Modeling spectral EEG interactions using graph-structured variational representation learning

Emotion recognition from electroencephalogram (EEG) signals remains a challenging problem due to the high dimensionality, nonlinearity, and complex spectral dependencies inherent in neural activity. Conventional deep learning approaches often treat EEG features independently, thereby limiting their ability to capture structured spectral relationships. In this work, we propose a graph-based representation learning framework that models frequency-domain EEG features as nodes within a structured graph and leverages a Graph Neural Network-Variational Autoencoder (GNN-VAE) to learn compact latent representations. Spectral adjacency is defined using k-ring neighborhood connectivity, enabling localized message passing across contiguous frequency bands. The learned latent embeddings are subsequently classified using recurrent and attention-based temporal models to capture sequential dependencies across spectral segments. Experiments conducted on an EEG emotion dataset comprising three affective states demonstrate that the proposed approach consistently outperforms traditional machine learning baselines and non-graph deep learning models, achieving an accuracy of [Formula: see text] 91% and F1-score of 0.903. Ablation analyses further confirm the contribution of graph-based encoding and variational regularization to improved generalization. While the current study focuses on fixed spectral connectivity and subject-dependent evaluation, the results highlight the potential of graph-structured latent modeling for EEG-based emotion recognition and provide a foundation for future extensions incorporating adaptive graph learning and explainable representations.