Semi-supervised classification method fusing local and global features

Abstract

The present disclosure relates to a semi-supervised classification method fusing local and global features, including the steps of preparing a data set, constructing an initial graph using the data set, predicting the initial graph using a local link prediction method and a global link prediction method respectively and performing symmetrization and normalization processing, fusing the processing results to obtain a probability matrix, fusing the labeled information into the constructed probability transition matrix and normalizing, propagating, keeping the first L rows of data in the labeled matrix unchanged after propagation, judging whether the updated labeled matrix converges and completing classification or returning to continue execution according to the judgment result, etc. The classification method of the present disclosure solves the problems in the prior art that the classification method only considers part of the features of the data, cannot be close to the actual situation, the classification is not accurate, the classification efficiency is low, etc.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to a data classification method, in particular, to a semi-supervised classification method fusing local and global features. BACKGROUND

[0002] The existing data classification methods include supervised classification, semi-supervised classification, unsupervised classification and the like. The supervised classification method needs a large number of labeled samples to train the model, which limits its application scenarios; the unsupervised classification does not need the class information of the data and is widely applied, but the classification effect is not good due to the lack of class information. The semi-supervised classification only needs a small amount of labeled data, has low cost, and can obtain good classification effect by learning the data distribution of a large amount of unlabeled data, so it has a wide range of application scenarios.

[0003] The semi-supervised classification method based on graph often obtains good effect and gets extensive attention due to the full use of the relationship between the data samples. In the semi-supervised classification based on graph, a fully connected graph structure is usually constructed first, which represents the relationship between the data, i.e. the similarity matrix of the data, and the matrix can be further converted into a probability transition matrix for label propagation between the data, and then the class information of the labeled data is used for label propagation to obtain the classification result.

[0004] However, in the current semi-supervised classification problem based on graph, only part of the features is considered when constructing the graph, and the internal structure of the data is not accurately expressed due to the incomplete consideration of the features of the data, resulting in that the classification result is not close to the actual situation, the classification effect is not ideal, the accuracy is low, and the classification efficiency is also low. SUMMARY

[0005] In view of the above problems, the present disclosure provides a semi-supervised classification method fusing local and global features, which solves the problems of the existing classification methods that only consider part of the features of the data, the classification is not accurate, and the classification efficiency is low.

[0006] The classification method of the present disclosure fuses the local features and global features of the data when constructing the graph, and the class information of the data is used to calculate the similarity between two data points, which is closer to the actual situation. In addition, the link prediction method is applied to the graph construction, and in the case of considering the basic structure of the graph, new connection relationships are predicted for the very dense graph which will reduce the classification performance, so as to improve the quality of the graph, so that the constructed graph can accurately express the internal structure of the data, thereby achieving better classification effect and improving the classification accuracy and classification efficiency.

[0007] Specifically, the present disclosure provides a semi-supervised classification method fusing local and global features, comprising:

[0008] S100, preparing a data set X, the data set X including labeled data X L and unlabeled data X U Two parts, labeled data X L The label information of the labeled data X is F L The features of the data are described by data attribute information, and L represents the number of labeled data.

[0009] S200, extracting the basic relationship of the data from the data set X prepared in step S100, and constructing an initial graph G0(V, E0), wherein V represents the nodes in the graph, and E0 represents the edges in the graph G0.

[0010] S300, predicting edges that do not exist in the initial graph G0 by a local link prediction method, sorting the predicted edges, and selecting a certain proportion of the predicted edges to join the initial graph G0 to obtain a graph G1(V, E1), wherein V represents the nodes in the graph, and E1 represents the edges in the graph G1.

[0011] S400, symmetrizing the similarity matrix W1 of the graph G1 obtained in step S300, and normalizing the similarity matrix W1 of the symmetrized graph G1.

[0012] S500, predicting edges that do not exist in the initial graph G0 by a global link prediction method, sorting the predicted edges, and selecting a certain proportion of the predicted edges to join the initial graph G0 to obtain a graph G2(V, E2), wherein V represents the nodes in the graph, and E2 represents the edges in the graph G2.

[0013] S600, symmetrizing the similarity matrix W2 of the graph G2 obtained in step S500, and normalizing the similarity matrix W2 of the symmetrized graph G2.

[0014] S700, fusing the graph G3 obtained in step S400 and the graph G4 obtained in step S600 by using the formula P0=αG3+(1-α)G4 to obtain a probability matrix P0(N*N), wherein α is a hyperparameter, is a number between 0 and 1, and N represents the number of nodes.

[0015] S800, constructing a probability transition matrix based on the probability matrix P0 obtained in step S700, fusing the labeled information into the probability transition matrix, and normalizing to obtain a normalized probability transition matrix.

[0016] S900, propagating the label information of the nodes according to the probability defined in the normalized probability transition matrix obtained in step S800.

[0017] S1000, make the first L rows of data in the marked matrix consistent with the first L rows of data in the original marked matrix after propagation, to prevent the marked information from being contaminated;

[0018] S1100, judge the updated calculated marked matrix in step S1000, if the matrix has converged and no longer changes, execute step S1200; otherwise, return to continue executing step S800;

[0019] S1200, obtain a semi-supervised classification result graph G5 that fuses local and global features, so that the unmarked data Xu in the data set X is assigned with corresponding marked information Fu, and the classification is completed.

[0020] Compared with the prior art, the present disclosure has the following beneficial technical effects:

[0021] (1) The characteristics of the data are comprehensively considered, the local and global features of the data are fused during graph construction, and the category information of the data is used to calculate the similarity between two data points, which can be more close to the actual situation;

[0022] (2) The link prediction method is applied to graph construction, the new connection relationship of the very dense graph which will reduce the classification performance is predicted under the consideration of the basic structure of the graph, the quality of the graph is improved, the constructed graph can accurately express the internal structure of the data, so that better classification effect is achieved, and the classification accuracy and classification efficiency are improved. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 The flow chart of the semi-supervised classification method of fusing local and global features of the present disclosure is shown.

[0024] Figure 2 The accuracy rate comparison diagram of the classification methods of the present CN, katz method and the present disclosure on a plurality of different data sets. DETAILED DESCRIPTION

[0025] The specific process of the semi-supervised classification method of fusing local and global features of the present disclosure is described below in combination with the accompanying Figure 1

[0026] In one embodiment, a semi-supervised classification method of fusing local and global features is provided, comprising:

[0027] S100, preparing a data set X, the data set X including marked data X L and unmarked data X U Two parts, the marked information of the marked data X L is F L , the characteristics of the data are described by data attribute information, and L represents the number of marked data; ​

[0028] S200, extracting the basic relationship of the data set X prepared in step S100, constructing an initial graph G0(V, E0), wherein V represents the nodes in the graph, and E0 represents the edges in the initial graph G0;

[0029] S300, predicting the edges not existing in the initial graph G0 by a local link prediction method, sorting the predicted edges, and selecting a certain proportion of the predicted edges to be added to the initial graph G0 to obtain a graph G1(V, E1), wherein V represents the nodes in the graph, and E1 represents the edges in the graph G1;

[0030] S400, symmetrizing the similarity matrix W1 of the graph G1 obtained in step S300, and normalizing the similarity matrix W1 of the symmetrized graph G1;

[0031] S500, predicting the edges not existing in the initial graph G0 by a global link prediction method, sorting the predicted edges, and selecting a certain proportion of the predicted edges to be added to the initial graph G0 to obtain a graph G2(V, E2), wherein V represents the nodes in the graph, and E2 represents the edges in the graph G2;

[0032] S600, symmetrizing the similarity matrix W2 of the graph G2 obtained in step S500, and normalizing the similarity matrix W2 of the symmetrized graph G2;

[0033] S700, fusing the graph G3 obtained in step S400 and the graph G4 obtained in step S600 by using the formula P0=αG3+(1-α)G4 to obtain a probability matrix P0(N*N), wherein α is a hyperparameter, is a number between 0 and 1, and N represents the number of nodes;

[0034] S800, constructing a probability transition matrix based on the probability matrix P0 obtained in step S700, fusing the labeled information into the probability transition matrix, and normalizing to obtain a normalized probability transition matrix;

[0035] S900, propagating the label information of the nodes according to the probability defined in the normalized probability transition matrix obtained in step S800;

[0036] S1000, making the first L rows of data in the propagated label matrix consistent with the first L rows of data in the original label matrix to prevent the labeled information from being contaminated;

[0037] S1100, judging the label matrix updated in step S1000, if the matrix has converged and no longer changes, executing step S1200; otherwise, returning to continue executing step S800;

[0038] S1200, obtaining a semi-supervised classification result graph G5 of fused local and global features, so that the unlabeled data Xu in the data set X is assigned with corresponding labeled information Fu, and the classification is completed.

[0039] In this embodiment, one row of the data set X represents one data, and one column represents an attribute of the data.

[0040] The labeled information polluted in step S1000 refers to that after the processing in step S900, the label of the labeled data may be changed, resulting in an error of the label of the labeled data, which is referred to as label information pollution.

[0041] In this embodiment, the execution steps of the classification method proposed in the present disclosure are described in detail, including preparing a data set, constructing an initial graph using the data set, predicting the initial graph using a local link prediction method and a global link prediction method respectively and performing symmetrization and normalization processing, fusing the processing results to obtain a probability matrix, fusing the labeled information into the constructed probability transition matrix and performing normalization, propagation, keeping the first L rows of the labeled matrix unchanged after the propagation, judging whether the updated labeled matrix converges and completing the classification or returning to continue execution according to the judgment result, and the like. The classification method of the present disclosure considers local and global features, is closer to the actual situation, makes the classification result more accurate, and has higher classification efficiency.

[0042] In a preferred embodiment, in step S200, the initial graph is constructed by using a K-Nearest Neighbor (KNN) algorithm or a minimum spanning tree algorithm.

[0043] The K-Nearest Neighbor (KNN) algorithm and the minimum spanning tree algorithm are both algorithms commonly used in the prior art for classification problems; the K-Nearest Neighbor (KNN) algorithm is to find the k most similar samples in a data set under the premise of determining the similarity measure, and a commonly used formula for calculating the similarity matrix is where exp represents an exponential function, W ij represents the element in the i-th row and the j-th column of W, d(x i , x j ) represents the distance between the nodes x i and x j connected in the graph, and the similarity between the data points x i and x j is calculated using the Euclidean distance as a measure: d(x i , x j ) = ||x i -x j || 2where p is a hyper-parameter, the larger p is, the larger the local influence of the function is, and p is usually set to 3; a minimum spanning tree algorithm, a spanning tree of a connected graph with n nodes is a minimal connected subgraph of the original graph, and contains all n nodes in the original graph and has the least edges that keep the graph connected.

[0044] In this embodiment, the K-Nearest Neighbor (KNN) algorithm and the minimum spanning tree algorithm are selected to extract the basic relationship of the data set X prepared in step S100, and an initial graph G0(V, E0) is constructed. In practice, other common algorithms can also be selected to extract the basic relationship of the data.

[0045] In a preferred embodiment, in step S300, the local link prediction method is the common neighbor (CN) method.

[0046] The common neighbor (CN) method is that when two users have many common neighbors, it is considered that the two users are likely to establish a connection. The similarity of two users is represented by the number of their common neighbors: S xy = |Γ(x)∩Γ(y), where Γ(y) represents the neighbors of node x in graph G0, and Γ(y) represents the neighbors of node y in graph G0, and S xy represents the number of common neighbors between nodes x and y. In this embodiment, the local link prediction method is the common neighbor (CN) method, and the common neighbor (CN) calculation formula (S xy = |Γ(x)∩Γ(y)) is used to calculate the edges that do not exist in graph G0.

[0047] In a preferred embodiment, in steps S300 and S500, a certain proportion is selected according to a proportion of 5%-10%.

[0048] In this embodiment, the selected proportion can be determined according to actual conditions, and it is generally appropriate to select a proportion of 5%-10%, for example, a proportion of 10%.

[0049] In a preferred embodiment, in step S400,

[0050] In step S300, the similarity matrix W1 of graph G1 is calculated, and the calculation formula is a radial basis function:

[0051] where exp represents an exponential function, W1 ij represents the element in the i-th row and the j-th column of W1, d(x i , x j ) represents the connected nodes x iand x j The similarity between data points x i and x j is calculated by using Euclidean distance as a measure: d(x i , x j ) = ||x i -x j ||. 2 σ is a hyperparameter used to adjust the local action range of the radial basis function. The greater the value of σ, the greater the local influence. σ can be 3 or another value according to actual needs.

[0052] The similarity matrix W1 of the graph G1 is symmetrized, and the formula for symmetrization is: where W1 represents the similarity matrix of the graph G1, and has a size of N*N. W1' is the transpose matrix of W1, and has a size of N*N.

[0053] The similarity matrix W1 of the symmetrized graph G1 is normalized, and NW1 represents the normalized matrix of the similarity matrix W1, and has a size of N*N. The calculation formula for normalization is where NW1 ij represents the value of the i-th row and j-th column element in the normalized matrix, W1 ij represents the value of the i-th row and j-th column element in the similarity matrix W1, min(W1 i ) represents the minimum value in the i-th row of the similarity matrix W1, and max(W1 i ) represents the maximum value in the i-th row of the similarity matrix W1.

[0054] After normalization, the data of NW1 is in the range of [0, 1], and the graph G3(V, E3) corresponding to NW1 is obtained, where V represents the nodes in the graph, and E3 represents the edges in the graph G3.

[0055] In a preferred embodiment, the method of global link prediction in step S500 uses Katz algorithm to calculate the edges not existing in G0.

[0056] Katz algorithm is an existing algorithm based on paths of different lengths. It is directly added to the path set and exponentially decays according to the length. In this embodiment, Katz algorithm is selected to calculate the edges not existing in G0. Katz algorithm is an existing algorithm, which will not be described here.

[0057] In another preferred embodiment, the method of global link prediction in step S600 uses the following formula:

[0058] The similarity matrix W2 of the graph G2 obtained in step S500 is calculated, and the calculation formula is a radial basis function:

[0059] Where exp represents the exponential function, W2 ij d(x) represents the element in the i-th row and j-th column of W2. i x j ) indicates that there is a connected node x in G2. i and x j The distance to the data point x is calculated using Euclidean distance as the measure. i and x j Similarity between them: d(x) i x j )=||x i -x j || 2 σ is a hyperparameter used to adjust the local range of action of the radial basis function. The larger the value of σ, the greater the local influence. σ can be 3, or other values ​​can be selected according to actual needs.

[0060] The similarity matrix W2 of graph G2 obtained in step S500 is symmetricized. The symmetry symmetry formula is as follows: Where W2 represents the similarity matrix of graph G2, which is of size N*N, and W2′ is the transpose of W2, which is of size N*N;

[0061] Next, the similarity matrix W2 of the symmetric graph G2 is normalized. Let NW2 represent the normalized matrix of similarity matrix W2, and its size is N*N. The normalization calculation formula is as follows: NW2 ij W2 represents the value of the element in the i-th row and j-th column of the normalized matrix. ij Represents the value of the element in the i-th row and j-th column of the similarity matrix W2, min(W2 i ) represents taking the minimum value in the i-th row of the similarity matrix W2, max(W2) i () represents taking the maximum value in the i-th row of the similarity matrix W2;

[0062] After normalization, the data of NW2 are all in the range of [0, 1], and the graph G4(V, E4) corresponding to NW2 is obtained, where V represents the node in graph G4 and E4 represents the edge in graph G4.

[0063] In a preferred embodiment, step S800 specifically includes:

[0064] F containing labeled information L The dataset X is represented in matrix form, resulting in an initial label matrix F0(N*C), where N represents the number of nodes and C represents the number of feature categories;

[0065] Multiply the initial label matrix F0 by the probability matrix P0 to obtain the label matrix Ft And use the formula Q=P t +βF t *F t The labeled information is fused into the probability transition matrix, where P... t Let P represent the probability transition matrix. t The initial value is P0, t represents the number of iterations, and F t ′ represents F t The transpose of , where β is a hyperparameter with values ​​between (0, 1), and Q represents a matrix that incorporates labeled information but is not normalized;

[0066] According to the formula Normalize Q, where Q ij This represents the element in the i-th row and j-th column of matrix Q. Representing the sum of the elements in the i-th row, we obtain the normalized probability transition matrix P. t+1 .

[0067] In a preferred embodiment, in step S900, the propagation formula is: F t+1 =P t+1 ×F t , where F t+1 This represents the label matrix obtained after the t-th propagation.

[0068] In a preferred embodiment, in step S1000, the formula is used. Make the propagated label matrix F t+1 The first L rows of data in the middle are consistent with the first L rows of data in the original label matrix F0 to prevent the labeled information from being contaminated; among them This represents the first L rows of the initial label matrix F0. F represents t+1 The first L rows in the dataset represent the number of data points in dataset X that already have labeled information.

[0069] In this embodiment, in order to prevent the tags of the marked data from being contaminated after the processing in step S900, it is necessary to reassign values.

[0070] In a preferred embodiment, the marker matrix F calculated in step S1000 is updated. t+1 Make a judgment. If the matrix has converged and no longer changes, proceed to step S1200; otherwise, return to continue to step S800.

[0071] S1200 is the marker matrix F updated and calculated in step S1000. t+1After convergence, a semi-supervised classification result map G5 of fused local and global features is obtained, so that the unlabeled data Xu in the data set X is assigned with corresponding label information Fu, so that the unlabeled data also has corresponding classification label information, and the classification process of the original data set is completed.

[0072] Matrix F after convergence t+1 Each row of data in F represents the category of the data, if F t+1 The element F ij in F is 1, then the category of the i-th data is j, so that the category corresponding to each data can be obtained from the F t+1 matrix, and the classification result of the original data set is obtained.

[0073] If it is judged that F t+1 has not converged, then it is necessary to return to step S800 to continue the classification process.

[0074] The above is a detailed description of the semi-supervised classification method of fusing local and global features of the present disclosure. According to the method, the data is classified, the local and global features are considered, the actual situation is closer, the classification result is more accurate, and the classification efficiency is also higher.

[0075] Experiment:

[0076] In order to verify the advantages of the semi-supervised classification method of fusing local and global features of the present disclosure compared with the classification method of the prior art, an experimental comparison verification is carried out.

[0077] In the experiment, for a plurality of different data sets (such as Figure 2 indicated by the abscissa), the CN+LPA (CN common neighbor+label propagation), katz+LPA (katz+label propagation), and the method FLG+LPA (fusion of local and global features+label propagation) of the present disclosure are used to compare the data classification, and the data set is selected from six commonly used semi-supervised classification test data sets g241c, g241n, USPS, digit1, COIL1, and COIL2. The samples and features of each test data set are shown in Table 1:

[0078] Table 1

[0079] Test dataset Number of samples Number of feature attributes Class g241c 1500 241 2 g241n 1500 241 2 USPS 1500 241 2 digit1 1500 241 2 COIL1 1500 241 2 COIL2 1500 241 2

[0080] Figure 2 The comparison shows the accuracy of the classification results of the above three methods on different data sets, and in the figure, ACC represents the classification accuracy of different methods. From Figure 2 it can be found that the classification result obtained by using the classification method (FLG+LPA) of the present disclosure is better than the prior art CN+LPA or katz+LPA, and the classification accuracy and efficiency are higher.

[0081] Although the embodiments of the present application have been described above with reference to the accompanying drawings, the present application is not limited to the above-described specific embodiments and areas of application, and the above-described specific embodiments are merely illustrative and instructive, but not restrictive. Many modifications can be made by those skilled in the art under the teachings of the present specification and without departing from the scope of the present application as defined by the claims.

Claims

1. A semi-supervised classification method of fusing local and global features executed on a computer, comprising: S100, preparing a data set X, the data set X including labeled data X L and unlabeled data X U Two parts, labeled data X L The label information of the labeled data is F L The features of the data are described by data attribute information, and L represents the number of labeled data; wherein the data set in the step S100 selects one of USPS, digit1, COIL1, and COIL2 data sets; S200, extracting the basic relationship of the data set X prepared in step S100 to construct an initial graph G0(V, E0), wherein V represents the nodes in the graph, and E0 represents the edges in the initial graph G0; S300, predicting the edges that do not exist in the initial graph G0 by a local link prediction method, sorting the predicted edges, and selecting a certain proportion of the predicted edges to join the initial graph G0 to obtain a graph G1(V, E1), wherein V represents the nodes in the graph, and E1 represents the edges in the graph G1; S400, symmetrizing the similarity matrix W1 of the graph G1 obtained in step S300, and normalizing the similarity matrix W1 of the symmetrized graph G1; S500, predicting the edges that do not exist in the initial graph G0 by a global link prediction method, sorting the predicted edges, and selecting a certain proportion of the predicted edges to join the initial graph G0 to obtain a graph G2(V, E2), wherein V represents the nodes in the graph, and E2 represents the edges in the graph G2; S600, symmetrizing the similarity matrix W2 of the graph G2 obtained in step S500, and normalizing the similarity matrix W2 of the symmetrized graph G2; S700, fusing the graph G3 obtained in step S400 and the graph G4 obtained in step S600 by the formula P0=αG3+(1-α)G4 to obtain a probability matrix P0(N*N), wherein α is a hyperparameter, is a number between 0 and 1, and N represents the number of nodes; S800, constructing a probability transition matrix based on the probability matrix P0 obtained in step S700, fusing the labeled information into the probability transition matrix, and normalizing to obtain a normalized probability transition matrix; S900, propagating the label information of the nodes according to the probability defined in the normalized probability transition matrix obtained in step S800; S1000, making the first L rows of data in the propagated label matrix consistent with the first L rows of data in the original label matrix to prevent the labeled information from being contaminated; S1100, judging the label matrix updated in step S1000, if the matrix has converged and no longer changes, executing step S1200; otherwise, returning to continue executing step S800; S1200, obtaining a semi-supervised classification result graph G5 of fusing local and global features, so that the unlabeled data Xu in the data set X is assigned with corresponding label information Fu, and the classification is completed; The method can improve the accuracy and efficiency of classification. 2.The method of claim 1, wherein in step S200, the initial graph is constructed by a K-Nearest Neighbor (KNN) algorithm or a minimum spanning tree algorithm. 3.The method of claim 1, wherein in step S300, the local link prediction method is a common neighbor (CN) method.

4. The method of claim 1, wherein the certain proportion in steps S300 and S500 is 5%-10%.

5. The method of claim 1, wherein in step S400, A similarity matrix W1 is calculated for the graph G1 obtained in step S300, and the calculation formula is a radial basis function: where exp represents an exponential function, represents an element in the i-th row and the j-th column of W1, represents a distance between nodes connected in G1 and , and the similarity between data points and is calculated using the Euclidean distance as a measure: , is a hyperparameter used to adjust the local range of the radial basis function, The greater the value, the greater the local influence. The similarity matrix W1 of the graph G1 is symmetrized, and the formula of the symmetrization is: wherein W1 represents the similarity matrix of the graph G1, with a size of N*N, is the transpose matrix of W1, with a size of N*N; The similarity matrix W1 of the symmetrization-processed graph G1 is normalized, and NW1 represents the normalized matrix of the similarity matrix W1, which has a size of N*N. The calculation formula of the normalization processing is wherein NW1 ij represents the value of the i-th row and j-th column element in the normalized matrix, W1 ij represents the value of the i-th row and j-th column element in the similarity matrix W1, min(W1 i ) represents the minimum value in the i-th row of the similarity matrix W1, and max(W1 i ) represents the maximum value in the i-th row of the similarity matrix W1. After normalization, the data of NW1 are all in the range of [0, 1], and the graph G3(V, E3) corresponding to NW1 is obtained, wherein V represents a node in the graph, and E3 represents an edge in the graph G3.

6. The method of claim 1, wherein in step S500, the method of global link prediction is to calculate the edges not existing in G0 by using Katz algorithm.

7. The method of claim 1, wherein in step S600, A similarity matrix W2 is calculated for the graph G2 obtained in step S500, and the calculation formula is a radial basis function: where exp represents an exponential function, represents an element in the i-th row and the j-th column of W2, represents the distance between nodes connected in G2 and The similarity between data points and is calculated using the Euclidean distance as a measure: , is a hyperparameter used to adjust the local range of the radial basis function, The greater the value, the greater the local influence. The similarity matrix W2 of the graph G2 is symmetrized, and the symmetrization formula is where W2 represents the similarity matrix of the graph G2, with a size of N*N, is the transpose matrix of W2, with a size of N*N; The similarity matrix W2 of the symmetrization-processed graph G2 is normalized, and NW2 represents the normalized matrix of the similarity matrix W2, which has a size of N*N. The calculation formula of the normalization processing is wherein NW2 ij represents the value of the i-th row and j-th column element in the normalized matrix, W2 ij represents the value of the i-th row and j-th column element in the similarity matrix W2, represents the minimum value in the i-th row of the similarity matrix W2, max(W2 i ) represents the maximum value in the i-th row of the similarity matrix W2. After normalization, the data of NW2 are all in the range of [0, 1], and the graph G4(V, E4) corresponding to NW2 is obtained, wherein V represents a node in the graph G4, and E4 represents an edge in the graph G4.

8. The method of claim 1, wherein step S800 is specifically: The data set X containing the labeled information F L is represented in the form of a matrix to obtain an initial labeled matrix F0(N*C), wherein N represents the number of nodes, and C represents the number of feature categories; F0 is multiplied by the probability matrix P0 to obtain the label matrix F t and the formula is used to fuse the labeled information into the probability transition matrix, where P t represents the probability transition matrix, P t has an initial value of P0, and t represents the number of iterations, represents the transpose matrix of F t , is a hyperparameter and takes a value between 0 and 1, and Q represents a matrix with fused labeled information but not normalized. According to the formula Q is normalized, where denotes the element in the i-th row and j-th column of the matrix Q, denotes the sum of the elements in the i-th row, resulting in the normalized probability transition matrix P t+1 .

9. The method of claim 8, wherein the propagation formula in step S900 is: where F t+1 represents the label matrix obtained after the tth propagation.

10. The method of claim 9, wherein in the step S1000, the formula is used The marked matrix F after propagation is made to be t+1 The first L rows of data are consistent with the first L rows of data in the original marked matrix F0, preventing the marked information from being contaminated; wherein The first L rows of data in the initial marked matrix F0 are represented as F0(L), The first L rows of data in the marked matrix F are represented as F(L), t+1 L represents the number of data in the data set X that initially have marked information.

Images