Large-scale multi-view clustering method based on global anchor diagram
By adopting global anchor map and iterative adaptive learning methods in multi-view clustering, the calculation and storage problems of large-scale data sets are solved, clustering performance and stability are improved, and effective clustering of large-scale data is achieved. kind.
Patent Information
- Application Number
- CN202410836915.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-26
- Publication Date
- 2024-11-26
AI Technical Summary
Existing multi-view clustering algorithms face problems of high computational costs and storage requirements when processing large-scale data sets, and heuristic anchor selection strategies may lead to suboptimal anchor selection, affecting clustering performance.
A multi-view clustering method based on global anchor map is adopted to obtain the optimal global binary graph matrix through iterative adaptive learning, spectral embedding is performed to obtain clustering results, and a k-means algorithm is combined to improve clustering performance.
It effectively reduces the storage space and computational complexity of data, significantly improves clustering performance and stability, can be applied to large-scale data sets, and significantly improves the accuracy and scalability of clustering results.
Smart Images

Figure CN119027699A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to view data processing, and in particular to a large-scale multi-view clustering method based on a global anchor graph. Background Art
[0002] With the rapid innovation of information technology and the surge in data volume, the amount of data collected from various fields and perspectives is increasing. For example, a 3D object can be described from different perspectives or from different sources. It is the complexity of this data that has prompted most researchers to think deeply about multi-view learning methods. Among these methods, multi-view clustering (MVC) is a key technique that utilizes datasets from multiple perspectives to improve clustering performance. The complementary information provided by different perspectives is considered to be a key factor in improving clustering quality because it reflects the characteristics of different aspects of the same data. However, existing multi-view clustering algorithms face severe challenges when processing large-scale datasets, especially in terms of computational cost and storage requirements.
[0003] Traditional multi-view subspace clustering (MVSC) methods attempt to reveal the intrinsic structure of the data by discovering a shared or a set of related subspaces in each view. These methods usually adopt a self-expressive framework to represent each data point as a linear combination of other points in the same view. Although these methods have achieved some success in multi-view clustering problems, the computational complexity and memory requirements of the global self-expressive framework increase rapidly with the size of the dataset, severely limiting their scalability and practicality.
[0004] In recent years, in order to solve these problems, some studies have begun to explore anchor-based strategies, selecting a small number of representative data points as anchors or landmarks to effectively approximate and reconstruct the original data space. In addition, some studies have proposed using bipartite graphs to simplify multi-view clustering problems, effectively reducing the dimension and computational cost of the problem by connecting data points and views. However, these methods usually rely on heuristic anchor selection strategies such as k-means or random sampling, which may lead to suboptimal anchor selection, thereby affecting clustering performance. Summary of the invention
[0005] The technical purpose of the present invention is to propose a large-scale multi-view clustering method based on a global anchor graph, which is used to solve the problem of excessive computation and storage burden of existing multi-view clustering methods.
[0006] In order to solve the above technical problems, the large-scale multi-view clustering method based on the global anchor graph proposed in this case includes the following steps: based on the multi-view clustering model, using iterative adaptive learning to obtain the optimal global bipartite graph matrix; based on the optimal global bipartite graph matrix, obtaining spectral embedding, and then obtaining clustering results; the multi-view clustering model is as follows:
[0007]
[0008] Where: st represents the constraint condition, α=[α1,α2,…,α v ] T , α i is the view coefficient corresponding to the i-th view, X i represents the data matrix of the i-th view, W i is the projection matrix corresponding to the i-th view, A is the global anchor graph matrix, G is the global bipartite graph matrix, λ is the preset regularization parameter, v is the number of views, k is the preset number of categories, I is the identity matrix, ||·|| F is the Frobenius norm.
[0009] In one implementation of the above technical solution, in one iteration, a four-step alternating optimization algorithm is used to sequentially obtain the current optimal projection matrix, the current optimal global anchor graph matrix, the current optimal global bipartite graph matrix, and the current optimal view coefficient.
[0010] In one implementation of the above technical solution, the optimal global bipartite graph matrix is subjected to singular value decomposition to calculate the spectral embedding.
[0011] In one implementation of the above technical solution, the spectrum embedding is used as an input of a clustering algorithm to obtain a final clustering result.
[0012] In one implementation of the above technical solution, a k-means algorithm is used to obtain clustering results.
[0013] In one implementation of the above technical solution, the current optimal projection matrix, the acquisition step includes converting the multi-view clustering model into the following form:
[0014]
[0015] In the formula, Tr means finding the trace of the matrix; C i Perform singular value decomposition to obtain the left singular value U and the right singular value V; by calculating To obtain the optimal projection matrix corresponding to the i-th view in the current iteration round. In addition, in the first iteration calculation, A and G are the initial values, and in the intermediate iteration, A and G are the calculated values in the previous iteration round.
[0016] In one implementation of the above technical solution, the current optimal global anchor graph matrix, the obtaining step includes converting the multi-view clustering model into the following form:
[0017]
[0018] in: Tr represents the trace of the matrix; D is subjected to singular value decomposition to obtain the left singular value U and the right singular value V, and the optimal global anchor graph matrix in the current iteration round is identified. And, W i is the optimal projection matrix corresponding to the i-th view obtained in the current iteration round, G and α i In the first iteration, it is the initial value; in the intermediate iterations, it is the value calculated in the previous iteration.
[0019] In one implementation of the above technical solution, the current optimal global bipartite graph matrix, the step of obtaining includes converting the multi-view clustering model into the following form:
[0020]
[0021] in, By executing the quadratic programming solver on each row of G, the optimal global bipartite graph matrix in the current iteration round is obtained. And, W i is the optimal projection matrix corresponding to the i-th view obtained in the current iteration round, A is the optimal global anchor map matrix obtained in the current iteration round, α i In the first iteration, it is the initial value; in the intermediate iterations, it is the value calculated in the previous iteration.
[0022] In one implementation of the above technical solution, the current optimal view coefficient adopts an analytical solution: Where: i =||X i -W i AG|| F .
[0023] According to the above method, this case also proposes a corresponding large-scale multi-view clustering system based on a global anchor graph. The system includes an acquisition module, an optimal global bipartite graph matrix acquisition module, and a clustering module; wherein: the acquisition module is configured to acquire v view data matrices; the optimal global bipartite graph matrix acquisition module is configured based on a multi-view clustering model, and uses iterative adaptive learning to obtain the optimal global bipartite graph matrix; the clustering module is configured based on the optimal global bipartite graph matrix to obtain spectral embedding and then obtain clustering results; wherein, the multi-view clustering model is as follows:
[0024]
[0025] Where: st represents the constraint condition, α=[α1,α2,…,α v ] T , α i is the view coefficient corresponding to the i-th view, X irepresents the data matrix of the i-th view, W i is the projection matrix corresponding to the i-th view, A is the global anchor graph matrix, G is the global bipartite graph matrix, λ is the preset regularization parameter, v is the number of views, k is the preset number of categories, I is the identity matrix, ||·|| F is the Frobenius norm.
[0026] The beneficial technical effects of this case are as follows: (1) Through the multi-view clustering model, the underlying anchor graphs in multiple views are adaptively learned by iteration, and the optimal anchor set is automatically determined. (2) The global bipartite graph obtained in this case automatically has a low-rank property, which can not only retain the main features of the original data and achieve effective clustering of large-scale data sets, but also reduce the storage space and computational complexity of the data. (3) By designing constraints for the bipartite graph, the clustering performance and stability are significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.
[0028] Figure 1 , one Schematic diagram of the multi-view clustering model structure in this implementation method.
[0029] Figure 2 , one Schematic diagram of the workflow of multi-view clustering in this implementation mode.
[0030] Figure 3 , one A data set information table used in the embodiment.
[0031] Figure 4 , one Clustering performance table in this implementation manner. DETAILED DESCRIPTION
[0032] In the field of multi-view subspace clustering, mainstream methods rely on the principle of self-expression to express the position of each data point in the global context by considering the entire data population. Although this approach deeply explores the comprehensive relationship between data points and has gained wide acceptance, it inevitably brings huge computational and storage burdens. These challenges become particularly severe when we face the vast field of big data in a multi-view setting, and the redundancy inherent in comprehensive self-expression often outweighs its benefits.
[0033] To address these challenges, this case proposes a method for large-scale multi-view clustering, which is a multi-view clustering method based on a global anchor graph. This method believes that there is a shared global latent data distribution space for multi-view data, and the anchors interact collaboratively in this latent space to reflect the shared information and intrinsic associations across views. Therefore, a multi-view clustering model is designed in this method, in which global anchor selection, bipartite graph construction, and clustering processes are unified. By iteratively adaptively learning the underlying anchor graphs in multiple views, the optimal anchor set is automatically determined, thereby more accurately capturing the underlying structure of the data. The global bipartite graph obtained by this method automatically has a low-rank property, which can not only retain the main features of the original data and achieve effective clustering of large-scale data sets, but also reduce the storage space and computational complexity of the data. By designing constraints for the bipartite graph, the clustering performance and stability are significantly improved.
[0034] The anchor graph mentioned above is a special graph structure that maps data points to a high-dimensional space and establishes connections based on the similarities between data points, while the anchor point is the center point of the cluster.
[0035] The following will be combined with the attached Figure 1-4 , which clearly and completely describes how to implement large-scale view clustering based on the global anchor graph in this case. Obviously, the described implementation is only a part of the implementation of this case, not all of the implementations. Based on the implementation in this case, all other implementations obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0036] (I) Multi-view clustering model
[0037] like Figure 1 As shown in the figure, it is a schematic diagram of the multi-view clustering model in this case. The model inputs v view feature data and records it as Each view corresponds to a learnable view coefficient, denoted as
[0038] The main part of the model is the projection matrix W i (i=1,2,…,v), the global anchor graph matrix A and the global bipartite graph matrix G, the determination process of these three is a process of mutually promoting learning. i=1,2,…,v, v is the number of views. Through the learning of these three, the reconstruction error of the input data is measured until the preset conditions are met, the learning is stopped, and the optimal global bipartite graph is obtained. Then, the singular value decomposition (SVD) is performed on the basis of the optimal global bipartite graph to obtain the spectral embedding H and clustering results.
[0039] The multi-view clustering model is as follows:
[0040]
[0041] In the above formula, st represents the constraint condition. α=[α1,α2,…,α v ] T α i is the view coefficient corresponding to the i-th view. represents the data matrix of the i-th view, is the projection matrix corresponding to the i-th view. is the global anchor matrix, and Represents the global bipartite graph matrix. is a real number field, and its right superscript indicates its dimension size. k is the preset number of categories, and I is the unit matrix. i is the feature dimension of the i-th view, and n is the number of samples. ||·|| F is the Frobenius norm. is the field of real numbers.
[0042] λ is the regularization parameter. It is a regularization component, which, combined with other appropriate constraints, can reduce the complexity of learning and overfitting, thereby making the model applicable to various data environments.
[0043] Next, we will describe how to solve the optimization problem. One possible approach is to use a four-step alternating optimization algorithm, which solves one variable while fixing the other variables.
[0044] Initially, A is a zero matrix, G = [I :,k ,0], W i is a zero matrix. ":,k" indicates the kth column of the matrix.
[0045] (I) When A, G and α are fixed i Solve for W in the case of i
[0046] In this case, the above optimization problem (1) can be simplified to:
[0047]
[0048] In order to eliminate the i Irrelevant terms, the F-norm is expanded into trace form, and the objective function of the simplified optimization problem is equivalent to:
[0049]
[0050] The optimization problem can then be further transformed into:
[0051]
[0052] in,
[0053] Finally, for C i Perform singular value decomposition (SVD) and calculate To obtain the optimal W i , where U and C i The left and right singular values of .
[0054] In the first iteration, A and G are initial values. In the iteration, A and G are the values calculated in the previous iteration.
[0055] (II) At a fixed W i , G and α i Solve A in the case of
[0056] In this case, the above optimization problem (1) can be transformed into:
[0057]
[0058] Similar to W i The above optimization problem (5) is equivalently reformulated as an optimization task that prioritizes trace maximization:
[0059]
[0060] Where D is defined as By performing singular value decomposition on D, we obtain the left singular value U and the right singular value V, and use their left and right singular vectors to identify the global anchor graph matrix A.
[0061] In this step, W i The value of is the value obtained in (I). G and α i In the first iteration round, it is the initial value; in the middle iteration round, it is the value calculated in the previous iteration round.
[0062] (XXI) At fixed W i , A and α i Solve G in the case of
[0063] In this case, updating the matrix G is achieved by considering an augmented objective that includes a regularization component. The optimization task is formulated as follows:
[0064]
[0065] The optimization process of G above can be restated as the following quadratic programming (QP) formula:
[0066]
[0067] The subscript “:,j” represents the jth column of the matrix. The definitions of the matrix M and the vector f are complex and include regularization effects. Specifically, we first calculate the matrix N as Then symmetrize it by averaging its transpose to form The vector f is determined by summing the contributions of all views, The subscript "i[:,j]" represents the j-th column of the i-th view matrix.
[0068] In this step, W i The value of is the value obtained in (I), and the value of A is the value obtained in (II). In the first calculation, In iterative calculation, it is the value of the previous calculation.
[0069] This iterative approach is implemented by executing the QP solver on each row of G (a k-dimensional vector).
[0070] It can be seen from the solved global bipartite graph G that it automatically has a low-rank property, which not only retains the main features of the original data, but also reduces the storage space and computational complexity of the data.
[0071] The entire bipartite graph G forms the basis for the spectral embedding process. For this purpose, the Gram matrix where Δ is a diagonal matrix with entries is equal to the sum of each row of G, where the subscripts i and j are row and column identifiers, and the two subscripts are row and column identifiers respectively. The matrix Δ is essentially a normalization of G, so the following theorem can be applied to achieve higher computational efficiency than the traditional eigendecomposition method to derive the spectral embedding H.
[0072] Theorem: The right singular vectors of G are contained in the singular value decomposition In the matrix V obtained in The eigenvectors are the same.
[0073] This theorem shows that the spectral embedding H can be effectively calculated by performing singular value decomposition on G, which is a key step that makes clustering analysis more efficient and accurate when dealing with large-scale data.
[0074] (IV) At a fixed W i , A and G to solve α i
[0075] In this case, the goal of the above optimization problem (1) is transformed into minimizing the weighted sum of square norms under the normalization constraint of the coefficients, that is, it is transformed into the following problem:
[0076]
[0077] Among them, ξi represents the Frobenius norm of the residual matrix of the i-th view, denoted as ξ i =||X i -W i AG|| F .
[0078] Using the Cauchy-Schwarz inequality, we can get the optimal α i The analytical solution is:
[0079]
[0080] In this step, W i , A and G are the values in the above steps in the current calculation.
[0081] So far, the above content is the process of solving the optimization problem (1). Each time W, A, G, α is calculated i It should be noted that when calculating A in S060, the W obtained in S040 is used. i When calculating G in S080, use the W obtained in S040 i , A obtained in S060. Calculate α in S100 i When using W obtained in S040 i , A obtained in S060, G obtained in S080. By calculating W i , A, G, α i The value of and the application of the solution of the previous parameter in the solution of the latter parameter can ensure that the result of the optimization problem of formula (1) has convergence, so that the optimal solution can be found. The detailed process of solving the optimal result of the optimization problem (1) is as follows Figure 2 As shown, the steps include:
[0082] S000, input data of v views As well as the number of categories k and the parameter λ.
[0083] S020, initialize W, A, G,
[0084] S040, solve W through formula (4) i Until convergence, the current optimal projection matrix is obtained.
[0085] S060. Solve A through equation (6) until convergence to obtain the current optimal global anchor graph matrix.
[0086] S080. Solve G through equation (8) until convergence to obtain the current optimal global bipartite graph matrix.
[0087] It is determined whether the preset termination condition is met, and if so, S120 is executed. Otherwise, S100 is executed.
[0088] S100, solve α through formula (10) i Until convergence, the current optimal view coefficient is obtained. Return to S040.
[0089] S120, obtaining the spectral embedding H by performing SVD on the optimal global bipartite graph matrix.
[0090] In step S120, the optimal global bipartite graph matrix finally obtained is recorded as S, and the spectral embedding H satisfies the following formula:
[0091]
[0092] The above function can be solved by performing SVD decomposition on S to obtain the spectral embedding H. According to the solution, the H matrix is the first k columns of the left singular value matrix U.
[0093] The spectral embedding H is used as the input of the clustering algorithm to obtain the final clustering result. The clustering algorithm is preferably k-means.
[0094] The following is a time complexity analysis of the calculations involved in the clustering method in this case.
[0095] (1) Solve W i complexity.
[0096] Calculating the Matrix The time complexity required is C i Performing singular value decomposition (SVD) requires time, because C i is a i ×k-sized matrix, the main cost is the k singular values and vectors. Matrix multiplication is then performed to update W i , considering that d i The product of ×k and k×k matrices requires operation.
[0097] (2) Solve the complexity of A.
[0098] Calculating the Matrix The time complexity required is The SVD required to update A has The computational complexity is because it is necessary to decompose a k×k matrix D. Therefore, the matrix multiplication involved in solving A requires additional time.
[0099] (3) Solve the complexity of G.
[0100] The complexity of solving the quadratic programming (QP) problem for G is This applies to every column for all samples, dominated by the inverse operations required during the optimization.
[0101] (4) Solving α i complexity.
[0102] Calculate α i is a direct operation that produces only The complexity of each step is summarized, and the total time cost T of the optimization phase can be expressed as:
[0103]
[0104] This sum simplifies to in is the dimension of all views. Assume k is significantly smaller than n and d i , the complexity can be approximated as A quasi-linear relationship with the sample size is shown.
[0105] (5) The complexity of obtaining the spectrum embedded in H in this case.
[0106] In this case, the optimal global bipartite graph matrix is decomposed by SVD to obtain the left singular matrix U. Considering that the dimension of the global bipartite graph matrix G is k×n, the time complexity of obtaining the left singular matrix U is
[0107] Assuming that convergence requires a fixed number of iterations, the complexity of k-means clustering applied to the H matrix is Therefore, the optimized process does not exceed the quasi-linear complexity observed in the optimization phase, thus improving the overall efficiency of the algorithm.
[0108] In summary, the proposed method for obtaining the spectral embedding H is generally more cubic than the conventional methods. It is an improvement and more scalable for large-scale clustering tasks.
[0109] (II) Experiment
[0110] (2.1) Experimental setup
[0111] Datasets: In the experimental evaluation of multi-view clustering algorithms, datasets of different scopes are key to comprehensive analysis. The datasets used in this study are Figure 3 shown.
[0112] Comparison algorithms: Nine mature multi-view clustering algorithms were selected for comparative analysis with this case. These nine algorithms are: partition-level multi-view subspace clustering (PMSC), multi-partition alignment clustering (mPAC), multi-view low-rank sparse subspace clustering (MLRSSC), parameter-free automatic weighted multi-graph learning (AMGL), scalable and parameter-free multi-view clustering bipartite graph fusion method (SFMC), multi-view k-means clustering tailored for big data (RMKM), binary multi-view clustering (BMVC), large-scale multi-view subspace clustering optimized for linear time (LMVSC), and fast parameter-free multi-view subspace clustering (FPMVS). Among these methods, PMSC, mPAC, and MLRSSC are classified as classic subspace-based multi-view clustering techniques. LMVSC, FPMVS, and the method proposed in this case are classified as anchor-based subspace methods. Finally, methods such as SFMC, RMKM, and BMVC adopt techniques such as bipartite graphs, k-means, or binarization to address the challenge of large-scale multi-view clustering.
[0113] (2.2) Experimental results
[0114] Clustering performance: Figure 4 Detailed description is given in the table. '-' indicates insufficient memory. ACC, NMI, Purity, and Fscore are four clustering performance evaluation indicators. The higher the value, the better the clustering result.
[0115] from Figure 4 As can be seen from the results, the proposed method has shown excellent performance in the field of multi-view clustering. This is especially evident on the Caltech101-20 dataset, where the proposed method achieved the highest scores in all indicators, enhancing its ability to identify and adapt to the intrinsic clustering structure of the data. In addition, when considering larger datasets such as CCV and NUS-WIDE-OBJ, the proposed method also maintained high performance, showing its robustness and adaptability to different data complexities. It is worth noting that due to the unique challenges of high dimensionality, the proposed method also achieved impressive results on the MNIST dataset, highlighting its effectiveness in distinguishing features that are critical to clustering tasks. The consistency of these results on multiple datasets demonstrates the versatility of the algorithm and confirms its practical application potential in various multi-view learning scenarios. The improvement in clustering performance of the proposed method, especially in ACC (accuracy), demonstrates that the proposed method has the ability to effectively utilize potential information between multiple views.
[0116] Through the description of the above embodiments, it can be clearly understood by those skilled in the art that a general clustering model is designed according to the disclosed method, by integrating learning-based anchor point selection and multi-view Figure 2The sub-graph not only realizes the multi-view clustering function enhancement and the mining of the underlying structure data, has a good clustering effect, but also reduces the amount of calculation and storage requirements. In the clustering model, the designed constraint strategy makes the model have good generalization ability and scalability, and the convergence of each part of the data in the model provides a guarantee for the clustering accuracy and robustness of the clustering performance.
[0117] Through the description of the above implementation modes, those skilled in the art can clearly understand that according to the method disclosed in the present invention, a corresponding large-scale multi-view clustering system based on a global anchor graph can be implemented.
[0118] Through the description of the above implementation modes, those skilled in the art can clearly understand that the method or system of the present disclosure can be implemented by means of software plus necessary general hardware, and of course can also be implemented by special hardware including special integrated circuits, special CPUs, special memories, special components, etc. In general, all functions performed by computer programs can be easily implemented by corresponding hardware, and the specific hardware structures used to implement the same function can also be various, such as analog circuits, digital circuits or special circuits, etc. However, for the present disclosure, software program implementation is a better implementation mode in most cases.
[0119] Although the embodiments of the present invention are described above in conjunction with the accompanying drawings, the present invention is not limited to the above specific embodiments and application fields, and the above specific embodiments are only illustrative and instructive, rather than restrictive. A person of ordinary skill in the art can also make many forms under the guidance of this specification and without departing from the scope of protection of the claims of the present invention, all of which belong to the protection of the present invention.
Claims
1. A large-scale multi-view clustering method based on global anchor graph, It is characterized in that The method comprises the following steps: Based on the multi-view clustering model, the optimal global bipartite graph matrix is obtained by iterative adaptive learning; Based on the optimal global bipartite graph matrix, the spectral embedding is obtained, and then the clustering results are obtained; The multi-view clustering model is as follows: Where: st represents the constraint condition, α=[α 1 , α 2 , …, α r ] T , α i is the view coefficient corresponding to the i-th view, X i represents the data matrix of the i-th view, W i is the projection matrix corresponding to the i-th view, A is the global anchor graph matrix, G is the global bipartite graph matrix, λ is the preset regularization parameter, v is the number of views, k is the preset number of categories, I is the identity matrix, ||·|| F is the Frobenius norm.
2. The method according to claim 1, It is characterized in that In one iteration, a four-step alternating optimization algorithm is used to obtain the current optimal projection matrix, the current optimal global anchor graph matrix, the current optimal global bipartite graph matrix, and the current optimal view coefficient in turn.
3. The method according to claim 1, It is characterized in that The optimal global bipartite graph matrix is subjected to singular value decomposition to calculate the spectral embedding.
4. The method according to claim 1, It is characterized in that The spectral embedding is used as the input of the clustering algorithm to obtain the final clustering results.
5. The method according to claim 1, It is characterized in that The k-means algorithm is used to obtain clustering results.
6. The method according to claim 2, It is characterized in that The current optimal projection matrix, the acquisition step includes converting the multi-view clustering model into the following form: In the formula, Tr means finding the trace of the matrix; C i Perform singular value decomposition to obtain the left singular value U and the right singular value V; By calculation To obtain the optimal projection matrix corresponding to the i-th view in the current iteration round; Furthermore, in the first iteration, A and G are initial values, and in the intermediate iteration, A and G are calculated values in the previous iteration.
7. The method according to claim 2, It is characterized in that The current optimal global anchor graph matrix is obtained by converting the multi-view clustering model into the following form: in: Tr means finding the trace of the matrix; Perform singular value decomposition on D to obtain the left singular value U and the right singular value V, and identify the optimal global anchor graph matrix in the current iteration round; And, W i is the optimal projection matrix corresponding to the th view obtained in the current iteration round, G and α i In the first iteration, it is the initial value; in the intermediate iterations, it is the value calculated in the previous iteration.
8. The method according to claim 2, It is characterized in that The current optimal global bipartite graph matrix is obtained by converting the multi-view clustering model into the following form: in, By executing the quadratic programming solver on each row of G, the optimal global bipartite graph matrix in the current iteration round is obtained; And, W i is the optimal projection matrix corresponding to the i-th view obtained in the current iteration round, A is the optimal global anchor map matrix obtained in the current iteration round, α i In the first iteration, it is the initial value; in the intermediate iterations, it is the value calculated in the previous iteration.
9. The method according to claim 2, It is characterized in that The current optimal view coefficient uses an analytical solution: Where: i =||X i -W i AG|| F .
10. A large-scale multi-view clustering system based on global anchor graph, Features: The system includes an acquisition module, an optimal global bipartite graph matrix acquisition module, and a clustering module; wherein: The acquisition module is configured to acquire v view data matrices; The optimal global bipartite graph matrix acquisition module is configured to obtain the optimal global bipartite graph matrix based on a multi-view clustering model using iterative adaptive learning; The clustering module is configured to obtain spectral embedding based on an optimal global bipartite graph matrix, and then obtain a clustering result; Among them, the multi-view clustering model is as follows: Where: st represents the constraint condition, α=[α 1 , α 2 , …, α r ] T , α i is the view coefficient corresponding to the i-th view, X i represents the data matrix of the i-th view, W i is the projection matrix corresponding to the i-th view, A is the global anchor graph matrix, G is the global bipartite graph matrix, λ is the preset regularization parameter, v is the number of views, k is the preset number of categories, I is the identity matrix, ||·|| F is the Frobenius norm.
