Learning Fair Representations via Distance Correlation Minimization.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Machine learning (ML) algorithms increasingly drive critical decisions, making algorithmic bias a significant challenge.
• Bias in ML, seen in areas like facial recognition and hiring, often involves sensitive attributes such as race and gender.
• Learning fair representations is a key strategy to mitigate bias in ML models.

Purpose of the Study:

• To propose novel algorithmic approaches, graph-fair and dist-fair, for learning fair representations.
• To reduce the dependence between sensitive attributes and learned representations in ML models.
• To offer simpler fairness constraints compared to existing methods, avoiding complex parameter tuning and adversarial networks.

Main Methods:

• Introduced graph-fair, utilizing graph Laplacian regularization to encode sensitive attribute information.
• Developed dist-fair, employing distance correlation as a fairness constraint for scale-invariant bias reduction.
• Theoretically established the link between graph regularization and distance correlation.

Main Results:

• Both graph-fair and dist-fair demonstrated effective reduction of bias related to sensitive attributes.
• Experimental results on real-world datasets showed superior trade-offs between fairness and utility compared to existing methods.
• The proposed methods simplify fairness constraints, eliminating the need for extensive parameter tuning or adversarial components.

Conclusions:

• Graph-fair and dist-fair offer effective and simplified solutions for achieving fairness in machine learning representation learning.
• These methods provide a promising direction for developing less biased and more reliable AI systems.
• The findings suggest that incorporating graph regularization and distance correlation can significantly enhance fairness without compromising model utility.