Fairness: three families, one impossibility
group fairness
demographic parity
P(Y=1 | A=a) equal
equalized odds
TPR/FPR equal across groups
conditional use accuracy equality
individual fairness
Dwork et al. 2012
|f(x) - f(x')| <= L * d(x, x')
Lipschitz; d task-specific
similar individuals
get similar decisions
counterfactual fairness
Kusner et al. 2017
invariant under attribute
counterfactual
needs causal DAG
backtracking (2024) sidesteps
impossibility + reconciliation
Chouldechova, KMR 2017: under unequal base rates, the three group criteria cannot all hold.
policy choice: demographic parity gives equal access; equalized odds preserves accuracy equity;
conditional use accuracy equality preserves predictive-value equity. each has a constituency.
2024 NeurIPS: CF-accuracy trade-off is bounded; model-agnostic conversion of optimal-unfair -> CF.
backtracking counterfactuals (arXiv:2401.13935): avoid intervening on protected attributes.
ICLR 2024 reconciliation: with explicit causal graphs, group and counterfactual are facets of one structure.
impossibility still holds on base rates; reconciliation is about what is being measured.