FaAlGrad: Fairness Through Alignment of Gradients Across Different Subpopulations

Abstract

The growing deployment of Machine Learning systems has increased interest in systems optimized for other important criteria along with the expected task performance. For instance, machine learning models often exhibit biases that lead to unfair outcomes for certain protected subpopulations. This work aims to handle the bias in machine learning models and enhance their fairness by aligning the loss gradients. Specifically, leveraging the meta-learning technique, we propose a novel training framework that aligns the gradients computed across different subpopulations for learning fair classifiers. Aligning the gradients enables our framework to regularize the training process, thereby prioritizing fairness over predictive accuracy. Our experiments on multiple benchmark datasets demonstrate significant improvements in fairness metrics without having any exclusive regularizers for fairness. Thus our work contributes to developing fairer machine learning models with broader societal benefits.