Semi-supervised EEG-based emotion recognition method by optimal transport distribution matching

Abstract Large volumes of labeled data are needed for accurate emotion recognition from Electroencephalography (EEG) signals, however labeling EEG signals needs careful trial recordings and professional analysis, which is costly and time-consuming. However, the majority of EEG-based deep learning methods for emotion recognition are fully supervised. To overcome this issue, Semi-Supervised Learning (SSL) emerges as a promising solution as it requires only a limited amount of labeled data and can learn additional information from a large set of unlabeled ones. This study introduces a new distribution consistency regularization between the labeled and unlabeled EEG segments based on optimal transport (OT) distance, which results in a robust semi-supervised EEG-based emotion algorithm. This regularization enhances the quality of pseudo-labels generated, enhancing the performance of the emotion classifier in an SSL paradigm. Moreover, the proposed method bridges the gap between EEG-based emotion recognition and image classification tasks by introducing appropriate augmentations for EEG signals. The DEAP and SEED datasets are used to evaluate the efficacy of our method for binary and multi-class emotion recognition tasks. The study reveals that the proposed strategy consistently outperforms peer methods by a large margin across various quantities of labeled data.