Fair Representation Learning for Continuous Sensitive Attributes Using Expectation of Integral Probability Metrics.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• AI fairness, or algorithmic fairness, seeks to prevent bias in AI systems.
• Fair Representation Learning (FRL) is a key approach, but current methods struggle with continuous sensitive attributes (e.g., age, income).

Purpose of the Study:

• To develop a novel Fair Representation Learning (FRL) algorithm capable of handling continuous sensitive attributes.
• To introduce a new metric for assessing fairness in representation spaces with continuous attributes.

Main Methods:

• Introduced the Expectation of Integral Probability Metrics (EIPM) to quantify fairness for continuous attributes.
• Developed a method to accurately estimate EIPM from finite samples.
• Proposed a new FRL algorithm, Fair Representation using EIPM with MMD (FREM), leveraging EIPM.

Main Results:

• Demonstrated that low EIPM values in representation spaces ensure fairness regardless of the prediction head.
• Showcased that FREM effectively handles continuous sensitive attributes.
• Experimental results indicate FREM outperforms existing baseline methods in AI fairness.

Conclusions:

• The proposed EIPM metric and FREM algorithm offer a robust solution for achieving AI fairness with continuous sensitive attributes.
• FREM provides a significant advancement over existing FRL techniques, enabling broader application of fair AI.