The Minimum Spanning Tree Based Emotion Classification Methods Of EEG Data

Emotion processing is one of the advanced cognitive functions of the human brain.As a complex,mysterious and interdisciplinary research field,affective cognitive mechanism and related research has attracted extensive attention of researchers in various disciplines.In emotional cognitive activities,the brain regions are interrelated,coordinated and cooperate with each other to form a complex network of brain functions.In recent years,many new discoveries and meaningful results have been obtained by using complex networks,graph theory and other methods to analyze brain networks.The minimum spanning tree represents the key information flow in the network and avoids many controversial issues such as the selection of traditional network threshold.It shows potential advantages in the field of brain network research.EEG signals can accurately describe the active state of the brain,and have the advantages of high time resolution,low cost,non-invasiveness and convenient operation.EEG has gradually become the primary tool adopted in emotional research at this stage.As the classification and prediction model of deep learning has injected new vitality into the field of computer vision and pattern recognition,and it has a good performance in the research of emotion classification.However,most of the current reseraches directly input the original EEG into the deep learning model,ignoring the large amount of noise and redundant information contained in the original EEG.Extracting suitable feature information from EEG signals and then sending it into deep learning model may be beneficial to further improve the accuracy of emotion classification.This dissertation presents a new method for emotion classification of EEG signals based on minimum spanning tree.First construct the minimum spanning tree and calculate its characteristics.Then,fuse the EEG frequency domain feature with the minimum spanning tree characteristics as emotion on the basis of the convolutional neural network.After feature fusion,the classification accuracy has increased evidently.Main tasks are as follow.(1)The EEG electrodes are used as brain network nodes,and the phase lag index measures the strength of the correlation between the nodes to generate the correlation matrix.Construct a minimum spanning tree based on the phase lag index and calculate its characteristics.Kruskal algorithm is used to construct the minimum spanning tree of connectivity brain network.The characteristics values of minimum spanning tree network were calculated and analyzed by repeated measurement ANOVA and paired t-test.It is found that the characteristics differences between different emotions are mainly concentrated in the? band.Among the characteristics,there were significant differences in mean phase lag index,maximum degree,leaf fraction,diameter and eccentricity.The results show that the topological structure of high arousal trees tends to be more star-shaped than that of low valence trees.At the same high arousal level,the topological structure of the low valence tree tends to be more star-shaped than that of the high valence tree.This result provides sufficient theoretical support for the cognitive mechanism of the brain in the emotional state from the perspective of the brain network.(2)Fusion of minimum spanning tree characteristics and frequency domain features through Fisher scoring.In this dissertation,Welch algorithm is used to calculate the power spectral density features of EEG signals.The F-score of the minimum spanning tree attribute is calculated by Fisher score,which is arranged in descending order and merged with the power spectral density feature.Finally,the F-score is transformed into a grayscale image as the input of the neural network.From the electrode positions corresponding to the minimum spanning tree properties,it is found that the electrode positions with higher F-score are located in the apical area,central area,frontal area and a small part of the occipital area.Compared with similar studies the use of,fisher scores can find more emotionally active brain regions and better reveal the activation state of the brain regions.(3)Research on emotion classification based on convolutional neural network.After fusion of frequency domain features and minimum spanning tree characteristics,a new EEG emotion classification method is proposed in combination with convolutional neural network.Compared with the power spectrum density feature matrix as the input of the convolutional neural network alone,the accuracy 82.33% and 75.46%.After fusion with the minimum spanning tree characteristic,The classification accuracy of arousal and valence can be improved by 3.37% and 4.46% respectively.The results show that the proposed fusion method of minimum spanning tree attributes and traditional frequency domain features is better than the current similar research,and this method can recognize emotion better.