The How of Explainable AI: Post-modelling Explainability

Currently AI models are often developed with only predictive performance in mind. Thus, the majority of the XAI literature is dedicated to explaining pre-developed models. This bias of focus along with the recent popularity of XAI research has resulted in development of numerous and diverse post-hoc explainability methods. It's challenging to understand this vast body of literature because of the numerous approaches to XAI. In order to make sense of the post-hoc explainability methods, we propose a taxonomy or a way of breaking down these methods that shows their common structure, organized around four key aspects: the target, what is to be explained about the model; the drivers, what is causing the thing you want explained; the explanation family, how the explanation information about the drivers causing the target is communicated to the user; and the estimator, the computational process of actually obtaining the explanation. For instance, the popular Local Interpretable Model-agnostic Explanations (LIME) approach provides explanation for an instance prediction of a model, the target, in terms of input features, the drivers, using importance scores, the explanation family, computed through local perturbations of the model input, the estimator.