A Study on Deep Learning Methods for Multivariate Time Series Classification
- Publication Type:
- Thesis
- Issue Date:
- 2024
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Multivariate time series classification presents a significant challenge with wide-ranging applications in finance, medicine, and engineering, necessitating the consideration of temporal patterns and inter-variable relations. While deep learning has emerged as a potent tool in various domains, its application to multivariate time series classification remains constrained. Existing methods often fall short in exploiting the unique characteristics of such data, lacking in their ability to fully capture multi-level temporal dependencies, adapt to diverse sequence structures, incorporate frequency information, and provide interpretability. In this thesis, novel deep learning approaches are proposed to address these limitations, encompassing feature-map-wise attention mechanisms, dynamic architectures, explanation modules, utilization of 2D representations with frequency data, and exploration of positional embedding techniques. Evaluation across multiple benchmark datasets spanning diverse domains and data characteristics demonstrates the superiority of the proposed methods over traditional and recent deep learning techniques, affirming their efficacy in enhancing multivariate time series classification accuracy.
Please use this identifier to cite or link to this item: