Fast and Accurate Explanations of Distance-Based Classifiers by Uncovering Latent Explanatory Structures