A dynamic risk assessment framework is proposed to address the complex challenge of concurrent logistics disruptions and policy compliance issues in cross-border supply chains. The framework integrates the Light Gradient Boosting Machine (LightGBM), the Multi-Scale Graph Attention (MGAT) network, and the Temporal Convolutional Network (TCN). This integration effectively captures the nonlinear coupling relationship between the evolving network topology and multimodal data. Utilizing order data from a Brazilian retailer, the model extracts node embedding features and time-series features through an incremental learning mechanism. Furthermore, hierarchical Shapley value analysis identifies critical risk transmission paths, and a deep reinforcement learning strategy engine optimizes risk mitigation solutions. The experimental results show that the model achieves a prediction accuracy of 92.5%, significantly outperforming traditional methods. The proposed strategy engine reduces the risk response time to 15 min, effectively mitigating risks while maintaining operational cost control.
Xi et al. (Thu,) studied this question.