Problem definition: Online Food Delivery (OFD) platforms have witnessed rapid global expansion, partly driven by shifts in consumer behavior during the COVID-19 pandemic. These platforms enable customers to order food conveniently from a diverse array of restaurants through their mobile phones. A core functionality of these platforms is the algorithmic matching of drivers to food orders, which is the focus of our study as we aim to optimize this driver-order matching process.
Methodology/results: We formulate real-time matching algorithms that take into account uncertain food processing times to strategically ‘delay’ the assignment of drivers to orders. This intentional delay is designed to create a ‘thicker’ marketplace, increasing the availability of both drivers and orders. Our algorithms utilize machine learning techniques to predict food processing times, and the dispatching of drivers is subsequently determined by balancing costs for idle driver waiting and for late deliveries. In scenarios with a single order in isolation, we show that the optimal policy adopts a threshold structure. Building on this insight, we propose a new k-level thickening policy with driving time limits for the general case of multiple orders. This policy postpones the assignment of drivers until a maximum of k suitable matching options are available. We evaluate our policy using a simplified model and identify several analytical properties, including the quasi-convexity of total costs in relation to market thickness, indicating the optimality of an intermediate level of market thickness. Numerical experiments with real data from Meituan show that our policy can yield a 54% reduction in total costs compared to existing policies.
Managerial implications: Our study reveals that incorporating food processing times into the dispatch algorithm remarkably improves the efficacy of driver assignment. Our policy enables the platform to control two vital market parameters of real-time matching decisions: the number of drivers available to pick up and deliver an order promptly, and their proximity to the restaurant. Based on these two parameters, our algorithm matches drivers with orders in real-time, offering significant managerial implications.
