[Paper Review] An Intersectional Definition of Fairness

James R. Foulds, Shimei Pan+2 more|arXiv (Cornell University)|Jul 1, 2018

Ethics and Social Impacts of AI33 references19 citations

TL;DR

This paper introduces an intersectional fairness framework for AI and machine learning that extends traditional fairness definitions by accounting for overlapping social identities such as race, gender, and class. It proposes three novel fairness metrics—differential fairness (DF), DF bias amplification, and differential fairness with confounders (DFC)—and a learning algorithm that enforces these criteria, demonstrating improved fairness for minority groups compared to subgroup fairness, especially in real-world datasets like COMPAS and Adult.

ABSTRACT

We propose definitions of fairness in machine learning and artificial intelligence systems that are informed by the framework of intersectionality, a critical lens arising from the Humanities literature which analyzes how interlocking systems of power and oppression affect individuals along overlapping dimensions including gender, race, sexual orientation, class, and disability. We show that our criteria behave sensibly for any subset of the set of protected attributes, and we prove economic, privacy, and generalization guarantees. We provide a learning algorithm which respects our intersectional fairness criteria. Case studies on census data and the COMPAS criminal recidivism dataset demonstrate the utility of our methods.

Motivation & Objective

• To address the limitations of existing fairness definitions in machine learning that fail to account for overlapping social identities.
• To develop fairness criteria grounded in intersectionality theory, which recognizes systemic oppression across race, gender, class, and other dimensions.
• To ensure fairness metrics are equitable across all subgroups, especially underrepresented minorities, by avoiding the inequity of subgroup fairness metrics.
• To provide theoretical guarantees for privacy, economic fairness, and generalization in the proposed framework.
• To empirically validate the method on real-world datasets, showing superior performance over subgroup fairness in protecting minority groups.

Proposed method

• Proposes differential fairness (DF) as a fairness metric that ensures equitable treatment across all intersectional subgroups defined by protected attributes.
• Introduces DF bias amplification to measure the extent to which an algorithm increases bias compared to the original data distribution.
• Develops differential fairness with confounders (DFC) to adjust for external factors that alter outcome distributions, improving fairness in real-world settings.
• Designs a learning algorithm that optimizes for DF, DF bias amplification, and DFC, using constrained optimization to enforce fairness constraints during model training.
• Employs a Gini coefficient-based metric to quantify inequity in fairness measurement across subgroups, enabling comparison of fairness distributional equity.
• Applies the framework to the COMPAS and Adult census datasets, using both real data and trained logistic regression models to evaluate fairness across multiple protected attribute combinations.

Experimental results

Research questions

• RQ1 How can fairness in AI be redefined to account for the intersection of multiple protected attributes such as race and gender?
• RQ2 Does the proposed differential fairness (DF) metric ensure equitable fairness measurement across subgroups of all sizes, including underrepresented minorities?
• RQ3 To what extent does the γ-SF fairness metric fail to preserve intersectionality, and how does it compare to ϵ-DF in terms of fairness distribution equity?
• RQ4 Can the proposed fairness metrics and learning algorithm achieve better fairness for minority groups than subgroup fairness, particularly in real-world datasets with complex demographic intersections?
• RQ5 How do confounding variables affect fairness, and can DFC effectively adjust for their influence on outcome distributions?

Key findings

• The ϵ-DF metric demonstrated significantly lower Gini coefficient (0.099 on Adult, 0.151 on COMPAS) than γ-SF (0.256 and 0.376 respectively), indicating more equitable distribution of fairness consideration across subgroups.
• In the COMPAS dataset, ϵ-DF maintained intersectionality by ensuring that fairness values did not increase when adding more protected attributes, while γ-SF violated this property (e.g., γ for gender > γ for race×gender).
• The DF-Classifier improved per-group fairness across all group sizes under both ϵ-DF and γ-SF, whereas the SF-classifier failed to improve γ-SF fairness for small groups.
• On the Adult dataset, ϵ-DF achieved a Gini coefficient of 0.2177 for nationality, while γ-SF scored 0.0045, indicating that γ-SF disproportionately downweights minority groups.
• The ϵ-DF metric was observed to treat all groups equally in fairness measurement, with high per-group ϵ values across both large and small subgroups, unlike γ-SF which favored larger groups.
• The study confirmed that subgroup fairness (γ-SF) can lead to 'fairness gerrymandering' by ignoring minority groups in overall fairness assessment, while the proposed DF framework avoids this by ensuring inclusive fairness evaluation.