RI: Medium: Foundations of Recourse Verification in Machine Learning

University Of California-San Diego, La Jolla CA

Investigators

Abstract

Machine learning models now automate decisions that affect millions of individuals in the United States, assigning predictions to decide who will receive a loan, a job interview, or a public service. Modern approaches for building such models do not account for actionability - i.e., how individuals can modify the features used by a model to determine their predictions. As a result, models in domains like lending and hiring can assign predictions that are fixed - meaning that individuals who are denied a loan or an interview may be permanently locked out from access to credit and employment. This project will develop new methods to ensure that models assign predictions that individuals can change through their actions in feature space. These methods will allow practitioners to build models that protect the right to access in applications like lending, hiring, and the allocation of public services. The project will develop methods that can be used to ensure access at various stages of the modern machine learning lifecycle. This includes methods for (1) confinement detection, i.e., to identify regions of feature space where individuals may be unable to change their features due to actionability constraints; (2) model-specific verification, i.e., to check that a model can provide recourse in model development or deployment; (3) learning with recourse guarantees, i.e., to train a model whose predictions can be changed through a well-defined set of actions in feature space. The methods will draw on the research teams’ expertise in using modern optimization techniques to promote fairness, robustness, and reliability in machine learning, and be refined through real-world applications in lending and hiring. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.