A novel multi-source contrastive learning approach for robust cross-subject emotion recognition in EEG data

• Published in: Biomedical signal processing and control Vol. 97; p. 106716
• Main Authors: Deng, Xin, Li, Chenhui, Hong, Xinyi, Huo, Huaxiang, Qin, Hongxing
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.11.2024
• Subjects: Brain–computer interface; Contrastive learning; Domain adaptation; Electroencephalograph(EEG); Emotion recognition; Domain adaptation; Emotion recognition; Electroencephalograph(EEG); Brain–computer interface; Contrastive learning
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2024.106716

Emotion Brain Computer Interface (BCI) based on Electroencephalography (EEG) is a significant branch in the field of affective computing. However, the variability in emotional feedback among subjects facing the same emotional stimuli and the potential presence of noisy labels in the data significantly limit the effectiveness and generalizability of EEG-based emotion recognition models. In this paper, we propose the Multi-Source Contrastive Learning Transfer Learning Model (MSCL) to address the emotion recognition in cross-subject scenarios. The MSCL model consists of three components: Firstly, the unsupervised and supervised contrastive learning are utilized to learn the differences and commonalities among different individuals. Secondly, domain adaptation methods and feature learning in multiple domains enhance the interaction of information between source and target domains, thereby improving the model’s generalization across different individuals. Thirdly, in order to mitigate the detrimental effects of noisy labels on the model, this research dynamically allocate the weights to source domains similar to the target domain based on behavioral characteristics within the target domain and employ the corresponding noise learning methods. In our experiments, MSCL achieves excellent cross-subject emotion recognition performance on the CEED Dataset and the SEED Dataset. Our research highlights the significant capability of MSCL model in addressing the issues associated with inter-subject variability and label noise within EEG-based emotion BCI systems. •Employs a unique combination of domain adaptation and contrastive learning to manage inter-subject variability.•Utilized noise learning to minimize the impact of noisy labels.•Designed and collected emotion experiment data to ensure the reliability of model training and validation.