Long-term Fairness with Selective Labels
Researchers have developed a new framework to address long-term fairness in machine learning models, particularly in scenarios where labels are selectively revealed. The proposed method decomposes fairness into observed and predicted label bias, allowing for true fairness to be estimated using confidence in a label predictor. This theoretical advancement has led to a novel reinforcement learning algorithm designed for effective long-term fair decision-making with selective labels, demonstrating comparable performance to an oracle with true label access in semi-synthetic environments. AI
IMPACT Introduces a novel approach to long-term fairness in ML, crucial for applications like hiring and lending where data is selectively revealed.