Time is Not Enough: Time-Frequency based Explanation for Time-Series Black-Box Models

Time-series explanations need to consider both time and frequency information: Traditional explainability methods for time-series models focus only on the time-domain, but this paper demonstrates that incorporating frequency-domain analysis through the SpectralX framework provides more complete insights into how AI models make decisions on time-series data.

A new explanation method (FIA) improves efficiency and clarity: The paper introduces Feature Importance Approximations, which uses insertion, deletion, and combination techniques to generate explanations more efficiently while providing class-specific insights—meaning explanations can be tailored to explain why a model made a particular prediction rather than offering only generic explanations.

The framework is designed for practical adoption: SpectralX is built as a flexible "plug-in" system that works with existing pre-trained time-series models without requiring modifications, making it easier for practitioners to add explainability to models already in use rather than requiring complete redesign.

User studies validate real-world usefulness: The authors conducted user studies to confirm that their time-frequency explanations are actually practical and understandable to people, not just theoretically sound—an important step in ensuring explainability methods serve their intended purpose of making AI decisions more transparent.

Performance improvements demonstrated across multiple datasets: Experiments on both synthetic and real-world UCR time-series datasets show that FIA outperforms existing perturbation-based explanation methods in the time-frequency domain, suggesting broader applicability across different time-series classification tasks.