Universal ring-of-abusers detection via multi-modal heterogeneous graph learning

As fraudulent and abusive activities performed by groups continue to plague e-commerce stores, we realize that detecting groups of abusers, or Ring-of-Abusers (RoAs), has become crucial. Unlike existing works about abuser detection on e-commerce stores that merely consider the individual features of abusers or the relationships among abusers, we design a Universal Ring-Of-Abusers Detection framework (abbreviated as U-ROAD) that integrates multi-modal features (e.g., numerical feature, text, and image) and the rich relationships among entities (i.e., seller, buyer, and product) on e-commerce stores. Especially, the U-ROAD framework designs three meta-paths to depict the high-order heterogeneous semantic relationships among entities. Then it leverages graph neural networks (GNNs) for obtaining the semantic node embeddings, which are fused with an aggregation layer and fed to an MLP classifier to detect abusers (e.g., abusive sellers and buyers) in different scenarios. As a result, (i): Flexible: The U-ROAD framework can be easily customized with different training labels to detect different types of abusers at e-commerce stores (e.g., rank abusers, financial abusers, and competitor abusers). (ii): Powerful and Timely: The framework is empirically powerful than some existing methods, which achieves an average 30% improvement in precision and 7.5% improvement in recall on four abuser (i.e., abusive seller and abusive buyer) detection tasks in two stores. Besides, our U-ROAD framework can detect abusers earlier than some existing methods on e-commerce stores. (iii): Explainable: The U-ROAD framework develops a visualization tool to help investigators understand the RoAs detection results.