Spammer Group Detection Approach Based on Deep Reinforcement Learning

Detecting spammer groups is important for maintaining the normal operation of e-commerce platforms. Nevertheless, current spammer group detection methods ignore the overlapping between spammer groups. Moreover, handcrafted indicators-based methods lack universality. Aiming at these problems, we propose a deep reinforcement learning-based spammer group detection approach. First, we model the review dataset as a user-item bipartite graph that serves as the interaction environment of the agent, and utilize the BiNE model to get the initial user vector representations. Then, we model the generation of candidate groups in the bipartite graph as the Markov decision process and introduce the self-attention mechanism to aggregate node features. Meanwhile, we use a deep Q-network to obtain the overlapped candidate group sequences. Finally, we apply an adversarial autoencoder to detect spammer groups. Experiments on Yelp Miami, Yelp New York, YelpCHI and Amazon datasets reveal that our method performs better than three baselines.