Proceedings of the Annual Meeting of the Cognitive Science Society

Capturing shared cognitive processes across individuals is crucial for cross-subject EEG emotion recognition. Existing studies overlook the subjective experiential component introduced by individual cognitive biases, causing redundant information to be entangled with the extracted emotion-related features. To this end, we propose a cross-subject EEG emotion recognition method named CDDP (Cognitive Debiasing via Disentangled Pre-training). Specifically, during pre-training, cognitive disentanglement and contrastive learning are leveraged to extract subject-invariant intrinsic features (essentially emotion-relevant features) and subject-specific bias features (subjective experiential knowledge) from EEG signals. Meanwhile, a variational autoencoder (VAE) is introduced to generate diverse bias features in the latent space, alleviating overfitting caused by limited source domain EEG data. During fine-tuning, an emotion classifier is jointly trained with the pre-trained subject-invariant intrinsic feature encoder to capture discriminative emotion representations. CDDP achieves accuracies of 88.62% and 75.38% on the SEED and SEED-IV datasets, respectively, demonstrating state-of-the-art performance.