Deceptive Review Spam Detection via Exploiting Task Relatedness and Unlabeled Data

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1817–1826, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Deceptive Review Spam Detection via Exploiting Task Relatedness and Unlabeled Data Zhen Hai† Peilin Zhao‡ Peng Cheng§ Peng Yang∗ Xiao-Li Li † Guangxia Li¶ †Institute for Infocomm Research, A*STAR, Singapore, {haiz,xlli}@i2r.a-star.edu.sg ‡Ant Financial, Hangzhou, China, peilin.zpl@alipay.com §SCSE, Nanyang Technological University, Singapore, pcheng1@ntu.edu.sg ∗Tencent AI Lab, Shenzhen, China, henryppyang@tencent.com ¶SCST, Xidian University, Xi’an, China, gxli@xidian.edu.cn Abstract Existing work on detecting deceptive reviews primarily focuses on feature engineering and applies off-the-shelf supervised classification algorithms to the problem. Then, one real challenge would be to manually recognize plentiful ground truth spam review data for model building, which is rather difficult and often requires domain expertise in practice. In this paper, we propose to exploit the related- ness of multiple review spam detection tasks and readily available unlabeled data to address the scarcity of labeled opinion spam data. We first develop a multi-task learning method based on logistic regression (MTL-LR), which can boost the learning for a task by sharing the knowledge contained in the training sig- nals of other related tasks. To leverage the unlabeled data, we introduce a graph Lapla- cian regularizer into each base model. We then propose a novel semi-supervised multi- task learning method via Laplacian regular- ized logistic regression (SMTL-LLR) to fur- ther improve the review spam detection per- formance. We also develop a stochastic al- ternating method to cope with the optimiza- tion for SMTL-LLR. Experimental results on real-world review data demonstrate the benefit of SMTL-LLR over several well-established baseline methods. 1 Introduction Nowadays, more and more individuals and organi- zations have become accustomed to consulting user- generated reviews before making purchases or on- line bookings. Considering great commercial ben- efits, merchants, however, have tried to hire peo- ple to write undeserving positive reviews to promote their own products or services, and meanwhile to post malicious negative reviews to defame those of their competitors. The fictitious reviews and opin- ions, which are deliberately created in order to pro- mote or demote targeted entities, are known as de- ceptive opinion spam (Jindal and Liu, 2008; Ott et al., 2011). By formulating deceptive opinion spam detection as a classification problem, existing work primarily focuses on extracting different types of features and applies off-the-shelf supervised classification algo- rithms to the problem (Jindal and Liu, 2008; Ott et al., 2011; Feng et al., 2012; Chen and Chen, 2015). Then, one weakness of previous work lies in the de- mand of manually recognizing a large amount of ground truth review spam data for model training. Unlike other forms of spamming activities, such as email or web spam, deceptive opinion spam, which has been deliberately written to sound authentic, is more difficult to be recognized by manual read. In an experiment, three undergraduate students were (randomly) invited to identify spam reviews from nonspam ones in hotel domain. As shown in Table 1, their average accuracy is merely 57.3% (Ott et …

Content truncated to 500 words