Multisource Wasserstein Adaptation Coding Network for EEG emotion recognition

• Published in: Biomedical signal processing and control Vol. 76; p. 103687
• Main Authors: Zhu, Lei, Ding, Wangpan, Zhu, Jieping, Xu, Ping, Liu, Yian, Yan, Ming, Zhang, Jianhai
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.07.2022
• Subjects: Deep learning; Domain adaptation; EEG; Emotion recognition; Transfer learning; Deep learning; Domain adaptation; Emotion recognition; Transfer learning; EEG
• DOI: 10.1016/j.bspc.2022.103687
• ISSN: 1746-8094

•A novel model is proposed to adapt multisource domain distribution.•The proposed model obtains better results than existing methods.•The model shows good performance on the most challenging domain adaptation tasks. Emotion recognition has an important application in human–computer interaction (HCI). Electroencephalogram (EEG) is a reliable method in emotion recognition and is widely studied. However, since the individual variability of EEG, it is difficult to build a generic model between different subjects. In addition, the EEG signals will change in different periods, which has a great impact on the model. Therefore, building an effective model for cross-sessions, cross-subjects, cross- subjects and sessions has become challenging. In order to solve this problem, we propose a new emotion recognition method called Multisource Wasserstein Adaptation Coding Network (MWACN). MWACN can simplify output data and retain important information of input data by Autoencoder. It also uses Wasserstein distance and Association Reinforcement to adapt marginal distribution and conditional distribution. We validated the effectiveness of the model on SEED dataset and SEED-IV dataset. In cross-sessions experiment, the accuracy of our model achieves 92.08% in SEED and 76.04% in SEED-IV. In cross-subjects experiment, the accuracy of our model achieves 87.59% in SEED and 74.38% in SEED-IV. In cross-subjects and sessions experiment, the accuracy of the MWACN is improved by 3.72% in SEED and 5.88% in SEED-IV. The results show that our proposed MWACN outperforms recent domain adaptation algorithms.