ELIME: Exact Local Interpretable Model-Agnostic Explanation

This paper presents exact local interpretable model-agnostic explanation (ELIME) algorithm for explainable machine learning, which provides a comprehensible explanation of the decision-making process and predictions of machine learning models. Building upon existing model-agnostic interpretation methods, our approach enhances feature importance evaluation through single-feature sensitivity analysis and introduces a weighted distance metric based on sensitivity values. This sensitivity information is utilized for both calculating distances and generating training data for model fitting, improving the quality and reliability of the explanations. The enhanced ELIME algorithm is particularly effective for tabular classification domains, offering explanations that closely resemble the decision boundaries of the model. Comparative analysis with local interpretable model-agnostic explanation (LIME), deterministic LIME (DLIME), and active learning-based DLIME (AL-DLIME) demonstrates that while ELIME achieves superior fidelity and accuracy compared to DLIME and AL-DLIME, its stability is lower. However, ELIME outperforms LIME ac