Dynamic heterogeneous graph neural networks for anomaly detection in supply chains

Supply chain anomaly detection is a critical task in supply chain management, aiming to enhance system stability and resilience by identifying potential anomalies in the supply chain network in a timely manner. However, the dynamic changes in nodes and relationships, along with the heterogeneous characteristics of the network, pose significant challenges for anomaly detection. Existing approaches struggle to capture the complex multi-relational interactions in dynamic networks. Furthermore, most studies focus on detecting node anomalies while underemphasizing edge anomalies, thus failing to fully reflect the complexity of both node and relational aspects in the supply chain. This paper proposes a dynamic heterogeneous graph neural network-based approach for supply chain anomaly detection, which jointly models the temporal dynamics and multi-relational heterogeneity of the supply chain network to achieve comprehensive anomaly detection for both nodes and edges. The approach extracts multi-relational features from the network via a message-passing module, incorporates temporal evolution information through a dynamic feature update module, and calculates anomaly scores for nodes and edges using a dynamic-heterogeneous joint reconstruction module. To help users interpret the detection results, the paper also introduces feature-level and temporal-level contribution analysis, with visual analytics for anomalous nodes and edges as auxiliary support. Results demonstrate that, on publicly available supply chain network datasets, the proposed dynamic heterogeneous graph neural network approach significantly outperforms baseline models. This study offers a novel and practical solution for supply chain management.