SLIDE: A surrogate fairness constraint to ensure fairness consistency

Area of Science:

• Artificial Intelligence
• Machine Learning
• Algorithmic Fairness

Background:

• AI algorithms play a critical role in social decision-making, necessitating both accuracy and fairness.
• Minimizing empirical risk with fairness constraints is a key approach in fair AI.
• Existing methods often use surrogate constraints due to computational challenges with 0-1 loss.

Purpose of the Study:

• To investigate the validity of current surrogate fairness constraints.
• To propose a new, computationally feasible, and asymptotically valid surrogate fairness constraint named SLIDE.
• To ensure learned models satisfy fairness constraints and achieve rapid convergence.

Main Methods:

• Investigated the theoretical validity of existing surrogate fairness constraints.
• Developed a novel surrogate fairness constraint, SLIDE.
• Conducted numerical experiments on benchmark datasets to evaluate SLIDE's performance.

Main Results:

• Identified limitations in the validity of existing surrogate fairness constraints.
• Demonstrated that SLIDE is computationally feasible.
• Showed that SLIDE is asymptotically valid, satisfying fairness constraints with fast convergence.
• Numerical experiments confirmed SLIDE's effectiveness across various datasets.

Conclusions:

• SLIDE offers a computationally efficient and theoretically sound approach to achieving fairness in machine learning.
• The proposed SLIDE constraint advances the development of fair AI algorithms.
• SLIDE provides a practical solution for building accurate and fair AI models for social decision-making.