Multiple class transfer learning framework with source label adaptive correction for EEG emotion recognition

• Published in: Biomedical signal processing and control Vol. 104; p. 107536
• Main Authors: Zhu, Lei, Xu, Mengxuan, Huang, Aiai, Zhang, Jianhai, Tan, Xufei
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.06.2025
• Subjects: Affective computing; Domain adaptation; Electroencephalography; Emotion recognition; Transfer learning; Domain adaptation; Affective computing; Emotion recognition; Electroencephalography; Transfer learning
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2025.107536

•A multiple class domain adaptation module is designed to reduce distribution differences between source and target domain.•A source label adaptive correction module is proposed to mitigate the negative impact of noisy source labels.•We maximized the nuclear norm of the target domain output matrix to enhance prediction accuracy of target data.•We conducted experiments in cross-subject and cross-session scenarios to showcase our model's generalization capability. The field of Electroencephalogram (EEG)-based emotion recognition has attracted considerable research attention. However, EEG signals exhibit individual differences, and some emotion labels within the EEG data may be inaccurately annotated. These factors may adversely affect the performance of emotion recognition tasks. To overcome these challenges, we propose a Multiple Class Transfer Learning framework with Source Label Adaptive Correction (MCTL-SLAC) for EEG emotion recognition. In this model, firstly, a multiple class domain adaptation module is designed to enable more precise measurement and mitigation of distribution discrepancies between source and target data. Secondly, a prototype representation network based on target data is constructed to enable the adaptive transformation of source labels, thereby reducing the negative effects of noisy labels. Finally, within the target label prediction module, the nuclear norm of the target domain output matrix is maximized to enhance prediction accuracy. We conducted experiments on the SEED, SEED Ⅳ, and FACED datasets, comparing our method with existing domain adaptation approaches. Additionally, to assess the generalization capability of our model, we performed experiments under cross-session and cross-subject conditions. Under the cross-subject condition, our approach achieved classification accuracies of 87.30%, 72.61%, and 42.28% on the three datasets, respectively. Under the cross-session condition, the classification accuracies were 93.57% on the SEED dataset and 75.44% on the SEED IV dataset. The results of the experiments highlight the efficacy of our proposed model in tasks related to emotion recognition based on EEG signals.