Enhancing EEG-based emotion recognition using Semi-supervised Co-training Ensemble Learning

Emotion recognition based on Brain-Computer Interface is crucial in deepening our understanding of humans' emotions and decision-making process. The enhanced precision in emotion measurement allows for more rigorous analysis of mental disorders and therapy effectiveness. Our method aims to solve two challenges in this field. First, existing models' features often fail to comprehensively capture the multiple dimensions of information in EEG signals, like temporal or frequency domains. We propose two models: a CNN model trained on temporal features and a DNN model trained on differential entropy features, and ensemble their predictions with weighted voting. Second, labels can be uncertain, where data is unconfidently labelled. This is due to emotions' subjectivity causing a lack of clear ground truth in EEG. The proposed method aims to mitigate this by using a semisupervised method that utilises data with uncertain labels as unlabelled data. Co-training is used to allow the two models to learn from each other. Our combined model achieves higher accuracy than temporal and spectral models by 8.51% and 6.87% respectively for the SEED dataset. For the MER dataset, its accuracy outperforms temporal and spectral models by 24.96% and 2.70% respectively for arousal classification, and 11.18% and 49.21% respectively for valence classification.