This paper considers how an online food delivery platform can improve last-mile delivery services’ performance using multi-source data. The delivery time is one critical but uncertain factor for online platforms that also regarded as the main challenges in order assignment and routing service. To tackle this challenge, we propose a data-driven optimization approach that combines machine learning techniques with capacitated vehicle routing optimization. Machine learning methods can provide more accurate predictions and have received increasing attention in the operations research field. However, different from the traditional predict-then-optimize paradigm, we use a new smart predict-then-optimize framework, whose prediction objective is constructed by decision error instead of prediction error when implementing machine learning. Using this type of prediction, we can obtain a more accurate decision in the following optimization step. Efficient mini-batching gradient and heuristic algorithms are designed to solve the joint order assignment and routing problem of last-mile delivery service. Besides, this paper considers the mutual effect between routing decision and delivery time, and provides the corresponding solution algorithm. In addition, this paper conducts a computational study and finds that the proposed method’s performance has an approximate 5% improvement compared with other methods.

本文考虑了在线食品配送平台如何使用多源数据改善最后一英里配送服务的绩效。交付时间是在线平台的一个关键但不确定的因素，在线平台也被视为订单分配和路由服务的主要挑战。为了解决这一挑战，我们提出了一种数据驱动的优化方法，该方法将机器学习技术与功能强大的车辆路线优化相结合。机器学习方法可以提供更准确的预测，并且在运筹学领域受到越来越多的关注。但是，与传统的“预测然后优化”范式不同，我们使用了一个新的智能“预测然后优化”框架，该框架的预测目标是在实现机器学习时由决策错误而不是预测错误构造的。使用这种类型的预测，我们可以在以下优化步骤中获得更准确的决策。设计了有效的小批处理梯度和启发式算法，以解决最后一英里交付服务的联合订单分配和路由问题。此外，本文还考虑了路由决策与交付时间之间的相互影响，并提出了相应的求解算法。此外，本文进行了计算研究，发现与其他方法相比，该方法的性能提高了约5％。本文考虑了路由决策与交付时间之间的相互影响，并提出了相应的求解算法。此外，本文进行了计算研究，发现与其他方法相比，该方法的性能提高了约5％。本文考虑了路由决策与交付时间之间的相互影响，并提出了相应的求解算法。此外，本文进行了计算研究，发现与其他方法相比，该方法的性能提高了约5％。