A Model-Agnostic Interpretability Approach Based on Enhanced Hierarchical Clustering

The Local Interpretable Model-Agnostic Explanations (LIME) is a relatively novel technique for explaining black-box models. LIME generates perturbed instances of the model based on a random instance, and then approximates the original model with a linear model derived from the weighted perturbed instances, thereby providing an explanation for individual predictions. While LIME effectively addresses the opacity of black-box models, the randomness in instance generation often leads to unstable and inaccurate predictions. For a given prediction, different explanations may be generated, posing critical issues, particularly in security-sensitive domains. In response to these challenges, we propose a stable version of LIME, named DDLIME(Density and Distance-based LIME). DDLIME leverages improved hierarchical clustering and KNN algorithms to select instances that are most similar to the data, thus reducing the impact of perturbations. Additionally, the enhanced hierarchical clustering, incorporating density and distance metrics, enables DDLIME to address the limitations of traditional hierarchical clustering in clustering non-spherical clusters. Consequently, DDLIME can effectively capture complex structures and patterns within the data. The experimental results on three binary public datasets from the UCI repository demonstrate that DDLIME provides more stable explanations compared to LIME. Furthermore, compared to both baseline models, DDLIME exhibits superior performance across various evaluation metrics.