FairDRO: Group fairness regularization via classwise robust optimization.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Algorithmic Fairness

Background:

• Existing group fairness methods are limited, falling into either re-weighting or regularization categories.
• These approaches often show performance limitations in specific scenarios.

Purpose of the Study:

• To propose FairDRO, a novel approach unifying re-weighting and regularization for enhanced group fairness.
• To develop an efficient optimization framework for achieving better accuracy-fairness trade-offs.

Main Methods:

• Introduced a classwise group distributionally robust optimization (DRO) framework named FairDRO.
• Incorporated a group fairness metric as regularization within the DRO objective.
• Developed an iterative algorithm with two variants based on surrogate loss selection.

Main Results:

• Derived theoretical results including closed-form re-weights, surrogate loss justifications, and convergence analysis.
• Experimental results demonstrate state-of-the-art performance in accuracy-fairness trade-offs.
• Showcased scalability and broad applicability across multiple benchmarks.

Conclusions:

• FairDRO effectively combines re-weighting and regularization strategies for group fairness.
• The proposed method offers superior and broadly applicable accuracy-fairness performance compared to existing techniques.