Year: 2025
Type: dissertation
Abstract: As machine learning systems permeate high-stakes fields such as healthcare, finance, and public policy, the demand for transparent and interpretable algorithms has intensified, giving rise to Explainable Artificial Intelligence (XAI) techniques that shed light on opaque “black‐ box” models. Among these, LIME (Local Interpretable Model‐Agnostic Explanations) stands out for generating simplified surrogate models around individual predictions, yet its dependence on randomized perturbations compromises its reliability: slight variations in random seeds or sampling strategies can produce divergent explanations for the same input. To quantify this instability, we designed a unified experimental framework evaluating LIME’s numerical consistency on four tabular datasets, Titanic, Iris, Wine Quality, and California Housing, using Random Forests for classification and regression tasks. We applied LIME repeatedly to identical instances under controlled conditions, assessing stability through both visual diagnostics (feature‐weight trajectories, boxplots, and frequency histograms) and quantitative measures (feature‐weight standard deviations, Top‐1 feature recurrence rate, and local approximation errors). Our results indicate that while LIME delivers stable, coherent feature attributions in simple classification scenarios, its explanations fluctuate markedly in complex regression contexts, with high variability in feature weights and approximation fidelity despite recurring top‐ranked predictors. By systematically documenting LIME’s context‐dependent variability and proposing a reproducible evaluation protocol, this thesis highlights a critical limitation of a popular XAI tool and underscores the necessity of rigorous stability checks before deploying local explanation methods in sensitive or regulated environments.
Author: Castilho, Tomás Martinho Filipe de Oliveira
Institution: Universidade Nova de Lisboa
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7113636966
