Research map / Learning theory and algorithms / Fairness
FROC: Building Fair ROC from a Trained Classifier
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.
Proposes FROC, a linear-time post-processing method that transports the group-wise ROC curves of a trained scorer within ε (L1) of each other, yielding a randomized classifier satisfying ε-Equalized ROC with minimal AUC loss.
Problem
Trained probabilistic binary classifiers can be unfair to one protected group across decision thresholds; existing group fairness notions check only a single threshold, and retraining is costly.
Approach
Defines ε_p-Equalized ROC (the Lp distance between group ROC curves bounded at all thresholds) and enforces it by geometrically shifting the upper group's ROC points toward the lower ROC using norm boundaries, then realizing the new ROC via randomized mixing of classifiers, with a claimed AUC-optimality result under assumptions.
Key results
- On ADULT with logistic regression, FROC at ε=0.05 reduces the maximum ROC gap from 0.08 with under 0.1% accuracy loss.
- Across experiments the authors report roughly 7-8% fairness improvement with at most about 2% accuracy drop.
- The paper reports that FROC is O(k) time for piecewise-linear inputs satisfying its assumption, and that its AUC loss decays to zero as ε loosens.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply FROC post-processing to Logistic Regression scores | Logistic Regression · Adult | accuracy loss: <0.1% accuracy loss for ε=0.05 | Helped |
| Apply FROC post-processing to classifier scores | Classifiers evaluated in the experiments · Adult and COMPAS | fairness: 7-8% improvement in fairness | Helped |
| Apply FROC post-processing to classifier scores | Classifiers evaluated in the experiments · Adult and COMPAS | accuracy: at most a 2% drop in accuracy | Hurt |
| Apply FROC to Weighted Ensemble L2 | Weighted Ensemble L2 · Adult | Disparate Impact: a 1% drop in Accuracy improves the Disparate Impact by 5% | Helped |
| Apply FROC to Random Forest Gini | Random Forest Gini · Adult | Disparate Impact: a 1% drop in Accuracy improves the Disparate Impact by 7% | Helped |
| Apply FROC to FNNC | FNNC · Adult | accuracy | Helped |
| Apply FROC to FNNC | FNNC · Adult | Disparate Impact: a 1% drop in Accuracy improves the Disparate Impact by 5% | Helped |
| Apply FROC to Weighted Ensemble L2 | Weighted Ensemble L2 · COMPAS | Disparate Impact: a 1% drop in Accuracy improves the Disparate Impact by 7% | Helped |
| Loosen the fairness constraint ε | FROC · Adult and COMPAS | AUC loss: decays to 0 | Helped |
Surprising findings
- FNNC and other baselines were reported to slightly beat FROC on accuracy at small ε, which the authors attribute to those methods overachieving the fairness target.
Limitations
- Optimality guarantees hold only under the stated assumptions (at most two ROC intersections, slope bounds, piecewise-linear scorer).
- Assumes non-intersecting or only low-FPR/high-TPR intersecting ROCs; intersecting ROCs are left to future work.
- Only binary protected attributes and L1 norm are fully analyzed in experiments.
Experimental setup
Datasets: ADULT COMPAS
Benchmarks: ADULT COMPAS
Metrics: accuracy AUC loss disparate impact mean equalized odds ROC distance (L1)
Baselines: FNNC FairProjection-CE FairProjection-KL
Builds on: equalized odds fair projection
Closest papers by meaning
- Preserving AUC Fairness in Learning with Noisy Protected Groups (ICML 2025 · similarity 0.86)
- Post-hoc bias scoring is optimal for fair classification (ICLR 2024 · similarity 0.86)
- Kernel-based Equalized Odds: A Quantification of Accuracy-Fairness Trade-off in Fair Representation Learning (NeurIPS 2025 · similarity 0.85)
- Fairness-aware Bayes Optimal Functional Classification (NeurIPS 2025 · similarity 0.84)
- OxonFair: A Flexible Toolkit for Algorithmic Fairness (NeurIPS 2024 · similarity 0.84)
- Efficient Fairness-Performance Pareto Front Computation (NeurIPS 2025 · 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.