EEG-based emotion recognition is a crucial task with significant implications for mental health monitoring, affective computing, and clinical decision support. While Graph Neural Networks (GNNs) have shown promise in modeling the complex spatio-temporal dynamics of EEG signals, existing methods fall short by only considering pairwise spatial dependencies and neglecting critical higher-order associations among functionally related brain regions. Moreover, these approaches often fail to achieve frequency-specific temporal learning, leading to suboptimal results. To overcome these limitations, we propose the Sub-band Embedded Spatio-Temporal Network (SESTN), a novel framework that jointly models multi-frequency spatial and temporal dependencies. In SESTN, EEG features are first projected into sub-band embedding spaces, enabling finer-grained frequency-specific representation. Our approach then uses neurophysiological priors to construct weighted hyperedges, capturing intra-regional spatial relationships more effectively. A multi-head, Mamba-based temporal module is subsequently utilized to extract frequency-specific temporal dynamics. Finally, a graph convolutional fusion strategy integrates these features across different frequency bands. Extensive experiments on two widely used EEG emotion recognition datasets, SEED and SEED-IV, demonstrate that SESTN significantly improves classification performance in both subject-dependent and subject-independent scenarios. These results underscore SESTN's potential as a robust and effective framework for real-world biomedical applications, particularly in affective state monitoring and mental health assessment.
Shi et al. (Thu,) studied this question.