Subgraph Patterns Enhanced Graph Neural Network for Fraud Detection

Abstract

Credit card fraud is a serious problem that causes significant losses for cardholders and issuing banks. Current detection methods often focus on spatial-temporal anomalies in transactions but tend to overlook the disguising techniques of fraudsters. We notice that covert credit card fraud activities frequently manifest localized clustering characteristics, which are particularly evident in different subgraph patterns. To address this, we propose Subgraph Patterns enhanced Graph Neural Network (SPGNN), a novel approach that effectively harnesses specific traits to significantly improve fraud detection capabilities. This method employs subgraph pattern features to more sharply distinguish between fraudulent and legitimate nodes, aiding in the identification of disguised fraudsters. In particular, we devise a probabilistic neighbor selector to assist nodes in selecting more similar minority class nodes, effectively balancing data distribution and filtering out disguised nodes. Furthermore, we introduce a reinforcement learning module for supervised similarity measurement, further filtering out disguised fraudsters. Extensive experiments on several benchmark datasets demonstrate that SPGNN surpasses state-of-the-art models in detecting fraudulent activities, achieving the most advanced performance.

Sheng Xiang

Ph.D. Candidate

My research interests include learning-based graph simulation for data management and spatio-temporal and uncertain data mining.