Bayes-Optimal Fair Classification with Linear Disparity Constraints via Pre-, In-, and Post-processing

2025-12-31 · Source: JMLR · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Xianli Zeng, Kevin Jiang, Guang Cheng, and Edgar Dobriban present methods for Bayes-optimal fair classification, published in 27(65):1−87, 2026. This research minimizes classification error while satisfying group fairness constraints, addressing disparate impacts on protected groups. The authors define linear and bilinear disparity measures, identifying popular metrics like demographic parity and equality of opportunity as bilinear. They derive Bayes-optimal fair classifiers under a single linear disparity measure, linking it to the Neyman-Pearson lemma, and provide explicit group-wise thresholding rules for bilinear measures. The work extends to multi-class protected attributes and equalized odds, even when protected attributes are not used in prediction. Based on these theoretical insights, the team developed practical methods: Fair Up- and Down-Sampling (pre-processing), Fair cost-sensitive Classification (in-processing), and a Fair Plug-In Rule (post-processing). These methods directly control disparity, achieve near-optimal fairness-accuracy tradeoffs, and show strong empirical performance, with their pre-processing method outperforming prior approaches at low disparity levels.

Key takeaway

For Machine Learning Engineers designing fair classification systems, you should consider these Bayes-optimal methods to directly control disparity while maintaining high accuracy. The explicit forms for linear and bilinear disparity measures, combined with the proposed pre-processing, in-processing, and post-processing techniques, offer a principled approach. You can achieve near-optimal fairness-accuracy tradeoffs. The pre-processing method shows superior accuracy at low disparity levels compared to prior methods. Implement these to improve your model's fairness guarantees.

Key insights

Bayes-optimal fair classifiers are derived for linear and bilinear disparity measures, enabling direct disparity control.

Principles

• Bayes-optimal fair classification minimizes error under fairness constraints.
• Neyman-Pearson lemma informs optimal fair classifier forms.
• Bilinear disparity measures yield explicit thresholding rules.

Method

Algorithms are designed for pre-processing (Fair Up- and Down-Sampling), in-processing (Fair cost-sensitive Classification), and post-processing (Fair Plug-In Rule) to learn fair Bayes-optimal classifiers.

In practice

• Implement group-wise thresholding for bilinear fairness.
• Apply Fair Up- and Down-Sampling for pre-processing.
• Use Fair Plug-In Rule for post-processing adjustments.

Topics

• Fair Classification
• Bayes-Optimal Classifiers
• Disparity Measures
• Algorithmic Fairness
• Pre-processing Methods
• Neyman-Pearson Lemma

Code references

• XianliZeng/Bayes-Optimal-Fair-Classification

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer

Related on AIssential

• Demystifying the Optimal Fair Classifier in Multi-Class Classification — This work characterizes and provides algorithms to achieve the optimal accuracy-fairness frontier in multi-class classification.
• Towards Provably Fair Machine Learning: Bayesian Approaches For Consistent and Transparent Predictions — The Fair Bayesian classifier ensures consistent, fair predictions across all subgroups by enforcing determinism and statistical consistency with principled abstention.
• Fairness Beyond Demographics: Optimizing Performance Across Appearance-Based Hidden Cohorts in Medical Imaging — Optimizing fairness across appearance-based hidden cohorts improves medical image analysis models without relying on demographic labels.
• No-Free-Fairness: Fundamental Limits and Trade-offs in Learning Systems — Unfairness in learning systems is fundamentally inherent, arising from problem structure, data limits, and model expressivity.
• Identification and Inference for Algorithmic Frontiers with Selective Labels — The paper identifies and infers fairness-accuracy frontiers despite selective label observation, using debiased machine learning.
• FairEnc: A Fair Vision-Language Model with Fair Vision and Text Encoders for Glaucoma Detection — FairEnc is a VLM pretraining method for glaucoma detection that debiases visual and text encoders against multiple sensitive attributes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by JMLR.