Long-term Fairness with Selective Labels

Long-term fairness algorithms aim to satisfy fairness beyond static and short-term notions by accounting for the dynamics between decision-making policies and population behavior. Most previous approaches evaluate performance and fairness measures from observable features and a label, which is assumed to be fully observed. However, in scenarios such as hiring or lending, the labels (e.g., ability to repay the loan) are selective labels as they are only revealed based on positive decisions (e.g., when a loan is granted). In this paper, we study long-term fairness in the selective labels setting and analytically show that naive solutions do not guarantee fairness. To address this gap, we then introduce a novel framework that leverages both the observed data and a label predictor model to estimate the true fairness measure value by decomposing it into the observed fairness and bias from label predictions. This allows us to derive sufficient conditions to satisfy true fairness from observable quantities by using the confidence in the predictor model. Finally, we rely on our theoretical results to propose a novel reinforcement learning algorithm for effective long-term fair decision-making with selective labels. In semisynthetic environments, the proposed algorithm reached comparable fairness and performance to an agent with oracle access to the true labels.

As artificial intelligence systems are increasingly common in our lives, it is critical that decisions are fair. However, measuring fairness is difficult because AI only sees part of the picture. For example, if a bank denies someone a loan, it never finds out whether that person would actually have repaid it. We only see outcomes for the people who are accepted. To tackle this challenge, we considered that a decision-maker will commonly try to predict an unseen outcome using prediction systems. Yet this system will often make incorrect predictions, at varying levels across different demographic groups. By considering how uncertain we are about the unobserved outcomes, we identified conditions necessary to ensure fairness in the long term and designed a decision-making algorithm to achieve them. Our results prove that it is possible to deploy fair AI in real-world scenarios, even when operating with incomplete data. Looking forward, researchers can combine our framework with new techniques for measuring uncertainty, paving the way for more equitable AI systems.