Spotlight
Prediction without Preclusion: Recourse Verification with Reachable Sets
Avni Kothari · Bogdan Kulynych · Tsui-Wei Weng · Berk Ustun
Machine learning models are often used to decide who receives a loan, a job interview, or a public benefit. Standard methods to learn such models use features about people but overlook their actionability. As a result, models can assign predictions that are fixed – meaning that consumers who are denied loans, interviews, or benefits are precluded from access to credit, employment, or assistance. In this work, we present a task called recourse verification to flag models that assign fixed predictions under a rich class of real-world actionability constraints. We develop methods to check if a model can provide recourse using reachable sets. We demonstrate how our tools can verify recourse in real-world lending datasets. Our results highlight how models can inadvertently assign fixed predictions that permanently bar access, and underscore the need to account for actionability in model development.