SCL-CSDG: Soft Contrastive Learning for Cross-Subject Domain Generalization in EEG Emotion Recognition

Electroencephalogram (EEG)-based emotion recognition has demonstrated significant potential in affective computing applications such as mental health monitoring, intelligent interaction, and adaptive learning. Unlike external behavioral cues, EEG signals provide an objective and real-time reflection of emotional states. However, substantial inter-subject variability in neural responses poses a major challenge for model generalization to unseen individuals. Most existing approaches rely on access to target subject data via transfer learning, which limits their practicality in real-world scenarios. To address this issue, we propose a Soft Contrastive Learning framework for Cross-Subject Domain Generalization (SCL-CSDG). The framework adopts a two-stage training strategy. In the pretraining stage, the model integrates data augmentation, attention-weighted encoding, and a hybrid mutual reconstruction module to enhance emotion-related representations and extract subject-invariant features. A soft contrastive learning module is further employed to relax strict binary constraints, enabling more flexible feature optimization. In the fine-tuning stage, a supervised emotion classifier is introduced to achieve task-specific adaptation. Experiments conducted on the SEED and SEED-IV datasets demonstrate that our method significantly improves recognition accuracy and cross-subject robustness.