Proceedings of the Annual Meeting of the Cognitive Science Society

Cross-subject electroencephalography (EEG) emotion recognition is challenged by pronounced inter-subject variability, which hinders generalization to unseen subjects in the domain generalization setting. Most existing approaches focus on aligning feature distributions, whereas the invariance of relational structure in latent representations has received far less attention. Inspired by representational similarity analysis (RSA), we revisit cross-subject EEG generalization from a structural perspective and propose an RSA-based structural consistency constraint (RSASC). This regularizer aligns representational dissimilarity matrices (RDMs) computed from different subjects' latent representations, encouraging the model to learn emotion-relevant representations that are more consistent across subjects. Experiments on the SEED and SEED-IV datasets show that the proposed regularizer yields stable improvements in cross-subject generalization: it achieves the best accuracy on SEED, and on SEED-IV it attains competitive performance with the lowest variance. These results suggest that structural invariance offers a novel perspective that complements distribution alignment for learning subject-invariant EEG representations.