Enhancing EEG Emotion Recognition with Parallel 3D Graph Convolution and Attention Multi-Scale Temporal Convolutional Network Modeling

In recent years, deep learning has provided new approaches for EEG-based emotion recognition, demonstrating significant advantages in exploring Spatial-Temporal emotional features. However, existing methods still face challenges in fully capturing the Spatial-Temporal features. Therefore, this method achieves efficient feature extraction and fusion through three core modules: the designed Para-G3D module learns the interactive representation of EEG cross Spatial-Temporal information in a parallel way to realize the optimized complementarity of cross Spatial-Temporal information; the constructed AT-MSTCN module dynamically adjusts the frequency band weights through the attention mechanism to learn emotion-relevant key frequency band information, and combines multi-scale temporal convolution to enhance the ability to capture temporal detailed features; the introduced Transformer module models the dependencies of long time series to learn the global information representation across time series. Subject-independent experiments are carried out on two public EEG emotion datasets, SEED and DEAP, and the Leave-One-Subject-Out (LOSO) cross-validation method is adopted to verify the model performance. The results show that the emotion recognition accuracy of PG3DAMT on the SEED dataset reaches 85.21%; the recognition accuracies on the valence and arousal dimensions of the DEAP dataset are 60.27% and 64.68% respectively, all superior to the comparison models, and the model shows good recognition stability on each dataset. This method realizes the unified modeling of Spatial-Temporal graph convolution for EEG signals, effectively fuses local Spatial-Temporal features and global temporal features, and provides a new and effective solution for EEG emotion recognition.