Combating Web-Based E-Commerce Fraud Syndicates: Fairness-Aware Hypergraph Contrastive Fraud Detection with Multi-Dimensional Reinforcement Rewards

Abstract

The proliferation of web-based e-commerce and digital payment platforms, while promoting financial inclusion, has inadvertently heightened the risks of collusive fraud, which disproportionately impacts vulnerable populations such as the elderly and low-income communities. This challenge underscores an urgent need for responsible web AI that ensures social fairness and transparency to protect the most susceptible users. To address this issue and contribute to building safer and more inclusive web ecosystems, we propose MHMRC, a novel semi-supervised model that integrates Multi-view Heterogeneous Hypergraph Contrastive learning (MHHC) and Multi-Dimensional Reinforcement reward driven Community detection (MDRC) for web-based e-commerce collusive fraud detection. First, MDRC quantifies intra-group node similarity using policy optimization guided by three specialized rewards: modularity, connectivity, and consistency. This process transforms isolated users into behaviorally correlated entities, enabling the precise identification of latent fraud syndicates even in the absence of explicit interaction links. This is particularly crucial for identifying coordinated attacks that target vulnerable groups. Second, MHHC integrates temporal, user-centric, and item-centric perspectives through hypergraph fusion. It applies contrastive objectives at node and hyperedge levels to preserve multimodal relationships while maximizing inter-group separation, thus effectively detecting cross-view camouflage by identifying inherent representational inconsistencies across heterogeneous dimensions. Extensive experiments on six real-world web financial transaction datasets demonstrate MHMRC's superiority over 14 SOTA models, achieving average improvements of 5.74% in AUC, 4.94% in F1, and 13.20% in AP.