ST-Tree with interpretability for multivariate time series classification

Area of Science:

• Machine Learning
• Data Science
• Time Series Analysis

Background:

• Multivariate time series classification is crucial but challenging.
• Deep learning models offer accuracy but lack interpretability.
• Traditional decision trees provide interpretability but lower accuracy.

Purpose of the Study:

• To develop an interpretable model for multivariate time series classification.
• To combine the strengths of Swin Transformer (ST) and decision trees.
• To enhance understanding of model decision-making in time series analysis.

Main Methods:

• Proposed ST-Tree model integrating Swin Transformer (ST) backbone with a neural tree.
• Leveraging ST's self-attention for local and global pattern recognition.
• Utilizing the neural tree for interpretable decision process visualization.

Main Results:

• ST-Tree demonstrated improved accuracy on 10 UEA datasets.
• The model successfully provided interpretable decision-making processes.
• Visualization of decisions offered clear insights into model behavior.

Conclusions:

• ST-Tree effectively balances accuracy and interpretability in multivariate time series classification.
• The model offers a valuable tool for gaining insights into complex time series data.
• This approach advances the field by providing explainable AI for time series tasks.