A SHAP-BASED INTERPRETABLE ETA DELAY PREDICTION FRAMEWORK FOR LAST-MILE LOGISTICS

Abstract

The precise estimation of the Estimated Time of Arrival (ETA) is a key aspect of last-mile logistics. This directly impacts customer satisfaction, efficiency, and decision-making in a transport environment. The classical ETA estimation models are distance-based and have limitations in capturing dynamic functional and real-world conditions such as congestion due to traffic, peak timing, multiple delivery issues, and weather. This work aims to devise a machine learning framework with strong predictive performance for the estimation of delivery delay severity (regression model) and the probability of delay occurrence (binary classification model) using a simulated dataset of logistics shipments. Five regression models – Linear Regression, Ridge Regression, LASSO Regression, Decision-boosting Regression, and Decision-forest Regression models – and five different classification models – Logistic Regression Classifiers, Decision-boosting Classifier, Decision-forest Classifier, RFB-Support Vector Machine Classifier, and K-Nearest Neighbour Classifier – are trained and tested on this dataset using multiple figures of merit. The performance on this dataset reveals that linear models are superior classifiers compared to decision-based models in estimating delay duration with a mean absolute error value of 4.8 minutes for all models. The delay probability classifiers perform extremely well on this dataset with Logistic Regression achieving well-balanced characteristics. In addition, SHAP-based Explainable AI Analysis integration is also included for understanding model output at a global and local perspective. Results and findings concerning Explainability Analysis show that traffic intensity, peak conditions, and stops are identified as top variables responsible for ETA Delay.