Study On Graph Learning Method For EEG Emotion Recognition

Emotion is important in human daily life,which has the function of information transmission and behavioral regulation.Endowing machines with the ability of recognizing emotions is of great significance in realizing the affective brain-computer interfaces.Affective computing is a current research hotspot and involves many disciplines,which aims to research emotion recognition model.EEG signals originating from the central nervous system are difficult to disguise and rich in emotion information.Thus,EEG signals are important data source for emotion recognition.Research on emotion recognition based on EEG signals has significant value.However,the EEG signals have non-stationary characteristics.Due to the subject factor and time factor,the training data and test data do not conform to the assumption of independent and identical distribution,which is the reason of low generalization ability of traditional machine learning methods.Usually,the performance of the model will increase as the number of samples increases,but labeling EEG signals needs to consume a lot of resources in practical situations.Inspired by existing graph learning methods,this thesis improves the EEG emotion recognition accuracy through the following two methods:(1)In order to solve the problem of differences in the distribution of EEG data,which leads to the reduction of model recognition ability.This thesis proposes an optimal graph coupled semi-supervised learning(OGSSL)model.Unlike traditional graph-based semi-supervised learning,OGSSL is an embedded model and unifies the adaptive graph learning and emotion recognition into a single objective.OGSSL not only projects the EEG data into the low-dimensional subspace,but also uses the subspace representation of the EEG data for constructing adaptive graph.Meanwhile,the projection matrix is added to the l2,1-norm,which ensures the projection matrix has the ability of feature selection.This thesis chooses the public EEG dataset SEEDⅣ to perform cross-session EEG emotion recognition experiments.In the three crosssession emotion recognition tasks,OGSSL achieves excellent average accuracies of 76.51%,77.08%and 81.29%,respectively.The experimental results show that OGSSL not only effectively improves the accuracy of EEG emotion recognition,but also has outstanding feature selection ability.(2)Aiming at the problem of high-dimensional data in EEG emotion recognition contains redundant features and noise features,this thesis proposes a discriminative subspace exploration with adaptive maximum entropy graph(DSMEG)model.The learning objective of DSMEG is to explore discriminative subspaces with increased inter-class scatter and decreased intra-class scatter.Unlike adaptive graph,DSMEG employs the maximum entropy constraint to avoid trivial solution problem.In DSMEG,discriminative subspace learning,adaptive maximum entropy graph and label prediction not only are combined into a unified framework,but also are coupled with each other.The discriminative subspace learning occupies the central position.DSMEG achieves the average recognition accuracies in three cross-session EEG emotion recognition tasks are 77.08%,78.80%and 81.88%,respectively.By comparing with the related semi-supervised learning model,it is verified that DSMEG can improve the recognition accuracy in the EEG emotion recognition task.Analyzing the projection matrix of the discriminant sub space can find EEG signals from the Gamma frequency band,the left/right temporal,prefrontal and(central)parietal lobes of the brain contribute more to emotion recognition.