“…To determine the informativeness of unlabeled data, such approaches use the model's predictive mean and variance to either model the uncertainty of unlabeled data [18,19,20,21,22,23] or estimate the expected model change affected by unlabeled data [24,25]. Instead of using a probabilistic model, others use deterministic models (e.g., support vector machines) and combine the classification results with some geometric rules for selecting queries [26,7,27,28]. Another body of work models active learning MD-17-1152 W. Chen, M. Fuge as a multi-armed bandit problem, automatically selecting queries to balance exploitation and exploration via a performance estimate [29,30].…”