An electroencephalogram emotion recognition method based on structured bipartite graph for quantifying migration

Abstract

The application discloses a kind of electroencephalogram emotion recognition methods for realizing quantization migration based on structured bipartite graph.The specific steps of the method include: first, collecting the electroencephalogram data of multiple subjects under different emotional states, second, pre-processing the data to obtain a sample matrix, then constructing a model with structured bipartite graph and a graph-based semi-supervised label propagation model to obtain the optimal bipartite graph and label matrix, and finally completing the emotional electroencephalogram recognition. This method represents the correlation degree between samples by constructing a source domain sample-target domain sample structured bipartite graph. The bipartite graph actively optimizes its structure according to the constraint term in the model and incorporates label information into the graph during the iteration process. At the same time, this method quantifies the transferability of source domain samples, enabling high transferability samples in the source domain to play a greater role in the label propagation process, and solving the problem of inaccurate emotion recognition caused by individual differences in cross-subject emotion recognition.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of electroencephalogram signal processing, and particularly relates to an electroencephalogram emotion recognition method based on structured bipartite graph for realizing quantitative migration. BACKGROUND

[0002] Emotion is a comprehensive psychological and physiological state of human heart under the condition of external stimulation, which is spontaneous and unconscious, and is often accompanied by physiological reactions of central nervous system and electrical activity of human cerebral cortex. The physiological signal produced by the electrical activity of human cerebral cortex, i.e. electroencephalogram (EEG), provides us with an emotion recognition data source. Emotion computing is a multi-disciplinary research field, involving computer science, brain cognitive science and psychology, aiming to study and develop theories, methods and systems capable of recognizing, interpreting, processing and simulating human emotions. Electroencephalogram emotion recognition is a new branch of emotion computing, and its research object is human electroencephalogram signal. Unlike common input signals including video, picture (limb movement, facial expression, etc.), voice, text, etc., human electroencephalogram signal belongs to physiological signal and is not easy to fake, so emotion recognition based on electroencephalogram signal can better reflect the true emotion of an individual.

[0003] How to make machines accurately distinguish human emotions like humans is an important goal of electroencephalogram emotion recognition research. In the field of cross-subject electroencephalogram emotion recognition, individual differences are inevitable. The electroencephalogram signals of a person in different states and at different times, as well as the electroencephalogram signals of different people, all have distribution differences, which makes it difficult for machine learning methods assuming independent and identically distributed to generalize. Therefore, the concept of transfer learning is proposed, and its core idea is to reduce the domain difference between source domain and target domain. However, common transfer learning methods such as maximum mean difference method and deep neural network transfer cannot intuitively quantify the transferability of samples, and the interpretability of the process of transfer learning is relatively poor.

[0004] The graph-based learning method generally constructs a graph correlation matrix through the similarity between samples, where the similarity between samples can be measured by methods such as Gaussian kernel and Euclidean distance. In the graph correlation matrix, the greater the edge between two samples, the greater the similarity between them, and the greater the probability of the same label. Therefore, the semi-supervised label propagation based on graph can be understood as the process of propagating a limited number of labeled samples to a large number of labeled samples. When constructing a structured bipartite graph, the electroencephalogram signal after data preprocessing and feature extraction needs to be taken as input, and the bipartite graph is composed of two sets of nodes, i.e. source domain and target domain. Figure OneThe one side is the source domain sample (labeled), the other side is the target domain sample (unlabeled), there is no connection in the layer, there is full connection between the layers, based on the structured graph theory, a structured bipartite graph which can well reflect the correlation degree of the source domain sample and the target domain sample is obtained. Based on the learned optimal structured bipartite graph, on the one hand, the label of the unlabeled sample with high accuracy can be obtained, and the internal mechanism is that the greater the correlation degree between samples, the greater the probability that the labels are the same; on the other hand, the transferability of the source domain sample can be quantified through the structured bipartite graph, so that the source domain sample with high transferability can play a greater role in the label propagation process, and thus the accuracy of emotion recognition is improved. SUMMARY

[0005] The present application aims at the deficiencies of the existing electroencephalogram emotion recognition technology, and proposes an electroencephalogram emotion recognition method based on structured bipartite graph for realizing quantitative transfer. The source domain sample-target domain sample structured bipartite graph Q and the label matrix F are jointly iteratively optimized to obtain a structured bipartite graph which can well describe the correlation degree of the source domain sample and the target domain sample, and thus the accuracy of electroencephalogram emotion recognition is improved. In the present application, on the one hand, the label of the unlabeled sample can be obtained according to the label matrix F, and the label is the emotional state of the subject corresponding to the sample at the acquisition time; on the other hand, the transferability of each source domain sample can be quantified through the structured bipartite graph Q, so that the source domain sample with the greatest influence in the label propagation process is found out, and the source domain sample with high transferability can play a greater role in the label propagation process.

[0006] An electroencephalogram emotion recognition method based on structured bipartite graph for realizing quantitative transfer, comprising the following steps:

[0007] Step 1: collecting electroencephalogram data of multiple subjects in different emotional states.

[0008] Step 2: pre-processing and feature extraction are performed on the electroencephalogram data collected in step 1 to obtain a sample matrix X of each subject d is the number of features, and k is the total number of samples; the sample matrix X of different subjects is selected as a source domain sample matrix X s =[s1,s2,…,s m ] T and a target domain sample matrix X t =[t1,t2,…,t n ] T .

[0009] Step 3: constructing an electroencephalogram emotion recognition model based on structured bipartite graph for realizing quantitative transfer, and the objective function is as follows:

[0010]

[0011] stQ≥0,Q=1,F≥0,F1=1

[0012] Among them, s i X is the source domain sample matrix s The sample in; t i For the target domain sample matrix X t The samples in the source domain; Q is a structured bipartite graph correlation matrix describing the degree of association between the source domain samples and the target domain samples; q ij Let represent the element in the i-th row and j-th column of the correlation matrix Q; γ is the regularization constraint term; α is a preset parameter with a value range of 1 to 10000. Tr(·) represents the trace operation of the matrix. L is the Laplace matrix; It is a soft label matrix; F u ∈R n×c It is the soft label matrix of the target domain samples; F l ∈R m×c is the soft label matrix of the source domain samples; c is the number of sentiment state categories.

[0013] Step 4: Based on the objective function established in Step 3, perform joint iterative optimization on the structured bipartite graph correlation matrix Q and label matrix F.

[0014] Step 5: Determine the transferability from source domain samples to target domain samples based on the correlation matrix Q; obtain the emotional state of the target subject during EEG collection based on the label matrix F.

[0015] Preferably, the preprocessing and feature extraction process in step 2 is as follows: EEG data is sampled at a sampling rate of 200Hz, and then noise and artifacts are filtered out using a bandpass filter ranging from 1Hz to 75Hz. In the frequency domain, EEG signals are divided into five frequency bands: Delta, Theta, Alpha, Beta, and Gamma. Emotional features of the EEG signals in different frequency bands are extracted, and the differential entropy of the EEG signals is calculated as the sample matrix X.

[0016] As a preferred option, the specific process of constructing the EEG emotion recognition model based on structured bipartite graphs to achieve quantization transfer in step 3 is as follows:

[0017] Step 3.1: Construct a structured bipartite graph model:

[0018]

[0019] In equation (1), γ is a regularization constraint term.

[0020] Calculating the square of the Frobenius norm of the incidence matrix Q yields:

[0021]

[0022] Substitute formula (2) into formula (1) to obtain:

[0023]

[0024] Step 3.2, establish a graph-based semi-supervised label propagation model:

[0025]

[0026] Optimize formula (4) to make the predicted label closer to the true label, and the optimized model is:

[0027]

[0028] In formula (5), l i ∈R c , u j ∈R c are soft labels of source domain and target domain samples respectively; y j ∈R c is the true label of the source domain sample, and c is the number of emotional state categories; λ ∞ is a positive value for constraining .

[0029] Convert formula (5) into a matrix form:

[0030] minTr(F T LF)+Tr((F-Y) T U(F-Y)) (6)

[0031] In formula (6), is a hard label matrix, where Y n ∈R n×c and Y m ∈R m×c are true labels of source domain and target domain samples respectively, and Y n =0. Wherein, U m is a diagonal matrix and the values of diagonal elements are all set to λ ∞ .

[0032] Step 3.3, combine the structured bipartite graph construction model and the graph-based semi-supervised label propagation model, construct a structured bipartite graph by forcing non-negativity and row normalization constraints, and perform label propagation to obtain the target function as follows:

[0033]

[0034] In formula (7), α is a preset parameter, and the value range is 1-10000.

[0035] As preferred, the specific process of the joint iterative optimization in step 4 is as follows:

[0036] Step 4.1, similarity between samples is calculated using Euclidean distance, and a structured bipartite graph association matrix Q is initialized; a label matrix F is initialized, so that all elements of the matrix F are equal to

[0037] Step 4.2, fixing the data label matrix F, updating the association matrix Q as follows:

[0038]

[0039] Solving equation (10), the updated association matrix Q is obtained.

[0040] Step 4.3, fixing the association matrix Q, updating the data label matrix F as follows:

[0041] minTr(F T LF) (12)

[0042] In the objective function equation (12), equation (13) is introduced to make the predicted label closer to the true label, and the updated objective function is shown in equation (14):

[0043] minTr((F-Y) T U(F-Y))(13)

[0044]

[0045] Solving equation (14), the updated label matrix F is obtained.

[0046] Step 4.4, repeating steps 4.2 and 4.3 alternately until the objective function converges. The joint iterative optimization of the association matrix Q and the label matrix F is completed.

[0047] As preferred, the identified emotion categories are four, which are happy, sad, fear and calm.

[0048] As preferred, if the transferability of the source domain sample to the target domain sample is less than a preset value, the source domain sample is replaced to re-execute steps 3 to 5; until the transferability of a source domain sample to the target domain sample is greater than or equal to the preset value, the emotional state of the target subject during the electroencephalogram acquisition is obtained according to the corresponding label matrix F.

[0049] The present application has the following beneficial effects:

[0050] 1. The electroencephalogram emotion recognition method based on structured bipartite graph for realizing quantitative migration provided in the application provides an effective tool with high accuracy and high reliability for electroencephalogram emotion recognition, and the label of the target domain sample can be continuously iteratively optimized through the establishment of a mathematical model, so that the emotional state of the target subject can be accurately recognized according to the electroencephalogram data.

[0051] 2. The application is aimed at the difficult cross-subject emotion recognition problem in the field of electroencephalogram research, and the structured bipartite graph constructed in the application can well describe the correlation degree between the source domain sample and the target domain sample, and the structured bipartite graph is dynamically adjusted in the process of label propagation, which makes the label information well integrated into the graph, so that the emotional state of the target subject can be accurately recognized.

[0052] 3. The electroencephalogram emotion recognition method based on structured bipartite graph for realizing quantitative migration provided in the application is different from the existing electroencephalogram emotion recognition method which only focuses on improving the performance of emotion recognition, and the application pays attention to mining the migratability of the electroencephalogram data of the subject. The correlation degree between the electroencephalogram sample of the source subject and the electroencephalogram sample of the target subject is obtained through the structured bipartite graph, the migratability of each sample of the source subject is quantified, the high-migratability sample of the source domain is prompted to play a greater role in the process of label propagation, and quantitative migration based on the structured bipartite graph is realized. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 The framework diagram of the electroencephalogram emotion recognition method based on structured bipartite graph for realizing quantitative migration

[0054] Figure 2 The flowchart of the electroencephalogram emotion recognition method based on structured bipartite graph for realizing quantitative migration DETAILED DESCRIPTION

[0055] The application will be further described below in combination with the drawings.

[0056] An electroencephalogram emotion recognition method based on structured bipartite graph for realizing quantitative migration is used for recognizing electroencephalogram samples with unknown emotional state labels.

[0057] The electroencephalogram signals of different subjects in the same emotional state should have a significant correlation degree, so it is expected to construct a bipartite graph representing the correlation degree of the source domain samples and the target domain samples, and to perform label propagation through the correlation between the samples and the distance between the samples. The conventional graph label propagation ignores the label information, and the graph constructed in the process of label propagation is not updated, so the label information cannot be well integrated into the graph. The bipartite graph constructed in the application is dynamically updated in the process of label propagation, so the label information can be well integrated into the bipartite graph, and the accuracy of emotion prediction is further improved. At the same time, the application quantifies the transferability of the source domain samples, can distinguish the transferability of each sample in the source domain, and promotes the source domain high-transferability samples to play a greater role in the label propagation process, solving the problem of inaccurate emotion recognition caused by individual differences in the cross-subject emotion recognition field.

[0058] As shown in Figure 1 and Figure 2 , the electroencephalogram emotion recognition method based on the structured bipartite graph realizes the quantification of the transfer, and specifically includes the following steps:

[0059] Step 1, respectively collecting electroencephalogram data of multiple subjects in different emotional states.

[0060] Step 2, pre-processing and feature extraction are performed on the electroencephalogram data collected in step 1. The electroencephalogram data is sampled at a sampling rate of 200Hz, and then a band-pass filter of 1Hz-75Hz is used to filter out noise and artifacts. In the frequency domain, the electroencephalogram signal can be divided into five frequency bands: Delta (1-4Hz), Theta (4-8Hz), Alpha (8-14Hz), Beta (14-31Hz) and Gamma (31-50Hz). Using signal processing technology, the emotional features of the electroencephalogram signal in different frequency bands are extracted, specifically using differential entropy (DE) as the feature, and calculating the differential entropy of the electroencephalogram signal as the sample matrix X. The sample matrix X has an emotional category label; in this embodiment, the emotional categories are four, namely happy, sad, fear and calm.

[0061] The sample matrix X corresponding to different subjects is divided into a source domain sample matrix X s =[s1,s2,…,s m ] T and a target domain sample matrix X t =[t1,t2,…,t n ] T, the source domain samples and the target domain samples come from different subjects. Step 3, constructing an electroencephalogram emotion recognition model based on a structured bipartite graph to realize quantitative migration, including a structured bipartite graph construction model and a graph-based semi-supervised label propagation model. The structured bipartite graph construction model and the graph-based semi-supervised label propagation model are combined to obtain a source domain sample-target domain sample structured bipartite graph correlation matrix and a label matrix, the former quantifies the migratability of the source domain sample, and the latter obtains the label of the target domain to complete emotion recognition.

[0062] Step 3.1, establishing a structured bipartite graph construction model:

[0063]

[0064] In formula (1), Q is a structured bipartite graph correlation matrix to be constructed to describe the correlation degree between the source domain samples and the target domain samples; q ij represents the element of the ith row and the jth column of the correlation matrix Q, which is used to describe the correlation degree between the ith sample of the target domain and the jth sample of the source domain; represents the square calculation of the 2-norm of the matrix; represents the square calculation of the Frobenius norm of the matrix; represents the smoothing constraint of Q, to avoid Q from having a large number of singular solutions; s i is a source domain sample matrix X s ; t i is a target domain sample matrix x t ; and represents the distance between the ith target domain sample and the jth source domain sample, and when is too large, the value of q ij will be reduced, so as to meet the requirement of minimizing the objective function; γ is a regularization constraint term.

[0065] The square calculation of the Frobenius norm of the matrix Q is performed to obtain:

[0066]

[0067] Substituting formula (2) into formula (1) obtains:

[0068]

[0069] Step 3.2, establishing a graph-based semi-supervised label propagation model:

[0070]

[0071] The formula (4) is an original label propagation model, and the formula (4) is optimized to obtain a more accurate label matrix F in the label propagation process, so that the predicted label is closer to the real label, and the optimized formula is as follows:

[0072]

[0073] In the formula (5), l i ∈R c , u j ∈R c are soft labels of source domain and target domain samples respectively; y j ∈R c is a real label of the source domain sample, and c is the number of emotional state categories; λ ∞ is a very large positive number, and the purpose is to constrain the value to be very small, so that the predicted label is close to the real label. In order to make the formula (5) easy to understand, it is converted into a matrix form as follows:

[0074] min Tr (F T LF) + Tr ((F-Y) T U (F-Y)) (6)

[0075] In the formula (6), L is a Laplace matrix; is a soft label matrix for label propagation, wherein F u ∈R n×c is a soft label matrix of the target domain sample, F l ∈R m×c is a soft label matrix of the source domain sample, and c is the number of emotional state categories; Tr(·) represents a trace operation of a matrix; is a hard label matrix, wherein Y n ∈R n×c and Y m ∈R m×c are real labels of source domain and target domain samples respectively, and since the target domain sample has no label, Y n = 0 is directly set in the model. Wherein, U m is a diagonal matrix and the values of diagonal elements are all set to λ ∞ .

[0076] Step 3.3, the joint structured bipartite graph construction model and the graph-based semi-supervised label propagation model, the structured bipartite graph is constructed and label propagation is carried out by forcing non-negativity and row normalization constraint, and the obtained target function is as follows:

[0077]

[0078] In formula (7), a is a preset parameter, and the value range is 1-10000.

[0079] Step 4, according to the objective function established in step 3, the structured bipartite graph correlation matrix Q and the label matrix F are jointly iteratively optimized.

[0080] Step 4.1, the similarity between samples is calculated using the Euclidean distance, the structured bipartite graph correlation matrix Q is initialized; the label matrix F is initialized, so that all elements of the matrix F are equal to

[0081] Step 4.2, fix the data label matrix F, update the source domain sample-target domain sample structured bipartite graph correlation matrix Q as follows:

[0082]

[0083] l in formula (8) i ∈R c , u j ∈R c are the labels of the source domain and the target domain samples respectively. Substitute formula (8) into formula (7) to obtain:

[0084]

[0085] Observing formula (9), it can be seen that the solution process of each i in formula (9) is independent. For the i-th sample, the problem can be simplified as:

[0086]

[0087] Let The above formula extracts q ij as a common factor, and substitutes d ij into the formula, and simplifies formula (10) to:

[0088]

[0089] Solve formula (11) by using a convex optimization package to obtain γ and q i . The obtained (i=1, 2, 3...n) is each row vector of Q, q i Adjust by merging the label constraint term, which means that in each iteration process, the label information is implicitly integrated into the construction of the graph.

[0090] Step 4.3, fix the correlation matrix Q, update the data label matrix F as follows:

[0091] minTr(F T LF) (12)

[0092] To obtain more accurate label matrix F in the process of label propagation, the equation (13) is introduced into the objective function (12) to make the predicted label closer to the real label, and the updated objective function is shown as equation (14):

[0093] minTr((F-Y) T U(F-Y)) (13)

[0094]

[0095] In equation (14), the partial derivative of F is taken and the result is assigned to zero, and then simplified to obtain:

[0096] (L+U)F=UY (15)

[0097] Because the affinity matrix Q is subject to the constraint of Q1 = 1, D n is a diagonal matrix, and its diagonal elements are the row sums of Q; therefore, D n = I n , where I n is an n x n identity matrix; Λ is a diagonal matrix, and its diagonal elements are the column sums of Q, thus obtaining the Laplacian combination matrix L:

[0098]

[0099] Substituting into equation (16) to obtain:

[0100]

[0101] Solving F l and F u in equation (17) to obtain:

[0102] F l = (ΛU m -Q T Q) -1 U m Y m (18)

[0103] F u = QF l (19)

[0104] Step 4.4, repeat steps 4.2 and 4.3 alternately until the objective function converges. At this point, the joint iterative optimization of the structured bipartite graph affinity matrix Q and the label matrix F is completed, and the optimal affinity matrix Q and the label matrix F are obtained.

[0105] Step 5, analyze the label matrix F uand the structured bipartite graph correlation matrix Q. According to the label matrix F u Obtain the emotional state of the target subject at the time of electroencephalogram acquisition, F u Each column corresponds to a category of emotional state, and each row corresponds to a sample emotional state prediction. Take each row of F u The category corresponding to the largest element in the row is the emotional state of the target subject at the time of electroencephalogram acquisition for the unlabeled sample.

[0106] Step 6, analyze the structured bipartite graph correlation matrix Q obtained in step 4. From Q1=1, it can be seen that the sum of the elements in each row of matrix Q is 1, and its specific meaning is that the sum of the correlation degrees of any target subject sample and all source domain samples is 1; the sum of the elements in each column of matrix Q can be expressed as: The specific meaning is that the sum of the correlation degrees of the ith source domain sample and all target domain samples.

[0107] Therefore, A i can be used to quantify the transferability of the source domain sample, A i The larger the A u , the higher the transferability of the source domain sample, so we can find samples with high transferability in the source domain, and these samples with high transferability play a more significant role in the process of graph label propagation. Based on this, the label matrix F u can be further updated to obtain more accurate target subject emotional categories, the specific process is as follows:

[0108] Select a number of samples with high correlation degrees from the source domain samples as the preferred data set; using the preferred data set, obtain the updated label matrix F u , and use the category corresponding to the largest element in the updated label matrix F u as the emotional category of the target subject.

Claims

1. A method for electroencephalogram emotion recognition based on quantized migration realized by a structured bipartite graph, characterized in that: Comprising the following steps: Step 1, collecting electroencephalogram data of a plurality of subjects in different emotional states; Step 2, pre-process and extract features from the collected electroencephalogram data in step 1 to obtain a sample matrix of each subject d is the number of features, k is the total number of samples; the sample matrix X of different subjects is selected as the source domain sample matrix X s =[s1,s2,…,s m ] T and the target domain sample matrix X t =[t1,t2,…,t n ] T ; Step 3, constructing an electroencephalogram emotion recognition model based on structured bipartite graph to realize quantitative migration, and the objective function is as follows: s.t. Q >= 0, Q1 = 1, F >= 0, F1 = 1 wherein s i is a sample in the source domain sample matrix X s s i is a sample in the target domain sample matrix X t t ij represents the ith row and jth column element of the association matrix Q; γ is a regularization constraint term; α is a preset parameter, and its value range is 1-10000; Tr(·) represents a trace operation of a matrix; L is a Laplace matrix; is a soft label matrix; F u ∈R n×c is a soft label matrix of the target domain sample; F l ∈R m×c is a soft label matrix of the source domain sample; c is the number of emotional state categories; Step 4, according to the objective function established in step 3, the joint iterative optimization of the correlation matrix Q and the label matrix F is carried out; Step 5, according to the correlation matrix Q, the migratability of the source domain sample to the target domain sample is judged; and according to the label matrix F, the emotional state of the target subject during the electroencephalogram collection is obtained.

2. The electroencephalogram emotion recognition method based on structured bipartite graph for quantization migration according to claim 1, characterized in that: The pre-processing and feature extraction process in step 2 is: sampling the electroencephalogram data with a sampling rate of 200 Hz, and then filtering out noise and artifacts through a band-pass filter of 1 Hz-75 Hz; in the frequency domain, the electroencephalogram signal is divided into 5 frequency bands: Delta, Theta, Alpha, Beta and Gamma; Extracting emotional features of electroencephalogram signals in different frequency bands, and calculating the differential entropy of electroencephalogram signals as the sample matrix X.

3. The electroencephalogram emotion recognition method based on structured bipartite graph for quantization migration according to claim 1, characterized in that: The specific process of constructing the electroencephalogram emotion recognition model based on structured bipartite graph to realize quantitative migration in step 3 is as follows: Step 3.1, establishing a structured bipartite graph construction model: In formula (1), γ is a regularization constraint term; The square of the Frobenius norm of the correlation matrix Q is calculated to obtain: Substitute formula (2) into formula (1) to obtain: Step 3.2, establishing a graph-based semi-supervised label propagation model: Optimizing formula (4) to make the predicted label closer to the true label, and the optimized model is: In formula (5), l i ∈R c , u j ∈R c are soft labels of source domain and target domain samples respectively; y j ∈R c is a true label of a source domain sample, and c is the number of emotion state categories; λ ∞ is a positive value for constraining . Convert formula (5) to matrix form: minTr(F T LF)+Tr((F-Y) T U(F-Y)) (6) In formula (6), is a hard label matrix, where Y n ∈R n×c and Y m ∈R m×c are the real labels of the source domain and the target domain samples, respectively, and Y n = 0. where U m is a diagonal matrix and the values of the diagonal elements are all set to λ ∞ ; Step 3.3, joint structured bipartite graph construction model and graph-based semi-supervised label propagation model, construct structured bipartite graph and perform label propagation through forced non-negativity and row normalization constraints, and the objective function is as follows: In formula (7), α is a preset parameter, and the value range is 1-10000.

4. The electroencephalogram emotion recognition method based on structured bipartite graph for quantization migration according to claim 1, characterized in that: The specific process of joint iterative optimization in step 4 is as follows: Step 4.1, compute the similarity between samples using Euclidean distance, initialize the structured bipartite graph affinity matrix Q; initialize the label matrix F such that all elements of the matrix F are equal to Step 4.2, fixing the data label matrix F, updating the correlation matrix Q as follows: Solving formula (10) to obtain the updated correlation matrix Q; Step 4.3, fixing the correlation matrix Q, updating the data label matrix F as follows: minTr(F T LF) (12) In the objective function formula (12), introduce formula (13) to make the predicted label closer to the true label, and the updated objective function is as shown in formula (14): minTr((F-Y) T U(F-Y)) (13) Solve formula (14) to obtain the updated label matrix F; Step 4.4, repeat steps 4.2 and 4.3 alternately until the objective function converges; complete the joint iterative optimization of the correlation matrix Q and the label matrix F.

5. The electroencephalogram emotion recognition method based on structured bipartite graph for quantization migration according to claim 1, characterized in that: The recognized emotional categories are four, namely happy, sad, fear and calm.

6. The electroencephalogram emotion recognition method based on structured bipartite graph for quantization migration according to claim 1, characterized in that: If the migratability of the source domain sample to the target domain sample is less than the preset value, replace the source domain sample and re-execute steps 3 to 5; until the migratability of a source domain sample to the target domain sample is greater than or equal to the preset value, the emotional state of the target subject during the electroencephalogram collection is obtained according to the corresponding label matrix F.

Images