A Novel Transfer Learning Model for Cross-Subject Emotion Recognition using EEGs

There have been many transfer learning models to solve the problem of individual differences in cross-subject emotion recognition using electroencephalogram (EEG) signals. However, the existing work consider little of the complexity of the class structure in the source domain, and may break the class structure in the target domain. In this paper, we propose a novel transfer learning model (CL-PSR-TL) based on the traditional domain-adversarial training of neural networks (DANN) in three aspects: 1) an inter-subject contrastive loss is additionally introduced in the source domain to extract the subject-irrelevant information; 2) a pairwise similarity mechanism with the effective pair selection is developed in the target domain to achieve a stable explore for the class structure; 3) a stepwise optimization strategy is applied to train the model. Then we evaluate the proposed model on two datasets (SEED and SEED-IV). Experimental results show that our proposed model achieves good performances compared with the state-of-the-art models.