A bayesian perspective to explainable AI

Abstract

The LIME(Local Interpretable Model-Agnostic Explanations) framework is a popular technique for providing simple to understand post-hoc model explanations. However, it su ers from the perplexing issue of inconsistent explanations. Attributed primarily to the sampling strategy used in the framework, this issue renders LIME completely useless for safety-critical domains like healthcare and robotics, where the notion of trustworthiness and consistency are of the utmost importance. In this work, a novel modi cation of LIME based on Gaussian process priors and Bayesian optimization based sampling(BO-LIME) is proposed. We illustrate the performance of the proposed technique on two real-world datasets, and demonstrate the superior stability of BOLIME using mean standard deviation of the explanations as the evaluation metric. Furthermore, we demonstrate that the proposed technique is able to generate faithful explanations using much fewer surrogate samples as compared to LIME.