Abstract
Emotion recognition methods based on Electroencephalogram (EEG) signals have become a research hotspot in affective computing due to their immense application potential. However, existing deep learning models often suffer from poor generalization ability in processing EEG signals, particularly in cross-dataset scenarios, mainly due to the high degree of individual variability in personal EEG signals. To enhance cross-dataset compatibility and transferability, we propose a novel graph contrastive model integrated with self-attentive conditional Wasserstein generative adversarial networks (SA-cWGAN) and graph contrastive learning for emotion recognition based on EEG signals. The SA-cWGAN module is employed to generate synthetic EEG samples, thus expanding the training dataset and improving the model robustness to inter- and intra-subject variations. The graph contrastive learning module learns invariant representations related to emotional states by maximizing the similarity of EEG signals in the same affective category while minimizing the similarity between different categories. On the DEAP dataset, the emotion recognition accuracy for the arousal and valence dichotomy tasks respectively reached 88.03% and 85.97%. On the MAHNOB-HCI dataset, the emotion recognition accuracy for the arousal and valence dichotomy tasks respectively reached 88.52% and 87.10%. These results demonstrate that the proposed method achieves competitive and stable performance across two distinct public datasets, highlighting its cross-dataset transferability and the effectiveness of the unified processing pipeline.