Tracking and Improving Information in the Service of Fairness
summary
Paper shows that disparities often come from unequal information content in calibrated risk scores, not from an unavoidable tradeoff between accuracy and parity. Concrete example: identical groups with true default rate 50% but one gets a constant 0.5 score in S while the other gets bimodal 0.75/0.25 in T, so a threshold 0.6 selects half of T and none of S. Formal results tie higher predictive information (calibrated, squared-error/variance view) to improved TPR/FPR/PPV and lender utility; gives audit tools and notes compatibility with differential privacy.
AI-written summary · about
chapters (4)
related talks
46mThe Measure and Mismeasure of Fairness
50mThe Measure and Mismeasure of Fairness
28mMeasuring Model Fairness
14mFair And Reliable Machine Learning For High-Stakes Applications:approaches Using Information Theory
53mGroups and Individuals in Fairness and Probability
42mFairness in Machine Learning: From Theory to Practice
30mBuilding fairer models for finance
43mAfternoon Session: Individuals, Groups, Pattern Recognition and Rich Data