The principal component analysis of linear kernel and XGBoost models are introduced to design electroencephalogram (EEG) classification algorithm of four emotional states under continuous audio-visual stimulation. In order to reflect universality, the traditional power spectral density (PSD) is used as the feature of EEG signal, and the feature importance measure under the weight index is obtained with XGBoost learning. Then linear kernel principal component analysis is used to process the threshold selected features and send them to XGBoost model for recognition. According to the experimental analysis, gamma-band plays a more important role than other bands in XGBoost model recognition; in addition, for distribution on channels, the central, parietal, and right occipital regions play a more important role than other brain regions. The recognition accuracy of this algorithm is 78.4% and 92.6% respectively under the two recognition schemes of subjects all participation (SAP) and subject single dependent (SSD). Compared with other literature, this algorithm has made a great improvement. The scheme proposed is helpful to improve the recognition performance of brain-computer emotion system under audio-visual stimulation.