Perturbation Effect: A Metric to Counter Misleading Validation of Feature Attribution

This paper provides evidence indicating that the most commonly used metric for validating feature attribution methods in eXplainable AI (XAI) is misleading when applied to time series data. To evaluate whether an XAI method attributes importance to relevant features, these are systematically perturbed while measuring the impact on the performance of the classifier. The assumption is that a drastic performance reduction with increasing perturbation of relevant features indicates that these are indeed relevant. We demonstrate empirically that this assumption is incomplete without considering low relevance features in the used metrics. We introduce a novel metric, the Perturbation Effect Size, and demonstrate how it complements existing metrics to offer a more faithful assessment of importance attribution. Finally, we contribute a comprehensive evaluation of attribution methods on time series data, considering the influence of perturbation methods and region size selection.