A data-driven preference learning approach for multi-objective vehicle routing problems in last-mile delivery

Abstract

Last-mile delivery service providers and drivers often choose routes deviating from the shortest distance, influenced by personal preferences and various business-level factors. This study introduces an innovative data-driven optimization approach for learning preferences of decision-makers (DMs) in multi-objective vehicle routing problems. Utilizing real-world historical data from a last-mile delivery logistics platform, we develop a machine-learning framework to learn DMs’ preferences in designing delivery routes. To design our approach we focus on a multi-objective capacitated vehicle routing problem with time windows and develop an integrated framework that combines supervised learning models, sampling techniques, and optimization methods to determine preference weights for objective functions based on selected features. We conduct extensive numerical experiments to test the proposed data-driven optimization approach. Our findings suggest that analyzing historical planned and actual routes reveals DMs’ preferences, such as prioritizing workload balance and minimizing fleet usage over travel distance alone. Furthermore, this study offers insights into key factors shaping last-mile delivery logistics, including workload distribution and deviations from pre-planned routes, enabling more informed and human-centered decision-making in logistics optimization.