Manifold sorting feature selection for emotion recognition

Emotion recognition based on electroencephalogram (EEG) has attracted extensive attention in the fields of human-computer interaction and psychological healthcare. However, the redundancy of EEG signals poses a challenge to emotion recognition. Therefore, linking feature selection in EEG-based emotion recognition with specific brain regions holds great promise for solving cross-domain transfer EEG emotion recognition. In this study, we propose a novel manifold sorting feature selection (MSFS) method, and its corresponding multi-source classification framework. MSFS selects channel features most similar to the T7 and T8 channels, which are highly related to each emotion category, using effective manifold-based similarity ranking. The selected features are combined and used for prediction and classification. Experimental results demonstrate that the average accuracy of the MSFS method combined with SVM on the SEED and SEED-IV is 77.63%, and 53.65%, which are 1.26% and 2.90% higher than the SVM without MSFS. Furthermore, the performance of MSFS is compared with other advanced feature selection methods, which proves the superiority of our method.