Self-Supervised Contrastive Representation Learning for Semi-Supervised Time-Series Classification.

Area of Science:

• Machine Learning
• Data Science
• Time Series Analysis

Background:

• Learning effective representations from unlabeled or sparsely labeled time-series data is challenging.
• Contrastive self-supervised learning has emerged as a powerful technique for representation learning from unlabeled data.
• Existing methods may not fully capture the temporal dynamics and contextual information inherent in time-series data.

Purpose of the Study:

• To propose a novel framework, Time-Series representation learning via Temporal and Contextual Contrasting (TS-TCC), for learning robust time-series representations from unlabeled data.
• To investigate the impact of time-series-specific data augmentation strategies in contrastive learning.
• To extend the framework to semi-supervised settings (Class-Aware TS-TCC) to leverage limited labeled data.

Main Methods:

• Developed TS-TCC, incorporating temporal contrasting and contextual contrasting modules.
• Introduced time-series-specific weak and strong data augmentations.
• Proposed Class-Aware TS-TCC (CA-TCC) utilizing pseudo-labels for class-aware contrastive loss in semi-supervised learning.
• Conducted a systematic study on time-series data augmentation selection.

Main Results:

• TS-TCC learns representations that achieve performance comparable to fully supervised training via linear evaluation.
• The proposed framework demonstrates high efficiency in few-shot learning and transfer learning scenarios.
• CA-TCC effectively improves representations by leveraging limited labeled data.

Conclusions:

• TS-TCC offers a powerful approach for self-supervised representation learning on time-series data.
• The framework provides significant advantages in scenarios with limited labeled data.
• The proposed methods advance the state-of-the-art in time-series representation learning.