Abstract
As corporate operations become increasingly networked and digitalized, fund flows, business interactions, and control relations intertwine across multiple entities to form complex transaction networks. Accounting fraud, therefore, exhibits more structured, collaborative, and concealed characteristics, which pose substantial challenges to traditional methods based on individual features or rule-driven detection. This study addresses accounting fraud detection in complex corporate transaction networks by modeling firms and their transaction relations within a unified graph structure. A graph neural network is introduced to jointly model node attributes and relational context. Fraudulent behavior is characterized from a structural perspective through its organizational patterns and risk propagation paths in the network. A multi-layer message passing mechanism integrates local transaction features with higher-order structural dependencies. The model can automatically learn latent anomaly patterns and relational cues without relying on manual rules. On this basis, a unified risk scoring framework maps structured representations to entity-level fraud risk intensity. Systematic identification of abnormal behavior in complex transaction environments is thus achieved. Comparative evaluation with representative methods shows that the proposed approach delivers more stable overall discrimination and risk differentiation. It effectively reduces misclassification and improves the detection of concealed fraudulent entities. Further analysis confirms the suitability of this structured modeling paradigm for complex relational scenarios. Integrating transaction network semantics with deep representation learning helps overcome the expressive limitations of traditional accounting analysis. This work provides a unified network-oriented modeling approach for intelligent auditing and financial risk analysis and offers a valuable reference for anomaly detection in complex economic systems.
