Open AccessImproving Session-to-session Transfer Performance of Emotion Recognition Using Adaptive Support Vector Machine
Author(s) -
Kai Yang,
Guangcheng Bao,
Ying Zeng,
Tong Li,
Jun Shu,
Bin Yan
Publication year - 2020
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1601/4/042028
Subject(s) - electroencephalography , computer science , session (web analytics) , support vector machine , artificial intelligence , pattern recognition (psychology) , occipital lobe , frontal lobe , speech recognition , psychology , neuroscience , world wide web
The non-stationarity of electroencephalograph (EEG) signals has been a barrier in real-life application of EEG-based emotion recognition. The features extracted from emotional EEG vary from one session to another and a model trained on a temporally-limited EEG dataset may generalize poorly to data recorded at a different time for the same individual. In this study, the EEG features that stably characterize the differences between emotions are firstly explored. Then the classic progressive transductive support vector machine (PTSVM) is extended to three-classes classifications by a new region labelling rule, furthermore K-nearest neighbour algorithm and iterative process are utilized to improve the confidence of the predicted labels. Experimental results indicate that high gamma band local activation difference features and network features derived from prefrontal lobe, the temporal lobe on both sides, the left occipital-parietal lobe, the right occipital lobe and the posterior occipital lobe are temporally stable features for distinguishing different emotions (positive, neutral and negative). And the proposed session-to-session transfer model achieves an averaged classification accuracy 63.56% for three emotions classification on our dataset recorded from 23 subjects in three different days, which is 4% higher than the existing best adaptive classification model. Hence, the proposed classification model can effectively improve the performance of EEG-based session-to-session emotion recognition.