Equalized Odds
Fairness criterion: A model satisfies equalized odds when True Positive Rate and False Positive Rate are equal across all protected groups.
Equalized Odds requires equal error rates (TPR/FPR) across all groups – stricter than Demographic Parity but mathematically often incompatible with it.
Explanation
Unlike demographic parity, equalized odds considers the actual label (ground truth). It requires: Given the same actual outcome, all groups should be treated equally. Relaxed version: Equal Opportunity (only TPR equal).
Marketing Relevance
Relevant fairness standard for high-risk decisions: credit scoring, hiring, medical diagnosis – where incorrect results should not vary by group.
Common Pitfalls
Mathematically incompatible with demographic parity (except with perfect model or equal base rates). Requires ground truth labels – which themselves can be biased.
Origin & History
Hardt, Price & Srebro defined Equalized Odds in 2016 (NeurIPS). The impossibility theorems (Chouldechova 2017, Kleinberg et al. 2016) showed that different fairness definitions cannot be satisfied simultaneously.
Comparisons & Differences
Equalized Odds vs. Demographic Parity
Demographic Parity ignores ground truth; Equalized Odds considers actual labels and requires equal error rates.
Equalized Odds vs. Calibration
Calibration requires equal probability meaning across groups; Equalized Odds requires equal error rates.