Handling Cold-Start Problem in Review Spam Detection by Jointly Embedding Texts and Behaviors

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 366–376 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1034 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 366–376 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1034 Handling Cold-Start Problem in Review Spam Detection by Jointly Embedding Texts and Behaviors Xuepeng Wang1,2, Kang Liu1, and Jun Zhao1,2 1 National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China 2 University of Chinese Academy of Sciences, Beijing, 100049, China {xpwang, kliu, jzhao}@nlpr.ia.ac.cn Abstract Solving the cold-start problem in review s- pam detection is an urgent and significant task. It can help the on-line review web- sites to relieve the damage of spammers in time, but has never been investigated by previous work. This paper proposes a nov- el neural network model to detect review spam for the cold-start problem, by learn- ing to represent the new reviewers’ review with jointly embedded textual and behav- ioral information. Experimental results prove the proposed model achieves an ef- fective performance and possesses prefer- able domain-adaptability. It is also appli- cable to a large-scale dataset in an unsu- pervised way. 1 Introduction With the rapid growth of products reviews at the web, it has become common for people to read reviews before making a purchase decision. The reviews usually contain abundant consumers’ per- sonal experiences. It has led to a significant in- fluence on financial gains and fame for business- es. Existing studies have shown that an extra half- star rating on Yelp causes restaurants to sell out 19% points more frequently (Anderson and Ma- gruder, 2012), and a one-star increase in Yelp rat- ing leads to a 5-9 % increase in revenue (Luca, 2011). This, unfortunately, gives strong incentives for imposters (called spammers) to game the sys- tem. They post fake reviews or opinions (called review spam) to promote or to discredit some tar- geted products and services. The news from BBC has shown that around 25% of Yelp reviews could be fake.1 Therefore, it is urgent to detect review s- 1http://www.bbc.com/news/technology-24299742 pam, to ensure that the online review continues to be trusted. Jindal and Liu (2008) make the first step to de- tect review spam. Most efforts are devoted to ex- ploring effective linguistic and behavioral features by subsequent work to distinguish such spam from the real reviews. However, to notice such pattern- s or form behavioral features, developers should take a long time to observe the data, because the features are based on statistics. For instance, the feature activity window proposed by Mukherjee et al. (2013c) is to measure the activity freshness of reviewers. It usually takes several months to count the difference of timestamps between the last and first reviews for reviewers. When the fea- tures show themselves finally, some major dam- ages might have already been done. Thus, it is important to design algorithms that can …

Content truncated to 500 words