On Demographic Group Fairness Guarantees in Deep Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Deep learning models often exhibit fairness disparities across demographic groups.
• Data distribution heterogeneity is a key factor contributing to these fairness gaps.
• Existing fairness guarantees may not adequately address distributional shifts.

Purpose of the Study:

• To develop a theoretical framework for analyzing the relationship between data distributions and fairness in deep learning.
• To derive theoretical bounds for fairness errors and convergence rates considering data heterogeneity.
• To propose and validate a practical method for improving equitable performance in deep learning models.

Main Methods:

• Developed a theoretical framework with novel bounds for fairness errors and convergence rates.
• Analyzed the impact of distributional differences on the fairness-accuracy trade-off.
• Conducted extensive experiments on diverse datasets across image, tabular, and text modalities.
• Proposed and implemented Fairness-Aware Regularization (FAR) to minimize inter-group feature discrepancies.

Main Results:

• Demonstrated that feature distribution differences across demographic groups significantly impact model fairness, especially for racial categories.
• Validated theoretical findings with empirical observations across multiple datasets and modalities.
• Showed that FAR consistently improves overall AUC, ES-AUC, and subgroup performance.
• Confirmed that distributional shifts are a fundamental limit to achieving fairness in deep learning.

Conclusions:

• Data distribution heterogeneity is a critical factor affecting AI fairness.
• Theoretical insights into feature distribution shifts can guide the development of more equitable algorithms.
• Fairness-Aware Regularization (FAR) is an effective practical approach to enhance fairness in deep learning.