| Abstract [eng] |
As machine learning (ML) models increasingly influence high-stakes decisions in healthcare, finance, and policy, the need for transparent and interpretable artificial intelligence (AI) systems becomes critical. Local Interpretable Model-Agnostic Explanations (LIME) offer a widely adopted approach for local model interpretability, yet their global aggregation remains methodologically limited. This project investigates and enhances global explanation strategies by aggregating local LIME explanations, addressing known limitations such as noise sensitivity and inconsistent feature relevance. To improve global interpretability, two novel strategies were tested: weighting explanations by their fidelity (R²) and kernel density estimation (KDE)-based aggregation. Experimental results confirmed that incorporating R² significantly reduced distributional divergence (JSD), improved stability, and mitigated performance degradation under feature removal. Although KDE achieved the highest rank correlation (Spearman = 0.91), it lagged in other metrics, indicating trade-offs. The enhanced aggregation methods outperformed existing literature baselines across several quality metrics, demonstrating their potential to strengthen the reliability and clarity of global explanations in classification models. |
