An Adaptive Graph-Based Oversampling Framework for Fraud Detection

Graph Neural Networks have demonstrated significant success in fraud detection tasks. However, most existing approaches, which rely on distinct message-passing mechanisms to aggregate node information, face two major challenges: the inherent label imbalance between fraudsters and the homophily-heterophily mixture in graph structural data. In this paper, we address the label imbalance by applying oversampling techniques to generate additional training nodes and edges. We propose an adaptive Oversampling Fraud Detection (OFD) framework that generates synthetic nodes and edges for the minority fraud classes. Furthermore, we employ an adaptive message aggregation technique to address the homophily-heterophily issue. Extensive experiments on three real-world datasets demonstrate that our method significantly improves detection accuracy and outperforms several state-of-the-art approaches, offering a robust solution to the label imbalance problem in fraud detection tasks.