Who Ate the Lunch? A Variational Bayes Algorithm for Identifying Refund Hunters

For online platforms such as UberEats and DoorDash, a central challenge in customer service is the lack of ground truth — when a customer reports that an order was never received, it is difficult for support agents to determine if the driver kept the food, a passer-by took the delivery, or the customer is falsely claiming a missing order. This fundamental uncertainty often results in platforms shouldering refunds and appeasement costs without holding either side of the market accountable. In this work, we propose a variational Bayesian (VB) algorithm for identifying strategic customers and drivers, considering both their frequency of platform use and the trustworthiness of the participants they had interacted with. When there is a large number of customers each with at least a few orders, we prove that the VB scores (i) recover the correct types on the driver side, and (ii) achieve the highest statistical power on the customer side (i.e., maximizing the true positive rate at any given false positive rate). Extensive experiments on both synthetic data and data from our industry collaborator — a major Southeast Asian platform — demonstrate that the proposed algorithm provides substantial and robust accuracy improvements over a number of benchmarks.