Emotion recognition method based on EEG transfer by combining semi-supervised regression and graph label propagation

Abstract

The present invention discloses an EEG transfer emotion recognition method combining semi-supervised regression and graph label propagation. The method collects EEG data from different subjects and extracts features as a sample matrix. Two sub-models, a dual mapping domain adaptation model and a semi-supervised regression and graph label propagation joint model, are constructed for joint optimization. During the optimization process, different variables are alternately iteratively optimized to continuously optimize the feature shared subspace, reduce the data distribution difference between the source domain and the target domain, learn more accurate target domain labels, and further reduce the difference between the two domains. This method can reduce the distribution difference of EEG data between subjects, and at the same time further suppress the influence of the differences between the subject data features on the emotion recognition process, solving the problem of poor recognition effect due to individual differences between subjects in the field of cross-subject emotion recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of electroencephalogram (EEG) signal processing, and specifically relates to an EEG migration emotion recognition method combining semi-supervised regression and graph label propagation. Background Art

[0002] In 1990, Salovey and Mayer proposed the concept of emotional intelligence, and believed that in addition to logical intelligence, emotional intelligence is also an important part of artificial intelligence. For machines, being able to automatically and accurately identify people's emotional states is the key to realizing machine emotional intelligence. Emotion is a state that integrates people's feelings, thoughts and behaviors. It not only includes people's psychological reactions to the external environment or their own stimuli, but also includes the physiological reactions that accompany such psychological reactions. In addition, people's emotional expressions, such as facial expressions and voices, will unconsciously disguise and change due to the influence of external environment, self-cognition and other factors. However, there is a direct correlation between EEG signals derived from central nervous system activities and people's real emotions, which cannot be disguised. Therefore, they are widely used in the field of objective emotion recognition and have become a research hotspot in this field. At the same time, with the continuous development of computer technology, biological sciences, neuroscience and other disciplines, EEG signals have played an increasingly important role in improving the accuracy and reliability of emotional state recognition in human-computer interaction.

[0003] Although EEG signals can be used to objectively identify emotional states because they cannot be disguised, there is an important problem with EEG signals, which is individual differences. EEG signals between people and EEG signals of a person at different times all have distribution differences, which makes it difficult to generalize machine learning methods that assume the same distribution. Therefore, the concept of transfer learning was proposed, and its basic idea is to use the knowledge of the auxiliary domain to promote the recognition task of the target domain. Common transfer learning methods can generally be divided into four categories: model-based, feature-based, sample-based, and relationship-based. Among them, feature-based methods are the most widely used transfer methods. The purpose of this type of method is to learn a shared feature representation and combine some distribution measurement strategies, such as the maximum mean difference, to map the target domain and source domain data into a shared subspace to reduce the difference between the two domains and achieve a state of basic approximate distribution. The calculation of the conditional distribution is closely related to the labels of the source domain and the target domain. Therefore, if the accurate target domain label can be learned, a better projection matrix can be obtained, which can better reduce the distribution difference between the source domain and the target domain to obtain a better target label, improve the model recognition accuracy, and ensure the reliability of emotional human-computer interaction. The mapping methods of the original data of the two domains can be divided into single mapping and dual mapping. Theoretically, compared with single mapping, dual mapping will consider the individual characteristics of the two domains in the mapping process, but in order to prevent the excessive reflection of the individual characteristics of the two domains in the shared subspace, it is also necessary to impose certain constraints on the mapping matrices of the two domains. Therefore, it is theoretically believed that single mapping should be included in dual mapping, while most of the existing technologies adopt a single mapping approach, ignoring the beneficial individual characteristics of the domain itself during the mapping process, and only retaining the common features that can be used for migration.

[0004] At the same time, due to the individual differences in EEG signals, the characteristic representations of EEG signals of the same emotion of different subjects will also be different. Most existing technologies simply perform cross-subject transfer recognition, which cannot well meet the requirements of human-computer interaction for emotion recognition accuracy. Therefore, it is necessary to combine certain learning methods to optimize the recognition process. Graph-based label propagation is a more suitable method, and its effect has been reflected in the existing technology (CN113157094A). However, when the undirected graph constructed is used for label propagation between two domains, it not only considers the association between the source domain and the target domain, but also considers the association between samples in the domain. In a dual mapping environment, this association may affect the final recognition result due to the retention of some individual characteristics of the two domains. Summary of the invention

[0005] In view of the shortcomings of the existing technology, this paper proposes a method for EEG emotion recognition based on joint semi-supervised regression and graph label propagation. By aligning the projection matrix P of the source domain data and the target domain data in the shared subspace, s ,P t, the sample-label mapping matrix W for semi-supervised regression, the sample-feature bipartite graph association matrix S, and the target domain sentiment label F t Perform joint iterative optimization, and obtain better aligned data by continuously iteratively optimizing the feature shared subspace, so as to better apply it to the learning of the sample-label mapping matrix W and the sample-feature association matrix S, so as to obtain a more accurate target domain sentiment label F t Application to learning in shared subspaces.

[0006] Step 1: Collect EEG data of multiple subjects in c different emotional states.

[0007] Step 2: Preprocess and extract features of the EEG data collected in step 1. Each sample matrix X consists of the EEG features of a subject, and the label vector y is the emotion label corresponding to the EEG features in the sample matrix X. Select two sample matrices X from different subjects as source domain data X, s and target domain data X t .

[0008] Step 3: Construct an EEG transfer emotion recognition model that combines semi-supervised regression and graph label propagation, including a dual-mapping domain adaptation model and a joint model of semi-supervised regression and graph label propagation.

[0009] Step 3.1: Establish a dual-map domain adaptation model:

[0010]

[0011]

[0012] in, are the mapping matrices that map the source domain data and the target domain data into the shared subspace, respectively. The superscript T represents transposition, p represents the shared subspace dimension, and d represents the original sample dimension. To augment the source domain data labels, Y s is the one-hot encoding of the source domain label, and c is the number of categories; To augment the target domain data labels, represents a column vector whose elements are all 1, F t is the one-hot encoding of the target domain label, n s ,n t are the number of samples of source domain data and target domain data, n=n s +n t Indicates the total number of samples; is a diagonal matrix, the diagonal matrix N dom The kth diagonal element in is the reciprocal of the number of samples of the kth source or target domain data, k = 1, 2, ..., c; is the center matrix, is the unit matrix, represents the square calculation of the Frobenius norm, ||P s -P t || 2,1 Mathematically, it is a row sparse constraint on the difference of the mapping matrix of the shared subspace of the two domains. Its significance is to prevent the excessive reflection of the individual characteristics of the two domains in the mapping process. 2,1 Represents the calculation of the 21-norm.

[0013] Step 3.2: Establish a joint model of semi-supervised regression and graph label propagation:

[0014]

[0015] stG≥0,G1=1,F≥0,F1=1,

[0016] in, is the sample-label mapping matrix, represents the sample label matrix, Represents X T The average of the difference between PW and F, represents the sample-feature bipartite graph association matrix, represents the sample-feature association matrix, 0 n ,0 p The dimensions are The all-zero matrix, represents the initialized graph incidence matrix constructed based on the current shared subspace data, is the sample-feature association matrix initialized based on the current shared subspace data, represents the sample and feature label matrix, is the graph Laplacian matrix, a diagonal matrix Its diagonal elements are the row sums of the matrix S, Tr(·) represents the trace of the matrix to be solved, and in formula (2) is the standard form of semi-supervised regression, is the row sparseness constraint on the matrix W, It is a bipartite graph construction and label propagation process.

[0017] Step 3.3: Combine the two models constructed in steps 3.1 and 3.2, integrate the dual-mapping domain adaptation model and the semi-supervised regression and graph label propagation joint model into a unified framework for joint optimization, and calculate the target domain sentiment label F t , the optimization model is:

[0018]

[0019] α,γ,β,λ are all adjustable parameters in the optimization process.

[0020] Step 4: According to the optimization model established in step 3.3, the projection matrix P of the source domain data and the target domain data aligned in the shared subspace is s ,P t , the sample-label mapping matrix W for semi-supervised regression, the sample-feature bipartite graph association matrix S, and the target domain sentiment label F t Perform joint iterative optimization.

[0021] Step 4.1: Initialize the target domain data label F t And the feature label F d The value of And construct the initial shared subspace mapping matrix P by dimensionality reduction s ,P t .

[0022] Preferably, the shared subspace mapping matrix is ​​constructed by principal component analysis (PCA).

[0023] Step 4.2: Fix the target domain data label F t and the shared subspace mapping matrix P s ,P t , update the sample-label mapping matrix W, the objective function is:

[0024]

[0025] Solve formula (4) to obtain the updated sample-label mapping matrix W′.

[0026] Step 4.3: Fix the sample-label mapping matrix W and the target domain data label F t , update the shared subspace mapping matrix P s ,P t The objective function is:

[0027]

[0028] Solve formula (5) to obtain the updated shared subspace mapping matrix P s ′,P t ′.

[0029] Step 4.4: Fix the shared subspace mapping matrix P s ,P t , target domain data label F t and feature label F d , update the sample-feature bipartite graph association matrix S and calculate the new feature label Fd , the objective function is:

[0030]

[0031] stG≥0,G1=1

[0032] Solve formula (6) to obtain the updated sample-feature bipartite graph association matrix S′ and feature label F d ′.

[0033] Step 4.5: Fixed sample-label mapping matrix W and shared subspace mapping matrix P s ,P t As well as the sample-feature bipartite graph association matrix S and feature label F d , update the target domain data label F t , the objective function is:

[0034]

[0035] stF≥0,F1=1

[0036] Solve formula (7) to obtain the updated target domain data label F t ′.

[0037] Step 4.6: Repeat steps 4.2, 4.3, 4.4, and 4.5 multiple times to complete the sample-label mapping matrix W and the shared subspace mapping matrix P. s ,P t , sample-feature bipartite graph association matrix S and target domain data label F t Joint iterative optimization.

[0038] Step 5: Substitute the result of the joint iterative optimization in step 4 into formula (3), and then use the EEG data feature X of the unknown emotional state as m As the target domain data X t Input formula (3) to solve the corresponding target domain data label F t , get the EEG data feature X m The corresponding emotional state at the time of collection.

[0039] The present invention has the following beneficial effects:

[0040] 1. The EEG transfer emotion recognition model combining semi-supervised regression and graph label propagation proposed in the present invention provides an effective tool with higher accuracy for emotional human-computer interaction. By continuously iteratively optimizing the target label through a mathematical model, the emotional state of the subject can be accurately identified based on the EEG data.

[0041] 2. The present invention is aimed at the difficult cross-subject scenario in the field of EEG research. It uses a dual-mapping domain adaptation model and a semi-supervised regression and graph label propagation joint model for iterative optimization. By aligning the projection matrix P of the source domain data and the target domain data in the shared subspace, s ,P t , sample-label mapping matrix W for semi-supervised regression, sample-feature association matrix S, and target domain sentiment label F t Perform joint iterative optimization, and obtain better aligned data by continuously iteratively optimizing the feature shared subspace, so as to better apply it to the learning of the sample-label mapping matrix W and the sample-feature association matrix S, thereby obtaining a more accurate target domain sentiment label F t Application to learning in shared subspaces.

[0042] 3. The present invention proposes two sub-models: a dual-mapping domain adaptation model and a semi-supervised regression and graph label propagation joint model. The dual-mapping pre-adaptation model adopts a method of mapping the source domain and the target domain to a shared subspace respectively, and combines the constraints of the mapping matrices of the two domains, which not only retains the common features between the domains, but also does not completely ignore the beneficial individual features of the two domains; the semi-supervised regression and graph label propagation joint model adopts a multi-model joint method to combine the semi-supervised regression model and the label propagation method based on the sample-feature bipartite graph to jointly perform target domain label recognition. The former is used to characterize the beneficial association within the domain, and the latter is based on the recognition that similar samples should have similar feature distributions. It is used to characterize the label propagation between the source domain and the target domain based on the feature, thereby ignoring the possible adverse effects of the sample association within the domain on the label propagation process. At the same time, it abandons the construction method of the conventional bipartite graph based on the association between samples, adds the feature level, and performs label propagation between the two domains from the perspective of the similarity of feature distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a framework diagram of the EEG transfer emotion recognition method;

[0044] Figure 2 Flow chart of the EEG transfer emotion recognition method;

[0045] Figure 3 It is a representation of constructing a sample-feature bipartite graph. DETAILED DESCRIPTION

[0046] The present invention uses a dual-mapping domain adaptation model and a semi-supervised regression and graph label propagation model to jointly iterate to identify samples with unknown emotional state labels. For the dual-mapping pre-adaptation model, the source domain and the target domain correspond to their own shared subspace mapping matrices P s ,P t, and at the same time, certain constraints are imposed on the differences that may be caused by the two domains in the mapping process; for semi-supervised regression and graph semi-supervised models, conventional bipartite graphs are constructed from the perspective of the association between samples. When labels are propagated, they are propagated through the association between samples, generally the distance between samples. We believe that samples of the same type should have similar feature distributions, so we expect to construct a bipartite graph that characterizes the sample-feature association to carry out label propagation. At the same time, since the bipartite graph completely ignores the association between samples in the domain, in order to further compensate for the possible beneficial effects of the association between the domains themselves, a multi-model joint approach is adopted to combine the semi-supervised regression model on the basis of semi-supervised graph label propagation to form a joint label recognition model. The present invention is further explained below in conjunction with the accompanying drawings;

[0047] like Figure 1 and Figure 2 As shown in FIG. 1 , the method for combining semi-supervised regression with graph label propagation for EEG transfer emotion recognition specifically includes the following steps:

[0048] Step 1: EEG data collection.

[0049] Human emotions do not appear to be very strong in daily situations. In order to collect strong emotional information, it is necessary to induce the subjects to a certain extent. This embodiment selects 4 film clips with obvious emotional tendencies and plays them to different subjects at different times. While watching the films, the EEG data of the subjects are collected by connecting the EEG cap leads to the corresponding brain areas as the original emotional EEG data set.

[0050] Step 2: Data preprocessing.

[0051] The EEG data collected in step 1 are sampled at a sampling rate of 200 Hz, and then the noise and artifacts are filtered out through a bandpass filter of 1 Hz to 75 Hz. The differential entropy (DE) is calculated as the sample matrix X in five frequency bands: Delta (1-4 Hz), Theta (4-8 Hz), Alpha (8-14 Hz), Beta (14-31 Hz) and Gamma (31-50 Hz):

[0052]

[0053] Among them, σ is the standard deviation of the probability density function; μ is the expectation of the probability density function.

[0054] It can be seen that the differential entropy feature is essentially the logarithmic form of the power spectrum density feature, that is, The preprocessing of EEG signals aims to improve the signal-to-noise ratio, thereby improving the preprocessing effect of the data and reducing interference.

[0055] Step 3: Establish an EEG transfer emotion recognition model that combines semi-supervised regression and graph label propagation, and integrate the dual-mapping domain adaptation model and the joint label recognition model: the former obtains a shared subspace to reduce the distribution difference between the source domain and the target domain, and the latter performs joint label recognition based on the obtained aligned data, and obtains the optimized target domain label for the former to better align the conditional distribution of the two domains, thereby obtaining the joint optimization objective function. The specific steps are:

[0056] Step 3.1: Establish a dual-map domain adaptation model:

[0057]

[0058]

[0059] in, are the mapping matrices that map the source domain data and the target domain data into the shared subspace, respectively. The superscript T represents transposition, p represents the shared subspace dimension, and d represents the original sample dimension. To augment the source domain data labels, Y s is the one-hot encoding of the source domain label, and c is the number of categories; To augment the target domain data labels, represents a column vector whose elements are all 1, F t is the one-hot encoding of the target domain label, n s ,n t are the number of samples of source domain data and target domain data, n=n s +n t Indicates the total number of samples; is a diagonal matrix, the diagonal matrix N dom The kth diagonal element in is the reciprocal of the number of samples of the kth source or target domain data, k = 1, 2, ..., c; is the center matrix, is the unit matrix, represents the square calculation of the Frobenius norm, ||·|| 2,1 represents the calculation of the 21 norm, ||P s -P t || 2,1 The row sparse constraint of the difference of the mapping matrix of the shared subspace of the two domains can be used to prevent the excessive reflection of the individual characteristics of the two domains in the mapping process.

[0060] Step 3.2: Establish a joint model of semi-supervised regression and graph label propagation:

[0061]

[0062] stG≥0,G1=1,F≥0,F1=1,

[0063] in, is the sample-label mapping matrix, represents the sample label matrix, Mathematically represent the semi-supervised regression X T The average of the difference between PW and F, represents the sample-feature bipartite graph association matrix, represents the sample-feature association matrix, 0 n ,0 p The dimensions are The all-zero matrix, represents the initialized graph incidence matrix constructed based on the current shared subspace data, is the sample-feature association matrix initialized based on the current shared subspace data, represents the sample and feature label matrix, is the graph Laplacian matrix, a diagonal matrix Its diagonal elements are the row sums of the matrix S, Tr(·) represents the trace of the matrix to be solved, and in formula (2) is the standard form of semi-supervised regression, is the row sparseness constraint on the matrix W, represents the square calculation of the 21 norm, It is a graph-based semi-supervised label propagation. The process first constructs an initialization graph based on the feature distribution of the shared subspace data, and then continuously iterates and optimizes the optimal sample-feature bipartite graph that is more suitable for the label propagation process based on the constructed initialization graph. In addition, since the bipartite graph that is expected to be constructed will not have a large difference from the feature distribution of the shared subspace data, that is, The target domain label is propagated based on the graph Laplacian matrix corresponding to the constructed optimal sample-feature bipartite graph to obtain the optimal target domain label.

[0064] Step 3.3: Combine the two models constructed in steps 3.1 and 3.2 and perform joint optimization according to the optimization model shown in formula (11):

[0065]

[0066] α,γ,β,λ are all adjustable parameters in the optimization process.

[0067] Step 4: According to the optimization model established in step 3.3, the projection matrix P of the source domain data and the target domain data aligned in the shared subspace is s ,P t, the sample-label mapping matrix W for semi-supervised regression, the sample-feature bipartite graph association matrix S, and the target domain sentiment label F t Perform joint iterative optimization.

[0068] Step 4.1: Initialize the target domain data label F t And the feature label F d The value of is 0.25, and the initial shared subspace mapping matrix P is constructed by principal component analysis (PCA) s ,P t .

[0069] Step 4.2: Fix the target domain data label F t and the shared subspace mapping matrix P s ,P t , update the sample-label mapping matrix W, the objective function is:

[0070]

[0071] The objective formula is an unconstrained optimization problem, which is calculated using the Lagrangian function. The following Lagrangian function is constructed:

[0072]

[0073] Among them, the matrix is a diagonal matrix with the i-th diagonal element being eps is a fixed minimum constant value. is the i-th row vector of the matrix, i is the row number, and j is the column number.

[0074] According to formula (12), we can derive the vector b and assign it to zero, and obtain:

[0075]

[0076] Then substitute formula (14) into formula (13) to perform elimination:

[0077]

[0078] According to formula (15), the matrix W is derived and assigned to zero to obtain:

[0079]

[0080] Step 4.3: Fix the sample-label mapping matrix W and the target domain data label F t , update the shared subspace mapping matrix P s ,P t , the objective function is:

[0081]

[0082] Defining the Matrix matrix is a matrix whose diagonal elements are (P s -P t ) is the inverse of the 2-norm of each row vector, then:

[0083] ||P s -P t || 2,1 =Tr(P T MP) (18)

[0084] definition:

[0085]

[0086]

[0087] Substitute equations (14), (18) and (19) into equation (17) to construct the Lagrangian function:

[0088]

[0089] According to formula (20), we can derive the matrix P and assign it to zero to obtain:

[0090] (XHX T ) -1 (T+M)P+P(αWW T )=(XHX T ) -1 (XHFW T ), (twenty one)

[0091] Use the Sylvester equation to solve equation (21).

[0092] Step 4.4: Fix the shared subspace mapping matrix P s ,P t , target domain data label F t and feature label F d , update the sample-feature bipartite graph association matrix S and calculate the new feature label F d , the objective function is:

[0093]

[0094] stG≥0,G1=1.

[0095] Splitting formula (22) yields:

[0096]

[0097] stG≥0,G1=1,

[0098] Solve equation (23) by row-by-row solution:

[0099]

[0100] in is the row vector of the sample label matrix F, is the feature label matrix F d The row vector of , let:

[0101]

[0102] Rewrite equation (24) as:

[0103]

[0104] Solving equation (26), we can get the updated sample-feature association matrix S′, and update the matrix D and matrix L accordingly. Then, by changing λTr(Y T LY) Use the Lagrangian function to transform the feature label matrix F d To update:

[0105] F d =(D2) -1 G T F. (27)

[0106] Step 4.5: Fixed sample-label mapping matrix W and shared subspace mapping matrix P s ,P t As well as the sample-feature bipartite graph association matrix S and feature label F d , update the target domain data label F t , the objective function is:

[0107]

[0108] stF≥0,F1=1.

[0109] Disassembling the expansion matrix F=[Y s ; F t ]as well as

[0110]

[0111] In formula (28), Converts to:

[0112]

[0113] In formula (28) Converts to:

[0114]

[0115] In formula (28), λTr(Y T LY) is converted to:

[0116]

[0117] Based on equations (28), (30), (31) and (32), the Lagrangian function is constructed:

[0118]

[0119] According to formula (33), the matrix F t Take the derivative and set it to zero:

[0120]

[0121] in:

[0122]

[0123] Based on KKT condition F t ⊙Φ=0, we get:

[0124]

[0125] in Represents the matrix Z s / t All negative values ​​in are replaced by 0. Represents the matrix Z s / t All positive values ​​in are replaced by 0 and the absolute value of negative values ​​is taken, that is,

[0126] Step 4.6: Repeat steps 4.2, 4.3, 4.4, and 4.5 multiple times to complete the sample-label mapping matrix W and the shared subspace mapping matrix P. s ,P t , sample-feature bipartite graph association matrix S and target domain data label F t Joint iterative optimization.

[0127] Step 5: Input the sample matrix X obtained in step 2 into the objective function after iterative optimization in step 4 to obtain the corresponding predicted value label, which is the emotional state of the subject corresponding to the sample at the time of collection. The prediction results of this method compared with the prior art are shown in Table 1:

[0128]

[0129] Table 1

[0130] The table compares the model recognition accuracy of this method with four conventional methods and the prior art mentioned in the background technology. Subject means the subject, "1-->2" means the source domain is subject 1, the target domain is subject 2, and so on. The values ​​in the table represent the model recognition accuracy. The accuracy of each group of experiments is retained to four decimal places, and the average accuracy of the methods is retained to six decimal places. It can be seen that the recognition accuracy of this method is significantly better than that of the other five methods.

Claims

1. A method for emotion recognition based on EEG transfer by combining semi-supervised regression and graph label propagation, characterized by: The method specifically comprises the following steps: Step 1: Collect EEG data of multiple subjects in c different emotional states; Step 2: Preprocess and extract features of the EEG data collected in step 1. Each sample matrix X consists of the EEG features of a subject, and the label vector y is the emotion label corresponding to the EEG features in the sample matrix X. Select two sample matrices X from different subjects as the source domain data X. s and target domain data X t ; Step 3: Construct an EEG transfer emotion recognition model that combines semi-supervised regression and graph label propagation, including a dual-mapping domain adaptation model and a semi-supervised regression and graph label propagation joint model; Step 3.1: Establish a dual-map domain adaptation model: Where, dom = s, t, are the mapping matrices that map the source domain data and the target domain data into the shared subspace, respectively. The superscript T represents transposition, p represents the shared subspace dimension, and d represents the original sample dimension. To augment the source domain data labels, Y s is the one-hot encoding of the source domain label, and c is the number of categories; To augment the target domain data labels, represents a column vector whose elements are all 1, F t is the one-hot encoding of the target domain label, n s ,n t are the number of samples of source domain data and target domain data, n=n s +n t Indicates the total number of samples; is a diagonal matrix, the diagonal matrix N dom The kth diagonal element in is the reciprocal of the number of samples of the kth source or target domain data, k = 1, 2, ..., c; is the center matrix, is the unit matrix, represents the square calculation of the Frobenius norm, ||P s -P t || 2,1 Mathematically, it is a row sparse constraint on the difference of the mapping matrix of the shared subspace of the two domains. Its significance is to prevent the excessive reflection of the individual characteristics of the two domains in the mapping process. 2,1 Indicates the calculation of 21 norm; Step 3.2: Establish a joint model of semi-supervised regression and graph label propagation: in, is the sample-label mapping matrix, represents the sample label matrix, Represents X T The average of the difference between PW and F, represents the sample-feature bipartite graph association matrix, represents the sample-feature association matrix, 0 n ,0 p The dimensions are The all-zero matrix, represents the initialized graph incidence matrix constructed based on the current shared subspace data, is the sample-feature association matrix initialized based on the current shared subspace data, represents the sample and feature label matrix, is the graph Laplacian matrix, a diagonal matrix Its diagonal elements are the row sums of the matrix S, Tr(·) represents the trace of the matrix to be solved, and in formula (2) is the standard form of semi-supervised regression, is the row sparseness constraint on the matrix W, It is a bipartite graph construction and label propagation process; Step 3.3: Combine the two models constructed in steps 3.1 and 3.2, integrate the dual-mapping domain adaptation model and the semi-supervised regression and graph label propagation joint model into a unified framework for joint optimization, and calculate the target domain data label F t , the optimization model is: Among them, α, γ, β, and λ are all adjustable parameters in the optimization process; Step 4: Initialize the target domain data label F t And the feature label F d , the principal component analysis method is used to construct the subspace mapping matrix P of the source domain and the target domain s ,P t ; Then, according to the joint optimization objective function obtained in step 3, by fixing the remaining variables and updating another variable, the sample-label mapping matrix W and the subspace mapping matrix P are sequentially s ,P t , sample-feature association matrix S and target domain data label F t Perform optimization and repeat the optimization process multiple times to achieve joint iterative optimization; Step 5: Substitute the result of the joint iterative optimization in step 4 into formula (3), and then use the EEG data feature X of the unknown emotional state as m As the target domain data X t Input formula (3) to solve the corresponding target domain data label F t , get the EEG data feature X m The corresponding emotional state at the time of collection.

2. The method for emotion recognition by combining semi-supervised regression and graph label propagation according to claim 1, characterized in that: In step 2, the collected EEG data are sampled at a frequency of 200 Hz, and then the sampled data are passed through a bandpass filter of 1 Hz to 75 Hz to filter out noise and artifacts; the data are then divided into five frequency bands of 1-4 Hz, 4-8 Hz, 8-14 Hz, 14-31 Hz and 31-50 Hz, and the differential entropy in each frequency band is calculated as the EEG feature in the sample matrix X.

3. The method for emotion recognition by combining semi-supervised regression and graph label propagation according to claim 1, characterized in that: Step 4 specifically includes the following steps: Step 4.1: Initialize the target domain data label F t And the feature label F d The value of And construct the initial shared subspace mapping matrix P by dimensionality reduction s ,P t ; Step 4.2: Fix the target domain data label F t and the shared subspace mapping matrix P s ,P t , update the sample-label mapping matrix W, the objective function is: Solve formula (4) to obtain the updated sample-label mapping matrix W′; Step 4.3: Fix the sample-label mapping matrix W and the target domain data label F t , update the shared subspace mapping matrix P s ,P t The objective function is: Solve formula (5) to obtain the updated shared subspace mapping matrix P s ′,P t ′; Step 4.4: Fix the shared subspace mapping matrix P s ,P t , target domain data label F t and feature label F d , update the sample-feature bipartite graph association matrix S and calculate the new feature label F d , the objective function is: Solve formula (6) to obtain the updated sample-feature bipartite graph association matrix S′ and feature label F d ′; Step 4.5: Fixed sample-label mapping matrix W and shared subspace mapping matrix P s ,P t As well as the sample-feature bipartite graph association matrix S and feature label F d , update the target domain data label F t , the objective function is: Solve formula (7) to obtain the updated target domain data label F t ′; Step 4.6: Repeat steps 4.2, 4.3, 4.4, and 4.5 multiple times to complete the sample-label mapping matrix W and the shared subspace mapping matrix P. s ,P t , sample-feature bipartite graph association matrix S and target domain data label F t Joint iterative optimization.

4. The method for emotion recognition by combining semi-supervised regression and graph label propagation according to claim 3, characterized in that: In steps 4.1, 4.3, and 4.4, the Lagrangian function is used to solve and optimize the sample-label mapping matrix W and the feature label F d , target domain data label F t The objective function of .

5. The method for emotion recognition by combining semi-supervised regression and graph label propagation according to claim 4, characterized in that: In step 4.2, the Lagrangian function and Sylvester equation are used to solve the optimized subspace mapping matrix P s ,P t .

6. The method for emotion recognition by combining semi-supervised regression and graph label propagation according to claim 4, characterized in that: In step 4.3, the objective function of the sample-feature association matrix S is simplified and optimized by solving it row by row, and then the solution is performed.

Images