Time Encoding Graph Attention Model for Financial Fraud Detection in Large-scale Financial Social Networks
Yuqi Wang, Hualin Zhan, Weinan Jiang · 2024
Graph data, ubiquitous across various domains, serves as a rich source of information. Anomaly detection within graphs has become a focal point of research, particularly in the financial sector, where fraud-related issues are escalating. However, detecting fraudulent activities in large-scale financial social networks presents a formidable challenge due to the vast network scale, complex user interactions, and dynamic changes. Traditional methods need to adapt to the evolving nature of fraudulent activities. This study proposes a graph attention model based on time-encoding functions in response to this challenge. This model harnesses both static and dynamic features of nodes within financial social networks. It effectively captures the temporal dynamics between nodes through time encoding functions and integrates the multi-head attention mechanism from Graph Attention Networks (GAT) to comprehensively aggregate neighbourhood information. Experimental results conducted on the real-world, large-scale financial social network dataset, DGraph, validate the model's outstanding performance in fraud detection tasks. This research introduces an innovative and practical approach, providing a new perspective for effectively addressing fraud issues within financial social networks.