A comprehensive analysis of perturbation methods in explainable AI feature attribution validation for neural time series classifiers
Abstract In domains where AI model predictions have significant consequences, such as industry, medicine, and finance, the need for explainable AI (XAI) is of utmost importance. However, ensuring that explanation methods provide faithful and trustworthy explanations requires rigorous validation. Fea...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Nature Portfolio
2025-07-01
|
| Series: | Scientific Reports |
| Subjects: | |
| Online Access: | https://doi.org/10.1038/s41598-025-09538-2 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract In domains where AI model predictions have significant consequences, such as industry, medicine, and finance, the need for explainable AI (XAI) is of utmost importance. However, ensuring that explanation methods provide faithful and trustworthy explanations requires rigorous validation. Feature attribution methods (AMs) are among the most prevalent explanation methods, as they identify decisive aspects that influence model predictions through feature importance estimates. Evaluating the correctness of AMs is typically done by systematically perturbing features according to their estimated importance and measuring the impact on the classifier’s performance. This paper extends our previous work which revealed flaws in the most commonly used metric for validating AMs when applied to time series data. In this work we introduce a novel metric, the Consistency-Magnitude-Index, which facilitates a faithful assessment of feature importance attribution. Additionally, we introduce an adapted methodology for robust faithfulness evaluation, leveraging a set of diverse perturbation methods. Our work includes an extended evaluation of AMs on time series data, that presents the influence and importance of perturbation methods and region size selection in relation to dataset and model characteristics. Based on the results of our extensive evaluation, we provide guidelines for future AM faithfulness assessments. Finally, we demonstrate our methodology through a concrete multivariate time series example. |
|---|---|
| ISSN: | 2045-2322 |