Electroencephalography (EEG) emotion recognition holds wide application potential in mental health diagnosis, human-computer interaction, brain-computer interfaces, and personalized user experiences. However, the nonlinear characteristics, low signal-to-noise ratio, and non-stationarity of EEG signals challenge traditional methods in extracting stable emotional features. To advance science-education integration, we designed an innovative teaching experiment centered on EEG emotion recognition using an SGC-Transformer network (SGCTNet). This architecture integrates graph neural networks and Transformers, leveraging graph convolutional networks' (GCN) strength in processing non-Euclidean spatial data and Transformers' capacity for capturing global dependencies. Additionally, to mitigate deep learning's reliance on large-scale datasets, we propose a data integration strategy enhancing inter-channel relationship modeling and generalization capability.

The proposed SGCTNet is a hybrid deep learning architecture fusing Simplified Graph Convolution (SGC) and Transformer modules for efficient EEG emotion recognition. First, the SGC module extracts topological spatial features between EEG channels by simplifying the GCN structure: removing intermediate nonlinear activation layers reduces model complexity while preserving rich spatial information. Second, the Transformer module employs a self-attention mechanism to comprehensively capture global long-range dependencies among channel nodes based on these topological features, strengthening channel information utilization efficiency. Furthermore, a data integration strategy improves generalization by incorporating EEG data from multiple historical experimental sessions into current training, maximizing existing data utility. Experiments utilized public datasets SEED and SEED-Ⅳ, employing control groups to systematically evaluate SGCTNet's performance across scenarios, validating model effectiveness and data strategy generalizability.

Experimental results demonstrate significant performance improvements with SGCTNet. On SEED-Ⅳ, the model achieved accuracies of 82.45%, 85.23%, and 87.62% across three sessions. On SEED, it attained 94.94%, 94.21%, and 96.87% accuracy, outperforming traditional CNNs, SVMs, Random Forests, and other deep learning models. Further analysis confirmed the data integration strategy substantially enhanced generalization: accuracy increased by 5.22% (Session 2) and 7.83% (Session 3) on SEEDⅣ, and by 3.58% (Session 2) and 3.72% (Session 3) on SEED.

SGCTNet integrates graph structure modeling and self-attention mechanisms, demonstrating strong modeling capability and excellent generalization in EEG emotion recognition. The developed experimental system possesses significant pedagogical value, facilitating the practical application of deep learning in EEG signal processing and supporting talent cultivation in this field.