TG ‐ RRNet : A Supply Chain Risk Perception Network Integrating Temporal Modeling and Generative Anomaly Detection

In dynamically evolving supply chain networks, identifying high‐risk nodes and abnormal behavior patterns is crucial for risk early warning and system stability. Existing methods mainly rely on static graph modeling or discriminative learning, which struggle to capture temporal evolution and often fail to detect “camouflaged normal” risk nodes under feature obfuscation. To address this, we propose the Temporal‐Generative Relation‐aware Risk Network (TG‐RRNet) to systematically tackle key challenges in dynamic high‐risk node identification. TG‐RRNet first constructs a time‐driven heterogeneous dynamic graph sequence, integrating three types of multimodal information including attribute similarity, historical interaction intensity, and historical risk factor to model the structural evolution process. A temporal‐aware graph neural network with temporal decay and graph attention extracts dynamic node representations and captures risk propagation paths. To model latent abnormal patterns, a generative anomaly detection module uses a variational autoencoder to learn latent representations and jointly measures potential risks through reconstruction errors and KL divergence. Finally, a multimodal cross‐attention mechanism dynamically fuses structured features, graph representations, and unstructured logs to generate unified risk representations for prediction. Experiments on real‐world supply chain datasets show that TG‐RRNet significantly outperforms state‐of‐the‐art methods in high‐risk node identification and anomaly detection, demonstrating strong practical value and generalization. Code is available at: https://github.com/PinmengLi/TG‐RRNet.git .