Time-series representation learning via Time-Frequency Fusion Contrasting

Area of Science:

• Machine Learning
• Data Science
• Signal Processing

Background:

• Labeling large time series datasets is expensive and time-consuming.
• Effective representation learning from unlabeled time series data is a significant challenge.
• Contrastive learning offers a promising approach for acquiring representations from unlabeled data.

Purpose of the Study:

• Propose a self-supervised time-series representation learning framework using Time-Frequency Fusion Contrasting (TF-FC).
• To learn effective time-series representations from unlabeled data by leveraging both time and frequency domains.
• Enhance the discriminative capacity of models for time series classification.

Main Methods:

• Developed a framework combining time-domain and frequency-domain augmentations for diverse sample generation.
• Time-domain augmentation included jitter, scaling, permutation, and masking.
• Frequency-domain augmentation involved Fast Fourier Transform (FFT) followed by filtering, frequency manipulation, and phase shifting, with kernel PCA for fusion.

Main Results:

• The TF-FC framework effectively extracts informative features by capturing both time and frequency domain characteristics.
• Experiments on SleepEEG, HAR, Gesture, and Epilepsy datasets demonstrated significant improvements in recognition accuracy.
• TF-FC outperformed other state-of-the-art (SOTA) methods in time series representation learning.

Conclusions:

• TF-FC provides a powerful self-supervised method for learning representations from unlabeled time series data.
• The fusion of time and frequency domain augmentations is key to enhancing feature extraction and model performance.
• This approach offers a cost-effective and efficient solution for time series analysis in various domains.