How Stable Are Post-Hoc Explanations? A Robustness Study of Feature Attribution in Data Mining
REST Journal on Data Analytics and Artificial Intelligence · 2025
DOI: 10.46632/jdaai/4/4/14 ↗Kurz gefasst
Fachartikel in der Zeitschrift REST Journal on Data Analytics and Artificial Intelligence, 2025. Verfasst von Rahul Raj und 1 weiteren Autor:innen. Die Arbeit nennt 23 Literaturangaben.
Abstract (Originalsprache)
Explainable artificial intelligence is increasingly used to justify the outputs of data mining models, but an explanation is only useful if it does not change every time the data, the model or the inputs change slightly. This paper studies the robustness of one widely used explanation, permutation-based feature importance, along three dimensions: resampling of the training data, choice of model family, and noise in the inputs being explained. Four model families were trained on two public datasets, a three-class wine classification task and a diabetes progression regression task, and their importance rankings were compared using Kendall's rank correlation. Rankings obtained from bootstrap retrained random forests agreed only moderately (mean tau 0.52 on wine and 0.58 on diabetes), rankings from different model families agreed at 0.41 and 0.67, and Gaussian input noise at a quarter of a feature's standard deviation reduced agreement to 0.64 and 0.78. Linear models gave the most stable rankings. In both datasets, removing the most important feature caused a closely related feature's importance to rise sharply, showing that credit is redistributed among related attributes. We recommend that explanations be reported with uncertainty and that single-run feature rankings not be presented as definitive.
Zitieren
APA
Raj, R., & Mishra, R. K. (2025). How Stable Are Post-Hoc Explanations? A Robustness Study of Feature Attribution in Data Mining. REST Journal on Data Analytics and Artificial Intelligence. https://doi.org/10.46632/jdaai/4/4/14
Vancouver
Raj R, Mishra RK. How Stable Are Post-Hoc Explanations? A Robustness Study of Feature Attribution in Data Mining. REST Journal on Data Analytics and Artificial Intelligence. 2025. doi:10.46632/jdaai/4/4/14
BibTeX
@article{Raj2025,
author = {Raj, Rahul and Mishra, Ranjan Kumar},
title = {How Stable Are Post-Hoc Explanations? A Robustness Study of Feature Attribution in Data Mining},
journal = {REST Journal on Data Analytics and Artificial Intelligence},
year = {2025},
doi = {10.46632/jdaai/4/4/14}
}
EUVIMED