Fakulta matematiky, fyziky
a informatiky
Univerzita Komenského v Bratislave

Doktorandské kolokvium KAI - Peter Anthony (7.4.2025)

v pondelok 7.4.2025 o 13:10 hod. v miestnosti I/9


02. 04. 2025 22.31 hod.
Od: Damas Gruska

Prednášajúci: Peter Anthony

Názov: A Comparative Assessment of the Trustworthiness of Post-Hoc Explanation Approaches for Black Box Models

Termín: 7.4.2025, 13:10 hod., I/9


Abstrakt:
As AI advances, the need for interpretable machine learning models increases, especially in critical fields like healthcare, cybersecurity, finance, and criminal justice. While black-box models often offer strong predictive performance, their lack of transparency has led to the development of post-hoc explanation methods to clarify decision-making. A key issue, however, is the fidelity of these explanations—how accurately they reflect the model’s behavior. This study introduces a fidelity-focused evaluation framework and evaluates three popular explainers—LIME, SHAP, and Anchor—across various datasets and classifier types (Random Forest, XGBoost, MLP, and SVM). Our analysis shows that while fidelity is moderate generally (not impressive), the effectiveness of explanation methods depends on the model and dataset, with SHAP and Anchor often outperforming LIME. These findings improve our understanding of explanation reliability and offer practical insights for selecting methods to boost interpretability in real-world scenarios. 

Stránka seminára