SigTime: Learning and Visually Explaining Time Series Signatures.

Area of Science:

• Biomedical research
• Data science
• Machine learning

Background:

• Time series pattern discovery is crucial for scientific discovery and decision-making, especially in biomedical research for improved diagnosis and patient outcomes.
• Existing methods struggle with computational complexity, interpretability, and capturing temporal structures.
• There is a need for advanced techniques to effectively analyze temporal patterns in time series data.

Purpose of the Study:

• To introduce a novel learning framework for time series pattern discovery.
• To develop an interpretable method for identifying meaningful temporal structures.
• To create a visual analytics system for exploring time series signatures.

Main Methods:

• A novel learning framework jointly trains two Transformer models.
• Complementary time series representations are used: shapelet-based for localized structures and feature engineering for statistical properties.
• A visual analytics system, SigTime, with coordinated views is developed for exploration.

Main Results:

• The learned shapelets act as interpretable signatures differentiating time series across classification labels.
• Quantitative evaluation on eight public and one proprietary clinical dataset demonstrates the framework's effectiveness.
• Demonstrated effectiveness through usage scenarios with domain experts on ECG and preterm labor data.

Conclusions:

• The proposed framework effectively captures meaningful temporal structures in time series data.
• The learned shapelets provide interpretable signatures for classification.
• The SigTime system facilitates exploration and insight generation from time series data.