A post-processing algorithm achieves distribution-free finite-sample group fairness guarantees with controlled excess risk for both group-aware and group-blind settings, shown minimax-optimal up to logs via lower bound.
(24) In Theorem 5, we compare the prediction error of ˆf G ˆλα with that of f ∗G ˜α , rather than f ∗G α
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
stat.ME 1years
2024 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Finite-Sample and Distribution-Free Fair Classification: Optimal Trade-off Between Excess Risk and Fairness, and the Cost of Group-Blindness
A post-processing algorithm achieves distribution-free finite-sample group fairness guarantees with controlled excess risk for both group-aware and group-blind settings, shown minimax-optimal up to logs via lower bound.