Cross-Subject EEG Classification via Soft Threshold Semi-Supervised Learning Framework

EEG-based brain-computer interface (BCI) systems, combined with deep learning techniques, have been widely applied across various domains. However, the limited availability of labeled EEG data restricts the generalization and effectiveness of existing deep learning models. Semi-supervised learning (SSL) offers a potential solution to this challenge. To address both of these issues, we propose a soft threshold semi-supervised learning framework (STSLF) for EEG-based BCIs. The STSLF includes three modules: a joint SSL module, a contrastive learning module and a domain alignment module. The joint SSL module use a truncated Gaussion function to make full use of pseudo-labels. The contrastive learning module introduces a contrastive loss to learn deep representation of unlabeled data. The domain alignment module mitigates individual differences by reducing uncertainty. We conducted experiments to verify the performance of the proposed STSLF using the MI4C, SEED and SEED-IV datasets. Experimental results on cross-subject tasks demonstrate that semi-supervised learning plays an important role in improving the performance of deep learning models for EEG signal analysis.