Intersectional fairness hits a resolution limit: SAFER penalizes only statistically certified gaps, sparing utility in sparse subgroups
Synopsis
In intersectional fairness auditing, finer subgroup definitions leave fewer samples per subgroup, so an observed gap may be nothing but sampling noise. SAFER attaches the fairness penalty to a differentiable soft-count likelihood-ratio evidence statistic, applying gradient only to subgroup gaps beyond what noise can explain, and keeps a small fixed-weight ungated term to prevent optimization stalling. On tabular benchmarks including Adult, COMPAS and German Credit plus purpose-built stress variants, it gives a better fairness-utility frontier where certified disparities exist and preserves the utility of the unconstrained model where none do.
Interpretation
Intersectional fairness training now separates statistically supported disparities from sampling noise instead of penalizing every observed subgroup gap as real bias. Earlier in-processing methods were certainty-blind: any observed subgroup disparity entered the objective regardless of whether it was distinguishable from sampling noise, producing fairness overfitting in sparse intersectional regimes where models sacrificed utility to fix spurious gaps. On the Adult Income, COMPAS and German Credit benchmarks and on stress variants built for each mechanism, single-knob ablations removing any SAFER component produce the failure the theory predicts on at least one stress variant; in sparse intersectional settings SAFER reaches a better fairness-utility Pareto frontier than the baselines.
The method first converts each subgroup's disparity into a continuous uncertainty-aware score, then uses that score to decide how much fairness gradient the subgroup receives during training. This score, the significant excess, keeps only the part of a subgroup's disparity exceeding what sampling variation can explain at a given confidence level; when the interval contains zero the subgroup's contribution is exactly zero. Training uses a two-sample deviance likelihood-ratio statistic computed on threshold-smoothed soft counts as the gate signal; the statistic needs no variance estimate, avoiding the boundary instability that arises when variance estimates collapse in sparse cells, while statistical certification stays with the held-out hard-count audit.
Beyond the gate, a small ungated term with a fixed weight keeps optimization from stalling under strong regularization before an active disparity is removed. The gated term's gradient decays exponentially once a subgroup's disparity falls below the threshold, and strong regularization pushes active subgroups back toward that boundary, stalling optimization; the ungated term carries no gate multiplier and so still supplies a descent direction. On the COMPAS 20/80 stress split, the variant without this dense-tail term plateaus in MSEE from a certain fairness weight onward, about ten standard errors from full SAFER; on German Credit the two coincide because the gate already keeps the hinge inactive.
Perspective
The demonstrated scope is tabular benchmarks including Adult Income, COMPAS, German Credit, ACS Coverage and Taiwan Credit, with intersection levels up to four and at most a few hundred audited subgroups, together with stress variants for low training budgets, deeper intersections and small validation budgets; claims are stated by regime, discussing frontier improvement only where certified disparities exist and utility preservation where none do.
A zero significant excess means only that no disparity is certifiable at the chosen confidence level and sample size; a real disparity in a small subgroup can remain uncertified, and SAFER then withholds the gradient that would reduce it. The training-time significance test does not yet include multiplicity corrections, and scaling to high-cardinality attribute spaces and to settings where subgroup membership is uncertain or partially observed remains open.