Are we really addressing fairness in machine learning algorithms?

Abstract

This study explores the effectiveness of neural network regularization functions in measuring fairness, particularly analyzing the disparity between true disparate impact and fairness regularization functions. Utilizing two standard fairness datasets, COMPAS and Adult, we investigate the gaps between fairness and actual outcomes, and assess the suitability of two fairness functions derived from literature. Through experiments, we find substantial variations in fairness outcomes as determined by disparate impact and probabilistic disparate impact, especially under varying the strength of the loss functions given to accuracy versus fairness. Results indicate that focusing solely on fairness metrics can lead to misleading outcomes, often disregarding individual merit and involvement. The findings highlight the complexity of implementing fairness in algorithmic decisions, suggesting a critical need for a deeper evaluation of fairness functions in predictive models.

Sandro Radovanović

Assistant Professor at University of Belgrade

My research interests include machine learning, development and design of decision support systems, decision theory, and fairness and justice concepts in algorithmic decision making.