Cross-subject electroencephalography-based emotion recognition using source-constrained graph representation learning

Abstract

Cross-subject emotion recognition from electroencephalography remains challenging because inter-subject variability alters spectral characteristics and spatial relationships across electrodes. This paper proposes a source-constrained graph representation learning method based on static differential entropy features. A band-aware encoder first refines the multi-band descriptor of each electrode, after which direct, local-graph, and global-graph paths capture complementary spatial information. For transductive adaptation to an unlabeled target subject, labeled source data remain the supervisory reference, while multi-source distribution alignment and source-prototype-referenced target alignment reduce cross-subject discrepancy. Gradient projection further limits transfer updates that conflict with source supervision. The method is evaluated on the SEED and SEED-IV datasets under leave-one-subject-out protocols, with all three recording sessions evaluated independently and experiments repeated using three random seeds. Averaged across sessions, the method achieves accuracies of 82.42\% and 68.69\% and macro-averaged F1 scores of 82.21\% and 68.18\% on SEED and SEED-IV, respectively.