Metric Learning for Individual Fairness
summary
Proposes learning a distance metric from human judgments by querying a small set of representative points, using cheap triplet (is A closer to R or B?) comparisons to order items and a few coarse real-valued distances to fix scale, then combining answers via the triangle inequality into a representative submetric that underestimates true distances by at most an additive alpha. Claims this cuts human cost from quadratic to roughly n log n cheap queries plus a handful of numeric queries, and can generalize if near/far threshold classifiers exist.
AI-written summary · about
chapters (13)
- 0:00Start
- 1:47Power of the Metric
- 6:10How we ask matters
- 7:41Mental Model
- 14:32Metrics and Submetrics
- 15:19Representative elements
- 17:05Query Strategy
- 20:40Intuition for Generalization: Threshold Functions
- 22:58Post-processing: an important tool
- 23:44How to Choose Representatives: Two Perspectives
- 25:00Summary
- 25:36Food for thought
- 26:51Selected Related Works
related talks
53mGroups and Individuals in Fairness and Probability
46mEliciting and Enforcing Subjective Individual Fairness
36mMachine Learning Fairness: Lessons Learned
47mA Multi-Group Perspective on Fairness
33mMachine Learning from Human Comparative Judgments
39mLocalization and Nonmetric Multidimensional Scaling via Paired Comparisons
31mMeasuring Model Fairness
28mMeasuring Model Fairness