Zhang Yong, Chen Wenyun, Cai Xingxing, Cheng Cheng
School of Information Engineering, Huzhou University, Huzhou, 313000, China; School of Computer and Information Technology, Liaoning Normal University, Dalian, 116081, China.
School of Information Engineering, Huzhou University, Huzhou, 313000, China.
Neural Netw. 2025 Oct;190:107614. doi: 10.1016/j.neunet.2025.107614. Epub 2025 May 29.
Electroencephalography (EEG) serves as a valuable technique for objective emotion recognition, with promising applications across diverse fields. However, a major obstacle to the general application of EEG-based emotion recognition systems is the lack of labeled data. To overcome this limitation, we propose a semi-supervised alignment self-knowledge distillation with a mixed contrastive learning model (SASD-MCL) for cross-subject EEG emotion recognition, addressing the issue of limited labeled data. Firstly, we utilize mixed contrastive data augmentation methods to enhance data diversity and richness. Secondly, we introduce semi-supervised similarity alignment techniques to effectively combine labeled and unlabeled data, thereby improving the model's generalization and robustness. Then, we utilize unsupervised self-knowledge distillation to convey intricate complex knowledge, expediting the adaptation process to the features of the target domain. Finally, we use semi-supervised multi-domain adaptation algorithms to successfully deal with data distribution disparities across various domains (labeled, unlabeled source and target domains), boosting the model's robustness and performance in cross-subject emotion recognition. Extensive experiments employing a semi-supervised cross-subject leave-one-subject-out validation methodology on the SEED and SEED-IV benchmark datasets demonstrate that our proposed model outperforms existing methods under various imperfect labeling scenarios. The model effectively resolves label scarcity issues in cross-subject emotion recognition using EEG, achieving average performance increases of 5.93% on SEED and 5.32% on SEED-IV.
脑电图（EEG）是一种用于客观情绪识别的重要技术，在各个领域都有广阔的应用前景。然而，基于EEG的情绪识别系统普遍应用的一个主要障碍是缺乏标注数据。为了克服这一限制，我们提出了一种带有混合对比学习模型的半监督对齐自知识蒸馏方法（SASD-MCL）用于跨被试EEG情绪识别，以解决标注数据有限的问题。首先，我们利用混合对比数据增强方法来提高数据的多样性和丰富性。其次，我们引入半监督相似性对齐技术，有效地结合标注和未标注数据，从而提高模型的泛化能力和鲁棒性。然后，我们利用无监督自知识蒸馏来传递复杂的知识，加速模型对目标域特征的适应过程。最后，我们使用半监督多域适应算法成功地处理不同域（标注、未标注源域和目标域）之间的数据分布差异，提高模型在跨被试情绪识别中的鲁棒性和性能。在SEED和SEED-IV基准数据集上采用半监督跨被试留一被试验证方法进行的大量实验表明，我们提出的模型在各种不完美标注场景下优于现有方法。该模型有效地解决了使用EEG进行跨被试情绪识别时的标签稀缺问题，在SEED上平均性能提升了5.93%，在SEED-IV上提升了5.32%。
