Research map / Learning theory and algorithms / Fairness
Fair Representation Learning with Controllable High Confidence Guarantees via Adversarial Inference
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 FRG, a fair representation learning framework that gives user-specified high-confidence bounds on demographic parity disparity for all downstream models, using an adversary, a t-test fairness test on held-out data, and a VAE-based candidate selection with Lagrangian constraint.
Problem
Fair representation learning typically only estimates upper bounds on downstream demographic parity disparity on training data, which can fail on unseen test data; there is no user-controllable high-confidence guarantee across all downstream tasks and models.
Approach
Split data into candidate-selection and fairness-test sets; train a VAE-style representation under a Lagrangian constraint on a confidence upper bound computed with an adversary predicting the sensitive attribute, then run a Student's t-based fairness test on held-out data that returns the representation only if the bound on demographic parity is at most epsilon with confidence 1-delta (otherwise
Key results
- FRG keeps the probability of violating the demographic parity constraint at or below delta=0.1 across datasets, while most baselines fail with probability at least 0.1 on test or adversarial tasks.
- FRG and FRG_supervised maintain a probability of returning a solution of at least 0.9 across all datasets.
- FARE's certificates are often several times larger than the desired epsilon, and it cannot be set to user-defined epsilon values.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FRG's fairness test with held-out data and adversarial inference | fair representation learning · Adult, Health, and Income downstream tasks | probability of satisfying ε-fairness: at least 0.9 | Helped |
| use FRG instead of the compared baseline methods | fair representation learning · Adult, Health, and Income downstream tasks | probability of violating ε-fairness: at least 0.1 | Helped |
| use FRG instead of baselines that satisfy fairness with high probability | downstream prediction · Adult, Health, and Income | AUC | Helped |
| add downstream-task supervision to FRG (ablation) | FRG · Adult and Health | AUC | No clear effect |
| increase δ (ablation) | FRG · target and adversarial tasks | prediction performance: ε=0.16 | Mixed |
| vary confidence-interval inflation α (ablation) | FRG · experiments with δ=0.1 | acceptance rate: δ=0.1 | No clear effect |
| use α closer to 1.0 rather than a larger α (ablation) | FRG candidate selection · ε=0.035, δ=0.01 | probability of returning NSF: ε=0.035; δ=0.01; α=1.0 | Hurt |
| use Hoeffding's inequality rather than Student's t-test for confidence bounds (ablation) | FRG confidence-bound procedure · the paper's experimental datasets | confidence-bound tightness | Hurt |
Surprising findings
- Supervised FRG yields only insignificant improvement over unsupervised FRG on Adult and Health, since more predictive candidates can expose more sensitive information.
- Analytic upper bounds on Delta_DP from prior work are typically loose and impractical for establishing guarantees while preserving utility.
Limitations
- Guarantees rely on i.i.d. data, a CLT-based Student's t-test requiring large samples, and an approximately optimal adversary rather than a true optimum.
- Evaluated only on binary-classification fairness with demographic parity in main experiments; extensions to other fairness notions are in appendix.
- Returns NSF when the fairness test fails, so the probability of returning a solution can be limited with small epsilon, delta, or data.
Experimental setup
Datasets: UCI Adult Heritage Health ACSIncome
Benchmarks: UCI Adult Heritage Health ACSIncome
Metrics: AUC demographic parity difference probability of violating constraint probability of returning solution (non-NSF) F1 accuracy equal opportunity difference equalized odds difference
Baselines: LAFTR ICVAE LMIFR CFAIR FCRL FARE
Builds on: FARE LAFTR VAE Seldonian algorithms
Closest papers by meaning
- Learning Adversarially Fair and Transferable Representations (ICML 2018 · similarity 0.91)
- On the Fairness ROAD: Robust Optimization for Adversarial Debiasing (ICLR 2024 · similarity 0.88)
- Kernel-based Equalized Odds: A Quantification of Accuracy-Fairness Trade-off in Fair Representation Learning (NeurIPS 2025 · similarity 0.88)
- Towards Poisoning Fair Representations (ICLR 2024 · similarity 0.87)
- On Optimal Steering to Achieve Exact Fairness (NeurIPS 2025 · similarity 0.87)
- DAFA: Distance-Aware Fair Adversarial Training (ICLR 2024 · similarity 0.87)
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.