Budgeted stream-based active learning via adaptive submodular maximization