Detecting collusive spammers on e-commerce websites based on reinforcement learning and adversarial autoencoder

Abstract

The collusive spamming behavior on e-commerce websites seriously affects the purchase decisions of consumers and disrupts the fair competition order among merchants. To address this problem, many approaches have been proposed to detect collusive spammers on e-commerce websites. However, the existing approaches rely on hand-crafted features for detecting collusive spammers, which is costly and time-consuming. With this limitation in mind, we propose a collusive spammer detection approach based on reinforcement learning and adversarial autoencoder. First, we model the given dataset as a user-product bipartite graph that is treated as the agent's interactive environment and use the modified reinforcement learning algorithm to obtain candidate groups. Second, we exploit the Doc2Vec model to generate the embedding of each candidate group and devise an adversarial autoencoder-based one-class classification model for detecting collusive spammers. The experimental results on the real-world review datasets show that the proposed approach has better detection performance than the state-of-the-art methods.