Demographic Parity
Fairness criterion: A model satisfies demographic parity when prediction rates (e.g., approval rate) are equal across all protected groups.
Demographic Parity requires equal prediction rates for all groups – the simplest fairness metric but blind to actual qualifications.
Explanation
Demographic Parity requires: P(Ŷ=1|A=a) = P(Ŷ=1|A=b) for all groups a, b. The prediction should be independent of the protected attribute. Easy to measure and communicate but ignores ground truth.
Marketing Relevance
Often used as a first fairness check. Particularly relevant when ground truth itself might be biased (e.g., historical hiring data).
Common Pitfalls
Ignores qualification differences (can lead to "reverse discrimination"). Mathematically incompatible with Equalized Odds. Can favor the "least qualified" of a group.
Origin & History
Demographic Parity has roots in US civil rights discourse and the EEOC 80% rule. Formalized in ML by Dwork et al. (2012). Impossibility theorems showed fundamental limitations.
Comparisons & Differences
Demographic Parity vs. Equalized Odds
Demographic Parity requires equal rates without considering ground truth; Equalized Odds requires equal error rates considering actual labels.
Demographic Parity vs. Individual Fairness
Demographic Parity is group-based; Individual Fairness requires that similar individuals are treated similarly.