This study is concerned with the design of a linear regression model that contributes to local explanations of black-box regression models and a realization of fuzzy residual rules for enhancing the local fidelity (accuracy) of the linear explanations. As one of the most popular explainable artificial intelligence (XAI) techniques, local interpretable model-agnostic explanations (LIMEs) help illuminate complex models by generating explanations that are locally faithful within the neighborhood of the data point to be explained. Linear regression has been extensively used in LIME due to its advantages of simplicity and interpretability. However, the fidelity of the linear regression model is significantly influenced by the sampling strategy, the selected kernel width, and some other factors. The model-agnostic system proposed in this study is composed of three components: a strategy for determining the optimal kernel size, a linear regression model that passes through the data point to be explained, and a collection of fuzzy rules that characterize the residuals (errors) of the local explanations. The fuzzy rules are constructed on the basis of the errors (residuals) produced by the local linear regression model. Furthermore, we perform feature selection by using the Lasso regression, such that the interpretability of the local model is further enhanced. Experimental studies show that the proposed architecture can produce explanations with higher precision, while the interpretable fuzzy rules lead to a more concise description of the characteristics of the black-box model.
Wu et al. (Thu,) studied this question.