HERA: a Hybrid Euler Reservoir Architecture for multivariate time series classification.

Area of Science:

• Machine Learning
• Data Science
• Time Series Analysis

Background:

• Multivariate time series classification (MTSC) is crucial for applications like human activity recognition and medical diagnosis.
• Existing methods struggle to capture complex nonlinear dynamics and global statistical properties effectively.
• Reservoir computing (RC) offers efficiency but requires extensive hyperparameter tuning.

Purpose of the Study:

• To propose HERA (Hybrid Euler Reservoir Architecture), a novel framework for MTSC.
• To improve classification accuracy by integrating diverse feature representations.
• To enable efficient hyperparameter optimization for RC models.

Main Methods:

• HERA employs a hybrid feature design integrating Euler State Network (EuSN) for dynamic inter-variable interactions and static statistical features.
• A self-optimization module using Covariance Matrix Adaptation Evolution Strategy (CMA-ES) is embedded for efficient hyperparameter tuning.
• The framework is evaluated on multiple public MTSC benchmark datasets.

Main Results:

• HERA achieves highly competitive classification accuracy on benchmark MTSC datasets.
• The proposed hybrid feature design effectively captures both dynamic and static sequence properties.
• The CMA-ES driven optimization module efficiently identifies optimal model configurations.

Conclusions:

• HERA offers a robust and accurate solution for MTSC tasks.
• The synergistic integration of dynamic and static features is key to HERA's performance.
• Automated hyperparameter optimization significantly enhances the practical applicability of RC models for MTSC.