Feature-Cost Sensitive Learning with Submodular Trees of Classifiers