Bucher Sahyouni

Published - 2026

DAP - Differential Adjusted Parity

A cooperative, non-adversarial objective for learning informative representations that remain consistent across sensitive domains.

Fair representation learningResponsible AINon-adversarial learningImbalanced classification
Download paper PDF View on arXiv

Overview

Differential Adjusted Parity (DAP) is a loss for producing unbiased, informative representations. It turns adjusted parity into a differentiable training objective that jointly considers downstream task accuracy and inconsistency across sensitive feature domains. The paper was published at the ICLR AFAA Workshop 2026.

Problem

Learned representations can encode sensitive attributes such as gender or race and carry those signals into downstream decisions. Adversarial debiasing can be unstable and computationally intensive, while existing non-adversarial approaches may satisfy a fairness objective by performing equally poorly across every sensitive domain.

Method

DAP replaces the hard accuracy term in adjusted parity with Soft Balanced Accuracy. The metric computes differentiable, per-class recall and gives equal weight to each class, making it suitable for imbalanced problems. Accuracy is measured independently across sensitive domains and combined with the variation between domains to form a single cooperative objective. In the experiments, DAP is trained alongside the standard task cross-entropy loss.

Key contributions

  • A differentiable extension of adjusted parity for an arbitrary number of sensitive domains.
  • Soft Balanced Accuracy, which preserves differentiability while accounting for class imbalance.
  • A cooperative, non-adversarial objective that rewards task accuracy and consistency across sensitive domains.
  • Evaluation on Adult and COMPAS against NODEBIAS, ALFR, CFair, and LAFTR.

Results / findings

On Adult, DAP achieved the lowest equalized odds and demographic parity differences in the reported comparison. At the strongest reported setting, equalized odds difference was 44.1% lower and demographic parity difference was 18.6% lower than the next-best baseline.

On COMPAS, DAP improved adjusted parity by 45.9% and reduced equalized odds difference by 12.4% relative to its nearest competitor. It also reported a 22.5% improvement in demographic parity difference and a 40.1% reduction in race classification accuracy relative to the corresponding second-best baselines. The paper notes sensitivity to its weighting hyperparameters as a limitation and identifies automatic calibration as future work.

Citation

Bucher Sahyouni, Matthew Vowels, Liqun Chen, and Simon Hadfield. “Differential Adjusted Parity for Learning Fair Representations.” ICLR AFAA Workshop, 2026. arXiv:2502.09765.