Exposing collaborative spammer groups through the review-response graph

Deceptive opinions of online merchandises, also known as review spams, cause great loss for consumers, manufacturers and even business-to-customer platforms. However, due to the weak supervision problem, especially the lack of ground-truth labels, identifying these untruthful reviews is challenging. What's even worse is that crowdsourcing workers out of manipulation campaigns always collaborate to distort an item's reputation, rendering the product together with its brand difficult to be rehabilitated. State-of-the-art solutions on spammer group recognition highlight co-reviewing behaviours or sentiment similarity to cluster reviews, which can only yield loosely-coupled candidates of reviewer sets. In this paper, we highlight the commenting interaction between reviews and model it as a bipartite graph and discover a new low-budget spam, i.e., responsive spam. Furthermore, we recognize strong-correlated groups of spam through a propagation technique upon two widely adopted spam indicators, i.e., text duplication and posting burstiness. Comparative results show that our approach is effective and outperforms state-of-the-art solutions with great significance.

Automated summary

Deceptive online merchandises, also known as review spams, result in significant losses for consumers, manufacturers, and business-to-customer platforms. However, their identification is challenging due to weak supervision and lack of ground-truth labels. The collaboration of crowdsourcing workers in manipulation campaigns further damages the reputation of products and brands. This paper proposes a novel approach using commenting interaction, bipartite graph modeling, and spam indicators to effectively and significantly recognize strong-correlated groups of spam reviewers, outperforming state-of-the-art solutions.