Safe Fairness Guarantees Without Demographics in Classification: Spectral Uncertainty Set Perspective.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Automated classification systems risk amplifying societal biases.
• Existing fairness methods often require demographic information, which is rarely available in practice.
• Robust optimization for fairness can be compromised by overly pessimistic uncertainty sets.

Purpose of the Study:

• To develop a fairness-aware classification method that does not require demographic group information.
• To address limitations of existing robust optimization techniques in fairness.
• To improve both fairness guarantees and overall performance in automated classification.

Main Methods:

• Introduction of SPECTRE, a minimax-fair method.
• Adjustment of the spectrum of a Fourier feature mapping.
• Constraining the deviation of worst-case distributions from empirical distributions.
• Theoretical analysis of computable bounds on worst-case error.

Main Results:

• SPECTRE achieves the highest average fairness guarantees.
• SPECTRE demonstrates the smallest interquartile range in fairness metrics.
• The method's effectiveness is validated on American Community Survey datasets across 20 states.
• SPECTRE outperforms state-of-the-art methods, including those with demographic data access.

Conclusions:

• SPECTRE offers a robust solution for achieving fairness in classification without demographic data.
• The method provides strong theoretical guarantees on worst-case error.
• SPECTRE represents a significant advancement in the field of fair machine learning.