Bias Mitigation in Machine Learning Models for Fair Decision-Making

The thesis research topic is AI Fairness (bias mitigation in machine learning models). We use Adult Census Income to compare the merits of reweighting, adversarial debiasing, and prejudice removal methods through experiments. Although the baseline models had high accuracies (84.1 percent), they had gender bias (males were 2.7 times more likely to be given favorable predictions). We achieved a greater reduction of up to 93.7 percent in statistical parity difference with an accuracy loss of only 0.7 percent using our optimized GerryFair implementation, in contrast to the other methods. Nevertheless, intersectional analysis demonstrated continued unfairness, especially to black females (4 times lower predictive rates than White males). The analysis has shown that the existing fairness methods must respond to implementation issues and intersectional dynamics more effectively to bring the much-desired equity to-real-life systems.