A Maximal Correlation Framework for Fair Machine Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Algorithmic fairness is a growing concern as machine learning (ML) models are increasingly adopted across industries.
• Existing methods for ensuring fairness often face computational challenges and may not offer flexible trade-offs between performance and fairness.

Purpose of the Study:

• To explore algorithmic fairness from an information-theoretic perspective.
• To introduce a novel framework for expressing and enforcing fairness constraints in ML algorithms.
• To develop computationally efficient optimization algorithms for achieving fairness criteria.

Main Methods:

• The study utilizes an information-theoretic approach, specifically the maximal correlation framework.
• Regularizers are derived from fairness constraints to enforce independence and separation criteria.
• Optimization algorithms are developed for both discrete and continuous variables, focusing on computational efficiency.

Main Results:

• The proposed algorithms are more computationally efficient than existing methods.
• The algorithms yield smooth performance-fairness tradeoff curves.
• Competitive performance is demonstrated on both discrete (COMPAS, Adult) and continuous (Communities and Crimes) datasets compared to state-of-the-art methods.

Conclusions:

• The maximal correlation framework provides an effective information-theoretic approach to algorithmic fairness.
• The developed optimization algorithms offer an efficient and competitive solution for enforcing fairness criteria in ML.
• This work contributes to the development of more equitable and trustworthy AI systems.