Temporal Opinion Spam Detection by Multivariate Indicative Signals

Online consumer reviews reflect the testimonials of real people, unlike advertisements. As such, they have critical impact on potential consumers, and indirectly on businesses. According to a Harvard study (Luca 2011), +1 rise in star-rating increases revenue by 5-9%. Problematically, such financial incentives have created a market for spammers to fabricate reviews, to unjustly promote or demote businesses, activities known as opinion spam (Jindal and Liu 2008). A vast majority of existing work on this problem have formulations based on static review data, with respective techniques operating in an offline fashion. Spam campaigns, however, are intended to make most impact during their course. Abnormal events triggered by spammers' activities could be masked in the load of future events, which static analysis would fail to identify. In this work, we approach the opinion spam problem with a temporal formulation. Specifically, we monitor a list of carefully selected indicative signals of opinion spam over time and design efficient techniques to both detect and characterize abnormal events in real-time. Experiments on datasets from two different review sites show that our approach is fast, effective, and practical to be deployed in real-world systems.