An XGBoost-enhanced fast constructive algorithm for food delivery route planning problem

Highlights

• Food delivery route planning problem is studied.

• XGBoost-enhanced fast constructive algorithm is proposed.

• Insertion-based heuristic with different sequencing rules is embedded.

• Adaptive selection mechanism using XGBoost is proposed.

• Effectiveness is validated on datasets from Meituan food delivery platform.

Abstract

As e-commerce booms, online food ordering and delivery has attracted much attention. For food delivery platforms, planning high-quality routes for drivers so as to accomplish the delivery tasks efficiently is of great importance. This paper addresses a food delivery route planning problem (FDRPP), which considers one driver delivering multiple orders from restaurants to customers. Due to the immediacy of the delivery tasks, very limited computational time is provided for generating satisfactory solutions. We mathematically formulate the FDRPP and propose an Extreme Gradient Boosting-enhanced (XGBoost-enhanced) fast constructive algorithm to solve the problem. To construct a complete route, an insertion-based heuristic with different sequencing rules is adopted, together with an acceleration strategy based on geographic information to speed up the insertion procedure. In order to avoid the waste of computational time, we design an adaptive selection mechanism to select sequencing rules for route construction. A classification model using XGBoost is established to predict the performance of different sequencing rules. Through analysis of the route construction procedure, three types of problem-specific features are designed to improve the performance of XGBoost. The effectiveness of the proposed algorithm is demonstrated by conducting experiments on datasets from Meituan food delivery platform, which shows that large amounts of computational time can be saved by our proposed algorithm, while guaranteeing the quality of solutions.

Introduction

With the development of e-commerce, modern life has become more and more convenient recently. People can enjoy a large diversity of services without walking outdoors, among which online food ordering and delivery service occupies a large portion. According to the report from Meituan and CFLP (2019), 70% of the online orders are generated by food delivery platforms. An earlier research by Morgan Stanley Research (2017) predicted that online food delivery was expected to grow by 16% annual compound rate in the next 5 years in the US. Taking Meituan, the largest food delivery platform in China, as an example, the maximal number of transaction orders in a single day has risen up to 30 million and the total number of transaction orders through 2019 has amounted to 18 billion (Meituan & CFLP, 2019). Worldwide food delivery companies such as Grubhub, Uber Eats, and Just Eat, are also developing fast with transaction amounting up to 94 billion dollars (Hirschberg, Rajko, Schumacher, & Wrulich, 2016). Stimulated by such a great market opportunity, online food ordering and delivery is developing and prospering very quickly.

While providing convenience and making considerable profits, food delivery platforms have to face with tremendous pressure. Not only the management of company staff, but also the performance of the dispatching system needs to accommodate to the increasing number of requests and service quality required by customers. Fig. 1 illustrates the general operation of the dispatching system in Meituan, where firstly customers use food ordering application on cellphone to select food they desire. After that, the dispatching system collects orders and pushes them to the restaurants and drivers. Restaurants prepare the food and drivers shuttle between restaurants and customers to finish the delivery tasks. To guarantee the well functioning of the dispatching system, order assignment and food delivery route planning need to be considered in decision-making. The former dispatches orders to appropriate drivers and the latter organizes a feasible route for every single driver. During peak hours of a day, explosive number of orders are generated, which brings difficulty in finding satisfactory solutions from enormous solution space. Meanwhile, due to the immediacy of delivery tasks, dispatching schemes need to be determined in a very short period of time. Therefore, it is necessary to design highly efficient algorithms for the dispatching system, considering both academic significance and practical needs.

In this paper, we study the food delivery route planning problem (FDRPP), which is more fundamental and essential than order assignment since it serves as an indispensable part to evaluate an order assignment scheme. According to statistics from Meituan, the food delivery route planning module is called most frequently in the whole dispatching system, nearly 5 billion times a day. Therefore, saving computational time of the food delivery route planning algorithm is of great significance. Besides, the solution quality of FDRPP also needs to be promised since it directly influences the satisfaction of both drivers and customers. A well-planned route can not only assign the priority of orders properly to optimize the customer satisfaction, but also shorten the travel distance of the route as much as possible to reduce the burden of the driver. Hence, FDRPP needs to be solved effectively and efficiently to improve the performance of the dispatching system.

However, most traditional algorithms are not suitable for solving FDRPP since they usually consume too much computational time to reach satisfactory results. Under this situation, we develop an Extreme Gradient Boosting-enhanced (XGBoost-enhanced) fast constructive algorithm, denoted as XGB-FCA, to solve the problem efficiently. XGB-FCA functions as a hybrid algorithm which combines the machine learning technology and traditional constructive heuristics. It can save large amounts of computational time while guaranteeing the quality of solutions. To be specific, we adopt an insertion-based heuristic with different sequencing rules to construct high-quality routes. An acceleration strategy based on geographic information is used to speed up the insertion procedure. To further save computational time and keep the solution quality at the same time, we design an adaptive selection mechanism to select sequencing rules for route construction. A classification model based on XGBoost is established to predict the performance of different sequencing rules. Moreover, problem-specific features are designed to improve the performance of XGBoost.

The remainder of the paper is organized as follows. In Section 2, we briefly review the related literature related to FDRPP. Section 3 introduces the notation and formulation of FDRPP. Section 4 describes the details of the proposed XGB-FCA. Experimental results on datasets from Meituan food delivery platform are shown in Section 5. Finally, we close the paper with some conclusions and future research ideas in Section 6.