Research map / Learning theory and algorithms / Fairness
FairICP: Encouraging Equalized Odds via Inverse Conditional Permutation
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 FairICP, an adversarial in-processing method for equalized odds with multi-dimensional or mixed-type sensitive attributes, using inverse conditional permutation (sampling A given Y via permutations weighted by Y|A density) and a KPC-based evaluation.
Problem
Enforcing equalized odds when the sensitive attribute is multi-dimensional or mixed-type (continuous, categorical), where estimating the conditional density A|Y is hard due to the curse of dimensionality.
Approach
Generate conditionally permuted copies of sensitive attributes by estimating the easier Y|A (one-dimensional response) instead of A|Y, via an inverse conditional permutation scheme, and plug these copies into the Fair Dummies adversarial framework; also use ICP for conditional permutation hypothesis tests.
Key results
- On Communities and Crime with 3 sensitive attributes, FairICP achieves MSE 0.418 with KPC 0.054 versus HGR's 0.596 loss and 0.068 KPC.
- On ACSIncome, FairICP achieves loss 0.215 and KPC 0.025 (power 0.80), outperforming HGR and GerryFair on the fairness-accuracy trade-off.
- In simulations, restricted TV distance of ICP permutations is much lower than CP's as the dimension of sensitive attributes grows.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FairICP instead of FairCP, with ICP-generated sensitive-attribute permutations (ablation) | Fairness-aware learning · Simulation studies | prediction loss–KPC trade-off | Helped |
| Use FairICP instead of FDL, replacing conditional randomization with ICP permutations (ablation) | Fairness-aware learning · Simulation studies | prediction loss–KPC trade-off | Helped |
| Use ICP rather than CP to generate conditional permutations (ablation) | Conditional permutation sampling · Simulation studies | restricted total variation distance | Helped |
| Increase the fairness penalty μ from 0 | FairICP and comparison methods · Simulation studies | KPC | Helped |
| Increase the fairness penalty μ from 0 | FairICP and comparison methods · Simulation studies | prediction loss | Hurt |
| Use FairICP rather than HGR | Fairness-aware learning · Crimes (all races) | prediction loss: 0.596(0.050) -> 0.418(0.047) | Helped |
| Use FairICP rather than HGR | Fairness-aware learning · ACS Income | prediction loss: 0.220(0.004) -> 0.215(0.003) | Helped |
| Use FairICP rather than Reduction | Fairness-aware learning · Adult | prediction loss: 0.161(0.004) -> 0.159(0.003) | Helped |
| Use FairICP rather than Reduction | Fairness-aware learning · Adult | KPC: 0.005(0.15) -> 0.010(0.13) | Hurt |
| Use FairICP rather than Reduction | Fairness-aware learning · COMPAS | KPC: 0.002(0.05) -> 0.008(0.04) | Hurt |
| Use FairICP rather than Reduction | Fairness-aware learning · COMPAS | prediction loss: 0.400(0.02) -> 0.402(0.02) | Hurt |
Surprising findings
- Reduction achieves the best loss and DEO on COMPAS in the table, so FairICP is not uniformly best across all tasks.
Limitations
- Adversarial training carries computational overhead, especially on large datasets.
- Requires access to sensitive attributes during training.
- Performance depends on accuracy of the estimated Y|A density.
Experimental setup
Datasets: Communities and Crime ACSIncome Adult COMPAS
Benchmarks: Communities and Crime ACSIncome Adult COMPAS simulation studies
Metrics: prediction loss (MSE / misclassification rate) KPC testing power DEO total variation distance
Baselines: FDL HGR GerryFair Reduction FairCP Oracle
Builds on: Fair Dummies Learning Conditional Permutation Holdout Randomization Test Masked Autoregressive Flow
Closest papers by meaning
- A Sequentially Fair Mechanism for Multiple Sensitive Attributes (AAAI 2024 · similarity 0.89)
- On the Fairness ROAD: Robust Optimization for Adversarial Debiasing (ICLR 2024 · similarity 0.88)
- 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)
- On the Alignment between Fairness and Accuracy: from the Perspective of Adversarial Robustness (ICML 2025 · similarity 0.87)
- Adversarial Fairness Network (AAAI 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.