Incomplete multi-view clustering method, device, system and storage medium based on graph completion and adaptive view weight allocation
Through the methods of graph completion and adaptive viewing weight allocation, the problem of data essence structure capture and viewing angle identification information imbalance in incomplete multi-view clustering is solved, and the clustering performance is improved.
Patent Information
- Application Number
- CN201910446754.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2019-05-27
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2039-05-27
AI Technical Summary
The existing incomplete multi-view clustering method is difficult to effectively capture the essential structure of the data, and it fails to fully consider the imbalance of the identification information of each perspective in the cluster, resulting in a degradation of clustering performance.
Using a method based on graph completion and adaptive perspective weight allocation, we can learn consistent low-dimensional representations by constructing and completing multi-view affine patterns, and introduce perspective weight identification factors to adaptively balance the various perspectives in cluster learning. importance.
Effectively capturing the essential structure of the data improves clustering performance, reduces the negative impact of the difference in information identification at the perspective, and improves the accuracy of clustering results.
Smart Images

Figure CN110135520B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of machine learning technology, and in particular to an incomplete multi-view clustering method, device, system and storage medium based on graph completion and adaptive view weight allocation. Background Art
[0002] In the field of machine learning, multi-view clustering can automatically divide massive unlabeled data into several categories by utilizing the multi-view features of the target. This technology has been widely used in scenarios such as image clustering and medical diagnosis. A variety of single-view clustering methods have been extended to multi-view situations, such as Multi-View Expectation Maximum Clustering (MVEM), Multi-View K-Means Clustering (MVKM), Multi-View Spectral Clustering (MVSC), Multi-View Nonnegative Matrix Factorization (MultiNMF), Multi-view Canonical Correlation Analysis (MCCA), etc. Among them, the core of multi-view clustering methods based on spectral clustering and non-negative matrix factorization is to learn consistent subspaces from different views, and then implement traditional single-view clustering methods (such as k-means) on the subspace to obtain the final clustering results. Since each perspective reveals the characteristics of the target from different aspects, different perspectives contain consistent category attribution information while also containing complementary identification information. Therefore, the clustering performance of using multiple perspectives can exceed that of a single perspective or simple splicing of multiple perspectives.
[0003] Due to some uncontrollable factors in the feature collection process, perspective loss often occurs. For example, different newspaper reports on the same news are regarded as different perspectives of the news. Since the news coverage of different newspapers is not exactly the same, perspective loss is a common situation when clustering news topics. Similarly, photos of an object taken with different devices or under different lighting conditions are regarded as different perspectives of the object. Due to damage to some equipment or substandard lighting conditions, perspective loss may also occur. In recent years, scholars have paid more and more attention to clustering methods in incomplete multi-perspective situations. Focusing on the challenge of perspective loss, scholars have proposed many methods to reduce the negative impact of missing perspectives. Some methods (Partial multi-View Clustering, PVC, Incomplete Multi-modality Grouping, IMG) use complete samples (samples exist in all views) to guide the learning of consistent subspaces. Since these methods require the existence of complete samples, their application scope is greatly limited; some methods (Multi-Incomplete-View clustering, MIC, Doubly Aligned Incomplete Multi-view Clustering, DAIMC) use the framework of non-negative matrix factorization (NMF) to give low or even zero weights to the reconstruction error terms of missing samples in the view. Such methods often fail to capture the essential structure of the data and are more sensitive to noise. Some methods take into account the essential structure of the data, such as using the affine graph of the complete view to estimate the similarity relationship of the samples of the missing view (Multiview Clustering with Incomplete Views, MCIC), but the actual multi-view datasets often do not have complete views. In order to overcome this problem, some methods directly estimate the similarity relationship of all missing samples (Incomplete multi-View Clustering, IVC, Incomplete Multiview Spectral Clustering with Adaptive Graph Learning, IMSC_AGL). However, inappropriate estimation often causes a significant deviation from the true result when the missing rate is large.
[0004] Although scholars have proposed a large number of incomplete multi-view clustering frameworks that can reduce the negative impact of missing perspectives to a certain extent, these methods generally have the following problems: The first defect is that existing methods cannot capture the essential structure of the data well. Aligning low-dimensional representations only for samples with complete perspectives or filling in missing samples and their similarity relationships unreasonably will cause deviations from the essential structure of the data, thereby affecting the clustering performance. The second defect is that existing methods do not consider the imbalance of clustering identification information from each perspective. Due to the differences in the identification information contained in the features represented by each perspective, the different degrees of noise interference of the perspective, and the differences in the degradation of identification performance due to different missing rates, the guiding role of each perspective on clustering learning is different. Existing methods generally use the identification information of each perspective to guide clustering in a balanced manner, resulting in a decline in clustering performance. Summary of the invention
[0005] The present invention provides an incomplete multi-view clustering method based on graph completion and adaptive view weight allocation, comprising the following steps:
[0006] Step 1, construction and completion of multi-view affine graphs: normalize the sample instances in each view, construct and complete the affine graphs of each view;
[0007] Step 2: Learning a low-dimensional representation that is consistent between different perspectives: Iteratively obtain a low-dimensional representation that is consistent between different perspectives;
[0008] Step 3, clustering step: normalize the global low-dimensional representation and use the traditional single-view clustering method to obtain cluster categories.
[0009] As a further improvement of the present invention, step 1, the construction and completion steps of the multi-view affine graph include:
[0010] Feature extraction and normalization steps: extract multiple features of the target object and regard each feature as a perspective: use Indicates the complete total sample set extracted under ideal conditions, and represents the sample set actually extracted from the vth perspective, where n and n v Respectively represent the total number of samples and the number of samples extracted from the vth perspective, d v represents the feature dimension of the v-th perspective, After extracting the various perspective features of the clustering target, according to the formula Normalize the samples in each perspective;
[0011] The steps of constructing and filling the affine map of each view:
[0012] First, at each viewing angle X (v) Medium Utilization Construct the affine graph W(v) ;
[0013] Secondly, using The affine graph is stretched to the same scale, where Represents the correspondence between the sample in the vth perspective and all samples. If the i-th sample in the v-th perspective represents the perspective feature of the j-th sample in the total samples, then is 1, otherwise it is 0;
[0014] Finally, the similar information about the missing samples in the affine graph after each viewing angle stretching obtained in the previous step is corrected: if the sample of the i-th sample in the total sample sequence is missing in the v-th viewing angle, then according to the formula Correct the similar information of the sample in the viewing angle, where Represents the k-th perspective affine map The i-th row of H∈R n×l Represents the existence relationship of each sample in the perspective. If the i-th sample in the total sample set has a corresponding sample in the k-th perspective, then H i,k is 1, otherwise it is 0; then according to the formula Symmetrize an affine graph.
[0015] As a further improvement of the present invention, step 2, the step of learning a consistent low-dimensional representation between each perspective includes:
[0016] Variable initialization step: weight ω of each perspective in guiding subspace learning (v) (v=1,...,l) are initialized equally, that is, ω (v) =1 / l(v=1,...,l); by solving the formula Initialize the low-dimensional representation Y of each view (v) (v=1,...,l), where The affine graph completed for the vth perspective Laplacian matrix of the global low-dimensional representation Y * No initialization is required. In the first iteration, the initial Y (v) ,ω (v) to seek;
[0017] Iteration step 1: Update the global low-dimensional representation Y * Specifically, first solve the matrix The eigenvalues and eigenvectors of , and then take the eigenvectors corresponding to the largest c eigenvalues Constructing a global low-dimensional representation
[0018] Iteration step 2: Update the low-dimensional representation Y of each view (v)(v=1,...,l). Specifically, for the vth viewing angle, first solve the matrix (λY * Y *T -L (v) ) and then take the eigenvectors corresponding to the largest c eigenvalues The low-dimensional representation Y that constitutes this perspective (v) ,Right now
[0019] Iteration step 3: Update the weighted discriminant factor ω of each view (v) (v=1,...,l). Specifically, for the vth viewing angle, use the formula Update the view weight discriminant factor ω (v) , where γ (v) =Tr(Y (v)T L (v) Y (v) )+λ(c-Tr(Y (v) Y (v)T Y * Y *T ));
[0020] Convergence condition determination step: Calculate the current t-th iteration objective function value If the number of iterations is less than the set value, the (t+1)th iteration is performed, that is, the first to third iterations are executed; if the convergence condition is met, the iteration is terminated and a consistent low-dimensional representation Y is output. * Otherwise, iterate continuously until the convergence condition is met, or the number of iterations exceeds the preset maximum value, the iteration is terminated and the global low-dimensional representation Y is output * .
[0021] As a further improvement of the present invention, in step 3, the clustering step, the low-dimensional representation Y * The i-th row can be regarded as the i-th sample We first normalize each line into a unit vector, and then use the traditional single-view clustering method (such as k-means) to divide the n samples into c clusters to obtain the sample clustering results.
[0022] The present invention also provides an incomplete multi-view clustering device based on graph completion and adaptive view weight allocation, comprising:
[0023] Multi-view affine graph construction and completion unit: used to normalize sample instances in each view, construct and complete the affine graph of each view;
[0024] Learning a low-dimensional representation unit that is consistent across perspectives: used to iteratively obtain a low-dimensional representation that is consistent across perspectives;
[0025] Clustering unit: used to normalize the global low-dimensional representation and obtain cluster categories using traditional single-view clustering methods.
[0026] As a further improvement of the present invention, the multi-view affine graph construction and completion unit includes:
[0027] Feature extraction and normalization module: extract multiple features of the target object and regard each feature as a perspective: Indicates the complete total sample set extracted under ideal conditions, and represents the sample set actually extracted from the vth perspective, where n and n v Respectively represent the total number of samples and the number of samples extracted from the vth perspective, d v represents the feature dimension of the v-th perspective, After extracting the various perspective features of the clustering target, according to the formula Normalize the samples in each perspective;
[0028] Construction and filling correction module of affine graphs of each view:
[0029] First, at each viewing angle X (v) Medium Utilization Construct the affine graph W (v) ;
[0030] Secondly, using The affine graph is stretched to the same scale, where Represents the correspondence between the sample in the vth perspective and all samples. If the i-th sample in the v-th perspective represents the perspective feature of the j-th sample in the total samples, then is 1, otherwise it is 0;
[0031] Finally, the similar information about the missing samples in the affine graph after each viewing angle stretching obtained in the previous step is corrected: if the sample of the i-th sample in the total sample sequence is missing in the v-th viewing angle, then according to the formula Correct the similar information of the sample in the viewing angle, where Represents the k-th perspective affine map The i-th row of H∈R n×l Represents the existence relationship of each sample in the perspective. If the i-th sample in the total sample set has a corresponding sample in the k-th perspective, then H i,k is 1, otherwise it is 0; then according to the formula Symmetrize an affine graph.
[0032] As a further improvement of the present invention, learning a low-dimensional representation module that is consistent between perspectives includes:
[0033] Variable initialization module: weight ω of each perspective in guiding subspace learning(v) (v=1,...,l) are initialized equally, that is, ω (v) =1 / l(v=1,...,l); by solving the formula Initialize the low-dimensional representation Y of each view (v) (v=1,...,l), where The affine graph completed for the vth perspective Laplacian matrix of the global low-dimensional representation Y * No initialization is required. In the first iteration, the initial Y (v) ,ω (v) to seek;
[0034] Iteration step 1: Update the global low-dimensional representation Y * Specifically, first solve the matrix The eigenvalues and eigenvectors of , and then take the eigenvectors corresponding to the largest c eigenvalues Constructing a global low-dimensional representation
[0035] Iteration step 2: Update the low-dimensional representation Y of each view (v) (v=1,...,l). Specifically, for the vth viewing angle, first solve the matrix (λY * Y *T -L (v) ) and then take the eigenvectors corresponding to the largest c eigenvalues The low-dimensional representation Y that constitutes this perspective (v) ,Right now
[0036] Iteration step 3: Update the weighted discriminant factor ω of each view (v) (v=1,...,l). Specifically, for the vth viewing angle, use the formula Update the view weight discriminant factor ω (v) , where γ (v) =Tr(Y (v)T L (v) Y (v) )+λ(c-Tr(Y (v) Y (v)T Y * Y *T ));
[0037] Convergence judgment module: used to calculate the current t-th iteration objective function value If the number of iterations is less than the set value, the (t+1)th iteration is performed, that is, the first to third iterations are executed; if the convergence condition is met, the iteration is terminated and a consistent low-dimensional representation Y is output. *Otherwise, iterate continuously until the convergence condition is met, or the number of iterations exceeds the preset maximum value, the iteration is terminated and the global low-dimensional representation Y is output * .
[0038] As a further improvement of the present invention, in the clustering unit, the low-dimensional representation Y * The i-th row can be regarded as the i-th sample We first normalize each line into a unit vector, and then use the traditional single-view clustering method (such as k-means) to divide the n samples into c clusters to obtain the sample clustering results.
[0039] The present invention also provides an incomplete multi-view clustering system based on graph completion and adaptive view weight allocation, comprising: a memory, a processor, and a computer program stored on the memory, wherein the computer program is configured to implement the steps of the method described in the present invention when called by the processor.
[0040] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the method described in the present invention when called by a processor.
[0041] The beneficial effects of the present invention are as follows: the present invention effectively captures the essential structure of data and improves clustering performance through the introduction of similar graph completion technology and perspective weight discrimination factors. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 It is a schematic diagram of the principle of the present invention.
[0043] Figure 2 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION
[0044] In view of the first defect of the background technology, the present invention uses graph completion technology to share the similarity relationship of samples between multiple perspectives, mines the essential structure of the data, and uses spectral clustering to further capture the nonlinear manifold structure of the data. In view of the second defect of the background technology, the present invention introduces a perspective weight discriminant factor to adaptively balance the importance of each perspective in clustering learning, thereby obtaining better clustering performance.
[0045] The present invention discloses an incomplete multi-view clustering method based on graph completion and adaptive perspective weight allocation, which can reasonably fill in the affine graph learned from the missing perspective and reduce the negative impact caused by the difference in identification information of each perspective. In order to facilitate the learning of a subspace that is as consistent as possible from multiple perspectives, we use the similar information of samples in each perspective to complete the affine graph learned from the missing perspective. Unlike the existing methods, the graph filling strategy fills the affine graph, does not require direct evaluation of the missing sample data, and the strategy can share complementary information between multiple perspectives, thereby greatly reducing the negative impact caused by the missing samples. In addition, considering the different feature identification capabilities represented by multiple different perspectives, the differences in sample missing rates, and the different degrees of noise interference, the present invention introduces an adaptively learned perspective weight identification factor to balance the role of different perspectives of identification information in clustering learning, thereby greatly improving the performance of clustering.
[0046] First, given an incomplete multi-view dataset d v and n v are the feature dimension and number of samples of the vth perspective respectively. At the same time, the total sequence containing all samples is given in We can obtain the perspective-sample indicator matrix H = [h1,h2,…,h l ]∈R n ×l , where the i-th column h i represents the existence of samples in the i-th view. If there is an instance of the j-th sample in the i-th view, then h i The corresponding element in is 1, otherwise it is 0. At the same time, we can get the corresponding relationship matrix between the vth view sample and all samples The kth column g k Represents the total sample sequence The kth sample in The instance correspondence in the vth view, if Corresponding to a sample in the vth perspective, such as the lth sample Then g k The lth element of is 1, and the other elements are 0.
[0047] In the present invention, first, at each viewing angle X (v) The Gaussian kernel function is used to construct the similarity matrix, that is, the formula Constructing a similarity matrix The parameter σ is used to control the rate at which the similarity between two samples decreases as their distance increases. We take σ = 1. Then, we use the correspondence matrix G (v)Expand the affine graph to n dimensions, that is at this time The similarity weights associated with the missing samples in the image are 0. In the next step, the similarity information of multiple perspectives is first used to estimate and correct the similarity weights of the missing samples in the affine image of each perspective, and then the iterative model is used to obtain the subspace representation consistent with each perspective.
[0048] During the affine image correction process, if the i-th sample in the total sample sequence is at the viewing angle X (v) If the corresponding instance is missing in , then use the formula The similarity weight of the missing sample is completed, that is, the similarity relationship of the missing instance of the perspective is estimated through the similarity relationship in other perspectives containing the sample instance, and then the similarity weight of the missing sample is completed. The symmetry of the affine image is ensured. The correction process of the affine image ensures that each view has a similar internal structure, and facilitates the mining of complementary information between views.
[0049] The traditional multi-view spectral clustering framework designs the following optimization problem:
[0050]
[0051] in is an affine graph The Laplace matrix, D (v) yes The degree matrix of It is calculated that all non-diagonal elements are 0. (v) ,Y * ∈R n×c They are the low-dimensional representation of the vth perspective and the consistent low-dimensional representation of all perspectives. c is the number of categories and needs to be specified in advance.
[0052] The traditional spectral clustering framework treats each view equally in the process of guiding subspace learning, which makes the view that is seriously polluted by noise or has weak feature discriminant information equal to the view with rich discriminant information in subspace learning, resulting in a decrease in clustering performance. In order to adaptively balance the role of different viewpoints of discriminant information in subspace learning, based on the affine graphs of each view completed above, the present invention proposes the following subspace learning model to learn the consistent low-dimensional representation Y between each view * :
[0053]
[0054] where ω (v) is the perspective weight discriminator for adaptive learning, which is used to balance the role of each perspective in subspace learning. The weight power r plays a smoothing role. λ is the balancing factor. (v)By introducing the subspace representation, the present invention can effectively reduce the negative impact caused by the differences in the identification information of each perspective, thereby guiding the model to learn a better subspace representation.
[0055] Model (2) contains multiple unknown variables, and it is difficult to find its optimal solution. We use the iterative update method to approximate the optimal solution of model (2), that is, when solving a certain variable, fix other unknown variables. (v) ,Y * ,ω (v) Solve them one by one.
[0056] 1) Solve for a consistent low-dimensional representation Y * :
[0057] Fixed variable Y (v) ,ω (v) , for a consistent low-dimensional representation Y * , model (2) degenerates into the following optimization problem:
[0058]
[0059] This optimization problem is an eigenvalue decomposition problem, so Y * The optimal solution is the matrix The eigenvectors corresponding to the largest c eigenvalues of are used as a matrix consisting of column vectors.
[0060] 2) Solve the low-dimensional representation Y of each perspective (v)
[0061] By observing model (2), we can find that the low-dimensional representations of different perspectives can be optimized separately, and the fixed variable Y * ,ω (v) , about Y (v) , the optimization problem (2) degenerates into the following optimization problem:
[0062]
[0063] This problem is also a typical eigenvalue decomposition problem, that is, Y (v) The optimal solution is the matrix (λY * Y *T -L (v) ) is the matrix consisting of the eigenvectors corresponding to the largest c eigenvalues.
[0064] 3) Solve the perspective discrimination factor ω (v)
[0065] Similarly, fixing Y (v) ,Y * , model (2) degenerates into the following formula:
[0066]
[0067] where γ (v) =Tr(Y (v)T L (v) Y (v) )+λ(c-Tr(Y (v) Y (v)T Y * Y *T )). In variable Y (v) ,Y * When γ is fixed, (v) is a fixed value, the solution to the optimization problem is
[0068]
[0069] By iteratively solving the above variables, we obtain the approximate solution of model (2). The iterative process is summarized as Algorithm 1:
[0070]
[0071]
[0072]
[0073] In summary, the specific implementation steps of the present invention are:
[0074] For any multi-view clustering problem, including complete multi-view data with completely corresponding samples and incomplete multi-view data with any missing cases, first normalize the sample instances in each view, build and complete the affine graph of each view, then learn the consistent low-dimensional representation between each view, and finally normalize the global low-dimensional representation and use the traditional single-view clustering method to obtain the clustering category. The following is a detailed description of the steps of the implementation process of the invention:
[0075] Step 1, construction and completion of multi-view affine graphs: normalize the sample instances in each view, construct and complete the affine graphs of each view;
[0076] Step 2: Learning a low-dimensional representation that is consistent between different perspectives: Iteratively obtain a low-dimensional representation that is consistent between different perspectives;
[0077] Step 3, clustering step: normalize the global low-dimensional representation and use the traditional single-view clustering method to obtain cluster categories.
[0078] Step 1: The construction and completion steps of the multi-view affine graph include:
[0079] (1) Feature extraction and normalization steps: For clustering tasks involving multiple perspectives, first flexibly extract multiple features of the target object and regard each feature as a perspective. For example, for web page data, its content and its hyperlink information can be extracted as two perspectives of the clustering task, or its text content and image content can be extracted as two perspectives. For clustering of image data, the elements of the original image can be used as features, and its HOG features can be extracted as other perspectives of the task. Indicates the complete total sample set extracted under ideal conditions, and represents the sample set actually extracted from the vth perspective, where n and n v Respectively represent the total number of samples and the number of samples extracted from the vth perspective, d v represents the feature dimension of the v-th perspective, After extracting the various perspective features of the clustering target, according to the formula Normalize the samples in each perspective;
[0080] (2) Construction and filling correction steps of affine images of each view:
[0081] First, at each viewing angle X (v) Medium Utilization Construct the affine graph W (v) ;
[0082] Secondly, using The affine graph is stretched to the same scale, where Represents the correspondence between the sample in the vth perspective and all samples. If the i-th sample in the v-th perspective represents the perspective feature of the j-th sample in the total samples, then is 1, otherwise it is 0;
[0083] Finally, the similar information about the missing samples in the affine graph after each viewing angle stretching obtained in the previous step is corrected: if the sample of the i-th sample in the total sample sequence is missing in the v-th viewing angle, then according to the formula Correct the similar information of the sample in the viewing angle, where Represents the k-th perspective affine map The i-th row of H∈R n×l Represents the existence relationship of each sample in the perspective. If the i-th sample in the total sample set has a corresponding sample in the k-th perspective, then H i,k is 1, otherwise it is 0; then according to the formula Symmetrize an affine graph.
[0084] Step 2: Learning a consistent low-dimensional representation across perspectives includes:
[0085] Variable initialization step: weight ω of each perspective in guiding subspace learning(v) (v=1,...,l) are initialized equally, that is, ω (v) =1 / l(v=1,...,l); by solving the formula Initialize the low-dimensional representation Y of each view (v) (v=1,...,l), where The affine graph completed for the vth perspective Laplacian matrix of the global low-dimensional representation Y * No initialization is required. In the first iteration, the initial Y (v) ,ω (v) to seek;
[0086] Iteration step 1: Update the global low-dimensional representation Y * Specifically, first solve the matrix The eigenvalues and eigenvectors of , and then take the eigenvectors corresponding to the largest c eigenvalues Constructing a global low-dimensional representation
[0087] Iteration step 2: Update the low-dimensional representation Y of each view (v) (v=1,...,l). Specifically, for the vth viewing angle, first solve the matrix (λY * Y *T -L (v) ) and then take the eigenvectors corresponding to the largest c eigenvalues The low-dimensional representation that constitutes this perspective
[0088] Iteration step 3: Update the weighted discriminant factor ω of each view (v) (v=1,...,l). Specifically, for the vth viewing angle, use the formula Update the view weight discriminant factor ω (v) , where γ (v) =Tr(Y (v)T L (v) Y (v) )+λ(c-Tr(Y (v) Y (v)T Y * Y *T ));
[0089] Convergence condition determination step: Calculate the current t-th iteration objective function value If the number of iterations is small (the number of iterations is less than the set value), such as t≤5, the (t+1)th iteration is performed, that is, the first to third iterations are executed; if the convergence conditions are met, such as |L t -L t-1 |<10 -5, then the iteration is terminated and a consistent low-dimensional representation Y is output * Otherwise, iterate continuously until the convergence condition is met, or the number of iterations exceeds the preset maximum value, such as t>100, the iteration is forced to terminate and the global low-dimensional representation Y is output * .
[0090] In step 3, the clustering step: the low-dimensional representation Y learned in step 2 is * The i-th row can be regarded as the i-th sample We first normalize each line into a unit vector, and then use the traditional single-view clustering method (such as k-means) to divide the n samples into c clusters to get the sample clustering result.
[0091] The present invention also discloses an incomplete multi-view clustering device based on graph completion and adaptive view weight allocation, comprising:
[0092] Multi-view affine graph construction and completion unit: used to normalize sample instances in each view, construct and complete the affine graph of each view;
[0093] Learning a low-dimensional representation unit that is consistent across perspectives: used to iteratively obtain a low-dimensional representation that is consistent across perspectives;
[0094] Clustering unit: used to normalize the global low-dimensional representation and obtain cluster categories using traditional single-view clustering methods.
[0095] The multi-view affine graph construction and completion unit includes:
[0096] Feature extraction and normalization module: extract multiple features of the target object and regard each feature as a perspective: Indicates the complete total sample set extracted under ideal conditions, and represents the sample set actually extracted from the vth perspective, where n and n v Respectively represent the total number of samples and the number of samples extracted from the vth perspective, d v represents the feature dimension of the v-th perspective, After extracting the various perspective features of the clustering target, according to the formula Normalize the samples in each perspective;
[0097] Construction and filling correction module of affine graphs of each view:
[0098] First, at each viewing angle X (v) Medium Utilization Construct the affine graph W (v) ;
[0099] Secondly, using The affine graph is stretched to the same scale, where Represents the correspondence between the sample in the vth perspective and all samples. If the i-th sample in the v-th perspective represents the perspective feature of the j-th sample in the total samples, then is 1, otherwise it is 0;
[0100] Finally, the similar information about the missing samples in the affine graph after each viewing angle stretching obtained in the previous step is corrected: if the sample of the i-th sample in the total sample sequence is missing in the v-th viewing angle, then according to the formula Correct the similar information of the sample in the viewing angle, where Represents the k-th perspective affine map The i-th row of H∈R n×l Represents the existence relationship of each sample in the perspective. If the i-th sample in the total sample set has a corresponding sample in the k-th perspective, then H i,k is 1, otherwise it is 0; then according to the formula Symmetrize an affine graph.
[0101] The modules for learning consistent low-dimensional representations between different perspectives include:
[0102] Variable initialization module: weight ω of each perspective in guiding subspace learning (v) (v=1,...,l) are initialized equally, that is, ω (v) =1 / l(v=1,...,l); by solving the formula Initialize the low-dimensional representation Y of each view (v) (v=1,...,l), where The affine graph completed for the vth perspective Laplacian matrix of the global low-dimensional representation Y * No initialization is required. In the first iteration, the initial Y (v) ,ω (v) to seek;
[0103] Iteration step 1: Update the global low-dimensional representation Y * Specifically, first solve the matrix The eigenvalues and eigenvectors of , and then take the eigenvectors corresponding to the largest c eigenvalues Constructing a global low-dimensional representation
[0104] Iteration step 2: Update the low-dimensional representation Y of each view (v) (v=1,...,l). Specifically, for the vth viewing angle, first solve the matrix (λY * Y *T -L (v)) and then take the eigenvectors corresponding to the largest c eigenvalues The low-dimensional representation that constitutes this perspective
[0105] Iteration step 3: Update the weighted discriminant factor ω of each view (v) (v=1,...,l). Specifically, for the vth viewing angle, use the formula Update the view weight discriminant factor ω (v) , where γ (v) =Tr(Y (v)T L (v) Y (v) )+λ(c-Tr(Y (v) Y (v)T Y * Y *T ));
[0106] Convergence judgment module: used to calculate the current t-th iteration objective function value If the number of iterations is less than the set value, the (t+1)th iteration is performed, that is, the first to third iterations are executed; if the convergence condition is met, the iteration is terminated and a consistent low-dimensional representation Y is output. * Otherwise, iterate continuously until the convergence condition is met, or the number of iterations exceeds the preset maximum value, the iteration is terminated and the global low-dimensional representation Y is output * .
[0107] In the clustering unit, the low-dimensional representation Y * The i-th row can be regarded as the i-th sample We first normalize each line into a unit vector, and then use the traditional single-view clustering method (such as k-means) to divide the n samples into c clusters to obtain the sample clustering results.
[0108] The present invention also discloses an incomplete multi-view clustering system based on graph completion and adaptive view weight allocation, comprising: a memory, a processor and a computer program stored on the memory, wherein the computer program is configured to implement the steps of the method described in the present invention when called by the processor.
[0109] The present invention also discloses a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the method described in the present invention when called by a processor.
[0110] The present invention belongs to the field of machine learning and is a new multi-perspective clustering method for incomplete perspective situations. Specifically, it is a technology that uses computer technology, digital image processing, text information processing and other technologies to achieve clustering. The present invention has a wide range of applications and can be used in various target recognition and clustering scenarios with multiple feature dimensions, such as security systems based on multi-feature recognition such as face and gait, enterprise user portrait analysis, search engine secondary clustering, text topic clustering, protein functional attribute prediction, etc.
[0111] The present invention focuses on capturing the essential structure of the data and adaptively assigns different weights to each perspective to balance the role of each perspective in clustering learning. Specifically, the present invention performs affine graph completion through the sample similarity structure of the shared perspective, and then implements joint spectral clustering to learn a consistent subspace on the affine graphs completed by each perspective to capture the essential structure of the data. At the same time, the present invention proposes an adaptive perspective weight discriminator to balance the role of each perspective in the subspace learning process. In summary, the present invention effectively captures the essential structure of the data and improves the clustering performance through the introduction of similarity graph completion technology and perspective weight discriminator.
[0112] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.
Claims
1. An incomplete multi-view clustering method based on graph completion and adaptive view weight allocation, characterized in that: The steps include: Step 1, construction and completion of multi-view affine graphs: normalize the sample instances in each view, construct and complete the affine graphs of each view; Step 2: Learning a low-dimensional representation that is consistent between different perspectives: Iteratively obtain a low-dimensional representation that is consistent between different perspectives; Step 3, clustering step: normalizing the global low-dimensional representation and using the traditional single-view clustering method to obtain cluster categories, wherein the cluster categories include classifying face data and text data; This incomplete multi-view clustering method is applied to various target recognition and clustering scenarios with multiple feature dimensions, including security systems based on face recognition, search engine secondary clustering, and text topic clustering; Step 1: The construction and completion steps of the multi-view affine graph include: Feature extraction and normalization steps: extract multiple features of the target object and regard each feature as a perspective: use Indicates the complete total sample set extracted under ideal conditions, and represents the sample set actually extracted from the vth perspective, where n and n v Respectively represent the total number of samples and the number of samples extracted from the vth perspective, d v represents the feature dimension of the v-th perspective, After extracting the various perspective features of the clustering target, according to the formula Normalize the samples in each perspective; The steps of constructing and filling the affine map of each view: First, at each viewing angle X (v) Medium Utilization Construct the affine graph W (v) ; Secondly, using The affine graph is stretched to the same scale, where Represents the correspondence between the sample in the vth perspective and all samples. If the i-th sample in the v-th perspective represents the perspective feature of the j-th sample in the total samples, then is 1, otherwise it is 0; Finally, the similar information about the missing samples in the affine graph after each viewing angle stretching obtained in the previous step is corrected: if the sample of the i-th sample in the total sample sequence is missing in the v-th viewing angle, then according to the formula Correct the similar information of the sample in the viewing angle, where Represents the k-th perspective affine map The i-th row of H∈R n×l Represents the existence relationship of each sample in the perspective. If the i-th sample in the total sample set has a corresponding sample in the k-th perspective, then H i,k is 1, otherwise it is 0; then according to the formula Symmetrize affine graphs; Step 2: Learning a consistent low-dimensional representation across perspectives includes: Variable initialization step: weight ω of each perspective in guiding subspace learning (v) , v=1,...,l are initialized equally, that is, ω (v) =1 / l, v=1,...,l; by solving the formula Initialize the low-dimensional representation Y of each view (v) , v = 1, ..., l, where The affine graph completed for the vth perspective Laplacian matrix of the global low-dimensional representation Y * No initialization is required. In the first iteration, the initial Y (v) ,ω (v) to seek; The first step of iteration: first solve the matrix The eigenvalues and eigenvectors of , and then take the eigenvectors corresponding to the largest c eigenvalues Constructing a global low-dimensional representation Iteration step 2: For the vth perspective, first solve the matrix (λY * Y *T -L (v) ) and then take the eigenvectors corresponding to the largest c eigenvalues The low-dimensional representation that constitutes this perspective Iteration step 3: For the vth perspective, use the formula Update the view weight discriminant factor ω (v) , where γ (v) =Tr(Y (v)T L (v) Y (v) )+λ(c-Tr(Y (v) Y (v)T Y * Y *T )); Convergence condition determination step: Calculate the current t-th iteration objective function value If the number of iterations is less than the set value, the (t+1)th iteration is performed, that is, the first to third iterations are executed; if the convergence condition is met, the iteration is terminated and a consistent low-dimensional representation Y is output. * Otherwise, iterate continuously until the convergence condition is met, or the number of iterations exceeds the preset maximum value, the iteration is terminated and the global low-dimensional representation Y is output * .
2. The incomplete multi-view clustering method according to claim 1, characterized in that: In step 3, the clustering step, the low-dimensional representation Y * The i-th row can be regarded as the i-th sample We first normalize each line into a unit vector, and then use the traditional single-view clustering method to divide the n samples into c clusters to obtain the sample clustering results.
3. An incomplete multi-view clustering device based on graph completion and adaptive view weight allocation, characterized in that: include: Multi-view affine graph construction and completion unit: used to normalize sample instances in each view, construct and complete the affine graph of each view; Learning a low-dimensional representation unit that is consistent across perspectives: used to iteratively obtain a low-dimensional representation that is consistent across perspectives; Clustering unit: used to normalize the global low-dimensional representation and obtain clustering categories using the traditional single-view clustering method, wherein the clustering categories include classifying face data and text data; The incomplete multi-view clustering device is applied to various target recognition and clustering scenarios with multiple feature dimensions, including security systems based on face recognition, search engine secondary clustering, and text topic clustering; The multi-view affine graph construction and completion unit includes: Feature extraction and normalization module: extract multiple features of the target object and regard each feature as a perspective: Indicates the complete total sample set extracted under ideal conditions, and represents the sample set actually extracted from the vth perspective, where n and n v Respectively represent the total number of samples and the number of samples extracted from the vth perspective, d v represents the feature dimension of the v-th perspective, After extracting the various perspective features of the clustering target, according to the formula Normalize the samples in each perspective; Construction of affine images for each viewing angle and filling correction module: First, at each viewing angle X (v) Medium Utilization Construct the affine graph W (v) ; Secondly, using The affine graph is stretched to the same scale, where Represents the correspondence between the sample in the vth perspective and all samples. If the i-th sample in the v-th perspective represents the perspective feature of the j-th sample in the total samples, then is 1, otherwise it is 0; Finally, the similar information about the missing samples in the affine graph after each viewing angle stretching obtained in the previous step is corrected: if the sample of the i-th sample in the total sample sequence is missing in the v-th viewing angle, then according to the formula Correct the similar information of the sample in the viewing angle, where Represents the k-th perspective affine map The i-th row of H∈R n×l Represents the existence relationship of each sample in the perspective. If the i-th sample in the total sample set has a corresponding sample in the k-th perspective, then H i,k is 1, otherwise it is 0; then according to the formula Symmetrize affine graphs; The modules for learning consistent low-dimensional representations between different perspectives include: Variable initialization module: weight ω of each perspective in guiding subspace learning (v) , v=1,...,l are initialized equally, that is, ω (v) =1 / l, v=1,...,l; by solving the formula Initialize the low-dimensional representation Y of each view (v) , v = 1, ..., l, where The affine graph completed for the vth perspective Laplacian matrix of the global low-dimensional representation Y * No initialization is required. In the first iteration, the initial Y (v) ,ω (v) to seek; The first step of iteration: first solve the matrix The eigenvalues and eigenvectors of , and then take the eigenvectors corresponding to the largest c eigenvalues Constructing a global low-dimensional representation Iteration step 2: For the vth perspective, first solve the matrix (λY * Y *T -L (v) ) and then take the eigenvectors corresponding to the largest c eigenvalues The low-dimensional representation Y that constitutes this perspective (v) ,Right now Iteration step 3: For the vth perspective, use the formula Update the view weight discriminant factor ω (v) , where γ (v) =Tr(Y (v)T L (v) Y (v) )+λ(c-Tr(Y (v) Y (v)T Y * Y *T )); Convergence judgment module: used to calculate the current t-th iteration objective function value If the number of iterations is less than the set value, the (t+1)th iteration is performed, that is, the first to third iterations are executed; if the convergence condition is met, the iteration is terminated and a consistent low-dimensional representation Y is output. * Otherwise, iterate continuously until the convergence condition is met, or the number of iterations exceeds the preset maximum value, the iteration is terminated and the global low-dimensional representation Y is output * .
4. The incomplete multi-view clustering device according to claim 3, characterized in that: In the clustering unit, the low-dimensional representation Y * The i-th row can be regarded as the i-th sample We first normalize each line into a unit vector, and then use the traditional single-view clustering method to divide the n samples into c clusters to obtain the sample clustering results.
5. An incomplete multi-view clustering system based on graph completion and adaptive view weight allocation, characterized by: include: A memory, a processor, and a computer program stored in the memory, wherein the computer program is configured to implement the steps of the method according to any one of claims 1 to 2 when called by the processor.
6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of any one of claims 1 to 2 when called by a processor.
Citation Information
Patent Citations
Learning method applied in multi-visual angle clustering robustness similarity graph expression
CN107784318A
Multi-view subspace clustering method based on implicit representation and self-adaptation
CN109002854A
An ultra-short-term wind power prediction method based on hybrid intelligent technology
CN109255726A
