Abstract
Recognizing emotions is a vital component of applications related to Brain-Computer Interfaces (BCI). Although electroencephalography (EEG) offers rapid and reliable measurements, it is inherently limited by a low signal-to-noise ratio and non-stationarity, leading to variability across subjects and recording sessions. Existing multi-source domain adaptive strategies usually assume that the inter-domain distributions are static, ignoring the intrinsic variations in feature distributions between domains. In order to overcome the above limitations, this paper proposes an unsupervised multi-source domain adaptive network based on a gating mechanism, extracting unique features for each target domain. Furthermore, the model employs different loss functions for distributional alignment in the source and target domains to minimize the marginal and conditional distributions between the domains. Cross-subject experiments on the SEED, SEED-IV, and DEAP datasets were performed, while cross-session investigations were executed on the SEED and SEED-IV datasets. The proposed model improved classification performance, achieving cross-subject accuracy of 92.54%, 85.86%, and 65.59% on the SEED, SEED-IV, and DEAP datasets, respectively, and cross-session accuracy of 95.88% and 82.09% on the SEED and SEED-IV datasets.