A Multi-Group Approach To Algorithmic Fairness
summary
Claims that multi-calibration yields predictors that are calibrated simultaneously over very rich, even exponential, families of overlapping subgroups and that such predictors exist and are computable by an iterative algorithm that post-processing fixes miscalibrated groups. Complexity links to agnostic learning. Applied to cardiac and osteoporotic risk models by enumerating 360 subgroups, the method substantially reduces miscalibration across groups on held-out data while not degrading overall accuracy.
AI-written summary · about
chapters (8)
- 0:00Introduction to fairness
- 5:15Causality and fairness
- 9:39Risk prediction problem
- 14:29Limitations of group fairness
- 19:45Multi-calibration framework
- 27:51Practical applications
- 32:24Research extensions
- 38:14Q&A session
related talks
47mA Multi-Group Perspective on Fairness
12mA Complexity-Theoretic Perspective on Fairness
01h 21mMulticalibration: Towards Fair Decision Making
53mGroups and Individuals in Fairness and Probability
43mAfternoon Session: Individuals, Groups, Pattern Recognition and Rich Data
19mHappyMap: A Generalized Multicalibration Method
33mOnline Multivalid Learning: Means, Moments, and Prediction Intervals
37mOmnipredictors