Multi-view clustering method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention provides a multi-view clustering method and device, electronic equipment and a storage medium, and the multi-view clustering method comprises the steps: carrying out the preprocessing of original data of a plurality of views, and obtaining the normalized data of the plurality of views; manifold learning processing is carried out on the normalized data of the multiple views, and a spectrum embedding feature matrix after fusion of the multiple views is obtained; respectively carrying out fuzzy clustering processing on the normalized data of the plurality of views to obtain respective spectral approximation feature matrixes of the plurality of views; and carrying out fusion clustering on the spectrum embedded feature matrixes of the plurality of views and all the spectrum approximate feature matrixes to obtain a clustering result of the plurality of views. Therefore, the purpose of improving the clustering efficiency and accuracy is achieved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a multi-view clustering method, device, electronic device and storage medium. Background Art

[0002] Multi-view data clustering is an important image processing technology in the field of machine learning and pattern recognition. Its function is to cluster similar data samples into clusters to reveal the intrinsic structure of the data. Traditional multi-view data clustering methods generally include K-means algorithm or spectral clustering algorithm, but these clustering methods are often difficult to effectively process high-dimensional, nonlinear and heterogeneous multimodal data. The main reasons are as follows:

[0003] First, due to the presence of a large amount of noise and redundant features in high-dimensional multimodal data, traditional multi-view data clustering methods find it difficult to extract low-dimensional features with physical significance from high-dimensional data, resulting in poor robustness of data clustering results.

[0004] Secondly, due to the complex nonlinear relationship between multimodal data, the traditional multi-view data clustering method mainly measures the similarity between data based on Euclidean distance. This measurement method cannot accurately describe the similarity between data, resulting in poor accuracy of data clustering results.

[0005] Thirdly, due to the heterogeneity of multimodal data, the importance of different feature subspaces of multimodal data is different, and traditional multi-view data clustering methods cannot adaptively fuse multimodal features, resulting in poor data clustering efficiency.

[0006] Therefore, there is an urgent need to provide a solution to improve the accuracy and efficiency of multi-view data clustering. Summary of the invention

[0007] In order to solve the technical problems of low clustering result accuracy and low clustering efficiency in traditional clustering algorithms for clustering high-dimensional, nonlinear and heterogeneous multi-view data, the purpose of the present invention is to provide a multi-view clustering method, device, electronic device and storage medium. The technical solution adopted in the embodiment of the present invention is as follows:

[0008] A first aspect of an embodiment of the present invention provides a multi-view clustering method, comprising:

[0009] Preprocessing the original data of the multiple views to obtain normalized data of the multiple views;

[0010] Performing manifold learning processing on the normalized data of the multiple views to obtain a spectral embedding feature matrix after the multiple views are fused;

[0011] Performing fuzzy clustering processing on the normalized data of the multiple views respectively to obtain spectral approximation feature matrices of the multiple views;

[0012] The spectral embedding feature matrices and all spectral approximation feature matrices of the multiple views are fused and clustered to obtain clustering results of the multiple views.

[0013] A second aspect of an embodiment of the present invention provides a multi-view clustering device, including:

[0014] A preprocessing module is configured to: preprocess the original data of the multiple views to obtain normalized data of the multiple views;

[0015] A manifold processing module is configured to: perform manifold learning processing on the normalized data of the multiple views to obtain a spectral embedding feature matrix after the multiple views are fused;

[0016] A fuzzy processing module is configured to: perform fuzzy clustering processing on the normalized data of the multiple views respectively to obtain spectral approximation feature matrices of the multiple views respectively;

[0017] The fusion processing module is configured to: fuse and cluster the spectrum embedding feature matrices and all spectrum approximation feature matrices of the multiple views to obtain clustering results of the multiple views.

[0018] According to a third aspect of an embodiment of the present invention, an electronic device is provided, comprising a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor can execute the machine executable instructions to implement the multi-view clustering method provided in the first aspect above.

[0019] According to a fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the multi-view clustering method provided in the first aspect is implemented.

[0020] The multi-view clustering method, device, electronic device and storage medium provided by the embodiments of the present invention, on the one hand, by performing manifold learning processing on the normalized data of multiple views, the obtained spectral embedding feature matrix can have the low-dimensional manifold structure of the multiple view data, that is, the extraction of the low-dimensional manifold structure of the view data can be achieved, which can enhance the robustness of clustering; on the other hand, by performing fuzzy clustering processing on the normalized data of multiple views, the obtained spectral approximation feature matrix can be used to establish the fuzzy similarity relationship between the data, which can better characterize the intrinsic structure of the data. On the other hand, by performing fusion clustering processing on the spectral embedding feature matrix obtained by manifold learning processing and all spectral approximation feature matrices obtained by fuzzy clustering processing, the synergistic effect of the two technologies of manifold learning processing and fuzzy clustering processing can be fully utilized, effectively improving the efficiency and accuracy of obtaining clustering results, and can improve the clustering performance of high-dimensional, nonlinear and heterogeneous multi-view data.

[0021] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0023] Figure 1 A structural block diagram of an electronic device provided by an embodiment of the present invention is shown;

[0024] Figure 2 A flowchart of a multi-view clustering method provided by an embodiment of the present invention is shown;

[0025] Figure 3 A schematic diagram of a process of extracting normalized data based on multiple views to obtain a spectral embedding feature matrix provided by an embodiment of the present invention is shown;

[0026] Figure 4 A schematic diagram of a process of obtaining a clustering result by clustering normalized data based on multiple views provided by an embodiment of the present invention is shown;

[0027] Figure 5 A schematic diagram showing a clustering result obtained by clustering using a multi-view clustering method provided by an embodiment of the present invention;

[0028] Figure 6 A functional module diagram of a multi-view clustering device provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0029] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0030] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0031] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0032] In order to solve the technical problems of low clustering result accuracy and low clustering efficiency in the clustering processing of high-dimensional, nonlinear and heterogeneous multi-view data by traditional clustering algorithms, the present invention provides a multi-view clustering method. On the one hand, by performing manifold learning processing on the normalized data of multiple views, the obtained spectral embedding feature matrix can have the low-dimensional manifold structure of the multiple view data, that is, the extraction of the low-dimensional manifold structure of the view data can be realized, which can enhance the robustness of clustering; on the other hand, by performing fuzzy clustering processing on the normalized data of multiple views, the obtained spectral approximation feature matrix can be used to establish the fuzzy similarity relationship between the data, which can better characterize the internal structure of the data. On the other hand, by performing fusion clustering processing on the spectral embedding feature matrix obtained by manifold learning processing and all spectral approximation feature matrices obtained by fuzzy clustering processing, the synergistic effect of the two technologies of manifold learning processing and fuzzy clustering processing can be fully utilized, effectively improving the efficiency and accuracy of obtaining clustering results, and improving the clustering performance of high-dimensional, nonlinear and heterogeneous multi-view data.

[0033] The multi-view clustering method provided by the present invention can be applied to electronic devices. Figure 1 , is a block diagram of the structure of an electronic device. The electronic device 100 includes a memory 110, a processor 120 and a communication module 130. The memory 110, the processor 120 and the communication module 130 are electrically connected to each other directly or indirectly to achieve data transmission or interaction. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0034] The memory is used to store programs or data. The memory may be, but is not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc.

[0035] The processor is used to read / write data or programs stored in the memory and execute corresponding functions.

[0036] The communication module is used to establish a communication connection between the electronic device and other communication terminals through a network, and is used to send and receive data through the network.

[0037] It should be understood that Figure 1 The structure shown is only a schematic diagram of the structure of the electronic device, and the electronic device may also include Figure 1 More or fewer components as shown, or with Figure 1 Different configurations are shown. Figure 1 Each component shown in the figure can be implemented by hardware, software or a combination thereof.

[0038] The following combination Figure 2 The multi-view clustering method provided by the embodiment of the present invention is described. Figure 2 is a flow chart of a multi-view clustering method provided by an embodiment of the present invention, the multi-view clustering method comprising:

[0039] In step S100, the original data of the multiple views are preprocessed to obtain normalized data of the multiple views;

[0040] In step S200, a manifold learning process is performed on the normalized data of the multiple views to obtain a spectral embedding feature matrix after the multiple views are fused;

[0041] In step S300, fuzzy clustering processing is performed on the normalized data of the multiple views respectively to obtain spectral approximation feature matrices of the multiple views respectively;

[0042] In step S400, the spectrum embedding feature matrices and all spectrum approximation feature matrices of the multiple views are fused and clustered to obtain clustering results of the multiple views.

[0043] When it is necessary to perform clustering processing on batch view data, in order to improve the efficiency and accuracy of clustering, the multi-view clustering method provided by the embodiment of the present invention can be used. First, the original data of the multiple views that need to be clustered can be input into the execution body for executing the multi-view clustering method provided by the embodiment of the present invention, and then the multi-view clustering method provided by the embodiment of the present invention is started.

[0044] In the operation process of the multi-view clustering method provided by the embodiment of the present invention, firstly, step S100 is used to pre-process the original data of multiple views to obtain normalized data of multiple views. The principle of the pre-processing technology can be referred to the related art, which will not be described in detail in the embodiment of the present invention. For the convenience of the following description, the normalized data of multiple views can be recorded as Where V represents the number of views, X (v) represents the vth view, represents the real number domain, N represents the number of data points, D (v) Indicates the number of feature dimensions of the v-th view.

[0045] After obtaining normalized data of multiple views through step S100, step S200 and step S300 may be executed serially or in parallel. It should be noted that when step S200 and step S300 are executed serially, there is no strict order between step S200 and step S300.

[0046] In the process of executing step S200, the normalized data of multiple views will be processed by manifold learning processing technology to obtain the spectral embedding feature matrix after the fusion of multiple views. It should be understood that the spectral embedding feature matrix is ​​usually a low-dimensional spectral embedding feature matrix, which is also the case in the embodiment of the present invention. The popular learning processing technology can adopt any one of the related technologies, but in order to further simplify the algorithm processing flow and improve the clustering efficiency, in some embodiments, the manifold learning processing technology can adopt random sampling technology or K-means algorithm technology. In this regard, the embodiment of the present invention also proposes a scheme for obtaining the spectral embedding feature matrix, that is, in the above step S200, the normalized data of the multiple views are subjected to manifold learning processing to obtain the spectral embedding feature matrix after the fusion of the multiple views, which may include:

[0047] In step S210, for each view, a set number of anchor points are determined for the view according to the normalized data of the view;

[0048] In step S220, for each view, a bipartite graph matrix of the view is determined according to the set number of anchor points of the view;

[0049] In step S230, a cross-view anchor graph including all bipartite graph matrices is constructed according to the bipartite graph matrices of the multiple views;

[0050] In step S240, the spectral embedding feature matrix is ​​obtained by processing according to the cross-view anchor map.

[0051] Thus, through step S210, a set number of anchor points, such as R anchor points, can be configured for each view according to actual scene requirements or clustering requirements. The value of R can be determined according to the number of data points N of the view. As an example, R can be taken from the range of [N / 2, N), but is not limited thereto. In addition, since the multiple views have been preprocessed through step S100 to obtain normalized data of the multiple views, that is, the data of the multiple views have been normalized, the number of data points contained in each view is the same. Therefore, in order to ensure the sampling consistency of the multiple views, in step S210, the number of anchor points set for each view is the same.

[0052] Next, step S220 may be performed to further process each view. A random sampling algorithm or a K-means algorithm may be used to search for R anchor points for each normalized view. Based on this, the set of R anchor points selected for the vth view of the multiple views may be expressed as 1≤v≤V, where V represents the total number of views of the multiple views, represents the i-th anchor point of the v-th view. Then, the bipartite graph matrix of the v-th view can be calculated Wherein, R represents the number of anchor points, and N represents the number of data points in the vth view. The calculation principle of the bipartite graph matrix can be referred to in the related art.

[0053] After obtaining the bipartite graph matrices of all views through step S220, step S230 may be performed to construct corresponding cross-view anchor graphs according to all bipartite graph matrices. As an example, V bipartite graph matrices may be constructed according to formula (1) to obtain a cross-view anchor graph W.

[0054]

[0055]

[0056] In formula (1), V has the same meaning as described above, that is, it represents the total number of views of the plurality of views; Z (1) , …, Z (V) A bipartite graph matrix representing V views.

[0057] After obtaining the cross-view anchor graph through step S230, step S240 may be performed to obtain a spectral embedding feature matrix according to the cross-view anchor graph, wherein the spectral embedding feature matrix may be obtained according to any implementation method in the related art according to the cross-view anchor graph. In order to further improve clustering efficiency and accuracy, in some embodiments, the multi-view clustering method provided in the embodiment of the present invention further proposes an implementation scheme for obtaining a spectral embedding feature matrix according to the cross-view anchor graph, that is, in the above step S240, the step of obtaining the spectral embedding feature matrix according to the cross-view anchor graph may include:

[0058] In step S241, a normalized Laplacian matrix of the cross-view anchor graph is calculated;

[0059] In step S242, the spectral embedding feature matrix is ​​calculated according to the spectral embedding matrix of the cross-view anchor graph and the normalized Laplacian matrix.

[0060] In step S241, the normalized Laplacian matrix L of the cross-view anchor graph can be calculated by formula (2):

[0061]

[0062] In formula (2), Δ represents a diagonal matrix, and its i-th diagonal element is I denotes the identity matrix and W denotes the cross-view anchor map.

[0063] Since the matrix reflects the low-dimensional manifold structure of the multi-view data It consists of the first K small eigenvectors of the Laplacian matrix L, so It can be regarded as the concatenation matrix of the spectral embedding matrix of the data point and the spectral embedding matrix of the anchor point, that is:

[0064]

[0065] In formula (3), is the spectral embedding matrix of the cross-view anchor graph, Represented as the spectral embedding matrix of the data point, the matrix It consists of the first K small eigenvectors of the Laplace matrix L. It is calculated Then, according to formula (3), we can obtain The spectral embedding matrix block corresponding to the data point in That is, the spectrum embedding feature matrix H required in step S240 and step S242. Figure 3 As shown, Figure 3 It is a schematic diagram of a process of extracting normalized data based on multiple views to obtain a spectral embedding feature matrix provided by an embodiment of the present invention.

[0066] In the process of executing step S300, the normalized data of multiple views will be processed separately by using fuzzy clustering processing technology to obtain the spectral approximation feature matrix of each view. This can accurately capture the fuzzy correlation of the data under multiple views, and enable the constructed spectral approximation feature matrix to carefully describe the relationship between the data. Relevant technologies can be used to implement the fuzzy clustering processing in step S300. However, in order to ensure clustering efficiency and accuracy, in some embodiments, the multi-view clustering method provided by the embodiment of the present invention also provides a corresponding fuzzy clustering processing scheme, that is, in the above step S300, the normalized data of the multiple views are fuzzy clustered separately to obtain the spectral approximation feature matrix of the multiple views, which can include:

[0067] In step S310, for each view, clustering processing is performed on the normalized data of the view by using a fuzzy clustering algorithm to obtain a membership matrix of the view;

[0068] In step S320, for each view, a fuzzy similarity matrix of the view is constructed according to the membership matrix of the view; the fuzzy similarity matrix is ​​a diagonal symmetric matrix in which all elements on the diagonal are 1;

[0069] In step S330, for each view, a spectral approximation feature matrix of the view is calculated according to the fuzzy similarity matrix of the view.

[0070] In the process of performing fuzzy clustering processing on each view, step S310 can be used to perform clustering operations on the normalized data of V views using a fuzzy clustering algorithm to obtain the membership matrix of each view. For example, for the vth view, its membership matrix can be obtained as

[0071] After obtaining the membership matrix of each view, step S320 may be performed to construct a fuzzy similarity matrix for each view. For example, for the vth view, its fuzzy similarity matrix may be obtained as The constructed A (v) Satisfy Diag(A (v) 1 N )=I N and That is, the fuzzy similarity matrix is ​​a diagonal symmetric matrix whose elements on the diagonal are all 1, where I N is the N×N dimensional identity matrix.

[0072] After obtaining the fuzzy similarity matrix of each view, step S330 can be performed to calculate the spectral approximation feature matrix of each view according to the fuzzy similarity matrix of each view. For example, for the vth view, its fuzzy similarity matrix A can be obtained by solving (v) The spectral approximation characteristic matrix of The solution principle can be found in the related technology.

[0073] After obtaining the spectral embedding feature matrix that fuses the features of multiple views and the spectral approximation feature matrix of all views through step S200 and step S300, step S400 can be executed to cluster the spectral embedding feature matrix and all spectral approximation feature matrices to obtain clustering results of all views. In this regard, in order to improve clustering efficiency and accuracy, in some embodiments, the embodiments of the present invention provide a scheme for clustering the spectral embedding feature matrix and all spectral approximation feature matrices, that is, in the above step S400, the step of fusing and clustering the spectral embedding feature matrices of the multiple views and all spectral approximation feature matrices to obtain the clustering results of the multiple views may include:

[0074] In step S410, a spectral approximation feature set formed by the spectral embedding feature matrix and all spectral approximation feature matrices is constructed;

[0075] In step S420, a consensus matrix, a feature matrix for implementing Laplace rank constraints, and a kinship matrix are determined according to the spectral approximation feature set;

[0076] In step S430, the kinship matrix is ​​updated according to the current consensus matrix and the current feature matrix to obtain the current kinship matrix;

[0077] In step S440, when the number of updates corresponding to the secondary kinship matrix reaches a set number threshold, the secondary kinship matrix is ​​used as the clustering result;

[0078] In step S450, when the number of updates corresponding to the secondary kinship matrix does not reach the set number threshold, the current consensus matrix and the current feature matrix are updated respectively, and the step of updating the kinship matrix according to the current consensus matrix and the current feature matrix is ​​returned to execute, that is, the above step S430 is returned to execute.

[0079] In the process of fusing and clustering the spectral embedding feature matrices and all spectral approximation feature matrices of the multiple views, for the convenience of calculation, the spectral embedding feature matrices and all spectral approximation feature matrices can be first stored as a set in step S410, and this set is the spectral approximation feature set. It can be seen that the spectral approximation feature set has a total of V+1 sub-matrices, and one sub-matrix is ​​one view. Based on this, the spectral approximation feature set can be expressed as {S (1) ,…,S (V) ,S (V+1) =H}.

[0080] After obtaining the spectral approximation feature set, step S420 can be executed to obtain parameters that affect the clustering results, including but not limited to: consensus matrix, feature matrix and kinship matrix. In addition, since the determination of the number of updates of the kinship matrix is ​​involved in steps S440 and S450, the counter iter used to calculate the number of updates can also be initialized to 0 in step S420. When the consensus matrix, feature matrix and kinship matrix are obtained for the first time, they can be obtained by initialization. As an example, the consensus matrix F can be initialized as The feature matrix Q is initialized to Q = F, and the kinship matrix Y is initialized to a matrix that satisfies the following conditions:

[0081] Y T 1 N =1 N ,0≤Y i,j ≤1,Y i,i =0

[0082] Among them, 1 N represents a scalar matrix where all elements on the diagonal are 1 and all elements other than the diagonal are 0; i,i represents the element in the i-th row and i-th column of the kinship matrix Y, Y i,j Represents the element in the i-th row and j-th column of the kinship matrix Y.

[0083] After obtaining the consensus matrix, the feature matrix and the kinship matrix through step S420, step S430 may be executed to update the kinship matrix to obtain the current kinship matrix. In this regard, in some embodiments, the embodiments of the present invention further provide a scheme for updating the kinship matrix, that is, in the above step S430, the step of updating the kinship matrix according to the current consensus matrix and the current feature matrix includes:

[0084] In step S431, a first parameter coefficient set for the current feature matrix is ​​obtained;

[0085] In step S432, a first auxiliary matrix is ​​calculated according to the current consensus matrix, the current feature matrix and the first parameter coefficients;

[0086] In step S433, the value of each column in the kinship matrix is ​​updated according to the first auxiliary matrix to obtain the current kinship matrix.

[0087] In the process of updating the kinship matrix, step S431 can be performed first to obtain the first parameter coefficient set for the feature matrix. When the first parameter coefficient is obtained for the first time, the first parameter coefficient β can be initialized in the above step S420; in addition, in order to balance the consensus matrix and the eigenvector difference matrix, the second parameter coefficient can also be obtained. Similarly, when the second parameter coefficient is obtained for the first time, the second parameter coefficient λ can be initialized together in the above step S420. The initial values ​​of the first parameter coefficient β and the second parameter coefficient λ can be determined based on experiments or experience, and the embodiment of the present invention does not limit this. During initialization, the first parameter coefficient β can be equal to the second parameter coefficient λ.

[0088] Next, step S432 may be performed to calculate a first auxiliary matrix using the first parameter coefficient β, the second parameter coefficient λ, the current consensus matrix, and the current feature matrix. In some examples, the first auxiliary matrix P may be calculated using the following formula:

[0089]

[0090] In formula (4), F represents the current consensus matrix, D Q represents the eigenvector difference matrix, which is calculated by the characteristic matrix Q, that is: [D Q ] i,j =||Q i -Q j || 2 .

[0091] After obtaining the first auxiliary matrix P, step S433 may be performed to update the kinship matrix Y using the first auxiliary matrix P. Since each column Y of the kinship matrix Y :,j They are independent of each other, so the value of each column of the kinship matrix Y can be calculated separately by the following first objective function:

[0092]

[0093] Among them, P :,j represents the jth column of the first auxiliary matrix P, and the secondary kinship matrix Y obtained by updating the above first objective function still needs to meet the following conditions:

[0094] Y :,j T 1 N =1,0≤Y i,j ≤1,Y i,i =0

[0095] Thus, the updated kinship matrix of the current time can be obtained through the above process. At this time, the value of the counter can be updated to iter=iter+1.

[0096] If the updated iter is greater than the maximum number of iterations, it means that the number of updates corresponding to the current kinship matrix has reached the set number threshold and the clustering accuracy requirement has been met. At this time, step S440 will be executed to use the current kinship matrix as the clustering result and output it. The relevant image processing process can be seen in Figure 4 , Figure 4 It is a schematic diagram of a process of obtaining clustering results by clustering normalized data based on multiple views provided by an embodiment of the present invention. On the contrary, if the updated iter is not greater than the maximum number of iterations, it means that when the number of updates corresponding to the secondary kinship matrix does not reach the set number threshold, the clustering accuracy requirement has not been met, so the secondary kinship matrix will continue to be iteratively updated. At this time, it is necessary to first update the current consensus matrix and the current feature matrix respectively, and then return to execute the above step S430 to update the kinship matrix again according to the updated consensus matrix and feature matrix. It can be seen that in the process of re-updating the kinship matrix, the consensus matrix and feature matrix used in the update should also be updated, so that the kinship matrix will have the effect of updating.

[0097] Therefore, before returning to execute step S430, the consensus matrix and the feature matrix are updated first. To this end, the embodiment of the present invention further proposes an update scheme for the consensus matrix and the feature matrix, as follows:

[0098] Consensus matrix update plan:

[0099] That is, in step S450, the process of updating the current consensus matrix may include:

[0100] In step S4511, the error matrix and view weight of each submatrix in the spectral approximation feature set are obtained;

[0101] In step S4512, for each submatrix in the spectral approximation feature set, a second auxiliary matrix corresponding to the submatrix is ​​calculated according to the current consensus matrix, the error matrix of the submatrix, and the view weight;

[0102] In step S4513, for each submatrix in the spectral approximation feature set, an error matrix of the submatrix is ​​updated according to the second auxiliary matrix of the submatrix;

[0103] In step S4514, for each submatrix in the spectral approximation feature set, the view weight of the submatrix is ​​updated according to the current consensus matrix, the submatrix and the error matrix after the submatrix is ​​updated;

[0104] In step S4515, for each submatrix in the spectral approximation feature set, an auxiliary variable for measuring the contribution of the submatrix to the current consensus matrix is ​​calculated according to the current consensus matrix and the updated error matrix and view weight of the submatrix;

[0105] In step S4516, a third auxiliary matrix is ​​calculated based on the spectral approximation feature set, the auxiliary variables of all sub-matrices and the updated error matrix, and the secondary kinship matrix;

[0106] In step S4517, the current consensus matrix is ​​updated according to the third auxiliary matrix.

[0107] In the process of first updating the consensus matrix, step S4511 realizes the acquisition of the error matrix and view weights, which are obtained by initialization, that is, the error matrix and view weights initially set for each submatrix in the spectral approximation feature set are obtained. The error matrix and view weights can be initialized together in step S420. For example, for the vth view, its error matrix ∑ (v) Can be initialized to Among them, I K represents the identity matrix of K×K dimensions, and its view weight ω (v) Can be initialized to It can be seen that the initialization error matrix of each view can be the same, and the initialization view weight of each view can also be the same. In the process of updating the consensus matrix for a non-first time, in order to ensure the accuracy of the updated consensus matrix, the error matrix and view weight obtained in step S4511 are the latest error matrix and view weight. For example, the error matrix and view weight obtained by respectively updating the consensus matrix in steps S4513 and S4514 can be obtained in the process of updating the consensus matrix last time.

[0108] After obtaining the error matrix and view weight of each submatrix, step S4512 may be performed to calculate the second auxiliary matrix corresponding to each submatrix. In some examples, the second auxiliary matrix of each submatrix may be calculated by the following formula:

[0109]

[0110] In formula (5), G (v) The second auxiliary matrix representing the v-th view (i.e., the v-th submatrix), ω (v) represents the view weight of the v-th sub-matrix, S (v) represents the vth sub-matrix, and F represents the current consensus matrix.

[0111] Next, step S4513 is performed to update the corresponding error matrix using the second auxiliary matrix of each sub-matrix. The error matrix of each sub-matrix can be updated by the following formula:

[0112]

[0113] Formula (6) is a formula for updating the value of the i-th diagonal element of the error matrix, where: represents the value of the i-th diagonal element in the error matrix of the v-th submatrix, The second auxiliary matrix G represents the vth submatrix (v) The element in the kth row and kth column of , similarly, The second auxiliary matrix G represents the vth submatrix (v) The element in the i-th row and i-th column of .

[0114] After the error matrix of each sub-matrix is ​​updated, step S4514 may be performed to update the view weight of each sub-matrix. The view weight of each sub-matrix may be updated by the following formula:

[0115]

[0116] In formula (7), ω (v) represents the updated view weight of v sub-matrices, p represents the parameter of the norm, α∈(0,+∞) represents the blur coefficient, and Among them, F represents the current consensus matrix, ∑ (v) is the error matrix after the vth sub-matrix is ​​updated, S (v) represents the vth sub-matrix.

[0117] After obtaining the updated error matrix and view weight of each submatrix, step S4515 may be performed to calculate an auxiliary variable used to measure the contribution of the submatrix to the current consensus matrix. The auxiliary variable of each submatrix may be calculated by the following formula:

[0118]

[0119] In formula (8), θ (v) is the auxiliary variable of the vth submatrix, F represents the current consensus matrix, p represents the parameter of the norm, α∈(0,+∞) represents the fuzzy coefficient, ∑ (v) represents the error matrix after the vth sub-matrix is ​​updated, S (v) represents the vth sub-matrix, ω (v) Represents the updated view weight of the vth sub-matrix.

[0120] After the auxiliary variables of each sub-matrix are calculated by formula (8), step S4516 can be executed to calculate the third auxiliary matrix. The third auxiliary matrix T can be calculated by the following formula:

[0121]

[0122] In formula (9), Y represents the current kinship matrix, θ (v) ,∑ (v) and S (v) The meaning of can be found in the relevant records above and will not be elaborated here.

[0123] Next, step S4517 is performed to update the consensus matrix F using the third auxiliary matrix T, wherein the consensus matrix F can be calculated by maximizing the following second objective function:

[0124]

[0125] F T =1 K

[0126] It can be seen that by using the third auxiliary matrix T and the constraint condition stFF T =1 K , the updated consensus matrix can be calculated by maximizing the second objective function, and the updated consensus matrix is ​​composed of the first K largest eigenvectors of the third auxiliary matrix T. Among them, the function Tr() is a trace function, which is used to calculate the trace of the matrix.

[0127] Update scheme of feature matrix:

[0128] That is, in step S450, the process of updating the feature matrix may include:

[0129] In step S4521, a Laplace matrix is ​​calculated based on the current kinship matrix;

[0130] In step S4522, an updated characteristic matrix is ​​obtained according to the Laplacian matrix.

[0131] In the process of updating the feature matrix, step S4521 can be executed to calculate the corresponding Laplace matrix L according to the current kinship matrix Y. Y , the calculation formula is as follows:

[0132]

[0133] Next, execute step S4525 to calculate the Laplace matrix L Y The updated feature matrix Q is calculated, where the feature matrix Q can be calculated by minimizing the following third objective function:

[0134]

[0135] QQ T =1 K

[0136] It can be seen that by following the Laplace matrix L Y and the constraint stQQ T =1 K , the updated feature matrix can be calculated by minimizing the third objective function. The updated feature matrix is ​​composed of the Laplace matrix L Y The first K small eigenvectors of .

[0137] Therefore, through the above scheme, an updated consensus matrix and feature matrix can be obtained. In the subsequent process, the updated consensus matrix and feature matrix can replace the original consensus matrix and feature matrix in step S430 as the latest current consensus matrix and current feature matrix, and the kinship matrix can be updated according to the same principle to obtain the current kinship matrix updated in this iteration.

[0138] In some embodiments, in order to enable the Laplacian matrix to more reasonably constrain the feature matrix based on the actual iteration situation to further improve the clustering efficiency and accuracy, the embodiment of the present invention further provides a solution for updating the first parameter system, that is, the multi-view clustering method provided by the embodiment of the present invention may also include:

[0139] In step S610, when the Laplace rank of the Laplace matrix is ​​greater than a set dimension, updating the value of the first parameter coefficient to twice the first parameter coefficient;

[0140] In step S620, when the Laplace rank of the Laplace matrix is ​​less than the set dimension, the value of the first parameter coefficient is updated to 1 / 2 of the first parameter coefficient;

[0141] In step S630, when the Laplace rank of the Laplace matrix is ​​equal to the set dimension, the value of the first parameter coefficient is kept unchanged.

[0142] The above steps S610 to S630 can be performed after the Laplace matrix is ​​calculated in step S4521 to update the first parameter coefficient. In this way, in each round of updating of the kinship matrix, the updated first parameter coefficient can be obtained in step S431, so that in step S432, the updated first parameter system is used to calculate the first auxiliary matrix.

[0143] Therefore, the kinship matrix is ​​iteratively updated continuously through the above scheme until the number of iterative updates reaches the set number threshold, and it can be considered that the clustering of multiple views has been completed. At this time, the kinship matrix obtained is the result of clustering multiple views. Please refer to Figure 5 , Figure 5 It is a schematic diagram of clustering results obtained by clustering using the multi-view clustering method provided by an embodiment of the present invention.

[0144] It is worth noting that the technical features or technical solutions in any of the above embodiments of the present invention can be combined with each other as long as there is no combination contradiction.

[0145] In order to execute the corresponding steps in the above multi-view clustering method embodiment and various possible methods, a multi-view clustering device is provided below. Optionally, the multi-view clustering device can adopt the above multi-view clustering method. Figure 1 For further information, please refer to Figure 6 , Figure 6 A functional module diagram of a multi-view clustering device provided in an embodiment of the present invention. It should be noted that the basic principle and technical effects of the multi-view clustering device provided in this embodiment are the same as those of the above-mentioned embodiment. For the sake of brief description, for the parts not mentioned in this embodiment, reference can be made to the corresponding contents in the above-mentioned embodiment. The multi-view clustering device 700 includes:

[0146] The preprocessing module 710 is configured to: preprocess the original data of the multiple views to obtain normalized data of the multiple views;

[0147] The manifold processing module 720 is configured to: perform manifold learning processing on the normalized data of the multiple views to obtain a spectral embedding feature matrix after the multiple views are fused;

[0148] The fuzzy processing module 730 is configured to: perform fuzzy clustering processing on the normalized data of the multiple views respectively to obtain spectral approximation feature matrices of the multiple views respectively;

[0149] The fusion processing module 740 is configured to: perform fusion clustering on the spectrum embedding feature matrices and all spectrum approximation feature matrices of the multiple views to obtain clustering results of the multiple views.

[0150] In some embodiments, the process of the manifold processing module 720 performing manifold learning processing on the normalized data of the multiple views to obtain the spectral embedding feature matrix after the fusion of the multiple views is configured as follows:

[0151] For each view, determining a set number of anchor points for the view according to normalized data of the view;

[0152] For each view, determining a bipartite graph matrix of the view according to the set number of anchor points of the view;

[0153] According to the bipartite graph matrices of the multiple views, a cross-view anchor graph including all the bipartite graph matrices is constructed;

[0154] The spectral embedding feature matrix is ​​obtained by processing according to the cross-view anchor map.

[0155] In some embodiments, the process of obtaining the spectral embedding feature matrix by the manifold processing module 720 according to the cross-view anchor graph is configured as follows:

[0156] Calculate and obtain a normalized Laplacian matrix of the cross-view anchor graph;

[0157] The spectral embedding feature matrix is ​​calculated according to the spectral embedding matrix of the cross-view anchor graph and the normalized Laplacian matrix.

[0158] In some embodiments, the process in which the fuzzy processing module 730 performs fuzzy clustering processing on the normalized data of the multiple views to obtain the spectral approximation feature matrices of the multiple views is configured as follows:

[0159] For each view, clustering the normalized data of the view is performed using a fuzzy clustering algorithm to obtain a membership matrix of the view;

[0160] For each view, a fuzzy similarity matrix of the view is constructed according to the membership matrix of the view; the fuzzy similarity matrix is ​​a diagonal symmetric matrix in which all elements on the diagonal are 1;

[0161] For each view, a spectral approximation feature matrix of the view is calculated according to the fuzzy similarity matrix of the view.

[0162] In some embodiments, the fusion processing module 740 performs fusion clustering on the spectrum embedding feature matrices and all spectrum approximation feature matrices of the multiple views to obtain the clustering results of the multiple views, which is configured as follows:

[0163] Constructing a spectral approximation feature set formed by the spectral embedding feature matrix and all spectral approximation feature matrices;

[0164] Determining a consensus matrix, a feature matrix for implementing Laplace rank constraints, and a kinship matrix according to the spectral approximation feature set;

[0165] According to the current consensus matrix and the current feature matrix, the kinship matrix is ​​updated to obtain the current kinship matrix;

[0166] When the number of updates corresponding to the secondary kinship matrix reaches a set number threshold, the secondary kinship matrix is ​​used as the clustering result;

[0167] When the number of updates corresponding to the secondary kinship matrix does not reach the set number threshold, the current consensus matrix and the current feature matrix are updated respectively, and the step of updating the kinship matrix according to the current consensus matrix and the current feature matrix is ​​returned to be executed.

[0168] In some embodiments, the fusion processing module 740 is configured to:

[0169] Get the first parameter coefficient set for the current feature matrix;

[0170] Calculate a first auxiliary matrix according to the current consensus matrix, the current feature matrix and the first parameter coefficients;

[0171] The value of each column in the kinship matrix is ​​updated according to the first auxiliary matrix to obtain the current kinship matrix.

[0172] In some embodiments, the process of updating the current consensus matrix by the fusion processing module 740 is configured as follows:

[0173] Obtaining an error matrix and a view weight of each submatrix in the spectral approximation feature set;

[0174] For each submatrix in the spectral approximation feature set, a second auxiliary matrix corresponding to the submatrix is ​​calculated according to the current consensus matrix, the error matrix of the submatrix, and the view weight;

[0175] For each submatrix in the spectral approximation feature set, updating an error matrix of the submatrix according to a second auxiliary matrix of the submatrix;

[0176] For each submatrix in the spectral approximation feature set, updating the view weight of the submatrix according to the current consensus matrix, the submatrix and the error matrix after the submatrix is ​​updated;

[0177] For each submatrix in the spectral approximation feature set, according to the current consensus matrix and the updated error matrix and view weight of the submatrix, an auxiliary variable for measuring the contribution of the submatrix to the current consensus matrix is ​​calculated;

[0178] A third auxiliary matrix is ​​calculated based on the spectral approximation feature set, the auxiliary variables of all submatrices and the updated error matrix, and the secondary kinship matrix;

[0179] The current consensus matrix is ​​updated according to the third auxiliary matrix.

[0180] In some embodiments, the process of updating the current feature matrix by the fusion processing module 740 is configured as follows:

[0181] A Laplace matrix is ​​calculated according to the current kinship matrix;

[0182] An updated feature matrix is ​​obtained according to the Laplace matrix.

[0183] In some embodiments, the fusion processing module 740 is further configured to:

[0184] When the Laplace rank of the Laplace matrix is ​​greater than a set dimension, updating the value of the first parameter coefficient to twice the first parameter coefficient;

[0185] When the Laplace rank of the Laplace matrix is ​​less than the set dimension, updating the value of the first parameter coefficient to 1 / 2 of the first parameter coefficient;

[0186] When the Laplace rank of the Laplace matrix is ​​equal to the set dimension, the value of the first parameter coefficient is kept unchanged.

[0187] Optionally, the above modules can be stored in the form of software or firmware. Figure 1 The memory shown in the figure or solidified in the operating system (OS) of the electronic device, and can be Figure 1 Meanwhile, the data and program codes required for executing the above modules can be stored in the memory.

[0188] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, a program segment or a part of a code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0189] In addition, the functional modules in the various embodiments of the present invention may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.

[0190] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0191] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A multi-view clustering method, characterized in that: include: Preprocessing the original data of the multiple views to obtain normalized data of the multiple views; Performing manifold learning processing on the normalized data of the multiple views to obtain a spectral embedding feature matrix after the multiple views are fused; Performing fuzzy clustering processing on the normalized data of the multiple views respectively to obtain spectral approximation feature matrices of the multiple views; The spectral embedding feature matrices and all spectral approximation feature matrices of the multiple views are fused and clustered to obtain clustering results of the multiple views.

2. The method according to claim 1, characterized in that The step of performing manifold learning processing on the normalized data of the multiple views to obtain a spectral embedding feature matrix after the multiple views are fused includes: For each view, determining a set number of anchor points for the view according to normalized data of the view; For each view, determining a bipartite graph matrix of the view according to the set number of anchor points of the view; According to the bipartite graph matrices of the multiple views, a cross-view anchor graph including all the bipartite graph matrices is constructed; The spectral embedding feature matrix is ​​obtained by processing according to the cross-view anchor map.

3. The method according to claim 2, characterized in that The step of obtaining the spectral embedding feature matrix according to the cross-view anchor graph comprises: Calculate and obtain a normalized Laplacian matrix of the cross-view anchor graph; The spectral embedding feature matrix is ​​calculated according to the spectral embedding matrix of the cross-view anchor graph and the normalized Laplacian matrix.

4. The method according to claim 1, characterized in that: The step of performing fuzzy clustering processing on the normalized data of the multiple views to obtain spectral approximation feature matrices of the multiple views respectively includes: For each view, clustering the normalized data of the view is performed using a fuzzy clustering algorithm to obtain a membership matrix of the view; For each view, a fuzzy similarity matrix of the view is constructed according to the membership matrix of the view; the fuzzy similarity matrix is ​​a diagonal symmetric matrix in which all elements on the diagonal are 1; For each view, a spectral approximation feature matrix of the view is calculated according to the fuzzy similarity matrix of the view.

5. The method according to claim 1, characterized in that The step of fusing and clustering the spectrum embedding feature matrices and all spectrum approximation feature matrices of the multiple views to obtain clustering results of the multiple views includes: Constructing a spectral approximation feature set formed by the spectral embedding feature matrix and all spectral approximation feature matrices; Determining a consensus matrix, a feature matrix for implementing Laplace rank constraints, and a kinship matrix according to the spectral approximation feature set; According to the current consensus matrix and the current feature matrix, the kinship matrix is ​​updated to obtain the current kinship matrix; When the number of updates corresponding to the secondary kinship matrix reaches a set number threshold, the secondary kinship matrix is ​​used as the clustering result; When the number of updates corresponding to the secondary kinship matrix does not reach the set number threshold, the current consensus matrix and the current feature matrix are updated respectively, and the step of updating the kinship matrix according to the current consensus matrix and the current feature matrix is ​​returned to be executed.

6. The method according to claim 5, characterized in that The step of updating the kinship matrix according to the current consensus matrix and the current feature matrix comprises: Get the first parameter coefficient set for the current feature matrix; Calculate a first auxiliary matrix according to the current consensus matrix, the current feature matrix and the first parameter coefficients; The value of each column in the kinship matrix is ​​updated according to the first auxiliary matrix to obtain the current kinship matrix.

7. The method according to claim 6, characterized in that The process of updating the current consensus matrix includes: Obtaining an error matrix and a view weight of each submatrix in the spectral approximation feature set; For each submatrix in the spectral approximation feature set, a second auxiliary matrix corresponding to the submatrix is ​​calculated according to the current consensus matrix, the error matrix of the submatrix, and the view weight; For each submatrix in the spectral approximation feature set, updating an error matrix of the submatrix according to a second auxiliary matrix of the submatrix; For each submatrix in the spectral approximation feature set, updating the view weight of the submatrix according to the current consensus matrix, the submatrix and the error matrix after the submatrix is ​​updated; For each submatrix in the spectral approximation feature set, according to the current consensus matrix and the updated error matrix and view weight of the submatrix, an auxiliary variable for measuring the contribution of the submatrix to the current consensus matrix is ​​calculated; A third auxiliary matrix is ​​calculated based on the spectral approximation feature set, the auxiliary variables of all submatrices and the updated error matrix, and the secondary kinship matrix; The current consensus matrix is ​​updated according to the third auxiliary matrix.

8. The method according to claim 7, characterized in that The process of updating the current feature matrix includes: A Laplace matrix is ​​calculated according to the current kinship matrix; An updated feature matrix is ​​obtained according to the Laplace matrix.

9. The method according to claim 8, characterized in that The method further comprises: When the Laplace rank of the Laplace matrix is ​​greater than a set dimension, updating the value of the first parameter coefficient to twice the first parameter coefficient; When the Laplace rank of the Laplace matrix is ​​less than the set dimension, updating the value of the first parameter coefficient to 1 / 2 of the first parameter coefficient; When the Laplace rank of the Laplace matrix is ​​equal to the set dimension, the value of the first parameter coefficient is kept unchanged.

10. A multi-view clustering device, characterized in that: include: A preprocessing module is configured to: preprocess the original data of the multiple views to obtain normalized data of the multiple views; A manifold processing module is configured to: perform manifold learning processing on the normalized data of the multiple views to obtain a spectral embedding feature matrix after the multiple views are fused; A fuzzy processing module is configured to: perform fuzzy clustering processing on the normalized data of the multiple views respectively to obtain spectral approximation feature matrices of the multiple views respectively; The fusion processing module is configured to: fuse and cluster the spectrum embedding feature matrices and all spectrum approximation feature matrices of the multiple views to obtain clustering results of the multiple views.

11. An electronic device, characterized in that: The invention comprises a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor can execute the machine executable instructions to implement the multi-view clustering method according to any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the multi-view clustering method according to any one of claims 1 to 9 is implemented.

Images