第3回思考院セミナー/ The 3rd Seminar of the School of Statistical Thinking (Hybrid)
- 【日時】
- 2026年10月15日（木）16:00〜17:00
参加無料 - 【場所】
- 統計数理研究所 セミナー室5 (D313・D314)
(zoomでの参加登録はこちらから) - 【講演者】
- Mohit Sharma (Tohoku Univ / ISM)
- 【演題】
- On Recovering Fair and Accurate Classifiers with Imperfect Distributions
- 【概要】
- In machine learning deployments, fairness and accuracy are frequently seen as conflicting goals. To understand this tradeoff, this talk explores the behavior of Bayes-optimal fair classifiers—the theoretical gold standard for classification under fairness constraints. We will examine the fairness-accuracy tradeoff from several angles. We investigate the conditions under which fairness constraints can actually help recover the true, unconstrained optimal classifier in the presence of data bias. Next, we bridge theory and practice by analyzing why theoretically equivalent implementations—whether applied via pre-processing, weighted risk minimization, or post-processing—often diverge and fail to mitigate unfairness in practical simulations. Finally, we introduce a framework to minimally steer biased distributions toward "ideal" distributions where classifiers are fair by default, alongside a model-free benchmarking technique to estimate the fairness-accuracy tradeoff using only class posterior probabilities.
