RIKEN AIP team proposes practical Bayes-optimal fairness-accuracy tradeoff estimation with soft labels

RatnRajiv · x · 2026-09-25

RIKEN AIP researcher Mohit Sharma (reposted by RatnRajiv) shared the paper "Practical Estimation of the Bayes Optimal Fairness-Accuracy Tradeoff with Soft Labels," co-authored with Okan Koç, Amit Deshpande, Gang Niu, and Prof. Masashi Sugiyama, done during his internship at RIKEN AIP.

The work tackles the core tension in ML fairness: fairness constraints typically cost accuracy, and the theoretical Bayes-optimal frontier of this tradeoff has been hard to estimate in practice. The paper proposes using soft labels to practically estimate that optimal tradeoff, offering an operational tool for quantifying how much accuracy fairness actually costs. The post is one item in a paper thread, so details live in the paper itself.

Original post →

More from Research

Research channel →