Research map / Learning theory and algorithms / Fairness
FairSplit: Decomposing the Embedding Space for Fair Classification
Andrii Shkabrii, Anna Beer, Ylli Sadikaj, Claudia Plant
Read the paperCheck your idea against 137,217 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 FairSplit, which decomposes the embedding space into a Fair Space used for prediction and a Sensitive Space that isolates sensitive attributes for fair classification.
Problem
Fair classification must keep high accuracy while limiting reliance on sensitive information in learned representations.
Approach
Learns a decomposition of the embedding space into a Fair Space used for prediction and a Sensitive Space that isolates sensitive information, trained jointly with the classification task.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| learn a mask-based Fair Space / Sensitive Space decomposition instead of applying HSIC to the full embedding (ablation) | FairSplit · CelebA-A | ΔDP: 26.6 -> 18.3 | Improved |
| add HSIC regularization on the embedding without decomposition (ablation) | ERM · CelebA-A | ΔDP: 52.8 -> 26.6 | Improved |
| add the learned mask before any loss acts on the Sensitive Space (ablation) | FairSplit · CelebA-A | ΔDP | Improved |
| regularize the Sensitive Space (lambda_s, rho_s > 0) so sensitive information moves into it (ablation) | FairSplit · CelebA-A | ΔDP | Improved |
| increase rho_s to shrink Fair Space dimensionality (ablation) | FairSplit · CelebA-A | ΔDP | Improved |
| increase the independence weight rho_f (ablation) | FairSplit · CelebA-A | ΔDP | Improved |
| vary the Fair Space compression weight lambda_f (ablation) | FairSplit · CelebA-A | ΔDP | No clear effect |
| change HSIC kernel type or bandwidth (ablation) | FairSplit · CelebA-A | ΔDP | No clear effect |
| use FairSplit instead of ERM when fine-tuning ViT-B/16 | ViT-B/16 · FairFace age>=30 | ΔDP: 23.0 -> 9.5 | Improved |
| use FairSplit instead of ERM when fine-tuning ViT-B/16 | ViT-B/16 · FairFace age>=30 | AP: 88.9 -> 86.4 | Hurt |
| transfer a FairSplit Fair Space to a new linear head on frozen CLIP ViT-L/14 | CLIP ViT-L/14 (frozen) · CelebA age | ΔDP: 21.4 -> 9.1 | Improved |
| 30% of sensitive labels missing during training (ablation) | FairSplit · CelebA-A | ΔDP: 52.8 (ERM) vs 28.0 | Mixed |
Closest papers by meaning
- FairNet: Dynamic Fairness Correction without Performance Loss via Contrastive Conditional LoRA (NeurIPS 2025 · similarity 0.88)
- Kernel-based Equalized Odds: A Quantification of Accuracy-Fairness Trade-off in Fair Representation Learning (NeurIPS 2025 · similarity 0.88)
- Constructing Fair Latent Space for Intersection of Fairness and Explainability (AAAI 2025 · similarity 0.88)
- Understanding Fairness and Prediction Error through Subspace Decomposition and Influence Analysis (NeurIPS 2025 · similarity 0.88)
- A Sequentially Fair Mechanism for Multiple Sensitive Attributes (AAAI 2024 · similarity 0.88)
- Efficient Fairness-Performance Pareto Front Computation (NeurIPS 2025 · similarity 0.88)
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.
