Explanation Space: A New Perspective into Time Series Interpretability

Time series require different explanation approaches than images or tables: The paper identifies that standard explainability methods from computer vision and tabular data don't translate well to time series because important features (like dominant frequencies) exist in different domains than the raw data, making them difficult for users to understand in their original form.

The "baseline problem" is unsolved for time series: While image classifiers can use black pixels and tabular models can use zero or mean values as baselines to show what "absence of a feature" looks like, time series lack a clear equivalent, limiting the application of popular explanation methods like SHAP or LIME.

Multiple explanation spaces can reveal different insights: The authors propose five different explanation spaces (beyond the standard time domain) that can each help clarify different aspects of model decisions, suggesting that a one-size-fits-all approach to time series interpretability is insufficient.

Existing models don't need to be retrained: The proposed method works with already-trained models and existing explanation tools, making it practical for real-world deployment without requiring organizations to rebuild their systems.

Better interpretability is critical for high-stakes applications: The paper emphasizes that human-understandable explanations of deep learning models are "essential for various critical and sensitive applications," highlighting why solving the time series explanation problem matters for responsible AI deployment.