Abstract: We present a novel approach for learning to generate counterfactual explanations of decisions made by classifiers. We show how to improve user scoring of explanations by using ranking data over pairs of diverse explanations. Applying this framework to a range of classification problems, we show how varying explanations can be learnt depending on their intended application. We empirically demonstrate the importance of context and that the precise form of a good explanation depends upon what it is to be used for.