MIFair: A Mutual-Information Framework for Intersectionality and Multiclass Fairness

Problem
This paper addresses the limitations of existing fairness frameworks in machine learning, particularly their inability to handle intersectionality and multiclass classification effectively. The authors highlight the lack of a universal definition of fairness and the challenges posed by context-specific bias metrics. MIFair is proposed as a solution to these gaps, providing a unified approach to bias assessment and mitigation. This work is presented as a preprint and has not yet undergone peer review.

Method
MIFair introduces a mutual-information-based framework for assessing and mitigating bias in machine learning models. The core technical contributions include a flexible metric template for fairness and an in-processing mitigation method inspired by the Prejudice Remover. The framework defines group fairness through the statistical independence between prediction-derived variables and sensitive attributes. MIFair’s information-theoretic foundation is reinforced by establishing equivalences with established fairness concepts such as independence and separation. The method supports intersectionality and complex subgroup structures, making it applicable to multiclass classification tasks. Regularization-based training is employed to reduce bias according to the selected fairness metric, allowing for a versatile application across various datasets.

Results
The authors conducted experiments on real-world tabular and image datasets to evaluate the effectiveness of MIFair. The results indicate that MIFair significantly reduces bias in multiclass and intersectional scenarios, outperforming traditional fairness methods. Specific performance metrics were not disclosed in the abstract, but the authors claim that MIFair maintains strong predictive performance while addressing previously unconsidered multi-attribute bias. The paper suggests that MIFair consolidates diverse fairness requirements into a single framework, facilitating consistent benchmarking across different applications.

Limitations
The authors acknowledge that while MIFair provides a comprehensive approach to fairness, it may still be limited by the complexity of defining sensitive attributes and the potential for overfitting in regularization-based training. Additionally, the framework’s reliance on mutual information may not capture all nuances of fairness in every context. The paper does not address the computational overhead that may arise from implementing the proposed methods in large-scale applications.

Why it matters
MIFair’s introduction of a mutual-information framework for fairness has significant implications for the field of machine learning, particularly in applications where intersectionality and multiclass classification are prevalent. By providing a unified and flexible approach to bias assessment and mitigation, MIFair can enhance the ethical deployment of machine learning systems across diverse domains. This work encourages further exploration of information-theoretic methods in fairness research and sets a foundation for future studies to build upon, potentially leading to more robust and equitable AI systems.

Authors: Jeanne Monnier, Thomas George, Frédéric Guyard, Christèle Tarnec, Marios Kountouris
Source: arXiv:2604.28030
URL: https://arxiv.org/abs/2604.28030v1