Abstract
To reduce the label dependency of traditional electroencephalogram(EEG) emotion recognition methods and address the limitations of existing contrastive learning approaches in modeling cross-stimulus emotional similarity, this paper proposes a group-level stimulus-aware self-supervised soft contrastive learning framework (GSCL) for EEG emotion recognition. GSCL constructs contrastive learning tasks based on the consistency of subjects' brain activities under identical stimuli and incorporates a soft assignment mechanism, which adaptively adjusts the weights of negative sample pairs according to inter-sample distances to enhance representation quality. Additionally, this study also designs a learnable shuffling-splitting data augmentation method to dynamically optimize data distribution via learnable shuffling parameters. Finally, on the public emotional dataset (DEAP), the proposed method achieves accuracies of 94.91%, 95.29%, and 92.78% for valence, arousal, and four-class classification tasks, respectively; while on the Shanghai Jiao Tong University Emotional EEG Dataset (SEED), its three-class classification accuracy reaches 95.25% as well. These results demonstrate that the proposed method yields higher classification accuracy, offering a new insight for self-supervised EEG emotion recognition.