Research map / Learning theory and algorithms / Fairness
FRAPPÉ: A Group Fairness Framework for Post-Processing Everything
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 FRAPPÉ, a framework that turns any regularized in-processing group-fairness method into a modular post-processing method by training an additive correction on base-model outputs, without needing sensitive attributes at inference.
Problem
Post-processing methods for group fairness are tailored to specific problem settings and fairness definitions and generally require sensitive attributes at inference, while in-processing requires retraining the full prediction model and access to the training pipeline.
Approach
Post-hoc module adds an additive correction T(x) to the frozen base model's outputs, trained with any in-processing fairness regularizer plus a divergence keeping outputs close to the base model; a GLM analysis shows this bi-level objective equals the in-processing objective up to a constant, so Pareto frontiers match.
Key results
- Compared with FairProjection on random forest base models, FRAPPÉ MinDiff reduces MEO by 53%, 37%, 32% and 50% on Adult, COMPAS, HSLS and ENEM respectively, without sacrificing more than 2% prediction error.
- On Adult, FRAPPÉ MinDiff computation for the full Pareto frontier takes 11.7 minutes versus 85.4 minutes for in-processing MinDiff (including base model training).
- With only 0.1% of training data annotated with sensitive attributes, in-processing MinDiff performs no better than a naive randomized baseline, while the post-processing variant maintains good error-fairness trade-offs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FRAPPÉ MinDiff for EqOdds instead of FairProjection | Random forest predictions · Adult | MEO: 53% | Helped |
| Use FRAPPÉ MinDiff for EqOdds instead of FairProjection | Random forest predictions · COMPAS | MEO: 37% | Helped |
| Use FRAPPÉ MinDiff for EqOdds instead of FairProjection | Random forest predictions · HSLS | MEO: 32% | Helped |
| Use FRAPPÉ MinDiff for EqOdds instead of FairProjection | Random forest predictions · ENEM | MEO: 50% | Helped |
| Train a FRAPPÉ post-hoc module instead of the full model | MinDiff on Adult · Adult | computation time: 85.4 -> 11.7 minutes | Helped |
| Train a FRAPPÉ post-hoc module instead of the full model | In-processing method on Adult · Adult | computation time: 113.4 -> 14.8 minutes | Helped |
| Use FRAPPÉ MinDiff rather than in-processing MinDiff when group labels are partial | MinDiff · Adult, partial group labels | prediction error at similar low FPR gap | Helped |
| Use FRAPPÉ MinDiff rather than early-stopped in-processing MinDiff, without early-stopping regularization | MinDiff · Adult, partial group labels | fairness-error trade-off | Helped |
Surprising findings
- In-processing MinDiff overfits the fairness regularizer under partial group labels: training FPR gap vanishes while test FPR gap grows during training.
- A linear-regression post-hoc module is sufficient to match in-processing Pareto frontiers built on 3-layer MLPs.
Limitations
- Theoretical equivalence is shown only for GLMs; for neural networks it is motivated but not proven.
- Assumes distribution shift between prediction and sensitive datasets is absent; left as future work.
- Only group fairness is studied; individual fairness and other equity notions are not addressed.
Experimental setup
Datasets: Adult COMPAS HSLS ENEM Communities & Crime
Benchmarks: Adult COMPAS HSLS ENEM Communities & Crime
Metrics: FPR gap (EqOpp) statistical parity difference mean equalized odds (MEO) HGR-infinity (EqOdds) prediction error
Baselines: MinDiff Reductions FairProjection Fairness Reprogramming Rejection-option classification
Builds on: MinDiff Reductions FairProjection Bregman divergence
Closest papers by meaning
- FFB: A Fair Fairness Benchmark for In-Processing Group Fairness Methods (ICLR 2024 · similarity 0.90)
- Enhancing Group Fairness in Online Settings Using Oblique Decision Forests (ICLR 2024 · similarity 0.88)
- Metric-Agnostic Continual Learning for Sustainable Group Fairness (AAAI 2025 · similarity 0.87)
- FairMT: Fairness for Heterogeneous Multi-Task Learning (NeurIPS 2026 · similarity 0.86)
- Fairness without Harm: An Influence-Guided Active Sampling Approach (NeurIPS 2024 · similarity 0.86)
- Some Optimizers are More Equal: Understanding the Role of Optimizers in Group Fairness (NeurIPS 2025 · similarity 0.86)
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.