "ref_id": "BIBREF11" }, { "start": 1476, "end": 1498, "text": "(Benczur et al., 2005)", "ref_id": "BIBREF0" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "2" }, { "text": "Before analyzing the social network in CQA, we introduce some definitions. We refer users on C-QA sites are someone who ask at least one question or answer at least one question. Moreover, users are divided into two categories: spammers and legitimate users. We define spammers as users who post at least one question or one answer intent to create spam.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Analysis on Social Network", "sec_num": "3" }, { "text": "A CQA site is particularly rich in user interactions. These interactions can be represented by Figure 1(a) , where a particular question has a number of answers associated with it, represented by an edge from the question to each of the answer. We also include vertices representing authors of question or answers. An edge from a user to a question means that the user asked the question, and an edge from an answer to a user means that the answer was posted by this user. In the example, a user U 1 asks a question Q 1 , while users U 4 , U 5 and U 6 answers this question. In order to observe the relation between users more clearly and directly, we summarize the relations between users as a graph shown in Figure 1 (b). This graph contains vertices representing the users and omits the actual questions and answers that connect the users. Question-answer relation:", "cite_spans": [], "ref_spans": [ { "start": 95, "end": 106, "text": "Figure 1(a)", "ref_id": null }, { "start": 710, "end": 718, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Analysis on Social Network", "sec_num": "3" }, { "text": "As shown in Fig- ure 2(a), U 4 answers U 1 's question.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Analysis on Social Network", "sec_num": "3" }, { "text": "We define that U 4 and U 1 have Question-answer relation. Furthermore, Question-answer relation can be divided into two disjoint sets: best-answer relation and non-best-answer relation.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Analysis on Social Network", "sec_num": "3" }, { "text": "Best-answer relation: U 1 selects U 5 's answer as the best answer. We define that U 1 and U 5 have best-answer relation. The solid lines in Figure 2 (b) express the best-answer relation.", "cite_spans": [], "ref_spans": [ { "start": 141, "end": 149, "text": "Figure 2", "ref_id": null } ], "eq_spans": [], "section": "Analysis on Social Network", "sec_num": "3" }, { "text": "Non-best-answer relation: U 1 does not select U 4 's answer as the best answer. We define that U 1 and U 4 have non-best-answer relation. The dashed lines in Figure 2 (c) express the non-best-answer relation.", "cite_spans": [], "ref_spans": [ { "start": 158, "end": 166, "text": "Figure 2", "ref_id": null } ], "eq_spans": [], "section": "Analysis on Social Network", "sec_num": "3" }, { "text": "From analyzing data crawled from CQA site, we present the following property about best-answer relation:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Best-answer Consistency Property", "sec_num": "3.1" }, { "text": "Best-answer consistency property: If U i selects U j 's answer as the best answer, the classes of users U i and U j should be similar.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Best-answer Consistency Property", "sec_num": "3.1" }, { "text": "We explain this property as follows: consider that a legitimate user is unlikely to select a spammer's answer as the best answer due to its low quality, while a legitimate user is unlikely to answer a spammer's question, so the possibility of a spammer selecting a legitimate user's answer will also be small. This means that two users linked via best-answer relation are more likely to share similar property than two random users.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Best-answer Consistency Property", "sec_num": "3.1" }, { "text": "Different from the general spammers, some spammers generate many fake best answers to obtain higher status in the community. We refer them as best answer spammers. In order to generate fake best answers, a spammer creates multiple user accounts first. Then, it uses some of the accounts to ask questions, and others to provide answers. Such spammers may post low quality answers to their own questions, and select those as the best by themselves. They may generate lots of fake best answers, which may highly impact the user experience.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Characteristics of Best Answer Spammer", "sec_num": "3.2" }, { "text": "Furthermore, when the spammer's intention is just advertising, we can easily identify signs of its activity: repeated phone numbers or URLs and then ignore them. However, when the spammer's intention is to obtain higher reputation within the community, the spam content may lack obvious patterns. Fortunately, there are still some clues that may help identify best answer spammers. Two characteristics are described as follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Characteristics of Best Answer Spammer", "sec_num": "3.2" }, { "text": "High best answer rate: Best answer rate is the ratio of answers selected as the best answer among the total answers. This kind of spammers have an incredible high best answer rate, compared to normal users. Specifically, in a possible best answer spammer pair, sometimes only one user has an incredible high best answer rate. Because normally one responses for asking and another for answering. So we calculate the best answer rate BR(i, j) for a user pair (u i , u j ) based on the maximum of their best answer rates:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Characteristics of Best Answer Spammer", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "BR(i, j) = M ax(BR(i), BR(j))", "eq_num": "(1)" } ], "section": "Characteristics of Best Answer Spammer", "sec_num": "3.2" }, { "text": "Where BR(i) is the best answer rate of u i . Time margin score: To be efficient, best answer spammers tend to answer their own ques-tion quickly. We consider the time margin score T ime(i, j) between a question posted and answered for u i and u j as an evidence.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Characteristics of Best Answer Spammer", "sec_num": "3.2" }, { "text": "T ime(i, j) = { 1, if T imeM argin(i, j) < \u03b5 0, otherwise", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Characteristics of Best Answer Spammer", "sec_num": "3.2" }, { "text": "(2) where T imeM argin(i, j) is the real time margin between u i asks a question and u j answers this question and \u03b5 = 30 minutes.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Characteristics of Best Answer Spammer", "sec_num": "3.2" }, { "text": "The best answer spammer score s(i, j) for a user pair (u i , u j ) can be calculated as the combination of these two scores:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Characteristics of Best Answer Spammer", "sec_num": "3.2" }, { "text": "s(i, j) = \u00b5BR(i, j) + (1 \u2212 \u00b5)T ime(i, j) (3)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Characteristics of Best Answer Spammer", "sec_num": "3.2" }, { "text": "\u00b5 is trade-off of two scores, here we simply set \u00b5 = 0.5. The value of s(i, j) is between 0 to 1. The higher s(i, j) is, the more likely u i and u j is a pair of the best answer spammers.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Characteristics of Best Answer Spammer", "sec_num": "3.2" }, { "text": "In this section, the framework of our proposed approach is presented. First, the problem is formally defined. Next, we build a baseline supervised predictor that makes use of a variety of textual features, and then the consistency property and best answer spammer characteristics are incorporated by adding regularization to the textual predictor, last we discuss how to effectively optimize it.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Spammer Detection on CQA Sites", "sec_num": "4" }, { "text": "On CQA sites, there are three distinct types of entities:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Statement", "sec_num": "4.1" }, { "text": "users U = {u 1 , ...u l+u }, answers A = {a 1 , ...a M }, and questions Q = {q 1 , ...q N }. The set of users U contains both U L = {u 1 , ...u l } of l labeled users and U U = {u l+1 , ...u l+u } of u un- labeled users.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Statement", "sec_num": "4.1" }, { "text": "We model the social network for U as a directed graph G = (U, E) with adjacency matrix A, where A ij = 1 if there is a link or edge from u i to u j and zero otherwise.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Statement", "sec_num": "4.1" }, { "text": "Given the input data {U L , U U , G, Q, A}, we want to learn a predictor c for a user", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Statement", "sec_num": "4.1" }, { "text": "u i . c(u i )\u2212 > {spammer, legitimate user} (4) Legitimacy score y i (0 \u2264 y i \u2264 1,i =1,2,...n)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Statement", "sec_num": "4.1" }, { "text": "is computed for all the users. The lower y i is, the more likely u i is a spammer.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Statement", "sec_num": "4.1" }, { "text": "In this subsection, we build a baseline predictor based on textual features in a supervised fashion.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "We regard the legitimacy scores as generated by combining textual features.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "We consider the following textual features.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "\u2022 The Length of answers: The length may to some extent indicate the quality of the answer. The average length of answers is calculated as a feature.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "\u2022 The ratio of Ads words in answers: Advertising of products is the main goal of a kind of spammers and they repeat some advertisement words in their answers.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "\u2022 The ratio of Ads words in questions: Some spammers will refer some Ads in questions in order to get attention from more users.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "\u2022 The number of received answers: The number of received answers can indicate the quality of the question.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "\u2022 Best answer rate: Best answer rate can show the quality of their answers.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "\u2022 The number of answers: It can indicate the authority of a user.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "\u2022 Relevance of question and answer: We measure the average content similarity over a pair of question and answer which is computed using the standard cosine similarity over the bag-of-words vector representation.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "\u2022 Duplication of answers: The Jaccard similarity of answers are applied to indicate the duplication of answers .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "With these features, suppose there are in total k features for each user u i , denoted as", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "x i . Then X = (x 1 , x 2 , ...x n )", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "is the k-by-n feature matrix of all users. Based on these features, we define the legitimacy score of each user as follows,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "y i = w T x i (5)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "where w is a k-dimensional weight vector. Suppose we have legitimate/spammer labels t i in the training set.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "t i = { 1, u i is labeled as legitimate user 0, u i is labeled as spammer", "eq_num": "(6)" } ], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "We will then define the loss term as follows,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "\u2126(w) = 1 l l \u2211 i=1 (w T x i \u2212 t i ) 2 + \u03b1w T w (7)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "Once we have learned the weight vector w, we can apply it to any user feature vector and predict the class of unlabeled users.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Text-based Spammer Prediction", "sec_num": "4.2" }, { "text": "In Section 4.2, each user is considered as a standalone item. In this subsection, we exploit social information to improve CQA spammer detection. In Section 3.1, the consistency property has been analyzed that users connected via bestanswer relation are more similar in property. So the property is enforced by adding a regularization term into the optimization model. The regularization is acted in a collection data set, including a small amount of labeled data(l users) and a large amount of unlabeled data(u users). Then the regularization term is formulated as:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Regularization for Consistency Property", "sec_num": "4.3" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "REG 1 (U ) = l+u \u2211 i,j A ij (y i \u2212 y j ) 2", "eq_num": "(8)" } ], "section": "Regularization for Consistency Property", "sec_num": "4.3" }, { "text": "Minimizing the regularization constraint will force users who have best-answer relation belong to the same class. We formulate this as graph regularization. The graph adjacency matrix A is defined as A ij = 1 if u j selects u i 's answer as the best answer, and zero otherwise. Then, Equation 8 becomes:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Regularization for Consistency Property", "sec_num": "4.3" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "REG 1 (w) = l+u \u2211 i,j A ij (w T x i \u2212 w T x j ) 2", "eq_num": "(9)" } ], "section": "Regularization for Consistency Property", "sec_num": "4.3" }, { "text": "With this regularization, then the objective function Equation 7 becomes:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Regularization for Consistency Property", "sec_num": "4.3" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "\u2126 1 (w) = 1 l l \u2211 i=1 (w T x i \u2212 t i ) 2 + \u03b1w T w +\u03b2 l+u \u2211 i,j A ij (w T x i \u2212 w T x j ) 2", "eq_num": "(10)" } ], "section": "Regularization for Consistency Property", "sec_num": "4.3" }, { "text": "In this subsection, we focus on best answer spammers. Since they cannot be easily detected by only textual features(Equation 7), we introduce an additional penalty score b i to each user u i which indicates the possibility of becoming a best answer spammer. With the penalty score b i , Equation 5 can be redefined as follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Regularization for Best Answer Spammer", "sec_num": "4.4" }, { "text": "y i = w T x i \u2212 b i (11)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Regularization for Best Answer Spammer", "sec_num": "4.4" }, { "text": "where b i is a non-negative score. In order to obtain b i , characteristics of best answer spammers are incorporated by adding graph regularization to the optimization problem. The regularization is also acted in a collection data set. Two kinds of regularization are presented as follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Regularization for Best Answer Spammer", "sec_num": "4.4" }, { "text": "Penalty for Best Answer Spammers in Pairs As described in Section 3.2, the score s(i, j) indicates the possibility of u i and u j becoming a pair of best answer spammers(Equation 3). We expect u i and u j , who create the spam together, should share this possibility together, as follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Regularization for Best Answer Spammer", "sec_num": "4.4" }, { "text": "b i + b j = e \u00d7 s(i, j)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Regularization for Best Answer Spammer", "sec_num": "4.4" }, { "text": ", where e is a penalty factor, we empirically set it to 0.5.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Regularization for Best Answer Spammer", "sec_num": "4.4" }, { "text": "Then we can also formulate this as graph regularization as:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Regularization for Best Answer Spammer", "sec_num": "4.4" }, { "text": "REG 2 (b) = l+u \u2211 i", "html": null, "num": null, "text": "Performance of our optimization methods with different regularization for comparison", "type_str": "table" } } } }