Generating Interpretable Counterfactual Explanations By Implicit Minimisation of Epistemic and Aleatoric Uncertainties

Counterfactual explanations (CEs) are a practical tool for demonstrating why\nmachine learning classifiers make particular decisions. For CEs to be useful,\nit is important that they are easy for users to interpret. Existing methods for\ngenerating interpretable CEs rely on auxiliary generative models, which may not\nbe suitable for complex datasets, and incur engineering overhead. We introduce\na simple and fast method for generating interpretable CEs in a white-box\nsetting without an auxiliary model, by using the predictive uncertainty of the\nclassifier. Our experiments show that our proposed algorithm generates more\ninterpretable CEs, according to IM1 scores, than existing methods.\nAdditionally, our approach allows us to estimate the uncertainty of a CE, which\nmay be important in safety-critical applications, such as those in the medical\ndomain.\n