Research map / Learning theory and algorithms / Fairness

Simple and Effective Specialized Representations for Fair Classifiers

NeurIPS 2025 · Learning theory and algorithms › Fairness · approach: Post-processing and transport for fairness · 0 citations (OpenAlex)

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 non-adversarial fair representation learning that matches conditional characteristic functions (FmCF) and a moment-based simplification (FmSS) to a Gaussian target, yielding provable fairness guarantees for logistic-regression classifiers.

Problem

Learning fair specialized representations that hide sensitive attributes while keeping task accuracy, without the instability of adversarial training or the cost of normalizing flows and VAE-based MMD objectives.

Approach

Match each conditional representation distribution Z|S independently to a standard Gaussian using a characteristic function distance penalty (FmCF); a truncated second-order moment relaxation (FmSS) reduces this to matching the per-group mean and variance, giving provable limits on logistic-regression leakage of S without needing S at inference.

Key results

• On Adult, FmCF reaches 85.01% accuracy versus 84.43% for FNF, with adversarial accuracy on a deep MLP of 56.64% (FNF 59.56%).
• On German (gender as sensitive attribute), FmCF attains 74.60% accuracy, above VFAE (72.00%) and CIAFL (69.50%), with adversarial accuracy 69.00%.
• FmCF and FmSS match or exceed FNF accuracy-fairness Pareto fronts on Crime, Health, COMPAS and Adult, with 95% confidence intervals over 10 seeds.

Measured effects reported in the paper

Surprising findings

• Lower latent dimensions make the penalties more effective, giving strong fairness even at 1-3 dimensions, which the authors note does not hurt downstream accuracy.
• Adversarial-learning methods reported on Adult with a deep MLP adversary show adversarial accuracy close to the chance level only for some baselines, while FNF's fairness collapses to 59.56% under the same evaluation.

Limitations

• Performance degrades in high-dimensional settings where the characteristic function distance loses effectiveness.
• The moment-based formulation is tailored to classification and does not directly extend to regression.
• Current fairness benchmarks are low-dimensional, limiting evaluation in complex settings.

Experimental setup

Datasets: Adult COMPAS German Crime Health Law School MNIST

Benchmarks: Adult German COMPAS Crime Health

Metrics: accuracy adversarial accuracy Pareto front accuracy-fairness

Baselines: Fair Normalizing Flows AdvForgetting MaxEnt-ARL LAFTR VFAE CIAFL IRwAL FRTrain CSN WassDB

Builds on: Fair Normalizing Flows Variational Fair Autoencoder Maximum Mean Discrepancy Learning Adversarially Fair and Transferable Representations

Closest papers by meaning

• Fair Classification with Adversarial Perturbations (NeurIPS 2021 · similarity 0.89)
• Efficient Fairness-Performance Pareto Front Computation (NeurIPS 2025 · similarity 0.87)
• Classes Are Not Equal: An Empirical Study on Image Recognition Fairness (CVPR 2024 · similarity 0.86)
• Kernel-based Equalized Odds: A Quantification of Accuracy-Fairness Trade-off in Fair Representation Learning (NeurIPS 2025 · similarity 0.86)
• Individual Fairness In Strategic Classification (NeurIPS 2025 · similarity 0.86)
• Fair Representation Learning with Controllable High Confidence Guarantees via Adversarial Inference (NeurIPS 2025 · similarity 0.85)