Graph-Based Deep Learning for E-Commerce Fraud Detection

Abstract

E-commerce growth has fueled increasingly sophisticated fraud schemes, including transaction manipulation and payment fraud. Traditional fraud detection methods struggle to adapt, leading to high false positive rates and ineffective detection of emerging fraud patterns. This study proposes a graph-based deep learning framework that models e-commerce transactions as a heterogeneous graph. It utilizes graph convolutional networks (GCN) and graph attention networks (GAT) for spatial fraud detection and temporal graph networks (TGNs) for tracking sequential fraud patterns. Semi-supervised and reinforcement learning mechanisms enhance adaptability to evolving fraud tactics. Experiments on real-world datasets show that the proposed model outperforms traditional methods, achieving higher accuracy and lower false positives. Its effectiveness in detecting multi-step fraud rings and synthetic transactions underscores the potential of graph-based deep learning in securing e-commerce platforms.