Fairness Definition Choice
Selecting which mathematical fairness criterion applies, given that the main criteria are provably incompatible and the choice is a value judgement.
There is no single measure of fairness, and this is not a gap in the literature but a proven result: the principal criteria cannot all be satisfied simultaneously except in degenerate cases.
The main ones ask different questions. Demographic parity — are outcomes distributed equally across groups? Equal opportunity — among those who should qualify, are they identified at the same rate across groups? Predictive parity — when the model predicts positively, is it right at the same rate across groups? Individual fairness — are similar individuals treated similarly?
Where base rates differ between groups, satisfying one of these necessarily violates another. The choice is therefore a judgement about what fairness means in this application, made by the people accountable for the decision — with legal and domain input — and not by the engineering team as an implementation detail.
What architecture must supply is the ability to make the choice visible and monitored: the model card records which criterion was chosen and why; monitoring reports the chosen metric disaggregated by group continuously, since fairness degrades with distribution shift; and the decision is revisited on retraining.
The position that fails is optimising accuracy and asserting fairness because protected attributes were not used as features — proxies are pervasive, and the absence of a variable does not remove its influence.