EEG Feature Selection for Emotion Recognition Based on Cross-subject Recursive Feature Elimination

The application of machine learning approaches to deal with the emotion recognition of physiological signals has received much attention on account of the objectivity of the electroencephalography (EEG) signals. However, the traditional feature selection methods are insufficient when building affective computing models between different subjects. In this paper, we propose a novel feature selection method termed as cross-subject recursive feature elimination (C-RFE) based on least square support vector machine to cope with this issue. This method is implemented through the absolute value of the component of the norm vector of the classification margin for all pairs of two subjects. The features are ranked in descending order of the importance via eliminating feature with the minimal contribution. The specific number of EEG feature subsets and emotion categories are operated in four machine learning models. The binary classification accuracy and F1-score of arousal and valence recognitions are achieved 0.6521, 0.6245, 0.6299 and 0.6295, respectively, for MAHNOB-HCI database, and 0.6461, 0.6176, 0.6529 and 0.6399, respectively for DEAP database.