Afternoon Session: Individuals, Groups, Pattern Recognition and Rich Data
Talkat Wrong at the Root: Racial Bias and the Tension Between Numbers and Words in Non-Internet Data · 2019-06-05..06-07·
summary
Distinguishes group vs individual fairness, arguing individual fairness requires a task-specific metric and showing a small number of queries to a fair arbiter can approximate that metric. Describes multi‑calibration and calibration on computationally defined sets as a bridge to group guarantees, plus evidence‑based ranking and a provable one‑sided test for dataset inadequacy. Warns that common preprocessing like PCA can increase reconstruction error unevenly (order‑of‑magnitude differences reported, roughly ten percent higher error for one subgroup), so feature choice and training data limit what scores can legitimately mean.
AI-written summary · about
chapters (14)
- 0:00Start
- 3:07Theory of Algorithmic Fairness • Definitions: Group vs Individual⚫ Group Fairness Examples
- 9:20Risk Scores
- 12:18The Tumor Example
- 16:03The Defining Problem of Al Scoring • We are exploring an approach with roots in complexity theory
- 20:12Starting Point: Bridging Groups vs Individuals
- 23:11Machine Learning's Unequal Representation
- 26:04But isn't this just a policy question?
- 27:49Profit, ethics, acceptable uses
- 30:48High-dimensional data mining
- 35:34Low dimensional approximations
- 37:59PCA on women and men
- 39:36Low-dimensional approximation on 2 groups
- 40:40Theorem
related talks
53mGroups and Individuals in Fairness and Probability
50mA Multi-Group Approach To Algorithmic Fairness
42mFairness in Machine Learning: From Theory to Practice
47mA Multi-Group Perspective on Fairness
50mThe Measure and Mismeasure of Fairness
46mEliciting and Enforcing Subjective Individual Fairness
36mMachine Learning Fairness: Lessons Learned
34mInterpretability and Algorithmic Fairness