Q-RoFormer: Quaternion Rotation-Based Transformer for Cross-Subject EEG Emotion Recognition

Electroencephalography (EEG)-based emotion recognition holds significant potential in affective brain–computer interfaces. However, substantial inter-subject variability—due to differences in skull anatomy, electrode placement, and emotional responses—leads to severe spatiotemporal distribution shifts, thus affects the model generalizability. To address this issue, this paper proposes a hypercomplex neural network model based on quaternion rotation, termed Q-RoFormer, which explicitly addresses cross-subject spatiotemporal distribution shifts from both spatial and temporal perspectives. Specifically, the quaternion Transformer module embeds multi-temporal-window EEG segments into a unified quaternion representation, leverages inter-window correlations in quaternion space to mitigate temporal distribution shifts, and applies quaternion rotation to align spatial distribution shifts across subjects. Subsequently, the quaternion long short-term memory network captures temporal dependencies across windows within the quaternion-valued sequence. Furthermore, Q-RoFormer adopts a pre-training–fine-tuning paradigm to enhance transfer learning capabilities. Extensive experiments on the SEED and SEED-IV datasets demonstrate the effectiveness of Q-RoFormer, achieving 93.47% and 81.92% accuracy, respectively. Ablation results confirm the role of spatial rotation in cross-subject generalization. The code is available at: https://github.com/liuyici/Q-RoFormer.