Towards a Rigorous Evaluation of XAI Methods on Time Series

XAI methods designed for images may not work equally well for time series data: The paper finds that explanation methods like DeepLIFT, LRP, and Saliency Maps perform inconsistently across different model architectures when applied to time series, suggesting that one-size-fits-all XAI approaches can be unreliable for temporal data.

SHAP emerges as the most dependable explanation method tested: Among the XAI methods evaluated, SHAP demonstrated robust performance across all model types in the experiments, making it a more trustworthy choice for explaining time series predictions compared to alternatives.

Most XAI methods themselves lack transparency: The paper highlights a critical paradox—the explainability methods used to make AI systems more interpretable are often "black-boxes themselves," meaning users cannot fully understand how these explanations are generated, undermining the goal of responsible AI.

Rigorous evaluation frameworks for XAI are lacking: The research introduces new verification techniques that account for the temporal dimension, indicating that existing evaluation methods for XAI are insufficient and may give false confidence in explanation quality without proper testing.

Visual interpretability alone is insufficient for validating explanations: The paper demonstrates that relying on visual inspection to prove XAI methods work—a common practice—is inadequate, particularly for time series where temporal relationships are critical to understanding model decisions.