Key points are not available for this paper at this time.
Global supply chains are increasingly exposed to multi-source uncertainties, ranging from geopolitical tensions to climate extremes, making the accurate and interpretable prediction of disruptions an urgent operational priority. Existing predictive models often rely on either shallow statistical learners, which struggle with high-dimensional interactions, or deep neural networks, which trade off interpretability for marginal performance gains. To address this gap, we propose an interpretable machine learning framework that couples a feature-attention mechanism with a gradient-boosted decision tree ensemble for early warning of shipment-level disruption events. First, a dedicated attention module is trained to assign importance weights to 14 heterogeneous risk factors, generating an interpretable feature ranking that highlights pivotal signals such as lead-time volatility and geopolitical risk. The reweighted features are then fed into a gradient boosting classifier, which effectively captures non-linear patterns and interaction effects. Evaluated on a publicly available dataset of 5000 international freight records available on Kaggle, the proposed framework achieves an AUC of 0.8213 (±0.0002 over three independent runs), matching the best-performing baseline (standard gradient boosting, 0.8212 ± 0.0001) and surpassing logistic regression (0.777), random forest (0.806), and a standalone feature-attention network (0.805). The attention module preserves full predictive accuracy while adding an interpretability layer that conventional black-box implementations lack. Notably, the framework preserves the predictive accuracy of gradient boosting while enhancing interpretability through attention-based feature ranking and dual-perspective importance analysis, achieving a precision of 0.770 and a balanced F1-score of 0.781. The convergence of attention-based interpretability and ensemble learning efficiency provides supply chain managers with a transparent decision-support tool—distinguishing genuine “tipping points” from “false alarms” and enabling targeted risk mitigation under deep uncertainty.
Wang et al. (Mon,) studied this question.