Mutual Information-Based Fair Active Learning

Machine learning (ML) has become central to data-driven decision making, thereby necessitating fairness to individuals and society. Quality and quantity of labeled data play a crucial role in realizing fair ML models as inadequate and unreliable labels can result in biased outcomes. In this context, active learning methods are regarded as promising pathways for efficiently collecting labeled data. In this paper, we propose a fair active learning method that characterizes the reduction in unfairness associated with ML models by quantifying the mutual information between the label predictions and sensitive attributes using efficient Bayesian sampling methods. Extensive experiments on multiple image and text datasets demonstrate that our method yields promising fairness-accuracy tradeoff when compared to existing deep learning-based methods.