TF-LIME : Interpretation Method for Time-Series Models Based on Time-Frequency Features.

Area of Science:

• Machine Learning
• Time Series Analysis
• Signal Processing

Background:

• Machine learning models are widely used for time series analysis.
• Existing interpretability methods often overlook time-frequency information, limiting insights into model behavior.
• There is a need for methods that explain model focus within the time-frequency domain.

Purpose of the Study:

• To develop a novel time-frequency domain-based interpretation method for time series analysis.
• To enhance the explainability of machine learning models by revealing their focus on time-frequency features.
• To introduce a new algorithm for precise segmentation of time-frequency data.

Main Methods:

• Extension of the Local Interpretable Model-agnostic Explanations (LIME) algorithm.
• Integration of Short-Time Fourier Transform (STFT) and inverse STFT.
• Development of a novel Time-Frequency Homogeneous Segmentation (TFHS) algorithm using peak detection and clustering.

Main Results:

• The TFHS algorithm demonstrated effectiveness in segmenting time-frequency matrices.
• The proposed TF-LIME method significantly improved the interpretability of time-series models in the time-frequency domain.
• Experiments on synthetic and real-world datasets (MIT-BIH) confirmed the method's efficacy and generalization capabilities.

Conclusions:

• The proposed time-frequency interpretation method offers a significant advancement in understanding machine learning models for time series.
• The TFHS algorithm provides precise segmentation, crucial for detailed time-frequency analysis.
• The method shows strong generalization and practical potential for diverse time series applications.