TV-Net: Temporal-Variable feature harmonizing Network for multivariate time series classification and interpretation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Multivariate time series classification (MTSC) is crucial across domains like healthcare and finance.
• Existing methods struggle to balance capturing complex temporal dependencies with providing interpretable results.

Purpose of the Study:

• To develop a deep learning framework for MTSC that excels in both predictive performance and interpretability.
• To address the challenge of learning discriminative patterns and their underlying feature dependencies.

Main Methods:

• A temporal-variable parallel deep learning framework integrating global and local feature mining.
• A graph attention mechanism with global awareness (GAT-g) to capture inter-node and node-to-graph contexts.
• Game theory-based Shapley values for quantifying feature combination utility and enhancing model interpretability.

Main Results:

• The proposed framework achieved superior performance on 11 out of 30 datasets from the UEA archive.
• Demonstrated improved accuracy in multivariate time series classification compared to 12 state-of-the-art methods.
• The interpretation module provided clear, instantiated explanations for predictions.

Conclusions:

• The novel framework offers a significant advancement in MTSC by achieving high performance and interpretability.
• The integration of graph attention and game theory provides a powerful approach for analyzing complex time series data.
• This method enhances the practical applicability of MTSC in real-world scenarios.