Explanations based on the Missing: Towards Contrastive Explanations with Pertinent Negatives.

In this paper we propose a novel method that provides contrastiveexplanations justifying the classification of an input by a black boxclassifier such as a deep neural network. Given an input we find what should beminimally and sufficiently present (viz. important object pixels in an image)to justify its classification and analogously what should be minimally andnecessarily \emph{absent} (viz. certain background pixels). We argue that suchexplanations are natural for humans and are used commonly in domains such ashealth care and criminology. What is minimally but critically \emph{absent} isan important part of an explanation, which to the best of our knowledge, hasnot been touched upon by current explanation methods that attempt to explainpredictions of neural networks. We validate our approach on three real datasetsobtained from diverse domains; namely, a handwritten digits dataset MNIST, alarge procurement fraud dataset and an fMRI brain imaging dataset. In all threecases, we witness the power of our approach in generating precise explanationsthat are also easy for human experts to understand and evaluate.