Semi-supervised multi-view clustering method based on double-constraint concept decomposition

Abstract

The invention discloses a semi-supervised multi-view clustering method based on double-constraint concept decomposition, which relates to the technical field of machine learning, combines point constraint with pairwise constraint, constructs a double-constraint algorithm based on prior label information of a complementary supervision mechanism, and realizes maximum utilization and progressive propagation of limited supervision information. And based on a concept decomposition framework, breaking through the limitation of a traditional method on non-negativity of data, processing complex data distribution in combination with a kernel method, obtaining a consensus matrix of multi-view data fusion as low-dimensional representation of views, performing clustering by adopting a k-means algorithm, and outputting a final category division result. According to the method, effective processing of complex multi-view data is realized, finally, the performance of a multi-view clustering task is comprehensively improved, and the bottlenecks in clustering precision, data adaptability and supervision information utilization efficiency in the prior art are broken through.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, and more particularly, to a semi-supervised multi-view clustering method based on double-constraint concept factorization. BACKGROUND

[0002] With the popularity of multi-modal data, multi-view clustering has become a research hotspot due to its ability to fuse complementary information across sources. Although non-negative matrix factorization (NMF)-based methods are widely used, they are limited to non-negative data processing and are difficult to capture nonlinear structures. Concept factorization (CF), as a generalization of NMF, breaks through the non-negative data constraint and is compatible with kernel methods, becoming a new direction for multi-view clustering (such as MVCC, MVDGCF, etc.). However, most existing CF methods are unsupervised learning methods and cannot effectively utilize the limited labeled information commonly found in real-world scenarios. Although some semi-supervised methods have been attempted, they still have problems such as insufficient utilization of supervision information and single constraint form, resulting in limited performance improvement in sparse labeling scenarios.

[0003] Existing multi-view clustering methods have significant shortcomings in utilizing limited supervision information to improve clustering performance. Specifically, most concept factorization (CF)-based multi-view clustering methods use unsupervised learning strategies and fail to effectively utilize the limited labeled information commonly found in real-world data, resulting in limited clustering accuracy. A few semi-supervised multi-view CF methods have a single form of supervision information utilization, only using simple methods such as label propagation to process labeled data, and fail to fully exploit the potential value of labeled information. At the same time, existing methods lack a complementary utilization mechanism for supervision information, neither directly ensuring the clustering consistency of known class samples through point constraints nor effectively propagating labeled information to unlabeled samples through pairwise constraints, resulting in limited coverage of supervision information and difficulty in guiding the overall clustering structure to converge to the true distribution. These problems limit the performance improvement of existing methods in sparse labeling scenarios and fail to meet the accuracy and robustness requirements of multi-view data clustering. Therefore, it is an urgent problem to overcome the deficiencies of existing multi-view clustering techniques in supervision information utilization and data processing and to provide an efficient semi-supervised multi-view clustering scheme. SUMMARY

[0004] To solve the above technical problems, the present application provides a semi-supervised multi-view clustering method based on double-constraint concept factorization, which maximizes the utilization and gradual diffusion of limited supervision information, fully utilizes the limited prior information, and improves the clustering accuracy in low labeling rate scenarios.

[0005] The present application provides a semi-supervised multi-view clustering method based on double-constraint concept factorization, which includes the following steps:

[0006] Obtaining a multi-view dataset containing a plurality of view samples, wherein part of the samples are provided with known labels, initializing a multi-view shared label constraint matrix, a pair constraint matrix, a representation matrix, a correlation matrix and a consensus matrix in the multi-view dataset;

[0007] Constructing an initial weight matrix based on a nearest neighbor relationship and a kernel function for each view, generating a final pair constraint matrix through a vertical and horizontal constraint propagation algorithm, updating the weight matrix to enhance the correlation of samples of the same type, and calculating a corresponding graph Laplacian matrix and a diagonal matrix;

[0008] In the alternating iteration, the correlation matrix is updated using the shared label constraint matrix, the projection matrix is updated using the graph Laplacian matrix and the diagonal matrix while fixing the correlation matrix, and the consensus matrix is updated based on the projection matrix of each view after each iteration;

[0009] After the iteration is completed, the final consensus matrix is taken as a low-dimensional representation of the view, a k-means algorithm is used for clustering, and a final class division result is output.

[0010] In the scheme, a multi-view dataset containing a plurality of view sample data is obtained, wherein part of the samples are provided with known labels, and a shared label constraint matrix, a representation matrix, a projection matrix and a consensus matrix of the view sample data in the multi-view dataset are initialized, specifically:

[0011] Input a multi-view dataset X = [X (1) ,X (2) ,…,X (v) ,…,X (nv) ], wherein X (v) represents a data matrix of the vth view, there are n views in the multi-view dataset, each view has m characteristic dimensions, the first p samples are provided with known class labels, the remaining n-p samples are unlabelled, and the number of classes is c;

[0012] Encoding the known class label information C p×c of the samples with known class labels, using an identity matrix I n-p to represent the initial free representation of the unknown label samples, constructing a multi-view shared label constraint matrix p×c according to the encoding representation C n-p of the known label part and the unknown label part I

[0013] Constructing a pair constraint based on the samples with known class labels, constructing a constraint relationship when the samples in a sample pair are of the same type, and an initial pair constraint matrix Z;

[0014] Introducing a projection matrix N (v) for each view, using the projection matrix N (v)and the shared label constraint matrix A generates a representation matrix V of each view (v) = AN (v) The low-dimensional representation of the sample is marked by the point constraint, and the sample is ensured to be consistent with the true label.

[0015] The initial correlation matrix represents the linear mapping relationship of the original data to the low-dimensional space, and the consensus matrix is generated by taking the mean of the representation matrix of each view.

[0016] In this scheme, the initial weight matrix is constructed based on the nearest neighbor relationship and the kernel function for each view sample data, and the final pairwise constraint matrix is generated by the vertical and horizontal constraint propagation algorithm, specifically:

[0017] The similarity between samples is calculated by the Gaussian kernel function for the data matrix of the vth view, and the sparse similarity matrix is constructed based on the relationship between each sample x i The nearest neighbors are obtained, and the sparse similarity matrix is constructed based on the relationship between the sample x i and the nearest neighbor sample, and the initial weight matrix is constructed by symmetric normalization;

[0018] The graph Laplacian matrix is calculated based on the initial weight matrix, and the vertical constraint propagation matrix F v and the horizontal constraint propagation matrix F h are set to random values between [-1, 1], and the propagation matrix of the pairwise constraint propagation is initialized;

[0019] The vertical constraint propagation matrix F v and the horizontal constraint propagation matrix F h are iteratively updated using the graph Laplacian matrix and the preset propagation parameters, and the extreme value is obtained after the vertical constraint propagation matrix F v and the horizontal constraint propagation matrix F h are iterated to the post-convergence, and the final pairwise constraint matrix is generated.

[0020] In this scheme, the weight matrix is updated to enhance the relevance of the same type of samples, and the corresponding graph Laplacian matrix and diagonal matrix are calculated, specifically:

[0021] The initial weight matrix and the final pairwise constraint matrix are obtained according to the update rule to obtain the updated weight matrix, enhance the same association, and calculate the diagonal matrix using the updated weight matrix, and construct the updated graph Laplacian matrix based on the diagonal matrix and the updated weight matrix;

[0022] The update rule is represented as:

[0023]

[0024] wherein represents the element in the updated weight matrix, s ijdenotes an element in the initial weight matrix, denotes an element in the final pairwise constraint matrix, if then increase strengthen the same class association, if then decrease weaken the same class association.

[0025] In the scheme, in the alternating iteration, the projection matrix is fixed to update the association matrix by using the shared label constraint matrix, and the association matrix is fixed to update the projection matrix by using the graph Laplacian matrix and the diagonal matrix, specifically:

[0026] By optimizing the objective function through alternating iteration, the projection matrix N (v) is fixed in the iteration, and each view association matrix W (v) is updated to minimize the reconstruction error and the point constraint loss, the known class label information is embedded into the association matrix through the shared label constraint matrix and the projection matrix, ensuring the low-dimensional representation consistency of the known class label samples, and the update formula of each view association matrix W (v) is expressed as:

[0027]

[0028] wherein, denotes an element in the association matrix, K (v) denotes the kernel matrix, and A denotes the shared label constraint matrix;

[0029] The association matrix W (v) is fixed in the iteration, and each view projection matrix N (v) is updated, the geometric information encoded by the Laplacian matrix is used to guide the unknown class label samples to gather towards the correct class, and the update formula of each view projection matrix N (v) is:

[0030]

[0031] wherein, denotes an element in the projection matrix, and α and β denote regularization parameters, denotes the updated weight matrix, V * denotes the updated consensus matrix, and A denotes the shared label constraint matrix.

[0032] In the scheme, after each iteration, the consensus matrix is updated based on the projection matrix of each view, specifically:

[0033] The updated projection matrix N (v) of each view and the shared label constraint matrix N (v) are obtained, the representation matrix of each view is calculated, the representation matrix of all views is averaged, the multi-view information is fused, the differences between views are eliminated, and the unified final consensus matrix V is obtained.* , represented as:

[0034]

[0035] Where n v represents the total number of views, and v represents the number of view items.

[0036] In this scheme, the objective function is optimized through alternating iterations, and the objective function is expressed as:

[0037]

[0038] Where X (v) W represents the data matrix of the v-th view. (v) N represents the association matrix of the v-th view. (v) V represents the projection matrix of the v-th view. * Let represent the final consensus matrix for multi-view fusion, where α and β represent regularization parameters. Let W represent the graph Laplacian matrix. (v) N (v) V * ≥0 indicates a non-negativity constraint, ensuring the decomposition results are interpretable. v The total number of views is represented by Tr, which represents the trace of the matrix and is used for scalarization of the manifold regularization term.

[0039] In this scheme, after the iteration is complete, the final consensus matrix is ​​used as a low-dimensional representation of the view. The k-means algorithm is then used for clustering to output the final category classification results, specifically:

[0040] Obtain the final consensus matrix V * and a preset number of clustering categories, for the final consensus matrix V * Normalization is performed, sample points are selected as initial cluster centers according to the preset number of cluster categories, the Euclidean distance between each sample and all initial cluster centers is calculated, and the sample is assigned to the nearest initial cluster center to generate a data cluster.

[0041] For each data cluster, the mean of all samples in the data cluster is recalculated to update the cluster center. The process stops when the change in the cluster center is less than the threshold or when the maximum number of iterations is reached. The final category label of each sample is obtained based on the final clustering result.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] The application proposes a semi-supervised multi-view clustering model based on double-constraint concept decomposition. By innovatively introducing a double supervision mechanism combining point constraint and pair constraint, on one hand, the point constraint is used to directly act on the labeled samples to ensure clustering consistency; on the other hand, the pair constraint is used to construct the'must link' and 'cannot link' relationships between samples, to propagate the labeled information to unlabeled data, capture the internal geometric structure of data, and expand the supervision range. Meanwhile, based on the concept decomposition framework, the negative value data and the kernel method are compatible, the effective processing of complex multi-view data is realized, and finally the performance of the multi-view clustering task is comprehensively improved, and the bottleneck of the prior art in clustering accuracy, data adaptability and supervision information utilization efficiency is broken through. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments or examples of the present application, the drawings needed to be used in the embodiments or example descriptions will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0045] Figure 1 A flowchart of a semi-supervised multi-view clustering method based on double-constraint concept decomposition is shown.

[0046] Figure 2 A framework diagram of a semi-supervised multi-view clustering method based on double-constraint concept decomposition is shown.

[0047] Figure 3 A block diagram of a semi-supervised multi-view clustering system based on double-constraint concept decomposition is shown. DETAILED DESCRIPTION

[0048] In order to more clearly illustrate the technical solutions in the embodiments or examples of the present application, the drawings needed to be used in the embodiments or example descriptions will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0049] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below.

[0050] Figure 1 A flowchart of a semi-supervised multi-view clustering method based on double-constraint concept decomposition is shown.

[0051] As Figure 1 shown, the present embodiment provides a semi-supervised multi-view clustering method based on double-constraint concept decomposition, comprising:

[0052] S102, obtaining a multi-view dataset containing a plurality of view samples, wherein part of the samples are provided with known labels, initializing a multi-view shared label constraint matrix, a pair constraint matrix, a representation matrix, a correlation matrix and a consensus matrix in the multi-view dataset;

[0053] S104, constructing an initial weight matrix based on a nearest neighbor relationship and a kernel function for each view, generating a final pair constraint matrix through a vertical and horizontal constraint propagation algorithm, updating the weight matrix to enhance the correlation of samples of the same type, and calculating a corresponding graph Laplacian matrix and a diagonal matrix;

[0054] S106, in the alternating iteration, updating the correlation matrix using the shared label constraint matrix while fixing the projection matrix, updating the projection matrix using the graph Laplacian matrix and the diagonal matrix while fixing the correlation matrix, and updating the consensus matrix based on the projection matrix of each view after each iteration;

[0055] S108, after the iteration is completed, taking the final consensus matrix as a low-dimensional representation of the views, using a k-means algorithm for clustering, and outputting a final class division result.

[0056] It should be noted that, as Figure 2 indicated, the present scheme combines point constraints and pair constraints to construct an LDC algorithm (double constraint algorithm based on prior label information) of a complementary supervision mechanism, realizes the maximum utilization and progressive propagation of limited supervision information; on the other hand, relying on the concept decomposition framework, the present scheme breaks through the limitation of traditional methods on non-negativity of data, combines kernel methods to process complex data distribution, and is applicable to multi-view data scenes such as news classification, fake news detection and user interest mining.

[0057] An input multi-view dataset X = [X (1) ,X (2) ,…,X (v) ,…,X (nv) ] is input, wherein X (v) represents a data matrix of the vth view, there are n views in the multi-view dataset, each view has m characteristic dimensions, the first p samples are provided with known class labels, the remaining n-p samples are unlabelled, and the number of classes is c;

[0058] The known class label information C p×c is encoded for the samples with known class labels, and the initial free representation I n-p is used to represent the unlabelled samples, and the known label part encoding representation C p×c and the unlabelled part representation I n-p are used to construct a multi-view shared label constraint matrix A ij = 1 indicates that the ith sample belongs to the jth class, and Aij =0 indicates other cases. The shared label constraint matrix uses known label information as point constraints, and unlabeled samples are gradually aligned with the true distribution through optimization.

[0059] Pair constraints are constructed based on samples with known category labels. The constraint relationship is established by determining whether the samples in a pair are of the same class. The initial pair constraint matrix Z provides the initial supervision signal for subsequent pair constraint propagation. Each element is defined as follows:

[0060]

[0061] Where M represents x i With x j If they share the same label, C indicates they belong to different labels, and 0 indicates other cases.

[0062] Introduce projection matrix N for each view (v) Using the projection matrix N (v) The shared label constraint matrix A generates the representation matrix V for each view. (v) =AN (v) By using point constraints to label the low-dimensional representation of the samples, we ensure that the samples are consistent with the true labels; we initialize the association matrix to represent the linear mapping relationship from the original data to the low-dimensional space, and find the following approximate representation X≈XWV in the concept decomposition. T Where X is the original data, W is the association matrix representing the degree of relevance of data points to the concept, and V is the representation matrix. Additionally, the mean of the representation matrices of each view is used to generate a consensus matrix, and all view information is integrated to obtain a unified low-dimensional representation.

[0063] It should be noted that, assuming Propagate pairwise constraint matrices, with absolute values ​​in |f ij | ≤ 1. When f ij >0, indicating (x i ,x j ) is a mandatory link constraint. ij <0 means the constraint cannot be linked, while |f ij | represents (x) i ,x j () is the confidence score for constraints that must be linked or cannot be linked. This is determined by defining its weight matrix. To compute the p-nearest neighbor graph

[0064]

[0065] For each data point x i Find its p-nearest neighbor (excluding itself). If x i For x jone nearest neighbor, the similarity between samples is computed by a Gaussian kernel function for the data matrix of the vth view, setting the weight s ij a(x i ,x j ) = exp(-‖x i -x j ‖ 2 / t) (where t is the kernel bandwidth parameter); otherwise s ij = 0. In addition, in order to make the weight matrix symmetric, set S = (S + S T ) / 2;

[0066] After obtaining the symmetric weight matrix, the graph Laplacian matrix L is calculated The matrix D involved is a diagonal matrix. Its construction rule is that the i-th diagonal element of D is determined by summing all elements s ij for all j, denoted as ∑ j s ij .

[0067] The elements of the vertical constraint propagation matrix F v and the horizontal constraint propagation matrix F h are set to random values between [-1, 1], and the vertical constraint propagation matrix F v and the horizontal constraint propagation matrix F h are iteratively updated using the graph Laplacian matrix and the preset propagation parameters. After the vertical constraint propagation matrix F v and the horizontal constraint propagation matrix F h are iterated to converge, the extremum is obtained to generate the final pairwise constraint matrix.

[0068] F v (t+1) = αLF v (t) + (1-α)Z

[0069]

[0070] where F v (t) is the limit of {F h (t)} and F * is the limit of {F * (t)}. F ij represents the final configuration of pairwise constraints established in the propagation process, reflecting their integrated representation. It is worth noting that F (v) represents a set of combined pairwise constraints and their associated confidence scores.

[0071] The initial weight matrix and the final pairwise constraint matrix are obtained. According to the update rule, the updated weight matrix is obtained, the same type of association is enhanced, and the updated weight matrix is used to calculate the diagonal matrix And based on the diagonal matrix and the updated weight matrix, an updated graph Laplacian matrix is constructed

[0072] The update rule is represented as:

[0073]

[0074] Wherein represents an element in the updated weight matrix, s ij represents an element in the initial weight matrix, represents an element in the final pairwise constraint matrix, if , increase Strengthen the same kind of association, if , reduce Weaken the same kind of association.

[0075] It should be noted that by alternating iterative optimization objective function, in the iteration fixed projection matrix N (v) , update each view association matrix W (v) , minimize reconstruction error and point constraint loss, by sharing label constraint matrix and projection matrix, embedding known class label information into the association matrix, ensure the consistency of the low-dimensional representation of known class label samples, the update formula of each view association matrix W (v) is represented as:

[0076]

[0077] Wherein, represents an element in the association matrix, K (v) represents the kernel matrix, A represents the shared label constraint matrix;

[0078] In the iteration fixed association matrix W (v) , update each view projection matrix N (v) , through the geometric information of Laplacian matrix coding, guide the unknown class label samples to the correct class aggregation, the update formula of each view projection matrix N (v) is represented as:

[0079]

[0080] Wherein, represents an element in the projection matrix, α, β represents the regularization parameter, represents the updated weight matrix, V * represents the updated consensus matrix, A represents the shared label constraint matrix.

[0081] Get each view updated projection matrix N (v) and shared label constraint matrix N (v), calculate each view representation matrix, take the average of the representation matrix of all views, fuse multi-view information, eliminate the difference between views, and obtain a unified final consensus matrix V * , is expressed as:

[0082]

[0083] where n v represents the total number of views, and v represents the number of view items.

[0084] It should be noted that by alternately iterating the optimization objective function, by minimizing the reconstruction error, maintaining geometric structural consistency (manifold regularization) and multi-view collaborative optimization (consensus constraint), a discriminative low-dimensional representation is learned, and the objective function is expressed as:

[0085]

[0086] where X (v) represents the data matrix of the vth view, W (v) represents the correlation matrix of the vth view, N (v) represents the projection matrix of the vth view, V * represents the final consensus matrix of multi-view fusion, and a and β represent regularization parameters, represents the graph Laplacian matrix, W (v) , N (v) , and V * ≥ 0 represent non-negativity constraints to ensure that the decomposition result is interpretable, n v represents the total number of views, and Tr represents the trace of the matrix, which is used for scalarization of the manifold regularization term.

[0087] By introducing a multi-view shared label constraint matrix A and an auxiliary matrix N (v) , a target function is obtained which applies point constraints under multi-view concept decomposition as a hard constraint, and the reconstruction error is measured using the Frobenius norm to ensure that the low-dimensional representation AN (v) can effectively reconstruct the original data X (v) . Pairwise constraint information is embodied through modification of the weight matrix, and data points in the same class have relatively high correlation values, while data points of different classes have lower correlation values. The double constraint algorithm based on prior labels combines point constraints and pairwise constraint propagation, fully utilizes limited prior label information, and is unified within the concept decomposition framework, and the manifold regularization term is The Laplacian matrix is used to constrain similar samples to have a close distance in the low-dimensional space. The multi- Figure 1 consistency term uses the Frobenius norm to measure the consistency error, and forces the low-dimensional representation AN(v) Approximating the final consensus matrix. Point constraints ensure consistency among known labeled samples, while pairwise constraints propagate supervisory information to unlabeled samples. Combining multi-view consensus matrices with complementary features effectively improves clustering accuracy.

[0088] It should be noted that obtaining the final consensus matrix V * and a preset number of clustering categories, for the final consensus matrix V * Normalization is performed, and sample points are selected as initial cluster centers according to the preset number of cluster categories. The Euclidean distance between each sample and all initial cluster centers is calculated, and the sample is assigned to the nearest initial cluster center to generate a data cluster. For each data cluster, the mean of all samples in the data cluster is recalculated to update the cluster center. The process stops when the change in cluster center is less than the threshold or the maximum number of iterations is reached. The final category label of each sample is obtained based on the final clustering result.

[0089] In a preferred embodiment of this application, the BBC news dataset includes four views (word frequency, TF-IDF, topic model features, and named entity statistics). The sample comprises 685 news articles, categorized into five types (politics, business, technology, entertainment, and sports). Approximately 68 articles (10% of the samples) are randomly selected to provide ground truth labels; the remainder are unlabeled. The first 68 rows correspond to known labels (e.g., political category sample A). ij =1, the rest A ij =0), the last 617 rows are an identity matrix. From the known labels, 50 pairs of "must-link" (similar samples) and 50 pairs of "cannot-link" (different samples) are randomly generated. Initial similarity is calculated based on the Gaussian kernel function, with a nearest neighbor count p=5, and then generated... Strengthen the association between similar samples.

[0090] Regularization coefficients α = 10, β = 100, and latent space dimension k = 5 (consistent with the number of classes). N is fixed. (v) Update W (v) Minimize reconstruction error, fix W (v) ,use Update N (v) This disseminates supervisory information to unlabeled samples and updates the consensus matrix. The objective function changes by less than 10 -6 Alternatively, the iteration can be terminated after 100 iterations. Then, the consensus matrix is ​​updated for clustering, 5 cluster centers are initialized, and the iteration continues until convergence, outputting the category labels for 685 news articles.

[0091] Figure 3 A block diagram of a semi-supervised multi-view clustering system based on dual-constraint concept decomposition is shown.

[0092] The second embodiment of the present application provides a semi-supervised multi-view clustering system based on double-constraint concept decomposition 3, comprising a memory 31 and a processor 32, wherein the memory comprises a semi-supervised multi-view clustering method program based on double-constraint concept decomposition, and when the semi-supervised multi-view clustering method program based on double-constraint concept decomposition is executed by the processor, the steps of the semi-supervised multi-view clustering method based on double-constraint concept decomposition are implemented.

[0093] The third aspect of the present application provides a computer readable storage medium, wherein the computer readable storage medium comprises a semi-supervised multi-view clustering method program based on double-constraint concept decomposition, and when the semi-supervised multi-view clustering method program based on double-constraint concept decomposition is executed by the processor, the steps of the semi-supervised multi-view clustering method based on double-constraint concept decomposition are implemented.

[0094] In several embodiments provided in the present application, it should be understood that the disclosed method and system can be implemented in other ways. The system embodiments described above are only illustrative, for example, the division of the units is only a logical function division, and actual implementation can have another division mode, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed components can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms. In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be realized in the form of hardware or hardware plus software functional unit.

[0095] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction related hardware, and the aforementioned program can be stored in a computer readable storage medium, and the program executes the steps including the above-mentioned method embodiments when executed; and the aforementioned storage medium includes mobile storage device, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and various storage program codes.

[0096] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A semi-supervised multi-view clustering method based on dual-concept decomposition, characterized in that, The method comprises the following steps: Obtain a multi-view data set comprising a plurality of view samples, wherein part of the samples are provided with known labels, and initialize a multi-view shared label constraint matrix, a pair constraint matrix, a representation matrix, a correlation matrix and a consensus matrix in the multi-view data set; For each view, construct an initial weight matrix based on a neighbor relationship and a kernel function, generate a final pair constraint matrix through a vertical and horizontal constraint propagation algorithm, update the weight matrix to enhance the correlation of samples of the same type, and calculate a corresponding graph Laplacian matrix and a diagonal matrix; In an alternating iteration, update the correlation matrix using the shared label constraint matrix while fixing the projection matrix, update the projection matrix using the graph Laplacian matrix and the diagonal matrix while fixing the correlation matrix, and update the consensus matrix based on the projection matrix of each view after each iteration; After the iteration ends, take the final consensus matrix as a low-dimensional representation of the view, perform clustering by using a k-means algorithm, and output a final class division result.

2. The dual constraint concept decomposition based semi-supervised multi-view clustering method according to claim 1, characterized in that, Obtain a multi-view data set comprising a plurality of view sample data, wherein part of the samples are provided with known labels, and initialize a shared label constraint matrix, a representation matrix, a projection matrix and a consensus matrix of the view sample data in the multi-view data set, specifically: Input a multi-view dataset X = [X (1) ,X (2) ,…,X (v) ,…,X (nv) ], where X (v) represents the data matrix of the vth view, there are n views in the multi-view dataset, each view has m feature dimensions, the first p samples have known class labels, the remaining n-p samples are unlabeled, and the number of classes is c; encoding known class label information C for samples with known class labels p×c using an identity matrix I n-p denotes an initial free representation of unknown label samples, the known label part encoding C is constructed from p×c and the unknown label part representation I n-p a multi-view shared label constraint matrix Construct a pair constraint based on samples with known class labels, construct a constraint relationship when samples in a sample pair are or are not samples of the same type, and initialize a pair constraint matrix Z; Introduce projection matrix N for each view (v) Using the projection matrix N (v) The shared label constraint matrix A generates the representation matrix V for each view. (v) =AN (v) By using point constraints to label the low-dimensional representation of the samples, we can ensure that the samples are consistent with the true labels. Initialize a correlation matrix to represent a linear mapping relationship of original data to a low-dimensional space, and generate a consensus matrix by taking an average of representation matrices of each view.

3. The dual constraint concept decomposition based semi-supervised multi-view clustering method according to claim 1, characterized in that, For each view sample data, construct an initial weight matrix based on a neighbor relationship and a kernel function, and generate a final pair constraint matrix through a vertical and horizontal constraint propagation algorithm, specifically: The data matrix of the vth view calculates the similarity between samples by a Gaussian kernel function, and calculates the initial weight matrix of the vth view by the following formula: i Obtains a preset number of nearest neighbors, and constructs a sparse similarity matrix based on the relationship between the sample x i and the nearest neighbor samples, and constructs an initial weight matrix by symmetric normalization; computing a graph Laplacian matrix based on the initial weight matrix, setting elements of a vertical constraint propagation matrix F v and a horizontal constraint propagation matrix F h to random values between [-1, 1], initializing the propagation matrices for pairwise constraint propagation; Iterative updating of the vertical constraint propagation matrix F using the tulip matrix and preset propagation parameters v and the horizontal constraint propagation matrix F h , according to the vertical constraint propagation matrix F v and the horizontal constraint propagation matrix F h Iterate to the post-convergence to get the extreme value to generate the final pair constraint matrix.

4. The dual constraint concept decomposition based semi-supervised multi-view clustering method according to claim 3, characterized in that, Update the weight matrix to enhance the correlation of samples of the same type, and calculate a corresponding graph Laplacian matrix and a diagonal matrix, specifically: Obtain an initial weight matrix and a final pair constraint matrix, obtain an updated weight matrix according to an update rule, enhance the correlation of the same type, calculate a diagonal matrix using the updated weight matrix, and construct an updated graph Laplacian matrix based on the diagonal matrix and the updated weight matrix; The update rule is represented as: where denotes an element in the updated weight matrix, s ij denotes an element in the initial weight matrix, denotes an element in the final pairwise constraint matrix, if then increase strengthening the same class association, if then decrease weakening the same class association.

5. The dual constraint concept decomposition based semi-supervised multi-view clustering method according to claim 1, characterized in that, In an alternating iteration, update the correlation matrix using the shared label constraint matrix while fixing the projection matrix, update the projection matrix using the graph Laplacian matrix and the diagonal matrix while fixing the correlation matrix, and specifically: By alternately iterating the optimization objective function, fixing the projection matrix N in iteration (v) , updating each view correlation matrix W (v) , minimizing the reconstruction error and the point constraint loss, embedding the known class label information into the correlation matrix through the shared label constraint matrix and the projection matrix, ensuring the low-dimensional representation consistency of the known class label samples, the update formula of each view correlation matrix W (v) is expressed as: wherein, denotes an element in the affinity matrix, K (v) denotes a kernel matrix, A denotes a shared label constraint matrix; Fixing the affinity matrix W in iteration (v) , updating each view projection matrix N (v) , the geometric information encoded by the Laplacian matrix guides the unknown class label samples to the correct class aggregation, and the update formula of each view projection matrix N (v) is: wherein, denote elements in the projection matrix, and a, b denote regularization parameters, denote the updated weight matrix, V * denote the updated consensus matrix, A denotes a shared label constraint matrix.

6. The dual constraint concept decomposition based semi-supervised multi-view clustering method according to claim 5, characterized in that, After each iteration, update the consensus matrix based on the projection matrix of each view, and specifically: Obtain the updated projection matrix N of each view (v) and the shared label constraint matrix N (v) Calculate the representation matrix of each view, take the average of the representation matrix of all views, fuse multi-view information, eliminate the difference between views, and obtain a unified final consensus matrix V * , which is represented as: where n v represents the total number of views, v represents the number of view items.

7. The dual constraint concept decomposition based semi-supervised multi-view clustering method according to claim 5, characterized in that, Optimize an objective function through an alternating iteration, and the objective function is represented as: where X (v) represents the data matrix of the vth view, W (v) represents the affinity matrix of the vth view, N (v) represents the projection matrix of the vth view, V * represents the final consensus matrix of the multi-view fusion, a, b represent regularization parameters, represents the graph Laplacian matrix, W (v) , N (v) , V * ≥ 0 represents the non-negativity constraint, ensuring the interpretability of the decomposition result, n v represents the total number of views, Tr represents the trace of the matrix, used for the scalarization of the manifold regularization term.

8. The dual constraint concept decomposition based semi-supervised multi-view clustering method according to claim 1, characterized in that, After the iteration ends, take the final consensus matrix as a low-dimensional representation of the view, perform clustering by using a k-means algorithm, and output a final class division result, and specifically: obtaining a final consensus matrix V * and a preset number of clustering categories, performing normalization processing on the final consensus matrix V * , selecting sample points as initial cluster centers according to the preset number of clustering categories, calculating the Euclidean distance between each sample and all initial cluster centers, and assigning to the nearest initial cluster center to generate data clusters; For each data cluster, recalculate the mean of all samples of the data cluster to update a cluster center, stop when the cluster center changes less than a threshold value or reaches a maximum number of iterations, and obtain a final class label of each sample according to the final clustering result.

Images