Groups and Individuals in Fairness and Probability
summary
Compares group fairness to individual fairness and argues individual fairness needs a task-specific metric; cites a metric-learning breakthrough that learns pairwise similarity with O(n log n) easy queries. Builds and analyzes multi-accuracy and multi-calibration scoring functions (constructible modulo agnostic learning hardness) and uses differential privacy to handle adaptively chosen sets; shows fair rankings map to multi-calibrated scores. Notes algorithmic constructions exist but can be computationally existential and flags small-circuit groups and differential expressiveness as unresolved practical limits.
AI-written summary · about
chapters (12)
- 0:00Intro
- 1:07Theory of Algorithmic Fairness
- 5:58What should the Metric Capture?
- 12:13On Individual Risk
- 14:22Individual vs Group Probabilities
- 26:16Individual Probabilities?
- 30:47The Tumor Example
- 35:59The Loon Example: Which Groups?
- 38:12Multi-Accuracy and Multi-Calibration
- 39:20Let's talk about C
- 42:29Domination Compatibility (Simplified)
- 44:17Perspective on Individual Probabilities
related talks
43mAfternoon Session: Individuals, Groups, Pattern Recognition and Rich Data
46mEliciting and Enforcing Subjective Individual Fairness
50mA Multi-Group Approach To Algorithmic Fairness
43mMetric Learning for Individual Fairness
47mA Multi-Group Perspective on Fairness
50mThe Measure and Mismeasure of Fairness
46mThe Measure and Mismeasure of Fairness
42mFairness in Machine Learning: From Theory to Practice