Statistical machine learning models have been widely deployed in businesses to assist automated decision making. However, as modern machine learning models (such as deep neural networks and SVMs) are becoming more and more complicated and working as black boxes, there is an urgent need for providing interpretable predictions. In order to tackle this challenge, we propose a novel approach to generate understandable counterfactual explanations. Our approach considers both the sparsity and plausibility, while existing methods only consider the sparsity but ignore the plausibility of the explanations. Specifically, we propose a framework that takes the plausibility of the counterfactual explanations into the design process by controlling the likelihood of the counterfactuals being d...[ Read more ]

Statistical machine learning models have been widely deployed in businesses to assist automated decision making. However, as modern machine learning models (such as deep neural networks and SVMs) are becoming more and more complicated and working as black boxes, there is an urgent need for providing interpretable predictions. In order to tackle this challenge, we propose a novel approach to generate understandable counterfactual explanations. Our approach considers both the sparsity and plausibility, while existing methods only consider the sparsity but ignore the plausibility of the explanations. Specifically, we propose a framework that takes the plausibility of the counterfactual explanations into the design process by controlling the likelihood of the counterfactuals being drawn from the same data distribution as the training dataset, while considers the sparsity at the same time. We demonstrate the value of this new approach with a case study from a real-world credit application classification task: classifying whether credit applicants will repay their credit accounts within 2 years. The prediction results would be then used to decide whether to approve or reject their credit applications. The empirical evaluations show that the explanations are sparse and thus easy to understand, as well as plausible. Further in-depth analysis shows that counterfactual explanations are helpful for model validation and model debugging.