Multicalibration: Towards Fair Decision Making
summary
Multicalibration requires predictions be calibrated not just overall but simultaneously over a computable class of subpopulations and supplies an algorithmic path to get there: it learns exponentially many subgroups from polynomial samples, it reduces to weak agnostic learning / boosting as the core computational primitive, and it outputs a post-processed predictor that serves as an omnipredictor for Lipschitz convex losses. The work proves sample bounds (VC-dimension driven), gives an auditing-based learning routine, and equates multicalibration with a notion of outcome indistinguishability.
AI-written summary · about
chapters (14)
- 0:00Start
- 1:18Announcements
- 1:52Michael Kim
- 5:07Risk Prediction
- 11:56Provable Certificate of Fairness
- 14:00Definitions
- 19:54Group Fairness
- 28:24Draw Inspiration from Learning Theory
- 35:55How Does Multi-Calibration Compare to Calibration across a Protected Attribute
- 49:51Auditing Predictors
- 53:48The Variance of the Estimator
- 56:56Loss Minimization
- 1:01:10Loss Independent Agnostic Learning
- 1:11:23Distinguisher Algorithm
related talks
12mA Complexity-Theoretic Perspective on Fairness
45mMulticalibration, Universal Adaptability and Causality
50mA Multi-Group Approach To Algorithmic Fairness
37mOmnipredictors
47mA Multi-Group Perspective on Fairness
48mOnline Adversarial Multicalibration And (Multi)Calibeating
19mHappyMap: A Generalized Multicalibration Method
33mOnline Multivalid Learning: Means, Moments, and Prediction Intervals