An unsupervised approach to detect review spam using duplicates of images, videos and Chinese texts

Abstract

Intuitively, image- or video-based recommendations seem to be more reliable than those containing plain text, and these types of recommendations have recently become widely encouraged and commonly seen across opinion sharing platforms. Considering their potential for manipulation, graphs (e.g., images and videos) are more vulnerable to spam than scripts. However, most state-of-the-art solutions for opinion spam detection are exclusively devoted to natural language parsing, and less work has been done concerning photos or videos. After investigating the top two business-to-customer websites, i.e., JD.com and TMALL.com, we propose an unsupervised approach to label suspected spam based on different types of duplication across images, videos and Chinese texts. Experiments verified the effectiveness of this approach and obtained several conclusions: 1) the situation of image spam is more severe than that of video and text spam; 2) for manipulation, borrowing something from a marketing page is less attractive than stealing from other reviewers; 3) in addition to using identical texts, spammers also use fictitious rare incidents to influence customers; and 4) overlapping duplications of images, videos and texts are common.

Introduction

People tend to believe things that are supported by photos. Due to their intuitiveness, graphs (including images and videos) have long been encouraged by e-business and opinion-sharing websites. Recently, they have become commonly seen on such sites. For instance, over 8% of the product reviews on TMALL.com (one of the top two largest business-to-consumer (B2C) retailers in China which is owned by Alibaba; referred as TMALL in this paper) have images or videos. On average, each review on JD.com (one of the top two largest B2C retailers in China; referred as JD in the following) has 1.05 photos and 0.5 videos. Moreover, graphic experiences are often promoted by headlines or recommended with a higher level of priority. Similar to the ways in which text is adopted, people share pictures to guide others, to express their intelligence, to connect with people or just to earn platform credits/coupons (Dellarocas and Narayan, 2006; Hennig-Thurau et al., 2004; Hu et al., 2011; Zhu and Zhang, 2010). Likewise, images could also be used to lie with the objective of misleading potential customers; this is a new type of opinion spam that is not electronic word-of-mouth. Even worse, since graphs always seem to be more reliable than plain texts, the expected probability that people will be fooled by this type of spam is greater than that of text manipulation. Over the last decade, opinion spam has drawn a considerable amount of attention, and there have been substantial achievements on the topic (Deborah and Baron, 1988; Dellarocas and Narayan, 2006; Heydari et al., 2015; Jindal and Liu, 2007, 2008; Jindal et al., 2010; Li et al., 2019; Lim, 2010, 2010; Liu and Pang, 2018; Mayzlin, 2006; Mukherjee et al., 2013; Ott et al., 2012, 2011; Paul Rayson, 2001; Savage et al., 2015; Somayeh, 2013; Xie et al., 2012; Zhang et al., 2019; Zhang et al., 2018); nevertheless, most of these work were based on natural language processing and thus cannot fit graphic experiences. Investigation along reviews hosted by JD and TMALL shows that either graph- or text-oriented duplication is common and reviews duplicate in multiple ways. For instance, spammers tend to borrow photos from introduction pages, copy-and-paste videos from other posts and/or refer to a specific scenario in their texts. To further unveil this case, in this paper, we propose an approach that can address duplication of texts, images and videos simultaneously; recognizing different kinds of duplication and labeling spam especially compound spam are the top two challenges.

To our knowledge, this is the first time that images or videos have been fully addressed in the context of review spam detection. Although fields like image forensics or video faking have covered the graph tampering problem for years with plentiful techniques, the manipulation in review systems is different and we cannot directly adopt these solutions to conquer it. 1) for profit efficiency, spammers opt to steal and post someone's images/videos without pixel manipulating or frame editing; and 2) spammers prefer to borrow marketing pictures from the item's webpage, which are carefully designed by sellers and always have backgrounds in pure white (0xFFFFFF) or black (0x000000). Specifically, the contribution of this paper is threefold: 1) We focus on both texts and graphs to uncover any review spam, 2) we introduce reasonable criteria by which to detect different types of duplication and 3) we find some interesting phenomena.

The remainder of this paper is organized as follows. First, we survey state-of-the-art studies in Section 2, and then we introduce our proposal in Section 3. After that, we conduct and discuss some experiments in Section 4. Finally, we conclude this paper in Section 5.