Alpha and Prejudice: Improving α-Sized Worst Case Fairness via Intrinsic Reweighting

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Worst-case group fairness often requires demographic data, which is frequently unavailable in real-world applications.
• Existing methods struggle with the practical limitations of data privacy and missing demographic information.

Purpose of the Study:

• To propose a novel framework for worst-case group fairness using a relaxed setting called alpha-sized worst-case fairness.
• To address the underexplored connection between group fairness and data privacy.
• To develop efficient and robust methods for training fair machine learning models.

Main Methods:

• Introduced a reweighting approach assigning sample weights based on fairness contributions.
• Developed a stochastic learning algorithm to efficiently handle global worst-case objectives.
• Proposed a robust variant to mitigate the impact of outliers.

Main Results:

• Demonstrated the relevance of alpha-sized worst-case fairness to data privacy.
• Showcased that the proposed reweighting method connects to existing fairness-through-reweighting techniques.
• Empirically validated superior performance against strong baselines on standard fairness benchmarks.

Conclusions:

• The proposed alpha-sized worst-case fairness framework offers a practical solution for achieving group fairness with limited demographic data.
• The developed methods provide an efficient, robust, and privacy-aware approach to machine learning fairness.
• This work bridges the gap between theoretical fairness objectives and practical implementation challenges.