Abstract: In order to enhance the performance of electroencephalogram (EEG)-based emotion
recognition and improve the accuracy of multi-classification, a spatial filtering algorithm using the
common spatial pattern (CSP) was proposed. Firstly, the traditional CSP method was used to design
the spatial domain filter. On this basis, three types of emotion recognition EEG signals (i.e., positive,
neutral, and negative) were linearly projected by this filter, so as to extract spatial features.
Furthermore, considering that the traditional joint approximation diagonalization (JAD) algorithm
using the “highest score eigenvalue” criterion may result in the failure to distinguish the
multi-classification emotional states, different eigenvalue selection methods were designed in terms
of the position of the eigenvalues with the highest scores. Under our lab environment, the
comparative experiments using the support vector model (SVM) as a classifier have been carried out.
The results show that the CSP-based spatial feature extraction method has an impressive accuracy of 87.54% on average in three-class emotion state recognition, proving the feasibility of the method in  the application of emotion recognition.

Key words: affective-brain computer interaction, common spatial pattern, joint approximation diagonalization, spatial filtering, emotion recognition