Intelligent audit decision support for enterprise related-party transactions based on knowledge graph and graph neural network.

Related-party transactions (RPTs) occupy a central position in contemporary corporate governance, serving as essential channels for resource allocation among affiliated business entities. The growing complexity of ownership arrangements and globalized operations has made RPT auditing increasingly challenging for financial professionals. Traditional audit methodologies struggle with intricate enterprise relationship networks that often span multiple jurisdictions and involve layered ownership through shell companies, rendering manual inspection approaches both time-consuming and incomplete. Despite advances in computational approaches, significant gaps persist in current research and practice. Existing audit systems inadequately capture the multi-dimensional relationships connecting enterprises, shareholders, executives, and transactions. Most machine learning methods treat transactions as independent observations, thereby ignoring the relational context that distinguishes legitimate RPTs from irregular ones. Furthermore, the dynamic nature of enterprise relationships demands adaptive models capable of tracking temporal evolution, yet static snapshots of relationship networks may produce outdated or misleading risk assessments. Four critical challenges must be addressed for effective RPT audit automation: modeling heterogeneous entity types linked through diverse relationship categories with different audit implications; capturing temporal dynamics as ownership structures change through acquisitions and personnel movements; providing interpretable outputs that support professional judgment and regulatory inspection; and integrating domain knowledge with data-driven learning in a coherent framework. This study develops an intelligent decision support system that bridges knowledge graph technology with graph neural network architectures to address these challenges. The proposed framework constructs a domain-specific knowledge graph capturing multi-dimensional enterprise relationships through entity resolution and temporal modeling. An enhanced graph neural network employs heterogeneous attention mechanisms alongside temporal fusion components to learn relationship-aware representations. Interpretable risk assessment emerges through attention weight visualization and risk propagation path analysis. Experimental evaluation on real-world enterprise data demonstrates substantial improvements over baseline methods, with practical deployment confirming utility for detecting undisclosed relationships and pricing anomalies in accounting firm engagements.