Interpretable Graph-Based Anomaly Detection for Related-Party Transactions in Auditing Systems

Abstract

In corporate auditing and risk management scenarios, related-party transactions often exhibit characteristics such as long relationship chains, numerous participating entities, and complex structural patterns. Furthermore, the scarcity and concealment of abnormal events make it difficult for traditional rule-based or single-point indicator-based methods to reliably detect high-risk clues. To address this, this paper constructs a heterogeneous relationship network for corporate related-party transaction data and proposes an interpretable subgraph representation and anomaly identification method for auditing. First, entity disambiguation and related-party alignment reduce pseudo-structural differences caused by naming fragmentation. Then, multi-scale candidate subgraphs are sampled from the transaction network to learn a subgraph representation that balances structure and attributes. Finally, an interpretable mechanism is used to output key evidence edges/core participating entities, forming an auditable risk evidence chain. Experimental results show that the proposed method outperforms mainstream baselines on multiple detection and interpretation metrics, significantly improving evidence readability and audit usability while maintaining detection performance. This provides a reliable technical path for uncovering corporate fraud clues and providing risk warnings.