LAUSR

Recently, electroencephalogram (EEG)-based emotion recognition has attracted increasing interests in research community. The weak, non-stationary, multi-rhythm and multi-channel properties of EEG data easily cause the extracted EEG samples and features contribute differently in recognizing emotional states. However, existing studies either failed to consider both the issues of sample and feature importance or only considered one of them. In this brief, we propose a new model termed sJSFE (semi-supervised Joint Sample and Feature importance Evaluation) to quantitatively measure the sample and feature importance by self-paced learning and feature self-weighting respectively. Experimental results on the SEED-IV data set show that the emotion recognition performance is greatly improved by mining both the sample and feature importance. Specifically, the average accuracy obtained by sJSFE across the three cross-session recognition tasks is 82.45%, which is respectively 3.72% and 7.21% and 10.47% and 18.82% higher than the results of traditional models. Besides, the feature importance vector depicts that the Gamma frequency band contributes the most, and the brain regions of prefrontal, left/right temporal and (central) parietal lobes correlate more to emotion recognition. The sample importance descriptor shows that continual transitions of video types in consecutive trials might weaken the feature-label consistency of the collected EEG data.