HappyMap: A Generalized Multicalibration Method
summary
A postprocessing method called HappyMap algorithm maps any predictor to satisfy a new fairness/robustness condition, s-Happy Multicalibration, by iteratively boosting weak learners (finding ct in the square-root family) to correct calibration errors. It proves coverage under covariate shift without knowing the test density provided the density ratio is within beta of the chosen c (coverage ≈ 1−δ with tolerance α+β), handles overlapping or infinite protected groups for equalized coverage, and a no-harm theorem shows interval length remains near-optimal when initialized close to the δ-quantile.
AI-written summary · about
chapters (10)
- 0:00Introduction to the problem
- 1:15Conformal prediction basics
- 2:51Challenges in prediction
- 4:45Fairness in group coverage
- 6:44Defining happy multi-calibration
- 8:41Versatility of the framework
- 11:10Happy map algorithm
- 13:00Applications and results
- 14:40No harm theorem
- 17:01Future work and conclusion
related talks
47mA Multi-Group Perspective on Fairness
01h 21mMulticalibration: Towards Fair Decision Making
50mA Multi-Group Approach To Algorithmic Fairness
45mMulticalibration, Universal Adaptability and Causality
37mOmnipredictors
31mUncertainty quantification: How much can you trust your machine learning model?
12mA Complexity-Theoretic Perspective on Fairness
48mOnline Adversarial Multicalibration And (Multi)Calibeating