# Improving LIME Stability via Density-Awareness: Evaluation and Comparison of AKDE-LIME

## TL;DR

This paper introduces AKDE-LIME, a more stable version of LIME for generating reliable local explanations in machine learning models.

## Contribution

AKDE-LIME introduces a density-aware weighting scheme to improve the stability of local explanations.

## Key findings

- AKDE-LIME produces significantly more stable explanations than standard LIME across various noise levels.
- AKDE-LIME's performance is comparable to or better than TreeSHAP and Anchor in robustness and stability.
- The method is effective on diverse tree-based models using real-world data.

## Abstract

This paper addresses the critical instability of Local Interpretable Model-agnostic Explanations (LIME). We introduce Adaptive Kernel Density Estimation LIME (AKDE-LIME), a novel approach that enhances local explanation stability by incorporating a density-aware weighting scheme. Unlike LIME’s standard proximity kernel, AKDE-LIME combines distance weighting with a Kernel Density Estimate (KDE) of the local sample distribution, assigning more representative weights to generated perturbations. We conduct a comprehensive evaluation of AKDE-LIME against LIME, TreeSHAP, and Anchor on five diverse tree-based models using a real-world dataset. Assessing performance on Stability and Robustness metrics across a matrix of noise levels (5% to 20%), our results consistently demonstrate that AKDE-LIME produces significantly more stable and robust explanations than standard LIME under all conditions. The performance of our method is often comparable to or better than state-of-the-art explainers like TreeSHAP. We conclude that AKDE-LIME is a promising and reliable alternative for generating trustworthy local explanations, addressing a key weakness of the original LIME algorithm.