Supervised Contrastive Learning for EEG-Based Cross-Subject Emotion Recognition

Cross-subject EEG-based emotion recognition presents significant challenges, and existing approaches primarily consist of domain adaptation (DA) and domain generalization (DG) methods. DG methods are preferred due to their ability to generalize across subjects without requiring access to test data. Nevertheless, current DG methods often overlook the inherent characteristics of EEG responses. In light of the shared patterns in EEG responses to emotional states, we propose the integration of contrastive learning into this domain to directly identify invariant emotion-relevant representations. However, to make contrastive learning better suited to our specific case, we extend the conventional strategy by enabling many-positive samples and incorporating label information. This extension allows us to effectively search for emotion-relevant invariant representations. In experiment, our proposed method is evaluated using publicly available benchmark datasets, DEAP, demonstrating its reliable performance in cross-subject EEG-based emotion recognition.