Class-Dependent Perturbation Effects in Evaluating Time Series Attributions

Attribution evaluation methods may be unreliable across different classes: The paper reveals that perturbation-based metrics—the standard tools for testing whether AI explanations are accurate—work well for some classes but fail to properly evaluate explanations for others, meaning current evaluation practices may give a false sense of confidence in explanation quality.

Class imbalance creates hidden evaluation problems: The research finds that class-dependent effects are particularly pronounced in imbalanced datasets, suggesting that AI systems trained on imbalanced data (common in real-world applications) may have explanations that appear trustworthy when they're actually not being properly validated.

Current evaluation metrics may reflect model bias rather than explanation quality: The authors demonstrate that perturbation-based metrics often reveal learned biases in the classifier itself rather than measuring whether the explanations are genuinely identifying important features, meaning practitioners may be validating the wrong thing.

A class-aware evaluation framework is needed: The paper proposes adding a penalty term that accounts for class-dependent effects, offering a practical improvement for organizations that need to evaluate explanations fairly across all classes in their datasets.

These problems likely extend beyond time series: Although the study focuses on time series data, the authors note that class-dependent evaluation effects probably occur across other types of structured data, suggesting the findings have broad implications for responsible AI deployment across multiple domains.