Incorporation of Multiple-Days Information to Improve the Generalization of EEG-Based Emotion Recognition Over Time.

Current studies have got a series of satisfying accuracies in EEG-based emotion classification, but most of the classifiers used in previous studies are totally time-limited. To produce generalizable results, the emotion classifier should be stable over days, in which the day-to-day variations of EEG should be appropriately handled. To improve the generalization of EEG-based emotion recognition over time by learning multiple-days information which embraces the day-to-day variations, in this paper, 17 subjects were recruited to view several video clips to experience different emotion states, and each subject was required to perform five sessions in 5 days distributed over 1 month. Support vector machine was built to perform a classification, in which the training samples may come from 1, 2, 3, or 4 days' sessions but have a same number, termed learning 1-days information (L1DI), learning 2-days information (L2DI), learning 3-days information (L3DI), and learning 4-days information (L4DI) conditions, respectively. The results revealed that the EEG variability could impair the performance of emotion classifier dramatically, and learning more days' information to construct a classifier could significantly improve the generalization of EEG-based emotion recognition over time. Mean accuracies were 62.78, 67.92, 70.75, and 72.50% at L1DI, L2DI, L3DI, and L4DI conditions, respectively. Features at L4DI condition were ranked by modified RFE, and features providing better contribution were applied to obtain the performances of all conditions, results showed that the performance of SVMs trained and tested with the feature subset were all improved for L1DI, L2DI (*p < 0.05), L3DI (**p < 0.01), and L4DI (*p < 0.05) conditions. It could be a substantial step forward in the development of emotion recognition from EEG signals because it may enable a classifier trained on one time to handle another.