Research Summary: When Post-Processing Fairness Constraints Help and When They Harm: Evidence from Eight Cross-Domain Evaluations

A cross-domain evaluation of machine learning fairness interventions demonstrates that static, single-instance fairness audits fail in production environments due to retraining dynamics and evolving user populations. The study introduces FAPE (Fairness Auditing for Production Environments), a four-stage framework testing Fairlearn's ThresholdOptimizer across eight distinct domains, finding that post-processing fairness constraints can either mitigate bias or introduce harm depending on domain-specific conditions.

Why it matters

One-off fairness certifications create significant strategic risk by masking model degradation and potential disparate impact over time. Algorithmic interventions that improve equity in one context may actively harm outcomes in another or deteriorate post-retraining. Organizations deploying predictive models must institute recurring, cross-domain governance to safeguard compliance and institutional credibility.

Key insights

• Production machine learning fairness audits commonly fail because they are executed only once at initial deployment and restricted to a single domain.
• Fairness metrics degrade post-deployment due to model retraining and shifts in the underlying user population.
• Algorithmic fairness interventions validated on one dataset rarely generalize predictably across heterogeneous deployment domains.
• The FAPE framework evaluates the ThresholdOptimizer post-processing intervention across eight sectors: criminal justice, income prediction, legal admissions, credit lending, agricultural lending, healthcare, education, and a multi-domain benchmark.
• Evaluations are benchmarked using demographic parity, equalized odds difference, and disparate impact ratios.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.26955