GCNFormNet: branched graph-transformer architecture for EEG-based emotion recognition

Existing EEG-based emotion recognition pipelines rely on complex preprocessing techniques, making it difficult to assess the true capability of the underlying architecture. To address this, we propose GCNFormNet, a hybrid architecture that models dynamic spatial and temporal dependencies in EEG signals. Our primary contribution is the dual-branched design, where GCN layers model graph-structured spatial relationships using a dynamically generated adjacency matrix, while Transformer blocks capture complex temporal dynamics using a Performer-based self-attention mechanism. We also replaced the traditional layer normalization of Transformers with DynamicTanh (DyT), an element-wise activation function that mimics the S-shaped mappings produced by normalization layers. We evaluate GCNFormNet using the unified EEGain framework on four benchmark datasets: SEED, SEED-IV, DEAP, and DREAMER. Our results demonstrate competitive performance, with the highest accuracy among compared methods on SEED-IV (0.46), supported by statistical significance tests. An interpretability analysis of the learned adjacency matrices revealed neurophysiologically meaningful connectivity patterns, including hemispheric asymmetry and prefrontal dominance discovered purely from data without anatomical priors. Finally, ablation and sensitivity analyses confirmed the synergistic contribution of GCN and Transformer components on three datasets, while revealing a dataset-specific dependency on DEAP where the full architecture was not optimal.