Enhancing financial fraud detection with hierarchical graph attention networks: A study on integrating local and extensive structural information

Abstract

In the highly competitive business environment, financial fraud detection plays a crucial role in safeguarding companies and investors. However, traditional methods face challenges when assessing the complex and multidimensional nature of financial data. Cutting-edge machine learning techniques, particularly the hierarchical graph attention network (HGAT), emerge as a promising approach for financial fraud detection. The proposed approach includes encoding adjacency matrices for capturing local relationships and utilizing multi-head self-attention to propagate structural attributes across multiple layers. Node embeddings are generated by the HGAT model, which integrates both local and extensive structural information through multihead self-attention. Through learning intricate inter-entity relationships, the HGAT model can effectively identify potential financial risks. Experiments performed on a publicly available financial report dataset demonstrate the superior performance of our model compared to the existing methods in detecting financial risks.