CNN-Based Interpretable Feature Extraction Methods Considering Pairwise Interactions.

Area of Science:

• Machine Learning
• Data Science
• Time Series Analysis

Background:

• Convolutional Neural Networks (CNNs) excel at multivariate time series analysis but struggle with detecting statistical interactions.
• Existing methods have limitations in capturing complex relationships between variables in time series data.

Purpose of the Study:

• To propose a novel framework that enhances multivariate time series classification performance.
• To enable objective assessment of individual variable influence and pairwise interaction effects.
• To overcome the structural limitations of CNNs in detecting statistical interactions.

Main Methods:

• Creative modification of convolutional filters and layer structures for feature extraction.
• Development of methods to capture pairwise interaction influences.
• Integration with interpretable models for causal analysis and feature importance calculation.

Main Results:

• Successfully extracted features that explain pairwise interactions in synthetic data.
• Demonstrated superior classification performance on real-world multivariate time series data compared to baseline methods.
• Quantified both individual and pairwise variable effects for in-depth causal analysis.

Conclusions:

• The proposed framework offers a practical and interpretable solution for multivariate time series classification.
• It is particularly effective in domains where variable interactions are crucial, such as healthcare, finance, and manufacturing.
• The method provides a significant advancement in analyzing complex time series data with interacting variables.