GRAPH NEURAL NETWORKS FOR FRAUD DETECTION: COMPARATIVE ANALYSIS OF METHODS

Abstract

The study focuses on the use of graph neural networks (GNN) to detect fraudulent transactions in the banking system. Modern methods of financial data analysis face the problem of low efficiency when processing large amounts of information, which makes it more difficult to identify abnormal and fraudulent transactions. The aim of the work is to evaluate the effectiveness of various GNN architectures in the context of banking transactions, as well as analyze their applicability to improve accuracy and reduce false alarms in security monitoring processes. In the course of the work, several models were investigated, among which the LGM-GNN architecture demonstrated the greatest efficiency, which showed the best results in terms of accuracy and completeness. The results confirm the possibility of using graph neural networks to solve the problem of fraud detection, which opens up prospects for their implementation in real financial monitoring systems. The conclusion of the paper includes recommendations for improving algorithms and directions for future research in the field of graph models for financial data analysis