On the Necessity of Multi-Domain Explanation: An Uncertainty Principle Approach for Deep Time Series Models

Single-domain explanations can be misleading: The paper demonstrates that explaining AI time series models using only time domain or only frequency domain attributions may miss important information, as these domains can highlight fundamentally different features that aren't direct translations of each other.

The uncertainty principle reveals explanation gaps: By applying the uncertainty principle from physics and signal processing, researchers can now mathematically identify when time and frequency domain explanations diverge significantly, indicating that both perspectives are necessary for a complete understanding of model decisions.

Multi-domain explanations improve transparency: To achieve genuinely comprehensive interpretability of time series AI models, practitioners should present explanations in multiple domains rather than selecting whichever domain produces the sparsest (simplest-looking) attribution, as this practice may obscure critical decision factors.

Current XAI best practices may be insufficient: The prevailing approach of presenting explanations in whichever domain appears most interpretable can create a false sense of understanding, potentially masking important model behaviors that only become visible when examining alternative domains.

Mathematical tools can guide responsible AI practices: The paper shows how concepts from physics and signal processing can be adapted to identify when AI explanations are incomplete, providing a principled method to ensure stakeholders receive more trustworthy and thorough model interpretations.