FinGuard-GNN: Dynamic Graph Neural Network Framework for Financial Fraud Detection

Abstract

With the significant increase in financial fraud incidents, financial fraud detection has become a critical research area. Complex financial relationship networks involving thousands or even millions of nodes present enormous challenges for fraud detection tasks. Although researchers have developed various graph-based methods to detect fraudulent behavior within these complex networks, existing approaches overlook two key issues in fraud graphs: the diversity of non-additive attributes and the distinguishability of grouped message passing from neighboring nodes. This paper proposes FinGuard-GNN (Financial Guardian Graph Neural Network), a novel dynamic graph neural network for financial fraud detection that addresses the aforementioned issues through innovative feature transformation strategies and a Cascaded Risk Diffusion (CRD) mechanism. For feature transformation, we implement Adaptive Tree Partitioning (ATP) encoding and Statistical Evidence Weighting (SEW) encoding to convert various types of non-additive node attributes into vector representations suitable for GNN aggregation operations, avoiding the generation of meaningless features while maintaining strong interpretability. For risk propagation, we design a feedback-based Cascaded Risk Diffusion strategy that enables dynamic accumulation and decay of risk information across the network. Additionally, we develop a Responsive Group Allocation (RGA) strategy that divides graph nodes into distinct groups followed by hierarchical aggregation, enhancing the distinguishability of fraudulent nodes. Experiments on two classic financial fraud datasets demonstrate that our proposed method achieves superior discriminative capability for fraudulent nodes compared to traditional graph algorithms and machine learning methods. The experimental results confirm the advantages of FinGuard-GNN in handling non-additive features in complex financial networks, improving node distinguishability, and capturing hierarchical risk propagation, providing a novel solution for the fi.