This paper proposes a novel pairwise learning framework with domain and category prototypes (PL-DCP) for emotion recognition in EEG-based affective brain-computer interfaces (aBCI). To address the inherent weaknesses of existing deep transfer learning-based emotion recognition methods, which suffer from dual dependencies on source and target domains and label noise, PL-DCP integrates the concepts of feature disentanglement and prototype inference. The feature disentanglement module extracts and separates domain and class features to compute dual prototype representations (domain and class prototypes). Domain prototypes capture inter-individual variation, while class prototypes capture commonalities across affect categories. The pairwise learning strategy mitigates the effects of mislabeling. Experimental results on the SEED, SEED-IV, and SEED-V datasets show that PL-DCP achieves accuracies of 82.88%, 65.15%, and 61.29%, respectively, outperforming existing state-of-the-art (SOTA) methods. In particular, it outperforms deep transfer learning methods that require both source and target data, while not using any target domain data during the learning process.
