Self-Supervised Learning for Time Series Analysis: Taxonomy, Progress, and Prospects.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Self-supervised learning (SSL) demonstrates strong performance in time series tasks.
• SSL reduces the need for extensive labeled datasets, enabling high performance with minimal labels via pre-training and fine-tuning.
• Existing surveys predominantly focus on computer vision and natural language processing, leaving a gap in time series SSL literature.

Purpose of the Study:

• To comprehensively review state-of-the-art self-supervised learning methods for time series data.
• To address the lack of a dedicated survey on time series SSL.
• To establish a structured overview and taxonomy of existing time series SSL approaches.

Main Methods:

• A systematic review of existing SSL and time series literature.
• Development of a novel taxonomy categorizing time series SSL methods into generative-based, contrastive-based, and adversarial-based approaches.
• Detailed analysis of ten subcategories, including their core principles, frameworks, and trade-offs.

Main Results:

• Categorization of time series SSL methods into generative, contrastive, and adversarial paradigms.
• Identification and summary of commonly used datasets for time series forecasting, classification, anomaly detection, and clustering.
• Discussion of the advantages and disadvantages of various SSL techniques in the time series domain.

Conclusions:

• The survey provides a foundational resource for understanding and advancing self-supervised learning in time series analysis.
• It highlights the potential of SSL to unlock insights from unlabeled time series data.
• Future research directions for SSL in time series are outlined, paving the way for further innovation.