Supply Chain Risk Mitigation Through the Integration of Supervised Learning Algorithms and Linear Programming

Abstract

Late delivery risk remains a critical operational challenge in modern supply chains, contributing to higher costs, degraded service levels, and diminished customer trust. This study proposes a hybrid predictive-prescriptive framework that integrates supervised machine learning with Linear Programming (LP) to support supply chain risk mitigation decisions. Following the CRISP-DM methodology, five supervised learning algorithms-Logistic Regression, Random Forest, XGBoost, LightGBM, and CatBoost-were developed and compared using the DataCo Smart Supply Chain dataset to estimate late delivery risk probability. The resulting probabilities were then embedded as risk parameters within an LP model to optimize inventory allocation, supplier selection, and shipment planning across four scenarios: Conventional, ML Only, LP Only, and the proposed Hybrid ML-LP approach. Random Forest achieved the best predictive performance (F1-score = 0.746; ROC-AUC = 0.836), outperforming the other four models, while SHAP and permutation importance analyses identified destination-related geographic features and abnormal order status as the dominant predictors of delivery delay. The optimization results show that LP-based scenarios reduced total operational cost by more than 90% relative to the Conventional and ML Only scenarios, and sensitivity analysis confirmed that the hybrid model's cost solution remained stable under ±10-30% parameter perturbations. An anomaly identified in the service-level and average-risk metrics for the LP-based scenarios, however, requires further verification before being reported as a definitive finding. Overall, this study demonstrates that combining supervised learning with mathematical optimization offers a promising, cost-effective, and replicable decision-support framework for proactive supply chain risk mitigation.