Explainable post-training bias mitigation with distribution-based fairness metrics

2025-04-01

AI Summary

This paper addresses the challenge of post-training fairness adjustment without model retraining. Methodologically, it introduces a novel distribution-driven differentiable fairness constraint optimization framework, designs an interpretable global bias metric family compatible with diverse models (e.g., gradient-boosted trees), and establishes a continuous, controllable post-processing paradigm for fairness calibration. Technically, it integrates distribution alignment constraints, interpretability regularization, and bias calibration. Evaluated on multiple benchmark datasets, the approach maintains high predictive accuracy while improving demographic parity (DP) and equalized odds (EO) by 30–50% over state-of-the-art post-processing methods. Moreover, it enables global bias attribution with human-interpretable explanations, offering both quantitative fairness enhancement and qualitative insight into bias sources.

Technology Category

• Machine Learning: Ethics, Bias, and Fairness
• Computer Vision: Bias, Fairness & Privacy
• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Social Networks and Social Media: Fairness and bias in social network and social media analysis

Abstract

We develop a novel optimization framework with distribution-based fairness constraints for efficiently producing demographically blind, explainable models across a wide range of fairness levels. This is accomplished through post-processing, avoiding the need for retraining. Our framework, which is based on stochastic gradient descent, can be applied to a wide range of model types, with a particular emphasis on the post-processing of gradient-boosted decision trees. Additionally, we design a broad class of interpretable global bias metrics compatible with our method by building on previous work. We empirically test our methodology on a variety of datasets and compare it to other methods.

Problem

• Mitigates post-training bias with fairness constraints
• Avoids retraining via explainable post-processing
• Applies to diverse models, especially decision trees

Innovation

• Optimization framework with fairness constraints
• Post-processing without retraining models
• Interpretable global bias metrics design