Fusion-based Spammer Detection Method by Embedding Review Texts and Weak Social Relations

Jie Wen, Jingyuan Hu, Hongbin Shi, Xin Wang, Chunyuan Yuan, Jizhong Han, Tao Guo

Abstract: Detecting spam reviews and spammers is a longstanding challenge, where review texts and social network information are two vital aspects used in many previous works. However, most of previous works treat multiple reviews as a whole for feature extraction, while ignoring the internal similarities and differences among them, which may cause the semantics of reviews disorder and cannot capture the most discriminative information. Furthermore, the social network information, which we call strong social relationship, is easy to be manipulated by spammers. Thus, using this strong relationship may discount the performance. In this work, we design a co-attention module and orthogonal decomposition module to automatically learn the similarities and differences of multiple reviews. To overcome the problems that the strong social relationship faced, we define the weak social relation graph based on the dynamic interactions. Then a general graph representation framework is presented to learn the interaction relationship information. Finally, we fuse the two aspects information to detect spammers. The experimental results significantly outperform the state-of-the-art approaches on two real-world benchmark datasets, which confirms the superiority of our model.