Exploring Multi-Task Learning for Fairness in Machine Learning Regression

Ensuring fairness in machine learning is a critical concern for high-stakes domains, yet most fairness-aware Multi-Task Learning (MTL) frameworks overlook regression problems in favor of classification. This work extends these techniques to regression, proposing a novel MTL framework that optimizes for equitable continuous outcomes across demographic subgroups. Our method dynamically reweights task-specific gradients during training to reduce disparities without compromising predictive accuracy. Evaluated on two real-world datasets, our approach, in the best scenarios, reduces subgroup disparity by up to 94.9% while also improving overall regression performance by up to 32.6%. These findings highlight the significant potential of fairness-aware MTL for creating more inclusive and responsible machine learning applications in sensitive domains.