MULTITRUST: GRAPH-BASED MULTI-PERSPECTIVE FRAUD DETECTION IN ONLINE MARKETPLACES

Abstract

The exponential growth of e-commerce platforms has revolutionized digital trade but has also amplified the risk of fraudulent transactions involving multiple participants such as buyers, sellers, and intermediaries. Traditional fraud detection methods often analyze transactions in isolation, ignoring inter-user dependencies and behavioral correlations, which limits their effectiveness against collusive or coordinated fraud. This study proposes MultiTrust, a GraphBased Multi-Perspective Fraud Detection Framework that integrates graph neural networks (GNNs) and multi-view learning to capture relational, behavioral, and contextual features of participants. By constructing an interaction graph among users, products, and transactions, MultiTrust identifies suspicious patterns through deep representation learning and relational inference. The framework fuses multiple data perspectives—transactional, temporal, and trust-based—to enhance detection accuracy and interpretability. Experimental results on benchmark e-commerce datasets demonstrate that MultiTrust significantly outperforms traditional supervised and anomalybased models in detecting collusive and multiparty fraud. This approach provides a scalable, explainable, and adaptive mechanism for safeguarding modern e-commerce ecosystems.