Research map / Learning theory and algorithms / Fairness
Unprocessing Seven Years of Algorithmic Fairness
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Empirically re-evaluates seven years of fairness methods against postprocessing via thousands of trained models on ACS and MEPS tabular data, introducing "unprocessing" and a relaxed-equalized-odds LP postprocessing package.
Problem
Whether fairness-aware methods (pre- and inprocessing) improve over the simple postprocessing baseline for equalizing error rates, given confounds from different unconstrained base models and different levels of constraint relaxation.
Approach
Unprocessing maps a fairness-constrained model back to an unconstrained model by optimizing group-specific thresholds to minimize loss, so methods using different base models can be compared on equal footing; an LP formulation gives postprocessing at any relaxation level r, and postprocessing the most accurate unprocessed model is shown to dominate the fairness-accuracy frontier.
Key results
- Postprocessing the single most accurate unprocessed model (a GBM) matches or dominates all ~1000 trained models per dataset across all fairness-constraint levels on the ACS datasets.
- Over 11,000 models trained across 11 evaluation scenarios on 6 datasets, no counterexample to postprocessing dominance was observed.
- Preprocessing methods (LFR, CR) achieved the weakest fairness-accuracy trade-offs, while EG and FairGBM performed best among fairness-aware methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Postprocess the highest-accuracy model with group-specific thresholds across constraint levels | Highest-accuracy model m* · ACSIncome, ACSTravelTime, ACSPublicCoverage, ACSMobility, and ACSEmployment | accuracy at matched equalized-odds violation | Helped |
| Postprocess the highest-accuracy model with group-specific thresholds across constraint levels | Highest-accuracy model m* · MEPS | accuracy at matched equalized-odds violation | Helped |
| Postprocess the highest-accuracy model with group-specific thresholds across constraint levels | Highest-accuracy model m* · ACS datasets with White and Black groups | accuracy at matched equalized-odds violation | Helped |
| Use LFR preprocessing | GBM-based fairness methods · ACS datasets | accuracy-fairness trade-off | Hurt |
| Use correlation remover preprocessing | GBM-based fairness methods · ACS datasets | accuracy-fairness trade-off | Hurt |
| Use exponentiated gradient or FairGBM inprocessing | GBM-based fairness methods · ACS datasets | Pareto-frontier area | Helped |
| Use only the two largest sensitive groups instead of four groups (ablation) | Fairness evaluation methods · ACS datasets | equalized-odds violation: 0.38 ... 0.16 | Helped |
| Unprocess fairness-constrained models before comparing their accuracy | EG and FairGBM models · ACSIncome | equalized-odds violation | No clear effect |
Surprising findings
- A fairness-aware method can yield the single most accurate unprocessed model, indicating its advantage came from better risk scores rather than constraint enforcement.
- Some fairness methods achieved higher test accuracy than unconstrained GBMs, which the authors partly attribute to compute advantages.
Limitations
- Restricted to tabular data; high-dimensional raw-feature data (images, text) not studied.
- Only evaluates methods achieving approximate error rate parity with race as the protected attribute; does not address whether equalizing error rates is desirable.
- Known counterexamples where postprocessing is suboptimal (non-squared loss, non-Bayes-optimal scores) were not observed empirically but may arise elsewhere; smaller datasets like MEPS show high variance.
Experimental setup
Datasets: ACSIncome ACSTravelTime ACSPublicCoverage ACSMobility ACSEmployment MEPS UCI Adult
Benchmarks: ACSIncome ACSTravelTime ACSPublicCoverage ACSMobility ACSEmployment MEPS
Metrics: accuracy equalized odds violation
Baselines: learned fair representations correlation remover exponentiated gradient reduction grid search reduction FairGBM
Builds on: equalized odds (Hardt et al. 2016) folktables fairlearn reductions
Closest papers by meaning
- FRAPPÉ: A Group Fairness Framework for Post-Processing Everything (ICML 2024 · similarity 0.86)
- Fairness without Harm: An Influence-Guided Active Sampling Approach (NeurIPS 2024 · similarity 0.85)
- Understanding challenges to the interpretation of disaggregated evaluations of algorithmic fairness (NeurIPS 2025 · similarity 0.84)
- Geometry of Relaxed Fair Regression: A Unified Framework for Aware and Unaware Settings (NeurIPS 2026 · similarity 0.84)
- AIM-Fair: Advancing Algorithmic Fairness via Selectively Fine-Tuning Biased Models with Contextual Synthetic Data (CVPR 2025 · similarity 0.84)
- The Fragility of Fairness: Causal Sensitivity Analysis for Fair Machine Learning (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.