Enhancing financial fraud detection with graph neural network and ensemble learning: insights from Related Party Transactions network

Related party transactions (RPTs) reflect business exchanges between firms and related entities built on trust and are susceptible to manipulation, therefore providing important signals for fraud detection. Owing to the structural complexity inherent in RPT ties, prior work has seldom incorporated these relations into fraud-detection tasks or modeled their higher-order network patterns. This study introduces the anomaly interaction of related parties to financial fraud detection, which provides a new perspective and improves the accuracy. Methodologically, this study proposes a stacking strategy to combine interaction features extracted from GNN with conventional financial indicators using base learners to improve detection performance. Based on data from 3793 listed companies, 68571 related companies, and 6072 related individuals, the empirical study constructs a multi-relational RPT network that reflects higher-order and cross-entity interactions beyond firm-level signals, the results show that including RPT information increases the AUC by 1.67% compared to using only financial data, and leveraging structure-aware representations of RPT further raises it by 3.04%. Robustness checks across different periods support the validity of the approach.