Improving fraud detection via imbalanced graph structure learning

Graph-based fraud detection methods have recently attracted much attention due to the rich relational information of graph-structured data, which may facilitate the detection of fraudsters. However, the GNN-based algorithms may exhibit unsatisfactory performance faced with graph heterophily as the fraudsters usually disguise themselves by deliberately making extensive connections to normal users. In addition to this, the class imbalance problem also causes GNNs to overfit normal users and perform poorly for fraudsters. To address these problems, we propose an Imbalanced Graph Structure Learning framework for fraud detection (IGSL for short). Specifically, nodes are picked with a devised multi-relational class-balanced sampler for mini-batch training. Then, an iterative graph structure learning module is proposed to iteratively construct a global homophilic adjacency matrix in the embedding domain. Further, an anchor node message passing mechanism is proposed to reduce the computational complexity of the constructing homophily adjacency matrix. Extensive experiments on benchmark datasets show that IGSL achieves significantly better performance even when the graph is heavily heterophilic and imbalanced.