Meaningful data sampling for a faithful local explanation method

Abstract

Data sampling has an important role in the majority of local explanation methods. Generating neighborhood samples using either the Gaussian distribution or the distribution of training data is a widely-used procedure in the tabular data case. Generally, this approach has several weaknesses: first, it produces a uniform data which may not represent the actual distribution of samples; second, disregarding the interaction between features tends to create unlikely samples; and third, it may fail to define a compact and diverse locality for the sample being explained. In this paper, we propose a sampling methodology based on observation-level feature importance to derive more meaningful perturbed samples. To evaluate the efficiency of the proposed approach we applied it to the LIME explanation method. The conducted experiments demonstrate considerable improvements in terms of fidelity and explainability.