文献类型：期刊文献

英文题名：PSGCL: Pseudo-siamese supervised graph contrastive learning for enhancing prior knowledge guidance in EEG classification

作者：Li, Guangqiang[1];Chen, Ning[1];Zhu, Hongqing[1];Niu, Yixiang[1];Xu, Zhangyong[1];Li, Jing[1];Zhu, Zhiying[1]

机构：[1]East China Univ Sci & Technol, Sch Informat Sci & Engn, Shanghai 200237, Peoples R China

年份：2026

卷号：343

外文期刊名：KNOWLEDGE-BASED SYSTEMS

收录：;EI(收录号：20262020729516);WOS:【SCI-EXPANDED(收录号:WOS:001753023600001)】；

基金：This work was supported by the National Natural Science Foundation of China [grant number 61771196, 61872143] .

语种：英文

外文关键词：EEG signal classification; Knowledge guidance; Pseudo-siamese network; Graph neural network; Supervised contrastive learning

摘要：Neurophysiological prior knowledge-based topological information is crucial for EEG-based classification tasks, such as emotion recognition (ER), auditory spatial attention detection (ASAD), and driver fatigue recognition (DFR). However, the guidance of these prior knowledges may be weakened, as EEG signals are nonstationary, susceptible to noise interference, and exhibit significant variability across subjects. To enhance the guidance of prior knowledge in the EEG classification tasks while improving its adaptability to the variation of EEG data, a pseudo-siamese supervised graph contrastive learning (PSGCL) model is proposed. Specifically, a contrastive learning (CL) strategy based on supervised contrastive loss (SupConsL) and similarity consistency loss (SimConsL) is constructed. It leverages relationships between samples of the same or different classes to optimize the adjacency matrix initialized with prior knowledge, thus enhancing its adaptability to EEG variations. Meanwhile, a pseudo-siamese network (PSN)-based learnable view augmentation strategy is designed to generate distinct but semantically consistent views to avoid the risk of downstream task-related information loss caused by arbitrary data augmentation. Additionally, a sparseness constraint loss (SCL) is incorporated to enforce the sparsity of adjacency matrix while preserving crucial brain-region connectivity relationships. Extensive experiments on six datasets for 3 tasks demonstrate that: (i) The proposed PSGCL model consistently outperforms state-of-the-art baselines with relatively low computational complexity in ER, ASAD, and DFR tasks; (ii) Both the SupConsL and SimConsL-based prior knowledge optimization and the SCL-based constraint strategy significantly contribute to the performance enhancement of the proposed model. (iii) PSGCL effectively enhances the clustering effect and discriminability of EEG representations.