Interpreting Classifiers by Multiple Views

Next to prediction accuracy, interpretability is one of the fundamental performance criteria for machine learning. While high accuracy learners have intensively been explored, interpretability still poses a difficult problem. To combine accuracy and interpretability, this paper introduces an framework which combines an approximative model with a severely restricted number of features with a more complex high-accuracy model, where the latter model is used only locally. Three approaches to this learning problem, based on classification, clustering, and the conditional information bottleneck method are compared.