Combining Machine Learning Algorithms and Exact Methods for Solving the Dynamic Vehicle Routing Problem

Abstract

Last-mile delivery is one of the costliest and most challenging stages of the modern supply chain, particularly as e-commerce demand grows and customer expectations for fast, reliable shipping become increasingly stringent. Understanding the patterns and determinants of delivery delay is an essential first step toward developing effective machine learning based prediction and optimization strategies for dynamic logistics operations. This study analyzes the DataCo Smart Supply Chain dataset, comprising 180,519 transaction records from 65,752 unique orders and 20,652 unique customers across five global market regions between January 2015 and September 2017. A descriptive and correlational analysis was conducted to examine sales trends, market and product-category performance, customer segmentation, order status, and delivery-risk patterns, followed by a Pearson correlation analysis of key operational variables. The results show that 54.83% of all orders were at risk of late delivery, with First Class and Second Class shipping modes exhibiting a counterintuitively higher delay risk (95.3% and 76.6%) than Standard Class (38.1%), suggesting that these labels reflect service-tier classification rather than delivery speed. Europe and LATAM emerged as the largest contributors to sales and profit, while outdoor-sport categories, particularly Fishing, Cleats, and Camping & Hiking, dominated both sales and profitability. Correlation analysis further revealed that profit was only weakly associated with sales and discount rate, indicating that profitability is driven more by cost structure than by revenue volume or pricing policy. These findings offer empirical evidence of the operational bottlenecks and behavioral patterns within the supply chain, providing a diagnostic foundation for developing machine learning prediction models and optimization frameworks for last-mile logistics.