Research map / Label-efficient and robust learning / Domain adaptation and generalization
Fairness without Demographics through Shared Latent Space-Based Debiasing
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 SLSD, which uses deep CCA over a shared latent space between a labeled source domain and an unlabeled target domain to estimate hidden protected groups, then debiases the target classifier via adversarial learning, with a relaxed variant R-SLSD using 1% protected labels.
Problem
Fair classification when protected attributes (e.g., gender, race) are unavailable in the target domain due to privacy or regulation, so group fairness cannot be directly enforced.
Approach
Transfers protected-group correlations from a labeled source domain to an unlabeled target domain by learning correlated latent representations via deep CCA with outcome-based sampling, refining group estimates with Cauchy-noise consistency training, and debiasing the classifier adversarially using the estimated groups.
Key results
- SLSD improves DPD on Adult by 52.4% over Baseline (DPD 0.078 vs 0.164) and DPR from 0.349 to 0.575.
- R-SLSD (1% protected labels) reaches DPD 0.080 on Default and 0.083 on ACSIncome, outperforming DRO and ARL on fairness metrics.
- SLSD lowers utility, e.g., Adult AUC min drops to 0.822 vs 0.888 for Baseline, a stated cost of fairness.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use shared latent-space group estimation and adversarial debiasing without target demographics | SLSD on Adult · Adult | DPR: 0.349±0.016 -> 0.575±0.158 | Helped |
| Use shared latent-space group estimation and adversarial debiasing without target demographics | SLSD on Adult · Adult | DPD: 0.164±0.012 -> 0.078±0.033 | Helped |
| Use shared latent-space group estimation and adversarial debiasing without target demographics | SLSD on Default · Default | DPD: 0.114±0.004 -> 0.084±0.029 | Helped |
| Use shared latent-space group estimation and adversarial debiasing without target demographics | SLSD on ACSIncome · ACSIncome | DPD: 0.123±0.004 -> 0.096±0.020 | Helped |
| Use 1% of target training examples with protected attributes (R-SLSD) | R-SLSD on Adult · Adult | DPR: 0.349±0.016 -> 0.559±0.047 | Helped |
| Use 1% of target training examples with protected attributes (R-SLSD) | R-SLSD on Default · Default | DPD: 0.114±0.004 -> 0.080±0.013 | Helped |
| Use 1% of target training examples with protected attributes (R-SLSD) | R-SLSD on ACSIncome · ACSIncome | DPD: 0.123±0.004 -> 0.083±0.006 | Helped |
| Apply DRO instead of the standard baseline | DRO · Adult | DPD: 0.164±0.012 -> 0.146±0.027 | Helped |
| Apply ARL instead of the standard baseline | ARL · Adult | DPD: 0.164±0.012 -> 0.170±0.004 | Hurt |
| Apply ARL instead of the standard baseline | ARL · Default | DPD: 0.114±0.004 -> 0.175±0.006 | Hurt |
| Transfer group estimation from Adult to COMPAS across distinct domains | SLSD · COMPAS | DPD: 0.235 -> 0.221 | Helped |
| Transfer group estimation from Adult to COMPAS across distinct domains using R-SLSD | R-SLSD · COMPAS | DPR: 0.642 -> 0.718 | Helped |
Surprising findings
- DRO and ARL often amplify the baseline's demographic bias on Adult and Default datasets.
- Fair learning methods relying on explicit fairness metric measurement struggle with noisy group estimates, motivating adversarial debiasing instead.
Limitations
- Performance drops in extreme cross-domain scenario (COMPAS target with Adult source), with reduced fairness gains.
- SLSD sacrifices AUC and balanced accuracy for fairness improvements.
- Currently restricted to binary classification and binary protected attributes (one-vs-all).
Experimental setup
Datasets: Adult ACSIncome Default COMPAS
Benchmarks: Adult ACSIncome Default COMPAS
Metrics: AUC overall AUC min Balanced accuracy overall Balanced accuracy min demographic parity difference demographic parity ratio
Baselines: Baseline DRO ARL ADM
Builds on: Deep CCA adversarial debiasing DRO ARL UDA consistency training
Closest papers by meaning
- Group Fairness by Probabilistic Modeling with Latent Fair Decisions (AAAI 2021 · similarity 0.89)
- Adversarial Latent Feature Augmentation for Fairness (ICLR 2025 · similarity 0.89)
- Fairness via Independence: A General Regularization Framework for Machine Learning (ICLR 2026 · similarity 0.89)
- On the Fairness ROAD: Robust Optimization for Adversarial Debiasing (ICLR 2024 · similarity 0.88)
- Fairness without Harm: An Influence-Guided Active Sampling Approach (NeurIPS 2024 · similarity 0.88)
- FFB: A Fair Fairness Benchmark for In-Processing Group Fairness Methods (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.