Differential Fairness: An Intersectional Framework for Fair AI

Area of Science:

• Computer Science
• Social Science
• Law

Background:

• Existing fairness metrics in AI/ML often fail to capture the complex, overlapping nature of discrimination.
• Intersectionality, a framework from social sciences and law, offers a lens to understand how multiple identity dimensions (e.g., race, gender, disability) interact to create unique experiences of oppression.
• Applying intersectional principles to AI fairness is crucial for mitigating nuanced forms of bias.

Purpose of the Study:

• To propose novel definitions of fairness in AI and machine learning systems grounded in the intersectionality framework.
• To develop theoretical guarantees (economic, privacy, generalization) for these intersectional fairness criteria.
• To create practical algorithms that operationalize and measure intersectional fairness in real-world applications.

Main Methods:

• Developed intersectional fairness criteria by extending existing fairness definitions to account for overlapping protected attributes.
• Proved theoretical guarantees, including economic, privacy, and generalization bounds, for the proposed fairness criteria.
• Designed and implemented two learning algorithms: a deterministic gradient method for theoretical analysis and a stochastic gradient method for scalability to big data, incorporating minibatch estimation.

Main Results:

• Demonstrated that the proposed fairness criteria are sensible across subsets of protected attributes and provide meaningful operationalization of AI fairness.
• Showcased the interpretability of the fairness measurements, analogous to differential privacy.
• Validated the utility of the developed algorithms through case studies on diverse datasets (census, COMPAS, HHP, HMDA).

Conclusions:

• The proposed intersectional fairness criteria offer a robust framework for developing more equitable AI systems.
• The developed algorithms effectively address the challenges of data sparsity in measuring intersectional fairness, enabling practical application.
• This work bridges theoretical fairness concepts with practical implementation, paving the way for fairer AI across various domains.