A Multi-Group Perspective on Fairness
summary
Claims that multi calibration guarantees exist for any class C and that small, efficient predictors can be learned from labeled outcomes; algorithm starts with 0.5 scores, iteratively finds subsets in C where calibration fails and adjusts their predicted value toward the empirical outcome. Runtime is roughly linear in |C| and polynomial in 1/alpha, sample complexity scales ~1/alpha^2, and the search step reduces to an agnostic learning oracle. Uses fresh samples or differential-privacy techniques to handle adaptivity and works as post-processing to improve existing predictors.
AI-written summary · about
chapters (19)
- 0:00Intro
- 1:04Algorithms Make and Inform Decisions
- 1:38Concern: Discrimination
- 3:04THE* Definition of Fairness?
- 6:17Prediction
- 6:58Learning from Outcomes
- 13:21Group Notions of Fairness
- 20:24Which Groups? A Computational Perspective
- 23:00Multicalibration [HKRR18]
- 26:45Example: Ageism in Healthcare
- 27:54Some Related Works
- 29:05Multicalibration: Flavor of Results
- 32:14Learning Multicalibrated Predictor
- 33:31The Algorithm
- 40:21Obtaining Statistics
- 41:18Time Complexity
- 44:34Where Does Metric Come From?
- 44:48Metric-Fair Machine Learning?
- 45:47Summary
related talks
50mA Multi-Group Approach To Algorithmic Fairness
12mA Complexity-Theoretic Perspective on Fairness
01h 21mMulticalibration: Towards Fair Decision Making
53mGroups and Individuals in Fairness and Probability
45mMulticalibration, Universal Adaptability and Causality
19mHappyMap: A Generalized Multicalibration Method
37mOmnipredictors
43mAfternoon Session: Individuals, Groups, Pattern Recognition and Rich Data