Intelligent Supply Chain Risk Identification via Heterogeneous Graph Learning and Multi-Entity Interaction Modeling

Abstract

To address the challenges of concealed abnormal behavior, complex risk correlation paths, and the inability of traditional methods to adequately characterize multi-entity interaction structures in intelligent supply chain operations, this paper proposes a heterogeneous graph neural network anomaly detection framework for risk identification tasks. This method uses core business objects such as orders, customers, products, and logistics as modeling units, unifying the representation of various types of entities and their relationships in the supply chain into a heterogeneous graph structure. Type-specific projection is used to achieve a unified mapping of the heterogeneous attribute space. Based on this, a relationship-aware message passing mechanism is combined to differentially aggregate neighborhood information under different semantic relationships. Furthermore, an adaptive relationship fusion strategy is utilized to enhance the expressive power of key risk correlations, thereby improving the model's ability to identify potential abnormal patterns in complex business scenarios. A complete methodological process is constructed around the supply chain anomaly detection task, including key steps such as heterogeneous graph construction, node representation learning, relationship-level information propagation, cross-relationship feature fusion, and anomaly probability output, enabling the model to mine risk features from both structural dependencies and semantic relationships. Validation is conducted using publicly available supply chain data. The results show that the proposed method exhibits good overall performance across multiple evaluation metrics, demonstrating that the framework can effectively improve the accuracy and stability of supply chain anomaly identification. This study provides a feasible approach for risk identification and intelligent analysis in complex supply chain environments, oriented towards multi-entity relationship modeling, and also provides new technical support for the application of graph learning methods in supply chain management scenarios.