Research map / Label-efficient and robust learning / Recognition and long-tail
FADES: Fair Disentanglement with Sensitive Relevance
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 FADES, a VAE-based fair disentanglement method that adds a sensitive relevant latent code and minimizes conditional mutual information so that target-sensitive correlations are absorbed by that code without domain-knowledge causal graphs.
Problem
Fair representation learning under data bias, where target labels and sensitive attributes are correlated, makes strict disentanglement into independent target and sensitive codes contradictory and unstable, and causal-graph methods need domain knowledge.
Approach
Introduce a sensitive relevant latent code z_R that captures information shared by target and sensitive attributes, and impose conditional independence of target and sensitive predictions given z_R by minimizing a conditional mutual information loss (with an information bottleneck regularizer), so that the remaining target and sensitive codes become independent without adversarial training or caus
Key results
- On CelebA, FADES reaches 0.918 accuracy with 0.032 EOD, versus 0.919 accuracy and 0.049 EOD for GVAE.
- On UTKFace, FADES reaches 0.802 accuracy with 0.057 EOD, while GVAE reaches 0.819 accuracy with 0.207 EOD.
- On C-MNIST reconstruction, FADES achieves 93.46% digit and 95.31% color recovery, above all baselines on digit accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FADES instead of the strongest reported baseline | CelebA downstream classification · CelebA | EOD: 0.038 -> 0.032 | Helped |
| Use FADES instead of the strongest reported baseline | UTKFace downstream classification · UTKFace | EOD: 0.271 -> 0.057 | Helped |
| Use FADES instead of the strongest reported baseline | Dogs and Cats downstream classification · Dogs and Cats | Acc: 0.727 -> 0.769 | Helped |
| Use FADES instead of the strongest reported baseline | Adult downstream classification · Adult | Acc: 0.851 -> 0.845 | Hurt |
| Use FADES instead of the strongest reported baseline | Unbiased C-MNIST reconstruction · Unbiased C-MNIST | Acc (Digit): 87.21 -> 93.46 | Helped |
| Use FADES instead of the strongest reported baseline | Unbiased C-MNIST reconstruction · Unbiased C-MNIST | Acc (Color): 99.31 -> 95.31 | Hurt |
| Add information-bottleneck regularization with β decreased through β = 0.5 (ablation) | FADES on UTKFace · UTKFace | accuracy and EOD | Helped |
| Increase information-bottleneck weight beyond β = 0.5 (ablation) | FADES on UTKFace · UTKFace | EOD | Hurt |
Surprising findings
- FairDisCo reaches 99.31% color accuracy on C-MNIST, higher than FADES, though it takes ground-truth color labels at the decoder.
Limitations
- Hyperparameters (lambda_CMI, lambda_TC, beta) require grid search, with the information bottleneck weight sensitive to tuning.
- Fairness violation increases when beta exceeds 0.5 in the ablation on UTKFace.
Experimental setup
Datasets: CelebA UTKFace Dogs and Cats Adult C-MNIST
Benchmarks: CelebA UTKFace Dogs and Cats Adult C-MNIST
Metrics: accuracy equalized odds demographic parity digit accuracy color accuracy
Baselines: GVAE FFVAE ODVAE FairDisCo FairFactorVAE
Builds on: beta-VAE FactorVAE FFVAE GVAE ODVAE
Closest papers by meaning
- FairNet: Dynamic Fairness Correction without Performance Loss via Contrastive Conditional LoRA (NeurIPS 2025 · similarity 0.86)
- Disentangling, Amplifying, and Debiasing: Learning Disentangled Representations for Fair Graph Neural Networks (AAAI 2025 · similarity 0.85)
- FairSIN: Achieving Fairness in Graph Neural Networks through Sensitive Information Neutralization (AAAI 2024 · similarity 0.85)
- FairICP: Encouraging Equalized Odds via Inverse Conditional Permutation (ICML 2025 · similarity 0.85)
- Causal Fairness under Unobserved Confounding: A Neural Sensitivity Framework (ICLR 2024 · similarity 0.85)
- Kernel-based Equalized Odds: A Quantification of Accuracy-Fairness Trade-off in Fair Representation Learning (NeurIPS 2025 · similarity 0.85)
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.