Universal representation learning for multivariate time series using the instance-level and cluster-level supervised...

Area of Science:

• Machine Learning
• Data Science
• Time Series Analysis

Background:

• Multivariate time series classification (MTSC) traditionally requires large labeled datasets for deep learning models.
• Acquiring extensive labeled data for MTSC is costly and time-consuming, especially in specialized fields like medicine.
• Insufficient data hinders model feature learning, leading to poor generalization in MTSC tasks.

Purpose of the Study:

• To introduce a novel supervised contrastive learning approach for time series classification (SupCon-TSC).
• To improve MTSC performance by learning discriminative low-dimensional representations from limited data.
• To enable interpretable outcomes through an end-to-end structure.

Main Methods:

• Employed supervised contrastive (SupCon) loss to capture the inherent structure of multivariate time series.
• Utilized strong and weak augmentation families to generate data for source and target networks.
• Implemented instance-level and cluster-level SupCon learning to capture contextual information and learn universal representations.

Main Results:

• SupCon-TSC demonstrated superior feature learning on small cardiopulmonary exercise testing (CPET) datasets.
• The model achieved better classification performance compared to existing methods on limited data scenarios.
• On the UEA Multivariate time series archive, SupCon-TSC outperformed state-of-the-art approaches.

Conclusions:

• Supervised contrastive learning is effective for multivariate time series classification, particularly with limited labeled data.
• SupCon-TSC offers a robust method for learning discriminative and universal representations in time series.
• The approach shows significant potential for real-world applications where data annotation is a bottleneck.