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

Enhancing fairness of trading environment: discovering overlapping spammer groups with dynamic co-review graph optimization

“…Existing methods for detecting spammer groups can be distinguished based on two classification criteria, i.e., the distinct methods of generating candidate groups and the features used in generating candidate groups. According to the distinct methods of generating candidate groups, existing detection algorithms can be classified into three categories such as the FIM-based algorithms (Mukherjee et al 2012;Xu et al 2013;Xu and Zhang 2015), the graph-based algorithms (Wang et al 2016(Wang et al , 2018aChoo et al 2015;Hu et al 2019;Zhang et al 2020Zhang et al , 2022aZhang et al , 2022bAkoglu et al 2013;Ye and Akoglu 2015;Zheng et al 2018;Zhu et al 2019;Shehnepoor et al 2021;Chao et al 2022), and the burst-based algorithms (Li et al 2017;Ji et al 2020;Liu et al 2018). According to the features employed to generate candidate groups, existing detection algorithms can also be divided into three categories: the ones based on group behavioral features (Mukherjee et al 2012;Xu et al 2013;Xu and Zhang 2015;Li et al 2017;Ji et al 2020) such as review content, timing, and ratings (which emphasizes the analysis of behavior patterns of the group), the ones based on relationship features (also called structural features) of the group (Choo et al 2015;Ye and Akoglu 2015;Zheng et al 2018;Zhu et al 2019;Shehnepoor et al 2021;Chao et al 2022), and algorithms that combine both group behavioral and structural features (Wang et al 2016(Wang et al , 2018a…”

“…In the following subsections, we focus on reviewing graph-based algorithms, as it is a prominent category within the first classification criterion and these algorithms have gained significant attention in spammer group detection in the e-commerce context. The graph-based approaches to spammer group detection can be further divided based on the construction of the graphs into two categories: homogeneous graph-based algorithms (Wang et al 2016(Wang et al , 2018aChoo et al 2015;Hu et al 2019;Zhang et al 2020Zhang et al , 2022a and heterogeneous graph-based algorithms (Akoglu et al 2013;Ye and Akoglu 2015;Zheng et al 2018;Zhu et al 2019;Shehnepoor et al 2021;Chao et al 2022;Zhang et al 2022b).…”

“…They then utilized the Canopy and K-means clustering algorithms to generate candidate groups and finally ranked the candidates using fake indicators to obtain the spammer groups. Zhang et al (2022b) proposed a spammer group detection algorithm called DCS-RLAA, which is based on reinforcement learning and adversarial autoencoders. They began by constructing a reviewerproduct bipartite graph, then employed an enhanced reinforcement learning algorithm, Sarsa, to obtain candidate groups.…”

A New Approach for Product Quality Check Based on Social Networks Opinions Analysis

“…Given this constraint, Zhang and al. [10] propose an innovative approach for detecting collusive spammers, utilizing reinforcement learning and an adversarial autoencoder. Our approach involves modeling the dataset as a userproduct bipartite graph, treating it as the agent's interactive environment, and employing a modified reinforcement learning algorithm to derive potential groups.…”