An Online Evolutive Framework for Cross-Subject EEG Emotion Recognition

Electroencephalography (EEG)-based emotion recognition has attracted growing attention due to its potential in healthcare and human-computer interaction. A major challenge in this domain lies in the high inter-subject variability of EEG signals, which significantly limits the generalization of models across unseen subjects. To address this issue, we propose a novel self-supervised framework that leverages multiple candidate models, each specialized in extracting discriminative representations from EEG data. These models are trained with pretext tasks combining reconstruction and contrastive objectives to capture both temporal dynamics and frequency-domain patterns. During inference, an online selection mechanism adaptively identifies the most suitable model based on reconstruction loss, thereby improving robustness against data drift. Experiments on public EEG emotion recognition datasets demonstrate that the proposed framework achieves superior performance under the leave-one-subject-out (LOSO) protocol, outperforming existing self-supervised methods. The results highlight the effectiveness of dynamically leveraging heterogeneous model capabilities, offering a promising direction for calibration-free and generalizable EEG emotion recognition.