EEG-Based Emotion Recognition Using Graph Neural Networks

Emotions are the psychological and physical states of human beings with consciousness in daily life.Emotion recognition is the basis and core technology for realizing human-machine emotional interaction.Due to the great success of artificial intelligence,emotion recognition has been successful in many fields such as computer science,cognitive science and so on.As the physiological signal of the cerebral cortex,Electroencephalography(EEG)signals can directly reflect the emotional state of human beings.Compared with external signals such as facial expressions and behaviors and postures,EEG signals are not deceptive.Therefore,the research of emotion recognition based on EEG signals has practical meaning.In this thesis,EEG signals are used as the basis for emotion recognition,and the CNN-LSTM network model and the model in the graph neural network are used to realize the emotion classification.The main research contents of this thesis are as follows:(1)In order to improve the accuracy of the emotion recognition of EEG,a CNN-LSTM model is proposed.The convolutional neural network(CNN)aims at independently learning some useful information from the preprocessed 62 channels of EEG.Then,the outputs of CNN are sequentially input into a long short-term memory(LSTM)network.The LSTM is used to extract multi-channel fusion emotional features,which are input to the fully connected layer immediately.Finally,the softmax classifier brings about the classification of the positive,neutral,and negative emotions.Emotion recognition experiments were carried out on the SJTU EEG emotion data set SEED,and an average recognition rate of 88.15% was obtained.In addition,an EEG-based emotion recognition model using LSTM is proposed.The input of this model is the emotional features of each channel of the EEG,and the LSTM network is used to extract the multi-channel fusion emotional features.Finally,the multi-channel fusion emotion features output by LSTM are input to the fully connected layer and softmax classifier to realize emotion classification.Finally,an average classification accuracy rate of 83.24% was achieved in the emotion classification experiment based on differential entropy feature.The results show that the CNN-LSTM model proposed is more effective,verifying the feasibility and effectiveness of the method.(2)An EEG-based emotion recognition model using graph convolutional neural network is proposed.The input of this model is the emotional features of each channel of EEG.The Chebyshev filter in the graph convolutional layer is used to extract the spatial structure of the EEG signal’s emotional features.The output feature of the graph convolutional layer is input to the fully connected layer.The softmax classifier brings about the emotion classification.There are three types of emotions,namely positive,neutral,and negative.Experiments were conducted on the SJTU EEG emotion data set SEED,and finally an average accuracy rate of 83.68% was achieved based on the differential entropy feature.Compared with the method based on the LSTM network,the experimental results obtained by the method based on the graph convolutional neural network are slightly better,indicating the feasibility of the graph convolutional neural network for modeling EEG signals.(3)In order to improve the accuracy of EEG signal emotion recognition,an EEG-based emotion recognition model using graph attention network is proposed.The model classifies the input EEG based on the emotional features,and the emotion categories are positive,neutral,and negative.The graph attention network designed in this paper contains two graph attention layers.The first layer is used to aggregate the features of neighbor nodes,and the second layer is used to realize emotion classification.Among them,the single-head attention mechanism and the four-head attention mechanism are respectively used in the first layer of the graph attention layer.Experiments were conducted on the SJTU EEG emotion data set SEED,and finally an average recognition rate of 90.22% was obtained in the experiment of the four-head graph attention mechanism based on the differential entropy feature,which is an increase of 2.07% compared to the experiment result based on the CNN-LSTM model,verifying the effectiveness of the method.