Interpretability of Machine Learning Models in Delivery Time Predictions

Abstract

Accurate prediction of delivery times is crucial for logistics companies to enhance customer satisfaction and optimize operations. This study employs Çnterpretable Çachine Çearning Çodels to predict delivery times using various features related to shipments. The ethodology Çncludes data preprocessing, feature extraction, outlier de-tection, Çodel training, and hyperparameter optimization. A particular focus s placed on model interpretability using both SHapley Additive exPlanations (SHAP) and Local Çnter-pretable Model-agnostic Çxplanations (LIME) analyses. The results demonstrate significant improvements in prediction accuracy with the sequential addition of feature groups, with Çandom Forest emerging as the bestperforming model after hyperparameter tuning. Feature Çmportance analysis highlights the ist influential factors, while SHAP and LIME analyses provide complementary insights Çnto how different features impact the predictions at both global and Çocal Çevels, enhancing the transparency of the model. comprehensive comparison of various machine learning and deep learning models is presented, along with practical Çmplications for ogistics anagement. This study contributes to the field by demonstrating the effectiveness of combining advanced machine learning techniques with interpretability methods nn logistics applications. © 2025 Elsevier B.V., All rights reserved.