Emotion recognition based on graph convolutional networks and EEG signals
Huang, Peizhou (2025)
Kandidaatintyö
2025
School of Engineering Science, Tietotekniikka
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe2025050536018
https://urn.fi/URN:NBN:fi-fe2025050536018
Tiivistelmä
Emotion recognition based on electroencephalogram (EEG) signals has become a key enabling technology for various human-centric applications, including affective computing, mental health assessment, and brain-computer interfaces. However, EEG signals themselves vary significantly across subjects and frequency bands, which complicates reliable and robust feature extraction. To address these challenges, based on dynamic graph convolutional neural network (DGCNN), we propose channel-band attention multiscale graph neural network (CBAM-GNN), a novel and interpretable EEG-based emotion recognition model that leverages the dual attention mechanism and graph neural network architecture.
In detail, our proposed model integrates a customized dual attention framework—containing channel and band attention modules—to dynamically recalibrate EEG features and selectively emphasize the most discriminative spatial and spectral information. To fully exploit the inherent graph structure of EEG signals, we use a Chebyshev polynomial-based graph convolution module that can effectively capture the spatial correlations between EEG electrodes. In addition, we introduce a multi-scale feature pooling strategy to aggregate EEG features across multiple graph scales to enhance the network processing spatial complexity.
We conduct extensive experiments on the SEED dataset, the most popular dataset in the field, with both subject-dependent and challenging subject-independent settings. The results show that our model achieves competitive recognition accuracy while maintaining interpretability and compactness. In addition, ablation experiments show that our carefully designed dual attention mechanism effectively alleviates the subject variability and signal complexity issues encountered by traditional models.
In summary, this work proposes an effective EEG-specific neural architecture, which provides a new innovative direction for attention-based feature recalibration strategies designed specifically for EEG data.
In detail, our proposed model integrates a customized dual attention framework—containing channel and band attention modules—to dynamically recalibrate EEG features and selectively emphasize the most discriminative spatial and spectral information. To fully exploit the inherent graph structure of EEG signals, we use a Chebyshev polynomial-based graph convolution module that can effectively capture the spatial correlations between EEG electrodes. In addition, we introduce a multi-scale feature pooling strategy to aggregate EEG features across multiple graph scales to enhance the network processing spatial complexity.
We conduct extensive experiments on the SEED dataset, the most popular dataset in the field, with both subject-dependent and challenging subject-independent settings. The results show that our model achieves competitive recognition accuracy while maintaining interpretability and compactness. In addition, ablation experiments show that our carefully designed dual attention mechanism effectively alleviates the subject variability and signal complexity issues encountered by traditional models.
In summary, this work proposes an effective EEG-specific neural architecture, which provides a new innovative direction for attention-based feature recalibration strategies designed specifically for EEG data.