Positive Unlabeled Learning for Deceptive Reviews Detection

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 488–498, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Positive Unlabeled Learning for Deceptive Reviews Detection Yafeng Ren Donghong Ji Hongbin Zhang Computer School Wuhan University Wuhan 430072, China {renyafeng,dhji,zhanghongbin}@whu.edu.cn Abstract Deceptive reviews detection has attract- ed significant attention from both business and research communities. However, due to the difficulty of human labeling need- ed for supervised learning, the problem re- mains to be highly challenging. This pa- per proposed a novel angle to the prob- lem by modeling PU (positive unlabeled) learning. A semi-supervised model, called mixing population and individual proper- ty PU learning (MPIPUL), is proposed. Firstly, some reliable negative examples are identified from the unlabeled dataset. Secondly, some representative positive ex- amples and negative examples are gener- ated based on LDA (Latent Dirichlet Al- location). Thirdly, for the remaining un- labeled examples (we call them spy ex- amples), which can not be explicitly iden- tified as positive and negative, two simi- larity weights are assigned, by which the probability of a spy example belonging to the positive class and the negative class are displayed. Finally, spy examples and their similarity weights are incorporated into SVM (Support Vector Machine) to build an accurate classifier. Experiments on gold-standard dataset demonstrate the effectiveness of MPIPUL which outper- forms the state-of-the-art baselines. 1 Introduction The Web has dramatically changed the way peo- ple express themselves and interact with others, people frequently write reviews on e-commerce sites, forums and blogs to achieve these purpos- es. For NLP (Natural Language Processing), these user-generated contents are of great value in that they contain abundant information related to peo- ple’s opinions on certain topics. Currently, on- line reviews on products and services are used extensively by consumers and businesses to con- duct decisive purchase, product design and mar- keting strategies. Hence, sentiment analysis and opinion mining based on product reviews have become a popular topic of NLP (Pang and Lee, 2008; Liu, 2012). However, since reviews infor- mation can guide people’s purchase behavior, pos- itive reviews can result in huge economic benefit- s and fame for organizations or individuals. This leaves room for promoting the generation of re- view spams. Through observations and studies of the predecessors (Jindal and Liu, 2008; Ott et al., 2011), review spams are divided into the following two classes: • Deceptive Reviews: Those deliberately mis- lead readers by giving undeserving positive reviews to some target objects in order to pro- mote the objects, or by giving unjust nega- tive reviews to some target objects in order to damage their reputation. • Disruptive Reviews: Those are non-reviews, which mainly include advertisements and other irrelevant reviews containing no opin- ion. Disruptive reviews pose little threat to peo- ple, because human can easily identify and ignore them. In this paper, we focus on the more chal- lenging ones: deceptive reviews. Generally, de- ceptive reviews detection is deemed to be a classi- fication problem (Ott et al., 2011; …

Content truncated to 500 words