Time-Series Representation Feature Refinement with a Learnable Masking Augmentation Framework in Contrastive...

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Time-series data present challenges in representation learning due to temporal dependencies and complex feature extraction.
• Existing methods often struggle to capture both global and local patterns effectively.

Purpose of the Study:

• To propose a novel framework for time-series representation learning.
• To enhance the learning of discriminative temporal features using self-supervised learning.
• To improve the robustness and context-awareness of learned representations.

Main Methods:

• Developed a novel framework integrating a learnable masking-augmentation strategy into contrastive learning.
• Employed a masking-based reconstruction approach within a self-supervised learning context.
• Utilized learnable masking as a dynamic augmentation technique to optimize contextual relationships.

Main Results:

• Achieved performance improvements of 2% (SleepEDF-78), 2.55% (SleepEDF-20), and 3.89% (UCI-HAR) in accuracy over baseline methods.
• Demonstrated significant performance gains on multiple time-series datasets, including Epilepsy.
• Highlighted the framework's ability to capture nuanced temporal dependencies and improve downstream task performance.

Conclusions:

• The proposed framework effectively learns robust and context-aware time-series representations.
• Learnable masking within contrastive learning is a promising strategy for advancing time-series analysis.
• The method offers significant improvements over existing approaches, paving the way for enhanced time-series understanding.