[Paper Review] InfoFair: Information-Theoretic Intersectional Fairness

Jian Kang, Tiankai Xie+3 more|arXiv (Cornell University)|May 24, 2021

Ethics and Social Impacts of AI37 references4 citations

TL;DR

This paper proposes InfoFair, a novel information-theoretic framework for intersectional fairness that minimizes mutual information between model predictions and a joint vector of multiple sensitive attributes to ensure statistical parity across all demographic subgroups. By leveraging a variational representation of mutual information, the method enables end-to-end, gradient-based debiasing in diverse learning tasks with minimal accuracy drop, demonstrating strong fairness improvements on real-world datasets.

ABSTRACT

Algorithmic fairness is becoming increasingly important in data mining and machine learning. Among others, a foundational notation is group fairness. The vast majority of the existing works on group fairness, with a few exceptions, primarily focus on debiasing with respect to a single sensitive attribute, despite the fact that the co-existence of multiple sensitive attributes (e.g., gender, race, marital status, etc.) in the real-world is commonplace. As such, methods that can ensure a fair learning outcome with respect to all sensitive attributes of concern simultaneously need to be developed. In this paper, we study the problem of information-theoretic intersectional fairness (InfoFair), where statistical parity, a representative group fairness measure, is guaranteed among demographic groups formed by multiple sensitive attributes of interest. We formulate it as a mutual information minimization problem and propose a generic end-to-end algorithmic framework to solve it. The key idea is to leverage a variational representation of mutual information, which considers the variational distribution between learning outcomes and sensitive attributes, as well as the density ratio between the variational and the original distributions. Our proposed framework is generalizable to many different settings, including other statistical notions of fairness, and could handle any type of learning task equipped with a gradient-based optimizer. Empirical evaluations in the fair classification task on three real-world datasets demonstrate that our proposed framework can effectively debias the classification results with minimal impact to the classification accuracy.

Motivation & Objective

• To address the gap in existing group fairness methods that focus on single sensitive attributes rather than intersections of multiple attributes.
• To develop a generalizable, end-to-end framework that directly optimizes statistical parity across all demographic groups formed by multiple sensitive attributes.
• To overcome limitations of surrogate constraints and unstable performance in prior methods by directly minimizing mutual information between predictions and joint sensitive attributes.
• To ensure fairness without sacrificing model accuracy, particularly in complex, real-world learning tasks.
• To extend the applicability of fairness methods beyond binary or categorical attributes to continuous-valued sensitive attributes in future work.

Proposed method

• The framework models all sensitive attributes as a joint vector to define fine-grained demographic groups for intersectional fairness.
• It formulates the fairness objective as minimizing the mutual information between model predictions and the joint sensitive attribute vector.
• A variational lower bound of mutual information is used, incorporating a variational distribution and density ratio estimation to enable differentiable optimization.
• The method leverages deep neural networks with gradient-based optimization, allowing integration into any differentiable learning task.
• The framework is generalizable to other fairness notions such as equalized odds and equal opportunity with minor modifications.
• It uses neural estimators like MINE or infoNCE for efficient mutual information estimation in high-dimensional spaces.

Experimental results

Research questions

• RQ1 Can mutual information minimization effectively enforce statistical parity across all intersectional demographic groups formed by multiple sensitive attributes?
• RQ2 How does the proposed variational mutual information framework compare to existing fairness methods in terms of fairness-accuracy trade-offs?
• RQ3 To what extent can the framework be generalized across different learning tasks and fairness definitions?
• RQ4 Does the method outperform surrogate-constrained approaches in mitigating bias when the model does not perfectly capture data-distribution relationships?
• RQ5 Can the framework be extended to handle continuous-valued sensitive attributes such as age or income?

Key findings

• The InfoFair framework achieves significant bias reduction across intersectional demographic groups in fair classification tasks with minimal impact on classification accuracy.
• Empirical evaluations on three real-world datasets show consistent fairness improvements across multiple sensitive attributes, including race, gender, and marital status.
• The method outperforms existing approaches that rely on surrogate constraints, particularly when model capacity is limited or data distributions are complex.
• The framework maintains strong performance even when the model does not perfectly capture the ground-truth data distribution, indicating robustness to approximation errors.
• The approach generalizes well beyond binary classification, demonstrating potential applicability to recommendation and other learning tasks.
• Preliminary results suggest the framework can be extended to continuous-valued sensitive attributes, though this remains an open direction for future work.