Research map / Learning theory and algorithms / Fairness
Fair Representation Learning with Maximum Mean Discrepancy Distance Constraint (Student Abstract)
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 a fair dimensionality-reduction framework that adds an MMD penalty between protected-group embedding distributions to the base loss (e.g., t-SNE KL), solved via Lagrangian gradient descent with a binary search over the multiplier.
Problem
Dimensionality reduction methods (PCA, t-SNE, autoencoders) can encode protected attributes such as gender in their representations, and there is no general framework that enforces fairness while preserving the underlying structure.
Approach
Constrain the embedding so that the MMD between the embedding distributions of the two protected groups is bounded by a threshold delta, and solve the resulting constrained problem by Lagrangian relaxation where a binary search selects the multiplier beta, with gradient descent per beta.
Key results
- With delta set to half of vanilla t-SNE's MMD, the gender separation becomes less apparent while income classes remain clustered.
- With delta set to 1/50 of vanilla t-SNE's MMD, males and females are scattered while income classes remain relatively separated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Set the MMD threshold δ to roughly 1/2 of vanilla t-SNE's MMD | Algorithm 1 with t-SNE · Census Income (1000 samples) | gender separation | Helped |
| Set the MMD threshold δ to roughly 1/2 of vanilla t-SNE's MMD | Algorithm 1 with t-SNE · Census Income (1000 samples) | income-class clustering | No clear effect |
| Set the MMD threshold δ to 1/50 of vanilla t-SNE's MMD | Algorithm 1 with t-SNE · Census Income (1000 samples) | gender separation | Helped |
| Set the MMD threshold δ to 1/50 of vanilla t-SNE's MMD | Algorithm 1 with t-SNE · Census Income (1000 samples) | income-class clustering | No clear effect |
Surprising findings
- Even under strong fairness constraint (delta at 1/50 of baseline MMD), income class structure remains relatively separated.
Limitations
- Problem is non-convex, so the binary search and gradient descent only find a heuristic solution.
- Evaluation is a small qualitative visualization on one 1000-sample subset with no quantitative fairness or utility metrics.
- Applicability to autoencoders and PCA is claimed but not experimentally shown.
Experimental setup
Datasets: Census Income (UCI Adult), 1000-sample subset
Benchmarks: Census Income subset (qualitative visualization)
Metrics: MMD between gender groups visual separation of gender and income classes
Baselines: vanilla t-SNE
Builds on: t-SNE Maximum Mean Discrepancy Zemel et al. 2013 fair representations
Closest papers by meaning
- Kernel-based Equalized Odds: A Quantification of Accuracy-Fairness Trade-off in Fair Representation Learning (NeurIPS 2025 · similarity 0.84)
- Flexibly Fair Representation Learning by Disentanglement (ICML 2019 · similarity 0.84)
- Self-Supervised Fair Representation Learning without Demographics (NeurIPS 2022 · similarity 0.84)
- Fair Representation Learning with Controllable High Confidence Guarantees via Adversarial Inference (NeurIPS 2025 · similarity 0.84)
- Stabilizing Sample Similarity in Representation via Mitigating Random Consistency (ICML 2025 · similarity 0.84)
- I-Con: A Unifying Framework for Representation Learning (ICLR 2025 · similarity 0.83)
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.