Modeling Practice of Graph Neural Networks in Related Party Transaction Recognition and Financial Fraud Detection

Detecting financial fraud in Related Party Transactions (RPTs) remains a critical challenge for financial institutions due to complex corporate structures, obfuscation tactics, and limitations of conventional rule-based detection systems. Traditional methods fail to capture hidden relational patterns and temporal dynamics, leading to undetected fraud and regulatory gaps. Hence, the research introduces a Heterogeneous Temporal Graph Neural Network for Related Party Transactions (HTGNN-RPT) to improve the detection and recognition of fraud in RPTs. The model operates in three key phases: 1) constructing a heterogeneous financial graph incorporating entities, transactions, and temporal attributes to represent RPT networks; 2) applying dynamic attention to capture time-sensitive relational dependencies; and 3) employing anomaly detection to flag suspicious subgraphs while ensuring interpretability for regulatory compliance. The experimental validation shows that HTGNN-RPT outperforms existing models with an increase of 8.3% accuracy, 8.57% in F1-score, 6.7% improvement in transaction volume, and 8.97% improvement in AUC-ROC for edge timestamp density. It also achieves 6.83% and 7.40% improvements in explainability across time windows and detection thresholds, respectively. Additionally, HTGNN-RPT surpasses RPT by 6.2% in edge count and 8.1% in time window size, demonstrating superiority in dynamic fraud detection scenarios. The research advances financial fraud detection with a scalable HTGNN-RPT model.