Graph signal representations for EEG analysis and machine learning classification

Abstract

This thesis investigates the dynamic characteristics of electroencephalography (EEG) signals using graph signal processing (GSP), focusing on modeling spatial and temporal variations in brain connectivity. Conventional approaches often rely on static graph structures, which do not adequately capture time-varying, task-specific, and subject-specific relationships among brain regions. To address these limitations, this thesis proposes a dynamic graph representation framework for EEG signals. This framework allows EEG signals to be represented with a broader set of spatial frequency components derived from multiple graphs. A unified spatiotemporal frequency representation is developed to reflect how spatial patterns evolve over time by combining temporal frequency components with graph-based spatial frequencies. The proposed methods are evaluated in the contexts of imagined speech classification and biometric identification. The effects of clustering strategies and eigenvector selection are systematically examined, and their influence on classification performance is demonstrated. The thesis also introduces a graph-based EEG data augmentation method designed to preserve channel-wise correlations while generating artificial EEG trials. The feasibility of using dynamic graph representations for biometric identification is further evaluated. Experimental results demonstrate that individual differences in EEG graph representations enable reliable subject identification, achieving 100% classification accuracy.