Time-Series Contrastive Learning Against False Negatives and Class Imbalance

Area of Science:

• Machine Learning
• Time-Series Analysis
• Data Science

Background:

• Self-supervised contrastive learning (SCL) advances time-series representation learning.
• Existing methods using InfoNCE loss often overlook false negatives and class imbalance.
• These issues can degrade model performance, particularly with imbalanced datasets.

Purpose of the Study:

• To theoretically analyze and address the problems of false negatives and class imbalance in SCL for time-series data.
• To propose a novel SCL framework that improves representation learning accuracy and robustness.
• To enhance minority-class representation with minimal annotation cost.

Main Methods:

• Modification of the SimCLR framework incorporating a multi-instance discrimination task to reduce false negatives.
• Introduction of a graph-based interactive projection head for enhanced feature learning.
• Implementation of semantic consistency regularization to improve minority-class representation.

Main Results:

• Consistent outperformance of state-of-the-art methods across six real-world time-series datasets.
• Achieved up to 3.96% higher accuracy and 10.73% improvement in F1-score.
• Demonstrated significant benefits for imbalanced data scenarios.

Conclusions:

• The proposed SCL approach effectively mitigates false negatives and class imbalance in time-series representation learning.
• The novel methods enhance model performance, particularly for underrepresented classes.
• This work offers a more robust and accurate solution for time-series analysis, especially in practical, imbalanced settings.