Using GNN to detect financial fraud based on the related party transactions network

Abstract

Financial fraud cases will greatly destroy the trust of market participants and affect the stability of the financial system. Existing research shows that the existence of related party transactions will increase the risk of financial fraud. Therefore, related party transactions may contain information that can help identify fraud. Existing studies have introduced simple statistics of related party transactions into fraud detection. However, with the increasing complexity of related party transaction networks, it is difficult for traditional methods to describe the whole picture of transactions, so the performance of fraud detection is limited. Therefore, this paper innovatively introduces a heterogeneous graph to connect listed companies and their related parties to form a complete picture of transactions, and then utilizes the graph neural network to make predictions. Experiments conducted on the data of Chinese listed companies in 2020 show that the introduction of related party transactions can improve the recognition effect. What's more, the related party transactions network can further improve the performance of fraud detection.