Supervised Learning for Arrival Time Estimations in Restaurant Meal Delivery

Abstract

Restaurant meal delivery companies have begun to provide customers with meal arrival time estimations to inform the customers’ selection. Accurate estimations increase customer experience, whereas inaccurate estimations may lead to dissatisfaction. Estimating arrival times is a challenging prediction problem because of uncertainty in both delivery and meal preparation process. To account for both processes, we present an offline and online-offline estimation approaches. Our offline method uses supervised learning to map state features directly to expected arrival times. Our online-offline method pairs online simulations with an offline approximation of the delivery vehicles’ routing policy, again achieved via supervised learning. Our computational study shows that both methods perform comparably to a full near-optimal online simulation at a fraction of the computational time. We present an extensive analysis on how arrival time estimation changes the experience for customers, restaurants, and the platform. Our results indicate that accurate arrival times not only raise service perception but also improve the overall delivery system by guiding customer selections, effectively resulting in faster delivery and fresher food.