Incomplete multi-view clustering method and system based on robust high-order graph learning
Through the robust high-order graph learning method, combined with robust principal component analysis and low-rank tensor learning, the problem of unconsidered high-order correlation between views in incomplete multi-view clustering is solved, and high-quality clustering performance is achieved .
Patent Information
- Application Number
- CN202411023166.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-29
- Publication Date
- 2024-11-22
AI Technical Summary
The existing incomplete multi-view clustering algorithm fails to effectively consider the higher-order correlation between different views, affecting the clustering effect and accuracy.
Using a method based on robust high-order graph learning, combining robust principal component analysis, low-rank tensor learning and adaptive weight allocation strategies, a high-quality complete similarity graph is constructed, and satisfactory clustering is obtained through spectral clustering. .
By integrating local similarity structures, basic divisions and complete similarity diagrams, fully consider the internal structure of the data, reduce noise, explore the local relationships of the samples, reveal high-order information between views, and achieve high-quality clustering results.
Smart Images

Figure CN119004152A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the technical field of incomplete multi-view clustering, and in particular to an incomplete multi-view clustering method and system based on robust high-order graph learning. Background Art
[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.
[0003] With the rapid development of multimedia technology, it has become easier to collect data from various resources, which often contain information from multiple views. For example, a piece of news can be reported from different angles such as text, images, and videos. Multi-view data can describe the information of things more comprehensively and accurately. Since multi-view data contains heterogeneous information, it has greater advantages than traditional single-view data. Therefore, multi-view learning has become an important research direction, covering many fields such as multi-view clustering and multi-view retrieval. However, due to equipment failure or other uncontrollable factors, multi-view data may be missing. It is challenging to use traditional clustering algorithms to handle these missing views.
[0004] As an effective method to solve these problems, the incomplete multi-view clustering algorithm has received widespread attention in multi-view learning. The incomplete multi-view clustering algorithm is a clustering method that processes multi-view data. Since it is difficult to process only two views in practical applications, it is also extremely difficult to manually obtain paired view samples. Therefore, the incomplete multi-view clustering method is more practical. At present, the existing incomplete multi-view clustering algorithms reveal the relationship between all samples by learning the similarity matrix of the view, assigning reasonable weights, and seeking a common graph. Although the previous methods have achieved good results, they did not take into account the high-order correlation between different views, which may affect the effect and accuracy of view clustering. Summary of the invention
[0005] In view of the above defects, the present invention proposes an incomplete multi-view clustering method and system based on robust high-order graph learning, which integrates robust representation, local similarity structure, basic partitioning, and complete similarity graph into a unified learning framework, combines robust principal component analysis and low-rank tensor learning to obtain a high-quality complete similarity graph and obtains satisfactory clustering performance through spectral clustering.
[0006] In order to achieve the above objectives, the present disclosure adopts the following technical solutions:
[0007] The first aspect of the present disclosure provides an incomplete multi-view clustering method based on robust high-order graph learning, comprising the following steps:
[0008] Adaptive graph learning is performed on the acquired incomplete multi-view dataset to obtain a local similarity graph, spectral clustering is performed on the local similarity graph of each view to obtain a pseudo label matrix for each view, the pseudo label matrix and the indicator matrix of the incomplete multi-view dataset are used to construct a first complete similarity graph, the first complete similarity graph is stacked and a low-rank tensor constraint is imposed on it to obtain high-order correlations between views, and an objective function is constructed based on the above operations;
[0009] Solve the objective function to obtain the second complete similarity graph;
[0010] The second complete similarity graph of each view is summed to obtain a public similarity graph, and spectral clustering is performed on the public similarity graph to obtain the final result.
[0011] As a further implementation method, an incomplete multi-view dataset is obtained, specifically:
[0012] A complete multi-view dataset is obtained, and random missing samples within the view are processed for the complete multi-view dataset to obtain an incomplete multi-view dataset and an indicator matrix.
[0013] As a further implementation method, before performing adaptive graph learning on the acquired incomplete multi-view dataset, it also includes performing robust principal component analysis on the incomplete multi-view dataset to obtain clean incomplete multi-view data, and performing adaptive graph learning on the clean incomplete multi-view data to obtain a local similarity graph.
[0014] As a further implementation method, the dimensions of the partial pseudo-label matrices under each view have different sizes. The pseudo-label matrix dimensions of each view are aligned by introducing an indicator matrix and a complete similarity graph is learned to obtain the first complete similarity graph. The objective function of the complete similarity graph learning is as follows:
[0015]
[0016] Among them, G (v) Represents the indicator matrix that records the index of existing samples in each view, S (v) represents the complete similarity graph of the v-th view, F (v) Represents the partial pseudo-label matrix of the v-th view.
[0017] As a further implementation method, an adaptive weight assignment strategy is introduced to provide reasonable weights for the complete similarity graph learning of each view. The objective function of the adaptive weight assignment strategy is:
[0018]
[0019] Among them, G (v) Represents the indicator matrix that records the index of existing samples in each view, S(v) represents the complete similarity graph of the v-th view, F (v) represents the partial pseudo label matrix of the vth view, w v is the weight factor of the vth view.
[0020] As a further implementation, the first complete similarity graph is stacked and a low-rank tensor constraint is imposed on it to obtain high-order correlation between views, specifically:
[0021] The first complete similarity graph of all views is organized into a third-order tensor, and the tensor nuclear norm is applied to the constructed tensor to seek low rank and explore high-order correlations between views. The objective function is:
[0022]
[0023] Among them, Φ[] represents the construction of a third-order tensor.
[0024] As a further implementation method, the objective function is constructed as:
[0025]
[0026] Among them, Z (v) represents the local similarity graph of the v-th view, U (v) represents the robust representation of the v-th view, E (v) represents the error matrix of the vth view, α and β represent weight parameters, and L Z(v) Representation and local similarity graph Z (v) The corresponding Laplacian matrix is, represents the i-th robust representation under the v-th view, Representation sample and The similarities between 1 represents the i-th vector of the local similarity graph, X (v) is the original data, λ is the weight parameter for balancing the robust representation and the error matrix, Tr represents the matrix trace operation, and I c Represents the identity matrix of dimension c.
[0027] A second aspect of the present disclosure provides an incomplete multi-view clustering system based on robust high-order graph learning, comprising:
[0028] The objective function construction module is configured to: perform adaptive graph learning on the acquired incomplete multi-view data set to obtain a local similarity graph, perform spectral clustering operation on the local similarity graph of each view to obtain a pseudo label matrix of each view, construct a first complete similarity graph using the pseudo label matrix and the indicator matrix of the incomplete multi-view data set, stack the first complete similarity graph and impose a low-rank tensor constraint on it to obtain high-order correlation between views, and construct an objective function based on the above operations;
[0029] The solving module is configured to: solve the objective function to obtain a second complete similarity graph;
[0030] The clustering module is configured to: sum the second complete similarity graph of each view to obtain a common similarity graph, and perform spectral clustering on the common similarity graph to obtain a final result.
[0031] A third aspect of the present disclosure provides a medium on which a program is stored. When the program is executed by a processor, the steps of the incomplete multi-view clustering method based on robust high-order graph learning described in the first aspect of the present disclosure are implemented.
[0032] The fourth aspect of the present disclosure provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of an incomplete multi-view clustering method based on robust high-order graph learning described in the first aspect of the present disclosure are implemented.
[0033] Compared with the prior art, the present invention has the following beneficial effects:
[0034] The incomplete multi-view clustering method based on robust high-order graph learning proposed in this paper integrates the tasks of learning local similarity structures, basic partitions, and complete similarity graphs into a unified learning framework. High-quality similarity graphs are learned by jointly using robust principal component analysis and low-rank tensors, and satisfactory clustering performance is achieved through spectral clustering.
[0035] The incomplete multi-view clustering method based on robust high-order graph learning proposed in the present disclosure uses robust principal component analysis to learn robust latent representations, which can fully consider the intrinsic structure of the data and reduce data noise. It explores the local relationship of available samples through an adaptive neighbor assignment strategy, and uses rank constraints to accurately divide samples into ideal clusters. It explores the local relationship of all samples through complete similarity graph learning, and reveals the hidden high-order information between different views by minimizing the tensor constructed by the complete similarity graphs of different views. Finally, the complete similarity graphs of different views are summed to obtain a consistent similarity graph, and the final clustering result is obtained through spectral clustering on the consistent similarity graph, obtaining satisfactory clustering performance.
[0036] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] The accompanying drawings constituting a part of the present disclosure are used to provide a further understanding of the present disclosure. The illustrative embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation on the present disclosure.
[0038] Figure 1 The process of the incomplete multi-view clustering method based on robust high-order graph learning disclosed in this paper is Figure 1 ;
[0039] Figure 2 The process of the incomplete multi-view clustering method based on robust high-order graph learning disclosed in this paper is Figure 2 . DETAILED DESCRIPTION
[0040] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.
[0041] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present disclosure belongs.
[0042] In the absence of conflict, the embodiments in the present disclosure and the features in the embodiments may be combined with each other.
[0043] Embodiment 1
[0044] like Figure 1-2 As shown, the first embodiment of the present disclosure provides an incomplete multi-view clustering method based on robust high-order graph learning, which mainly includes the following steps:
[0045] S1 performs adaptive graph learning on the acquired incomplete multi-view dataset to obtain a local similarity graph, performs spectral clustering on the local similarity graph of each view to obtain a pseudo label matrix for each view, constructs a first complete similarity graph using the pseudo label matrix and the indicator matrix of the incomplete multi-view dataset, stacks the first complete similarity graph and applies a low-rank tensor constraint to it to obtain high-order correlations between views, and constructs an objective function based on the above operations;
[0046] S2 solves the objective function and obtains the second complete similarity graph;
[0047] S3 sums the second complete similarity graph of each view to obtain a public similarity graph, and performs spectral clustering on the public similarity graph to obtain the final result.
[0048] First, S1 obtains an incomplete multi-view dataset, specifically:
[0049] A complete multi-view dataset is obtained, and random missing samples in the view are processed for the complete multi-view dataset to obtain an incomplete multi-view dataset and an indicator matrix.
[0050] Robust principal component analysis is performed on the acquired incomplete multi-view dataset to obtain clean incomplete multi-view data.
[0051] Robust principal component analysis is a method that can enhance the robustness of a model to outliers and noise in the data. Studies have shown that robust principal component analysis improves the generalization ability of the model by capturing the intrinsic structure of the data. In this field, imposing low-rank constraints is a key technology. Low-rank constraints assume that the data has a low-dimensional linear structure and simplify the feature representation, thereby reducing model complexity, improving generalization ability, and helping to extract more stable and interpretable features. In short, low-rank constraints help us extract concise and effective information from complex data, making the model more robust in the face of various challenges. Therefore, the objective function of robust principal component analysis is:
[0052]
[0053] stX( v )=U( v )+E( v )
[0054] Among them, U (v) represents the robust representation learned from the original data and E (v) represents the error in the original data, λ is the weight parameter for balancing the robust representation and the error matrix, and X (v) is the original data.
[0055] Adaptive graph learning on clean incomplete multi-view data to obtain local similarity graphs.
[0056] In order to explore the local structure inside the data, the clean data obtained by robust principal component analysis is used for adaptive graph learning. Therefore, in the present invention, the local similarity graph is learned, and the specific goals are as follows:
[0057]
[0058] Among them, Z ( v ) represents the local similarity graph learned from the robust representation, represents the i-th robust representation under the v-th view, α represents the weight parameter, Representation sample and The similarities between 1 represents the i-th vector of the local similarity graph.
[0059] In order to make the learned local similarity graph of each view have an ideal clustering structure, a low-rank constraint is imposed on it. The objective function is:
[0060]
[0061] Among them, α and β represent weight parameters. Representation and local similarity graph Z (v) The corresponding Laplacian matrix is, and is a diagonal matrix. Its i-th diagonal entry is expressed as However, directly solving the constraint rank of the above equation is a complex and challenging task. Representation Matrix The i-th smallest eigenvalue, we can constrain Convert to Obviously, when The rank constraint can be satisfied. v represents the number of existing samples in the vth view, and c represents the number of clusters.
[0062] Furthermore, by further transforming the constraints into spectral clustering, we can obtain the following objective function:
[0063]
[0064] F (v)T F (v) =I c .
[0065] in, represents the partial pseudo label matrix of the v-th view, is the Z in the vth view (v) The Laplace matrix, Z (v)T is the transpose of the local similarity graph, is a diagonal matrix and Tr represents the matrix trace operation, I c Represents the identity matrix of dimension c.
[0066] Considering that the dimensions of the partial pseudo-label matrices under each view of the learning are of different sizes, the connection between the missing samples and the existing samples cannot be established. By introducing the indicator matrix to align the dimensions of each view and perform complete similarity graph learning, the objective function of the complete similarity graph learning is as follows:
[0067]
[0068] Among them, G (v) Represents the indicator matrix that records the index of existing samples in each view, S (v) represents the complete similarity graph of the v-th view, F (v) Represents the partial pseudo-label matrix of the v-th view.
[0069] Considering the different importance of different views, treating each view equally may lead to suboptimal clustering results. Therefore, an adaptive weight allocation strategy is introduced, and its objective function is:
[0070]
[0071] Among them, G (v) Represents the indicator matrix that records the index of existing samples in each view, S (v) represents the complete similarity graph of the v-th view, F (v) represents the partial pseudo label matrix of the vth view, w v is the weight factor of the vth view.
[0072] Obviously, w v Depends on S (v) With G (v) F (v) F (v)T G (v) The above formula provides a reasonable adaptive weighting scheme for learning the complete similarity graph. When the difference is small, the weight of the corresponding view becomes larger. This makes reliable views more important in graph structure learning.
[0073] In order to consider the inter-view connections and ensure the global consistency of the complete similarity graph, the complete similarity graph of all views is organized into a third-order tensor. Then, the tensor nuclear norm is imposed on this constructed tensor to seek low rank and effectively explore high-order correlations between views.
[0074] The objective function is:
[0075]
[0076] Among them, Φ[] represents the construction of a third-order tensor.
[0077] By integrating robust principal component analysis, adaptive graph learning, spectral clustering, complete similarity graph learning, adaptive weight assignment strategy and low-rank tensor learning into a unified learning framework, the overall objective function of incomplete multi-view clustering based on robust high-order graph learning is obtained as follows:
[0078]
[0079] Among them, Z (v)represents the local similarity graph of the v-th view, U (v) represents the robust representation of the v-th view, E (v) represents the error matrix of the vth view, α and β represent weight parameters, Representation and local similarity graph Z (v) The corresponding Laplacian matrix.
[0080] The first three items aim to remove noise and outliers while exploring the local similarity structure of available samples in the original data, and the fourth item obtains a partial pseudo-label matrix for each view. The last two items explore the correlation between all samples through complete similarity graph learning and introduce the tensor nuclear norm to mine high-order connection relationships between views. This approach integrates robust principal component analysis, adaptive graph learning, spectral clustering, complete similarity graph learning, adaptive weight assignment strategy, and low-rank tensor learning into a unified framework and optimizes them in a mutually reinforcing manner. Finally, we calculate the public similarity graph as And spectral clustering is applied to obtain the final clustering results.
[0081] S2 solves the objective function and obtains the second complete similarity graph.
[0082] An iterative optimization strategy and augmented Lagrange multiplier method are used to solve the objective function of incomplete multi-view clustering based on robust high-order graph learning, and the second complete similarity graph is obtained.
[0083] Step 2.1: Introduce an auxiliary tensor To separate the overall objective function, it can be reformulated as:
[0084]
[0085] Furthermore, the augmented Lagrangian function of the objective function is as follows:
[0086]
[0087] Among them, μ>0 represents the penalty factor, and the matrix Y (v) and tensors denotes two Lagrange multipliers.
[0088] Step 2.2: Fix other variables and solve for the robust representation U (v) :
[0089]
[0090] By calculating U (V) The derivative of is set to zero, then the optimal solution is:
[0091]
[0092] Step 2.3: Fix other variables and solve the error matrix E (v) , we can get:
[0093]
[0094] E (v) The optimal solution is:
[0095]
[0096] Step 2.4: Fix other variables and solve the local similarity graph Z (v) , retain and Z (v) The relevant specific items are:
[0097]
[0098] because The equation can be transformed:
[0099]
[0100] make and You can get:
[0101]
[0102] let Can be further converted to:
[0103]
[0104] The closed-form solution can be obtained:
[0105]
[0106] Step 2.5: Discard irrelevant variables and solve the complete similarity graph S (v) , the above augmented Lagrangian function is expressed as follows:
[0107]
[0108] Let M (v) =G (v)T F (v) F (v)T G (v) , Q (v) =J (v)T -W (v) / μ, S (v) The derivative of is zero, so:
[0109] S (v) =(2w v M (v)+μQ (v) ) / (2w v +μ)
[0110] Step 2.6: Fix other variables and solve the partial pseudo-label matrix F (v) , F (v) The following updates are available:
[0111]
[0112] Ignore the second term w in the above equation v The following equation is obtained:
[0113]
[0114] Remove the irrelevant terms in the above formula (Tr(S (v)T S (v) )), the remaining items are substituted into the update to get:
[0115]
[0116] In the formula F (v) The optimal solution is D (v) The c eigenvectors corresponding to the c smallest eigenvalues.
[0117] Step 2.7: Fix other variables, about The problem can be stated as:
[0118]
[0119] in, The above model can be solved using the tensor tube reduction operator:
[0120]
[0121] in, represents the tensor singular value decomposition, represents a diagonal tensor, whose diagonal elements in the Fourier domain are expressed as
[0122] Step 2.8: Fix other variables and solve for the Lagrange multiplier Y (v) , and parameter muon problem, these three variables can be updated as:
[0123] Y (v) =Y (v) +μ(X (v) -U (v) -E (v) )
[0124]
[0125] μ=min(ρμ,μ0)
[0126] Among them, ρ and μ0 are constants.
[0127] S3 sums the second complete similarity graph of each view to obtain a public similarity graph, and performs spectral clustering on the public similarity graph to obtain the final result.
[0128] After convergence, the above solution obtains a complete similarity graph, establishing the connection between missing and existing samples. The complete similarity graphs of different views are summed to obtain a common similarity graph, and then spectral clustering is performed on the common similarity graph to obtain the final clustering result.
[0129] The performance of the method is evaluated using clustering tasks. Three commonly used clustering evaluation indicators are used to evaluate the clustering effect: clustering accuracy (ACC), normalized mutual information (NMI) and purity. Among them, ACC represents the proportion of the number of correctly predicted data samples to the total number of data samples; NMI is used to measure the degree of correlation between two random variables, that is, the amount of information about another random variable in one random variable; purity is used to calculate the proportion of correctly clustered samples to the total number of samples. The numerical range of these three evaluation indicators is between 0 and 1, and the larger the value, the better the performance.
[0130] Embodiment 2
[0131] Embodiment 2 of the present disclosure provides an incomplete multi-view clustering system based on robust high-order graph learning, including:
[0132] The objective function construction module is configured to: perform adaptive graph learning on the acquired incomplete multi-view data set to obtain a local similarity graph, perform spectral clustering operation on the local similarity graph of each view to obtain a pseudo label matrix of each view, construct a first complete similarity graph using the pseudo label matrix and the indicator matrix of the incomplete multi-view data set, stack the first complete similarity graph and impose a low-rank tensor constraint on it to obtain high-order correlation between views, and construct an objective function based on the above operations;
[0133] The solving module is configured to: solve the objective function to obtain a second complete similarity graph;
[0134] The clustering module is configured to: sum the second complete similarity graph of each view to obtain a common similarity graph, and perform spectral clustering on the common similarity graph to obtain a final result.
[0135] The more detailed steps are the same as those in the first embodiment and will not be repeated here.
[0136] Embodiment 3
[0137] Embodiment 3 of the present disclosure provides a medium on which a program is stored. When the program is executed by a processor, the steps of an incomplete multi-view clustering method based on robust high-order graph learning as described in Embodiment 1 of the present disclosure are implemented.
[0138] The more detailed steps are the same as those in the first embodiment and will not be repeated here.
[0139] Embodiment 4
[0140] Embodiment 4 of the present disclosure provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of an incomplete multi-view clustering method based on robust high-order graph learning as described in Embodiment 1 of the present disclosure are implemented.
[0141] The more detailed steps are the same as those in the first embodiment and will not be repeated here.
[0142] The above description is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. For those skilled in the art, the present disclosure may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.
Claims
1. Incomplete multi-view clustering method based on robust high-order graph learning, It is characterized in that The following steps are involved: Adaptive graph learning is performed on the acquired incomplete multi-view dataset to obtain a local similarity graph, spectral clustering is performed on the local similarity graph of each view to obtain a pseudo label matrix for each view, the pseudo label matrix and the indicator matrix of the incomplete multi-view dataset are used to construct a first complete similarity graph, the first complete similarity graph is stacked and a low-rank tensor constraint is imposed on it to obtain high-order correlations between views, and an objective function is constructed based on the above operations; Solve the objective function to obtain the second complete similarity graph; The second complete similarity graph of each view is summed to obtain a public similarity graph, and spectral clustering is performed on the public similarity graph to obtain the final result.
2. The incomplete multi-view clustering method based on robust high-order graph learning as claimed in claim 1, It is characterized in that Get an incomplete multi-view dataset, specifically: A complete multi-view dataset is obtained, and random missing samples in the view are processed for the complete multi-view dataset to obtain an incomplete multi-view dataset and an indicator matrix.
3. The incomplete multi-view clustering method based on robust high-order graph learning as claimed in claim 1, It is characterized in that Before performing adaptive graph learning on the acquired incomplete multi-view dataset, robust principal component analysis is performed on the incomplete multi-view dataset to obtain clean incomplete multi-view data, and adaptive graph learning is performed on the clean incomplete multi-view data to obtain a local similarity graph.
4. The incomplete multi-view clustering method based on robust high-order graph learning as claimed in claim 1, It is characterized in that The dimensions of the partial pseudo-label matrices under each view have different sizes. By introducing the indicator matrix to align the dimensions of the pseudo-label matrices of each view and performing complete similarity graph learning, the first complete similarity graph is obtained. The objective function of the complete similarity graph learning is as follows: Among them, G (v) Represents the indicator matrix that records the index of existing samples in each view, S (v) represents the complete similarity graph of the v-th view, F (v) Represents the partial pseudo-label matrix of the v-th view.
5. The incomplete multi-view clustering method based on robust high-order graph learning as claimed in claim 4, It is characterized in that An adaptive weight assignment strategy is introduced to provide reasonable weights for the complete similarity graph learning of each view. The objective function of the adaptive weight assignment strategy is: Among them, G (v) Represents the indicator matrix that records the index of existing samples in each view, S (v) represents the complete similarity graph of the v-th view, F (v) represents the partial pseudo label matrix of the vth view, w v is the weight factor of the vth view.
6. The incomplete multi-view clustering method based on robust high-order graph learning as claimed in claim 1, It is characterized in that The first complete similarity graph is stacked and a low-rank tensor constraint is imposed on it to obtain high-order correlations between views, specifically: The first complete similarity graph of all views is organized into a third-order tensor, and the tensor nuclear norm is applied to the constructed tensor to seek low rank and explore high-order correlations between views. The objective function is: S=Φ[S (1) ,S (2) ,...,S (v) ]. Among them, Φ[] represents the construction of a third-order tensor.
7. The incomplete multi-view clustering method based on robust high-order graph learning as claimed in claim 1, It is characterized in that The objective function is constructed as: Among them, Z (v) represents the local similarity graph of the v-th view, U (v) represents the robust representation of the v-th view, E (v) represents the error matrix of the vth view, α and β represent weight parameters, Representation and local similarity graph Z (v) The corresponding Laplacian matrix is, represents the i-th robust representation under the v-th view, Representation sample and The similarities between 1 represents the i-th vector of the local similarity graph, X (v) is the original data, λ is the weight parameter for balancing the robust representation and the error matrix, Tr represents the matrix trace operation, and I c Represents the identity matrix of dimension c.
8. Incomplete multi-view clustering system based on robust high-order graph learning, It is characterized in that include: The objective function construction module is configured to: perform adaptive graph learning on the acquired incomplete multi-view data set to obtain a local similarity graph, perform spectral clustering operation on the local similarity graph of each view to obtain a pseudo label matrix of each view, construct a first complete similarity graph using the pseudo label matrix and the indicator matrix of the incomplete multi-view data set, stack the first complete similarity graph and impose a low-rank tensor constraint on it to obtain high-order correlation between views, and construct an objective function based on the above operations; The solving module is configured to: solve the objective function to obtain a second complete similarity graph; The clustering module is configured to: sum the second complete similarity graph of each view to obtain a common similarity graph, and perform spectral clustering on the common similarity graph to obtain a final result.
9. A medium having a program stored thereon, It is characterized in that When the program is executed by a processor, the steps of the incomplete multi-view clustering method based on robust high-order graph learning as described in any one of claims 1 to 7 are implemented.
10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, It is characterized in that When the processor executes the program, the steps in the incomplete multi-view clustering method based on robust high-order graph learning are implemented as described in any one of claims 1 to 7.
