Abstract
INTRODUCTION: Emotion recognition based on electroencephalogram (EEG) signals has shown increasing application potential in fields such as brain-computer interfaces and affective computing. However, current graph neural network models rely on predefined fixed adjacency matrices during training, which imposes certain limitations on the model's adaptability and feature expressiveness. METHODS: In this study, we propose a novel EEG emotion recognition framework known as the Dynamic Graph Attention Network (DGAT). This framework dynamically learns the relationships between different channels by utilizing dynamic adjacency matrices and a multi-head attention mechanism, allowing for the parallel computation of multiple attention heads. This approach reduces reliance on specific adjacency structures while enabling the model to learn information in different subspaces, significantly improving the accuracy of emotion recognition from EEG signals. RESULTS: Experiments conducted on the EEG emotion datasets SEED and DEAP demonstrate that DGAT achieves higher emotion classification accuracy in both subject-dependent and subject-independent scenarios compared to other models. These results indicate that the proposed model effectively captures dynamic changes, thereby enhancing the accuracy and practicality of emotion recognition. DISCUSSION: DGAT holds significant academic and practical value in the analysis of emotional EEG signals and applications related to other physiological signal data.