Improve the generalization of emotional classifiers across time by using training samples from different days.

Electroencephalographic (EEG)-based emotion recognition has attracted increasing attention from the field of human-computer interaction (HCI). However, there are a number of challenges for machines to correctly recognize human emotional states. One problem is how to generalize the emotion model across time, since the brain may show different patterns of EEG for the same emotion experience at different time. This study investigated the feasibility of adding samples from different days to the training set to improve the generalization of the emotion classifier. Eight subjects participated in this experiment, and they were asked to watch different kinds of movie clips to produce neutral, positive or negative emotional states for different trials. Five sessions in five different days were conducted for each subject. EEG signals were recorded throughout the experiment. Support vector machine was used to perform a classification of the three emotional states, in which the training samples may come from 1, 2, 3 or 4 days' sessions but have a same number. The results showed that three categories were classified with average accuracies of 64.9%, 68.7%, 70.9%, and 73.0% respectively for 1-day, 2-day, 3-day and 4-day conditions. N-day condition represented the case in which data from N days were sent to train the SVM and the remaining 5-N days were used to form the testing set. Importantly, the accuracy had an increasing trend with the number of days in the training set for all subjects. Compared with 1-day condition, the accuracy of 4-day condition has about 10% improvement with a peak of 81.2% By analyzing the selected features in 4-day condition, we found the distributions of the selected features were relatively stable across days. The classifier did select emotion-relevant features and reject time-relevant features. These results suggested that incorporation of samples from different days to the training set could improve the generalization of an emotion classifier across time.