Multi-view clustering method and device, equipment, storage medium and program product

Abstract

The embodiment of the invention provides a multi-view clustering method and device, equipment, a storage medium and a program product, and relates to the field of financial science and technology. Obtaining a view data set corresponding to each view; based on the similarity between the feature elements in each view data set, constructing a similar matrix of each view data set; multiplying the similar matrixes of the plurality of view data sets to obtain a consistency matrix, and reconstructing the similar matrix of each view data set based on the consistency matrix to obtain a reconstructed similar matrix; and performing multi-view clustering based on the plurality of reconstructed similar matrixes to obtain a clustering result, the clustering result being used for indicating a clustering label corresponding to each sample. Through the above mode, the complementary information of the consistency matrix is utilized to maintain the stability of the clustering boundary, the misjudgment and missed judgment caused by the missing of the sample information in the view are reduced, and the accuracy of the multi-view clustering result is improved.

Description

Technical Field

[0001] This application relates to the field of financial technology, and in particular to a multi-view clustering method, apparatus, device, storage medium, and program product. Background Technology

[0002] With the rapid development of digital data collection methods, the same customer in financial scenarios often leaves diverse "digital footprints" across different channels: click logs from mobile banking apps, facial images from high-definition cameras at bank counters, voiceprint records from ATMs, RFID fingerprints from POS terminals, and even transaction time-series data from corporate online banking. These heterogeneous and heterogeneous features naturally constitute multiple views, comprehensively depicting customer behavior patterns and risk profiles. How to organically integrate these views and accurately segment customer groups without manual annotation has become a key aspect for banks to improve risk control, marketing, and operational efficiency. Multi-view clustering is therefore regarded as the alchemy of a "data gold mine."

[0003] Over the past decade, academia and industry have proposed numerous solutions for multi-view clustering. Typical approaches include: semi-supervised methods based on co-training, which iteratively optimize pseudo-labels by leveraging the complementarity between views; algorithms based on subspace learning, which map each view to a consistent latent low-dimensional space before implementing K-means; graph neural network-based techniques, which construct an adjacency graph for each view and obtain a unified representation through attention or tensor fusion; and self-supervised frameworks based on contrastive learning, which achieve representation alignment by maximizing mutual information between views.

[0004] However, the real-world production environment in banks is far more complex than experimental scenarios. Due to compliance requirements, equipment malfunctions, network fluctuations, or differences in customer authorization, situations frequently arise such as a branch's camera going offline on a given day, certain mobile phone models refusing to collect voiceprints, or some POS terminals losing radio frequency signals. This can lead to missing blocks or even entire columns in the corresponding views. Most existing methods assume that "each sample is fully available in all views," and once missing data occurs, the accuracy of multi-view clustering results decreases. Summary of the Invention

[0005] This application provides a multi-view clustering method, apparatus, device, storage medium, and program product to solve the technical problem that existing multi-view clustering methods rely on the fact that each sample is fully available in all views, but the accuracy of multi-view clustering results is reduced once a missing sample occurs.

[0006] In a first aspect, this application provides a multi-view clustering method, the method comprising:

[0007] Obtain the view dataset corresponding to each view, where each view dataset contains feature elements of multiple samples under the corresponding view;

[0008] Based on the similarity between feature elements within each view dataset, a similarity matrix is ​​constructed for each view dataset;

[0009] The similarity matrices of multiple view datasets are multiplied to obtain a consistency matrix, and the similarity matrix of each view dataset is reconstructed based on the consistency matrix to obtain a reconstructed similarity matrix;

[0010] Multi-view clustering is performed based on multiple reconstructed similarity matrices to obtain clustering results, wherein the clustering results are used to indicate the clustering label corresponding to each sample.

[0011] In one possible implementation, reconstructing the similarity matrix of each view dataset based on the consistency matrix to obtain a reconstructed similarity matrix includes:

[0012] Based on the consistency matrix, the similarity matrix of each view dataset is reconstructed to obtain the reconstructed similarity matrix;

[0013] Based on the multiple reconstruction similarity matrices, the consistency matrices of the multiple view datasets are obtained again, and based on the newly obtained consistency matrices and the one-time matrix before this reconstruction, it is determined whether the reconstruction termination condition is met.

[0014] If the condition is not met, the reconstruction similarity matrix of each view dataset is reconstructed based on the re-obtained consistency matrix until the reconstruction termination condition is met.

[0015] In one possible implementation, reconstructing the similarity matrix of each view dataset based on the consistency matrix to obtain a reconstructed similarity matrix includes:

[0016] For any column of similarity elements in each similarity matrix, the feature weight of the corresponding column of similarity elements is determined based on the difference between the similarity element and the reference element at the same position in the consistency matrix, wherein the feature weight is inversely proportional to the difference.

[0017] Based on all feature weights corresponding to each similarity matrix, determine the weight matrix of each similarity matrix;

[0018] Based on the weight matrix and corresponding row constraints of each similar matrix, a reconstruction operation of the similar matrix is ​​performed to obtain the reconstructed similar matrix.

[0019] In one possible implementation, determining whether the reconstruction termination condition is met based on the re-acquired consistency matrix and the one-time matrix before this reconstruction includes:

[0020] Summing the newly acquired consistency matrix and the consistency matrix before this reconstruction yields the position information value corresponding to the newly acquired consistency matrix, and the position information value corresponding to the consistency matrix before this reconstruction.

[0021] If the position information value corresponding to the consistency matrix obtained again is less than the position information value corresponding to the consistency matrix before this reconstruction, then the reconstruction end condition is met.

[0022] If the position information value corresponding to the consistency matrix obtained again is greater than or equal to the position information value corresponding to the consistency matrix before this reconstruction, then it is determined that the reconstruction termination condition is not met.

[0023] In one possible implementation, the multi-view clustering based on multiple reconstructed similarity matrices to obtain clustering results includes:

[0024] The unified graph matrix before iteration is determined based on multiple reconstructed similarity matrices;

[0025] Based on the unified graph matrix before iteration, determine the view weights of each of the reconstructed similarity matrices after iteration, and determine the unified graph matrix after iteration based on the multiple reconstructed similarity matrices and the multiple view weights after iteration;

[0026] Based on the unified graph matrix after this iteration and the unified graph matrix before this iteration, determine whether the iteration termination condition is met;

[0027] If not satisfied, the view weights of each reconstructed similarity matrix are iterated again based on the unified graph matrix after the iteration to obtain the view weights after the kth iteration and the corresponding unified graph matrix after the kth iteration, until the iteration termination condition is met, where k is the iteration number;

[0028] If satisfied, then apply rank constraints to the unified graph matrix after this iteration to obtain the clustering results.

[0029] In one possible implementation, determining whether the iteration termination condition is met based on the unified graph matrix after the current iteration and the unified graph matrix before the current iteration includes:

[0030] Determine the difference matrix based on the unified graph matrix before and after this iteration;

[0031] The F-norm is calculated based on the difference matrix, and the relationship between the F-norm and the convergence threshold is determined.

[0032] If the F-norm is less than the convergence threshold, then the iteration termination condition is determined to be met.

[0033] If the F-norm is greater than or equal to the convergence threshold, then the iteration termination condition is determined not to be met.

[0034] In one possible implementation, the step of applying rank constraints to the unified graph matrix after this iteration to obtain the clustering result includes:

[0035] Determine the diagonal matrix of the iterated unified graph matrix, and determine the Laplace matrix based on the diagonal matrix and the iterated unified graph matrix;

[0036] Given a preset number of clusters, apply a rank constraint to the Laplacian matrix;

[0037] If the Laplacian matrix satisfies the rank constraint, the iterated unified graph matrix is ​​split into multiple sub-matrices based on the number of clusters, and the clustering label of each sub-matrix is ​​determined.

[0038] Based on each submatrix and its corresponding clustering label, determine the clustering label for each sample, and determine the clustering result based on the clustering label for each sample.

[0039] Secondly, this application provides a multi-view clustering apparatus, comprising: an acquisition module, a construction module, and a clustering module, wherein:

[0040] The acquisition module is used to acquire the view dataset corresponding to each view, wherein each view dataset contains feature elements of multiple samples under the corresponding view;

[0041] The construction module is used to construct a similarity matrix for each view dataset based on the similarity between feature elements within each view dataset;

[0042] The construction module is further configured to multiply the similarity matrices of multiple view datasets to obtain a consistency matrix, and reconstruct the similarity matrix of each view dataset based on the consistency matrix to obtain a reconstructed similarity matrix;

[0043] The clustering module is used to perform multi-view clustering based on multiple reconstructed similarity matrices to obtain clustering results, wherein the clustering results are used to indicate the clustering label corresponding to each sample.

[0044] In one possible implementation, the construction module is further configured to:

[0045] Based on the consistency matrix, the similarity matrix of each view dataset is reconstructed to obtain the reconstructed similarity matrix;

[0046] Based on the multiple reconstruction similarity matrices, the consistency matrices of the multiple view datasets are obtained again, and based on the newly obtained consistency matrices and the one-time matrix before this reconstruction, it is determined whether the reconstruction termination condition is met.

[0047] If the condition is not met, the reconstruction similarity matrix of each view dataset is reconstructed based on the re-obtained consistency matrix until the reconstruction termination condition is met.

[0048] In one possible implementation, the construction module is further configured to:

[0049] For any column of similarity elements in each similarity matrix, the feature weight of the corresponding column of similarity elements is determined based on the difference between the similarity element and the reference element at the same position in the consistency matrix, wherein the feature weight is inversely proportional to the difference.

[0050] Based on all feature weights corresponding to each similarity matrix, determine the weight matrix of each similarity matrix;

[0051] Based on the weight matrix and corresponding row constraints of each similar matrix, a reconstruction operation of the similar matrix is ​​performed to obtain the reconstructed similar matrix.

[0052] In one possible implementation, the construction module is further configured to:

[0053] Summing the newly acquired consistency matrix and the consistency matrix before this reconstruction yields the position information value corresponding to the newly acquired consistency matrix, and the position information value corresponding to the consistency matrix before this reconstruction.

[0054] If the position information value corresponding to the consistency matrix obtained again is less than the position information value corresponding to the consistency matrix before this reconstruction, then the reconstruction end condition is met.

[0055] If the position information value corresponding to the consistency matrix obtained again is greater than or equal to the position information value corresponding to the consistency matrix before this reconstruction, then it is determined that the reconstruction termination condition is not met.

[0056] In one possible implementation, the clustering module is further configured to:

[0057] The unified graph matrix before iteration is determined based on multiple reconstructed similarity matrices;

[0058] Based on the unified graph matrix before iteration, determine the view weights of each of the reconstructed similarity matrices after iteration, and determine the unified graph matrix after iteration based on the multiple reconstructed similarity matrices and the multiple view weights after iteration;

[0059] Based on the unified graph matrix after this iteration and the unified graph matrix before this iteration, determine whether the iteration termination condition is met;

[0060] If not satisfied, the view weights of each reconstructed similarity matrix are iterated again based on the unified graph matrix after the iteration to obtain the view weights after the kth iteration and the corresponding unified graph matrix after the kth iteration, until the iteration termination condition is met, where k is the iteration number;

[0061] If satisfied, then apply rank constraints to the unified graph matrix after this iteration to obtain the clustering results.

[0062] In one possible implementation, the clustering module is further configured to:

[0063] Determine the difference matrix based on the unified graph matrix before and after this iteration;

[0064] The F-norm is calculated based on the difference matrix, and the relationship between the F-norm and the convergence threshold is determined.

[0065] If the F-norm is less than the convergence threshold, then the iteration termination condition is determined to be met.

[0066] If the F-norm is greater than or equal to the convergence threshold, then the iteration termination condition is determined not to be met.

[0067] In one possible implementation, the clustering module is further configured to:

[0068] Determine the diagonal matrix of the iterated unified graph matrix, and determine the Laplace matrix based on the diagonal matrix and the iterated unified graph matrix;

[0069] Given a preset number of clusters, apply a rank constraint to the Laplacian matrix;

[0070] If the Laplacian matrix satisfies the rank constraint, the iterated unified graph matrix is ​​split into multiple sub-matrices based on the number of clusters, and the clustering label of each sub-matrix is ​​determined.

[0071] Based on each submatrix and its corresponding clustering label, determine the clustering label for each sample, and determine the clustering result based on the clustering label for each sample.

[0072] Thirdly, this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0073] The memory stores computer-executed instructions;

[0074] The processor executes computer execution instructions stored in the memory to implement the method as described in any of the first aspects.

[0075] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any of the first aspects.

[0076] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the method described in any of the foregoing descriptions.

[0077] The multi-view clustering method, apparatus, device, storage medium, and program products provided in this application construct a similarity matrix based on the similarity between feature elements within the view dataset of each view. Multiple similarity matrices are multiplied to obtain a consistency matrix, and each similarity matrix is ​​reconstructed based on the consistency matrix to obtain a reconstructed similarity matrix. Subsequently, multi-view clustering is performed based on the multiple reconstructed similarity matrices to obtain clustering results indicating the clustering label corresponding to each sample. Through this method, the complementary information of the consistency matrix is ​​used to maintain the stability of the clustering boundaries, reducing misclassification and omissions caused by missing sample information in the view, and improving the accuracy of multi-view clustering results. Attached Figure Description

[0078] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0079] Figure 1 A flowchart illustrating a multi-view clustering method provided in this application embodiment. Figure 1 ;

[0080] Figure 2 A flowchart illustrating a multi-view clustering method provided in this application embodiment. Figure 2 ;

[0081] Figure 3 A flowchart illustrating a multi-view clustering method provided in this application embodiment. Figure 3 ;

[0082] Figure 4 This is a schematic diagram of the structure of a multi-view clustering device provided in an embodiment of this application;

[0083] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0084] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0085] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0086] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.

[0087] It should be noted that the multi-view clustering method, apparatus, device, storage medium, and program product provided in this application can be used in the fintech field, as well as in any other field. The application fields of the multi-view clustering method, apparatus, device, storage medium, and program product in this application are not limited. For ease of understanding, subsequent embodiments will be described using the fintech field as a background.

[0088] In the wave of banking digitalization, the heterogeneous "digital footprints" left by the same customer across multiple channels such as mobile banking, bank counters, ATMs, POS terminals, and corporate online banking naturally constitute multi-view data, becoming an invaluable goldmine for accurately portraying customer risk profiles. Over the past decade, multi-view clustering research has attempted to fuse these views through methods such as collaborative training, subspace learning, graph neural networks, and contrastive learning to achieve unsupervised customer segmentation. However, in real-world production environments, issues such as camera outages, missing voiceprints, and lost radio frequency signals due to compliance issues, equipment malfunctions, or authorization differences shatter the ideal assumption that "all samples are complete in all views," reducing the accuracy of multi-view clustering results.

[0089] This application provides a multi-view clustering method. A similarity matrix is ​​constructed based on the similarity between feature elements within the view dataset of each view. Multiple similarity matrices are multiplied to obtain a consistency matrix. Each similarity matrix is ​​then reconstructed based on the consistency matrix to obtain a reconstructed similarity matrix. Subsequently, multi-view clustering is performed based on these reconstructed similarity matrices to obtain clustering results indicating the cluster label corresponding to each sample. By utilizing the complementary information of the consistency matrices, the method maintains the stability of the cluster boundaries, reduces misclassification and omissions caused by missing sample information in the views, and improves the accuracy of multi-view clustering results.

[0090] Figure 1 A flowchart illustrating a multi-view clustering method provided in this application embodiment. Figure 1 .like Figure 1 As shown, the method includes:

[0091] S101. Obtain the view dataset corresponding to each view.

[0092] Each view dataset contains feature elements of multiple samples under the corresponding view.

[0093] In this step, with the rapid development of digital data collection methods, the same customer in a financial scenario often leaves different "digital footprints" on different channels: click logs of mobile banking apps, facial images from high-definition cameras at bank counters, voiceprint records from ATMs, RFID fingerprints from POS terminals, and even transaction time-series data from corporate online banking. These heterogeneous and heterogeneous features naturally constitute multiple views, comprehensively depicting the customer's behavioral patterns and risk profile. How to organically integrate these views and accurately segment customer groups without manual annotation has become a key link for banks to improve risk control, marketing, and operational efficiency. Multi-view clustering is therefore regarded as the alchemy of "data gold mine." Over the past decade, academia and industry have proposed numerous solutions for multi-view clustering. Typical approaches include: semi-supervised methods based on collaborative training, which iteratively optimize pseudo-labels using the complementarity between views; algorithms based on subspace learning, which map each view to a consistent latent low-dimensional space before implementing K-means; graph neural network-based techniques, which construct an adjacency graph for each view and obtain a unified representation through attention or tensor fusion; and self-supervised frameworks based on contrastive learning, which achieve representation alignment by maximizing mutual information between views. However, the real-world production environment in banks is far more complex than experimental scenarios. Due to compliance requirements, equipment failures, network jitter, or differences in customer authorization, situations often arise such as a branch's camera going offline on a given day, certain mobile phone models refusing to collect voiceprints, or some POS terminals losing radio frequency signals, resulting in missing blocks or even entire columns in the corresponding views. Most existing methods assume that "each sample is complete and usable in all views," which reduces the accuracy of multi-view clustering results once missing data occurs. Therefore, this embodiment constructs a consistency matrix based on the multi-view similarity matrix and reconstructs and completes each view in reverse, thereby improving the accuracy of the clustering results.

[0094] View dataset The set of features for all users in the v-th view, expressed mathematically as: , This refers to the feature dimension, specifically the number of feature elements for a single user in the v-th view. For example, the acoustic view contains 20 features such as speech rate, pitch, and emotional fluctuation. This represents the sample size, or the number of users. Let i be the feature vector of sample i in the v-th view, i.e., the feature elements. For example, the feature elements of the user in the acoustic view are: .

[0095] Specifically, raw data is extracted from multiple business systems of the bank, and the feature dimensions of the corresponding view are determined based on the business systems. Based on multiple raw data and defined feature dimensions Determine the view dataset for this view. For example, extract raw data from the call center system and determine the feature dimensions of the view corresponding to the call center system. The value is 20. Based on the original data and feature dimension 20, the first view, i.e., the acoustic view, is determined as follows: .

[0096] S102. Based on the similarity between feature elements within each view dataset, construct a similarity matrix for each view dataset.

[0097] Similarity matrix Let be the matrix representing the connection probabilities between two items for all samples in the v-th view. Similarity matrix The similarity element in the i-th row and j-th column indicates the similarity between sample i and sample j in the v-th view. The similarity is negatively correlated with the distance between the corresponding samples.

[0098] Specifically, the first optimization problem is solved for each view dataset to obtain the similarity matrix corresponding to that view dataset. The first optimization problem is:

[0099] .

[0100] The corresponding first constraint is:

[0101] .

[0102] in, The sum of the similarities of sample i with respect to all other samples is equal to 1, which means that the similarity distribution of each sample is forced to be normalized.

[0103] S103. Multiply the similarity matrices of multiple view datasets to obtain a consistency matrix, and reconstruct the similarity matrix of each view dataset based on the consistency matrix to obtain a reconstructed similarity matrix.

[0104] In this step, considering that the consistency matrix, through the solidification and reverse mapping mechanism of cross-view consensus, can supplement and correct the similarity information of missing views when some view data is missing, by utilizing the high-confidence associations jointly confirmed by the remaining views: its Hadamard product operation suppresses the accidental similarity generated only by a single view to near zero, while amplifying the strong associations jointly supported by all views, thereby introducing complementary cross-view information into the missing region; subsequently, using the consistency matrix as an alignment benchmark, the consensus strength is injected back into the missing view through the reconstruction step, so that the similarity distribution of the view is consistent with the overall cross-view structure, thus both filling the information gap and suppressing noise interference. Therefore, this embodiment supplements the missing information of each view through the consistency matrix.

[0105] Specifically, the similarity matrices of multiple datasets are multiplied to obtain a consistency matrix corresponding to the multiple similarity matrices. This allows for the extraction of consistent information between different similarity matrices, which is information present in every view. The specific formula is as follows: Consistency matrix elements in .in, Used to indicate the cross-view similarity between sample i and sample j in the m-th view.

[0106] Only when all views are assigned high similarity, such as , … The average value is relatively large, and the calculated value is... It is close to 1 if any view has low similarity, such as Calculated This strengthens consistency.

[0107] For example, the similarity between Sample 1 and Sample 2 in the first view (transaction view) is as follows: The similarity between Sample 1 and Sample 2 in the second view (acoustic view) is as follows: The similarity between Sample 1 and Sample 2 in the third view (behavioral view) is as follows: .

[0108] The elements in the consistency matrix used to indicate the cross-view similarity between sample 1 and sample 2 are:

[0109] .

[0110] After obtaining the consistency matrix, a second optimization problem is solved based on the feature elements in each column of each similarity matrix and the corresponding elements in the consistency matrix. This second optimization problem determines the feature weights of each feature element in the similarity matrix. The higher the degree of similarity between the feature element in the column and the corresponding element in the consistency matrix, the higher the feature weight; the lower the degree of similarity, the lower the feature weight. After obtaining the feature weights of each feature element, a weight matrix for the similarity matrix is ​​determined based on multiple feature weights. A reconstruction operation, i.e., a column-weighted operation, is then performed based on this weight matrix and the corresponding similarity matrix to obtain a reconstructed similarity matrix. This allows information completion of the view corresponding to the reconstructed similarity matrix to be achieved through the consistency matrix. The specific implementation process will be described in subsequent embodiments and will not be elaborated upon here.

[0111] S104. Perform multi-view clustering based on multiple reconstructed similarity matrices to obtain clustering results.

[0112] The clustering results are used to indicate the clustering label corresponding to each sample.

[0113] In this step, in a real-world banking production environment, equipment and environmental issues can cause unreliable noise or holes in the similarity matrices of some views. To prevent these views with large errors from dragging down the overall clustering result during fusion, the view weights of the similarity matrices for each view are determined through multiple iterations. Based on the view weights from the last iteration and multiple similarity matrices, a unified graph matrix is ​​determined to generate the final clustering result. The process is as follows: each similarity matrix is ​​assigned an initial weight, and a weighted integration operation is performed based on multiple similarity matrices and their corresponding initial weights to obtain the unified graph matrix before iteration. This embodiment does not impose any limitations on the initial weights. For example, the average weight can be determined based on the number of views and used as the initial weight; alternatively, historical multi-view aggregation records can be retrieved to determine the multiple view weights finally determined after multiple iterations for each view, and the initial weight of that view can be determined based on the average of the aforementioned multiple view weights. Then, an iterative operation is performed: the view weight of each view after this iteration is determined based on the unified graph matrix before this iteration. This process reduces the view weights of views with large errors and increases the view weights of views with small errors. In the first iteration, the unified graph matrix before this iteration is the unified graph matrix obtained by weighting according to the initial weights.

[0114] Furthermore, after obtaining the unified graph matrix used to determine the clustering results, a rank constraint is applied to the unified graph matrix to obtain the clustering results of multi-graph clustering. This rank constraint is used to specify the number of sub-matrices obtained by splitting the unified graph matrix, that is, the number of cluster labels.

[0115] This application provides a multi-view clustering method. A similarity matrix is ​​constructed based on the similarity between feature elements within the view dataset of each view. Multiple similarity matrices are multiplied to obtain a consistency matrix. Each similarity matrix is ​​then reconstructed based on the consistency matrix to obtain a reconstructed similarity matrix. Subsequently, multi-view clustering is performed based on these reconstructed similarity matrices to obtain clustering results indicating the cluster label corresponding to each sample. By utilizing the complementary information of the consistency matrices, the method maintains the stability of the cluster boundaries, reduces misclassification and omissions caused by missing sample information in the views, and improves the accuracy of multi-view clustering results.

[0116] Figure 2 A flowchart illustrating a multi-view clustering method provided in this application embodiment. Figure 2 This application provides a detailed description of the steps for reconstructing the similarity matrix of each view dataset based on the consistency matrix. For example... Figure 2 As shown, the method includes:

[0117] S201. For any column of similarity elements in each similarity matrix, determine the feature weight of the corresponding column of similarity elements based on the difference between the similarity elements and the reference elements at the same position in the consistency matrix.

[0118] The feature weights are inversely proportional to the differences.

[0119] In this step, the similarity elements in any column of the similarity matrix of the v-th view are... The similarity between sample j and the other n samples is recorded. The consistency matrix A is a matrix obtained across all views, with its j-th column referencing elements. It is considered as a reference similarity of sample j in the global view.

[0120] To achieve the compression of the noisy view and the enhancement of the reliable view, for each Corresponding The second optimization problem is constructed using Euclidean distance: with weighted least squares as the objective, minimize...

[0121] ,

[0122] And in the second constraint: Seeking a solution. (Here) Defined as "feature weights," their physical meaning is: the greater the difference between the integer similarity of sample j in view v and the consistency matrix, the better. The smaller the value, the more the contribution of the view to the sample is automatically diminished; the smaller the gap, the more... The larger the value, the more reliable the view's description of the sample.

[0123] S202. Based on all feature weights corresponding to each similarity matrix, determine the weight matrix of each similarity matrix.

[0124] In this step, after obtaining the feature weights corresponding to the similar elements in each column of each similarity matrix, a weight matrix is ​​constructed based on the multiple feature weights. The element in the j-th column of the feature matrix corresponding to the v-th view.

[0125] S203. Based on the weight matrix and corresponding row constraints of each similar matrix, perform a reconstruction operation on the similar matrix to obtain a reconstructed similar matrix.

[0126] In this step, for each similarity matrix, each similar element in the similarity matrix is ​​multiplied by its corresponding weight according to the weight matrix of the similarity matrix, resulting in: Then, based on the weighted columns and the corresponding original indices, the similarity matrix is ​​updated until all columns in the similarity matrix have been updated, resulting in a candidate similarity matrix. Next, row constraints are applied to the candidate similarity matrix, including forcing diagonal elements to 0. Negative elements are truncated to 0. ; standardization .

[0127] S204. Based on the multiple reconstruction similarity matrices, obtain the consistency matrix of the multiple view datasets again, and determine whether the reconstruction termination condition is met based on the newly obtained consistency matrix and the one-time matrix before this reconstruction.

[0128] In this step, the similarity matrix of the views is reconstructed multiple times to continuously enhance the consensus information among the multiple views. Therefore, whether the consensus information increases before and after reconstruction is also used as a termination condition to determine whether to end the reconstruction. Whether the consensus information increases is determined by the consensus strength index, namely, the position information value. The position information value indicates the sum of all non-zero elements in the current consensus matrix. If the position information value shows a monotonically increasing trend, it indicates that the consensus information among the multiple views is continuously enhancing, and the similarity matrix of each view needs to be reconstructed again; if the position information value decreases, it indicates that continuing reconstruction cannot enhance the consensus information among the multiple views, and the reconstruction of the similarity matrix needs to be stopped. This results in multiple reconstructed similarity matrices used for multi-view clustering. The specific implementation method is as follows:

[0129] Summing the newly acquired consistency matrix and the consistency matrix before this reconstruction yields the position information value corresponding to the newly acquired consistency matrix, and the position information value corresponding to the consistency matrix before this reconstruction.

[0130] If the position information value corresponding to the consistency matrix obtained again is less than the position information value corresponding to the consistency matrix before this reconstruction, then the reconstruction end condition is met.

[0131] If the position information value corresponding to the consistency matrix obtained again is greater than or equal to the position information value corresponding to the consistency matrix before this reconstruction, then it is determined that the reconstruction termination condition is not met.

[0132] S205. If not satisfied, the reconstruction similarity matrix of each view dataset is reconstructed based on the re-acquired consistency matrix until the reconstruction termination condition is met.

[0133] In this step, if the location information value corresponding to the consistency matrix obtained again is greater than or equal to the location information value corresponding to the consistency matrix before this reconstruction, it is determined that the reconstruction termination condition is not met, and it is necessary to continue to reconstruct the similarity matrix of each view.

[0134] Specifically, the reconstructed similarity matrix of each view dataset is reconstructed based on the re-obtained consistency matrix until the positional information value of the reconstructed consistency matrix is ​​less than the positional information value of the consistency matrix before the reconstruction. The specific process has been described in detail in the foregoing embodiments and will not be elaborated upon here.

[0135] For example, the initial consistency matrix The non-zero elements in the array are:

[0136] .

[0137] Calculate the initial consistency matrix The location information value is:

[0138] .

[0139] Consistency Matrix after First Reconstruction The non-zero elements in the array are:

[0140] .

[0141] Calculate the initial consistency matrix The location information value is:

[0142] .because If the reconstruction stop condition is not met, continue the reconstruction of the similarity matrix.

[0143] Consistency Matrix after Second Reconstruction The non-zero elements in the array are:

[0144] .

[0145] Calculate the initial consistency matrix The location information value is:

[0146] .

[0147] because If the reconstruction stop condition is not met, continue the reconstruction of the similarity matrix.

[0148] Consistency Matrix after Third Reconstruction The non-zero elements in the array are:

[0149] .

[0150] Calculate the initial consistency matrix The location information value is:

[0151] .

[0152] because The refactoring stop condition is determined to be met.

[0153] This application provides a multi-view clustering method. It determines the feature weight of each column based on the difference between the similarity element in the similarity matrix and the corresponding reference element in the consistency matrix, and constructs a weight matrix based on multiple feature weights. Then, the similarity matrix is ​​reconstructed based on the weight matrix and row constraints to obtain a reconstructed similarity matrix. Next, the reconstructed consistency matrix is ​​determined based on multiple reconstructed similarity matrices. The method then determines whether the reconstruction termination condition is met based on the positional information values ​​of the re-obtained consistency matrix and the initial matrix before reconstruction. If the condition is not met, the reconstructed similarity matrix for each view dataset is reconstructed again based on the re-obtained consistency matrix until the reconstruction termination condition is met. Through this method, a closed-loop reconstruction of "column-level error—feature weight—matrix reconstruction—consistency check" can continuously amplify the consensus information of multiple views and suppress single-view noise without introducing additional manual annotation. Errors caused by missing or abnormal views are gradually weakened through feature weights, while jointly confirmed strong associations are repeatedly strengthened, thereby improving the reliability of the clustering results.

[0154] Figure 3 A flowchart illustrating a multi-view clustering method provided in this application embodiment. Figure 2 This application provides a detailed description of the steps involved in performing multi-view clustering based on multiple reconstructed similarity matrices to obtain clustering results. For example... Figure 3 As shown, the method includes:

[0155] S301. Determine the unified graph matrix before iteration based on the multiple reconstructed similarity matrices.

[0156] In this step, in a real-world banking environment, due to factors such as camera outages, moisture in the voiceprint collector, or POS machine radio frequency interference, the similarity matrix corresponding to any view may contain unreliable noise or holes. To avoid such defective views negatively impacting the overall clustering results during the fusion stage, this embodiment uses an "iterative weight purification mechanism" to dynamically correct the view weight of each view, that is, the view weight of the corresponding reconstructed similarity matrix.

[0157] Since the view weight of each view is determined based on the difference between the unified graph matrix and the reconstruction similarity matrix of that view: the greater the difference, the smaller the view weight obtained after iteration; the smaller the difference, the greater the view weight obtained after iteration. Therefore, before the iteration operation, an initial unified graph matrix needs to be determined based on multiple reconstruction similarity matrices, that is, the unified graph matrix before iteration.

[0158] Specifically, each reconstructed similarity matrix is ​​assigned an initial weight. This initial weight can be either averaged based on the number of views or calculated by calling historical multi-view aggregation records, using the average of the view weights obtained from the final convergence of each iteration as the initial weight. Then, a weighted integration operation is performed to obtain the unified graph matrix before iteration. The formula is: .in, This is the unified graph matrix before the iteration, i.e., the unified graph matrix of the 0th iteration; Let V be the initial view weights for the v-th view, i.e., the view weights before the iteration. Let v be the reconstructed similarity matrix of the v-th view.

[0159] S302. Based on the unified graph matrix before iteration, determine the view weights of each of the reconstructed similarity matrices after iteration, and based on the multiple reconstructed similarity matrices and the multiple iterated view weights, determine the unified graph matrix after iteration.

[0160] In this step, when noise appears in a view, the Frobenius distance between the reconstructed similarity matrix of that view and the unified graph matrix before iteration increases, meaning the difference increases. Therefore, it is necessary to weaken the influence of this view during view fusion. Thus, this embodiment calculates the view weight after iteration based on the Frobenius distance between the unified graph matrix before iteration and each reconstructed similarity matrix. The closer the Frobenius distance between the unified graph matrix and the reconstructed similarity matrix, the larger the corresponding view weight after iteration; the farther the Frobenius distance between the unified graph matrix and the reconstructed similarity matrix, the smaller the corresponding view weight after iteration.

[0161] Specifically, based on the unified graph matrix before iteration, the difference between this unified graph matrix and the reconstruction similarity matrix of each view, i.e., the Frobenius distance, is calculated. Then, the view weights of each view are iterated based on the Frobenius distance, thus obtaining the iterated view weights. The formula for calculating the iterated view weights based on the unified graph matrix before iteration is as follows:

[0162] .

[0163] Where k is the number of iterations. This is the unified graph matrix before the k-th iteration.

[0164] After obtaining the view weights after each iteration of the reconstructed similarity matrix, the third optimization problem is solved based on multiple view weights: The corresponding third constraint is: The unified graph matrix after the k-th iteration is obtained. .

[0165] The third constraint is used to restrict each row in the unified graph matrix after iteration to zero negative values ​​and to normalize rows.

[0166] S303. Based on the unified graph matrix after this iteration and the unified graph matrix before this iteration, determine whether the iteration termination condition is met.

[0167] In this step, if the difference between the unified graph matrix before and after the iteration is small, it indicates that the iteration has not introduced new cross-view information; that is, the graph structure will not change significantly. Therefore, it can be determined that the consensus of the multi-view has converged. If the iteration continues, it will lead to redundant computation and will not indicate the purity of clustering based on the unified graph matrix. It is then determined that the iteration can be stopped, saving computational resources while ensuring robustness.

[0168] Furthermore, since the Frobenius distance can accurately indicate the difference between two matrices, this embodiment uses the Frobenius distance, i.e., the F-norm, to determine the difference between the unified graph matrices before and after iteration, and thus determine whether the iteration termination condition is met. The formula for calculating the F-norm of the unified graph matrices before and after iteration is: The corresponding iteration termination condition is: , The value of can be determined based on the actual application scenario. For example, It can be 10 -3 -10 -5 Any value between.

[0169] The specific process is as follows:

[0170] Determine the difference matrix based on the unified graph matrix before and after this iteration;

[0171] The F-norm is calculated based on the difference matrix, and the relationship between the F-norm and the convergence threshold is determined.

[0172] If the F-norm is less than the convergence threshold, then the iteration termination condition is determined to be met.

[0173] If the F-norm is greater than or equal to the convergence threshold, then the iteration termination condition is determined not to be met.

[0174] S304. If not satisfied, the view weights of each reconstructed similarity matrix are iterated again based on the unified graph matrix after the iteration to obtain the view weights after the k-th iteration and the corresponding unified graph matrix after the k-th iteration, until the iteration termination condition is met.

[0175] Where k is the number of iterations.

[0176] In this step, when the F-norm of the difference matrix is ​​greater than or equal to the convergence threshold, it indicates that the consensus information of the multiple views is still unstable. That is, the similarity of some sample pairs is still changing drastically, and the compensation for noisy views or missing regions has not been fully completed. The principle is that the view weights and the unified graph matrix are mutually causal: views with large errors need to have their view weights further reduced, while views with small errors need to have their view weights further increased. Each fine-tuning of the view weights, in turn, changes the edge weight distribution of the unified graph matrix, thereby correcting the errors of each view again. Only after a sufficient number of rounds of mutual calibration between "weights and the unified graph" will the F-norm of the difference matrix of the unified graph matrix before and after the iteration fall below the convergence threshold. At this point, the graph structure tends to be satisfactory, the multi-view consensus reaches its optimum, and the iteration can terminate.

[0177] Therefore, when it is determined that the iteration termination condition is not met based on the unified graph matrix before and after the iteration, the view weights of each reconstructed similarity matrix are iterated again based on the unified graph matrix after the iteration to obtain the view weights after the k-th iteration and the corresponding unified graph matrix after the k-th iteration, until the iteration termination condition is met.

[0178] S305. If satisfied, then apply rank constraints to the unified graph matrix after this iteration to obtain the clustering results.

[0179] In this step, since the number of zero eigenvalues ​​in the Laplacian matrix is ​​strictly equal to the number of connected components in the corresponding unified graph matrix, the unified graph matrix can only be resolved into submatrices with the same number of cluster labels when the number of zero eigenvalues ​​is exactly equal to the preset number of cluster labels. Each submatrix corresponds to one cluster label, and all samples within the submatrix are assigned the cluster label of that submatrix, thus directly determining the clustering result. The specific process is as follows:

[0180] Determine the diagonal matrix of the iterated unified graph matrix, and determine the Laplace matrix based on the diagonal matrix and the iterated unified graph matrix;

[0181] Given a preset number of clusters, apply a rank constraint to the Laplacian matrix;

[0182] If the Laplacian matrix satisfies the rank constraint, the iterated unified graph matrix is ​​split into multiple sub-matrices based on the number of clusters, and the clustering label of each sub-matrix is ​​determined.

[0183] Based on each submatrix and its corresponding clustering label, determine the clustering label for each sample, and determine the clustering result based on the clustering label for each sample.

[0184] Specifically, based on the unified graph matrix obtained after iteration, the diagonal matrix is ​​calculated using the following formula: Next, the Laplacian matrix is ​​calculated based on the diagonal matrix and the unified graph matrix, using the following formula: It should be noted that U is the unified graph matrix obtained after the iteration. Then, the rank constraint is determined based on the preset number of clusters and the total number of samples, as follows:

[0185] ( Where c is the number of clusters. Next, a rank constraint is applied to the Laplacian matrix. The process is as follows: verify whether the current Laplacian matrix satisfies the rank constraint, that is, determine the current connected component based on the current Laplacian matrix, and determine whether the current connected component is equal to the number of clusters.

[0186] Furthermore, if it is determined that the current connected component is not equal to the number of clusters, the regularization coefficient is increased or decreased. Then, by solving the fourth optimization problem, the unified graph matrix and embedding matrix are updated until the Laplacian matrix satisfies the applied rank constraint. Specifically, if the current connected component is greater than the number of clusters, the regularization coefficient is decreased; if the current connected component is less than the number of clusters, the regularization coefficient is increased.

[0187] The fourth optimization problem that needs to be solved is: The corresponding fourth constraint is: .in, The regularization coefficient is . This is the embedding matrix.

[0188] It should be noted that when When large enough, force This ensures that the unified graph matrix has a cluster size of *c* connected components, i.e., *c* connected components. for The i-th smallest eigenvalue.

[0189] After confirming that the Laplacian matrix satisfies the rank constraint, it indicates that the unified graph matrix already contains the number of connected components corresponding to the number of clusters. The unified graph matrix is ​​then split into multiple sub-matrices along these connected components, with the number of sub-matrices equal to the number of clusters. Each sub-matrix is ​​then assigned a cluster label, and all samples covered by that sub-matrix are assigned the corresponding cluster label. Finally, the labels of all samples are aggregated to form the complete clustering result.

[0190] This application provides a multi-view clustering method. A unified graph matrix is ​​determined before iteration based on multiple reconstructed similarity matrices and initial weights. Then, the view weights after iteration are determined based on the difference between the unified graph matrix and each reconstructed similarity matrix. The unified graph matrix after iteration is then determined based on the unified graph matrices before and after iteration. If the iteration termination adjustment is not satisfied, the view weights and unified graph matrix are iterated until the iteration termination adjustment is satisfied. If satisfied, a rank constraint is applied to the Laplacian matrix of the final unified graph matrix. When the rank constraint is satisfied, the unified graph matrix is ​​split into multiple sub-matrices, and each sub-matrix is ​​assigned a corresponding clustering label. Then, all samples covered by each sub-matrix are assigned the corresponding clustering label to obtain the clustering result. The above method can continuously weaken unreliable views and amplify reliable views in scenarios with missing views, noise, or anomalies, until cross-view consensus converges. Then, by using rank constraints, the unified graph matrix is ​​precisely split into the target number of sub-matrices and directly assigned clustering labels. This avoids error accumulation and eliminates the need for post-processing segmentation, significantly improving the robustness and clustering accuracy of multi-view clustering results.

[0191] Figure 4 This is a schematic diagram of a multi-view clustering device provided in an embodiment of this application. Figure 4 As shown, the multi-view clustering device 40 includes: an acquisition module 401, a construction module 402, and a clustering module 403, wherein:

[0192] The acquisition module 401 is used to acquire the view dataset corresponding to each view, wherein each view dataset contains feature elements of multiple samples under the corresponding view;

[0193] The construction module 402 is used to construct a similarity matrix for each view dataset based on the similarity between feature elements within each view dataset;

[0194] The construction module 402 is further configured to multiply the similarity matrices of multiple view datasets to obtain a consistency matrix, and reconstruct the similarity matrix of each view dataset based on the consistency matrix to obtain a reconstructed similarity matrix;

[0195] The clustering module 403 is used to perform multi-view clustering based on multiple reconstructed similarity matrices to obtain clustering results, wherein the clustering results are used to indicate the clustering label corresponding to each sample.

[0196] In one possible implementation, the construction module 402 is further configured to:

[0197] Based on the consistency matrix, the similarity matrix of each view dataset is reconstructed to obtain the reconstructed similarity matrix;

[0198] Based on the multiple reconstruction similarity matrices, the consistency matrices of the multiple view datasets are obtained again, and based on the newly obtained consistency matrices and the one-time matrix before this reconstruction, it is determined whether the reconstruction termination condition is met.

[0199] If the condition is not met, the reconstruction similarity matrix of each view dataset is reconstructed based on the re-obtained consistency matrix until the reconstruction termination condition is met.

[0200] In one possible implementation, the construction module 402 is further configured to:

[0201] For any column of similarity elements in each similarity matrix, the feature weight of the corresponding column of similarity elements is determined based on the difference between the similarity element and the reference element at the same position in the consistency matrix, wherein the feature weight is inversely proportional to the difference.

[0202] Based on all feature weights corresponding to each similarity matrix, determine the weight matrix of each similarity matrix;

[0203] Based on the weight matrix and corresponding row constraints of each similar matrix, a reconstruction operation of the similar matrix is ​​performed to obtain the reconstructed similar matrix.

[0204] In one possible implementation, the construction module 402 is further configured to:

[0205] Summing the newly acquired consistency matrix and the consistency matrix before this reconstruction yields the position information value corresponding to the newly acquired consistency matrix, and the position information value corresponding to the consistency matrix before this reconstruction.

[0206] If the position information value corresponding to the consistency matrix obtained again is less than the position information value corresponding to the consistency matrix before this reconstruction, then the reconstruction end condition is met.

[0207] If the position information value corresponding to the consistency matrix obtained again is greater than or equal to the position information value corresponding to the consistency matrix before this reconstruction, then it is determined that the reconstruction termination condition is not met.

[0208] In one possible implementation, the clustering module 403 is further configured to:

[0209] The unified graph matrix before iteration is determined based on multiple reconstructed similarity matrices;

[0210] Based on the unified graph matrix before iteration, determine the view weights of each of the reconstructed similarity matrices after iteration, and determine the unified graph matrix after iteration based on the multiple reconstructed similarity matrices and the multiple view weights after iteration;

[0211] Based on the unified graph matrix after this iteration and the unified graph matrix before this iteration, determine whether the iteration termination condition is met;

[0212] If not satisfied, the view weights of each reconstructed similarity matrix are iterated again based on the unified graph matrix after the iteration to obtain the view weights after the kth iteration and the corresponding unified graph matrix after the kth iteration, until the iteration termination condition is met, where k is the iteration number;

[0213] If satisfied, then apply rank constraints to the unified graph matrix after this iteration to obtain the clustering results.

[0214] In one possible implementation, the clustering module 403 is further configured to:

[0215] Determine the difference matrix based on the unified graph matrix before and after this iteration;

[0216] The F-norm is calculated based on the difference matrix, and the relationship between the F-norm and the convergence threshold is determined.

[0217] If the F-norm is less than the convergence threshold, then the iteration termination condition is determined to be met.

[0218] If the F-norm is greater than or equal to the convergence threshold, then the iteration termination condition is determined not to be met.

[0219] In one possible implementation, the clustering module 403 is further configured to:

[0220] Determine the diagonal matrix of the iterated unified graph matrix, and determine the Laplace matrix based on the diagonal matrix and the iterated unified graph matrix;

[0221] Given a preset number of clusters, apply a rank constraint to the Laplacian matrix;

[0222] If the Laplacian matrix satisfies the rank constraint, the iterated unified graph matrix is ​​split into multiple sub-matrices based on the number of clusters, and the clustering label of each sub-matrix is ​​determined.

[0223] Based on each submatrix and its corresponding clustering label, determine the clustering label for each sample, and determine the clustering result based on the clustering label for each sample.

[0224] This embodiment provides a multi-view clustering device that can execute the multi-view clustering method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0225] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 5 As shown, the electronic device 50 includes a processor 501 and a memory 502 communicatively connected to the processor 501. Optionally, the electronic device 50 further includes a communication component 503. The processor 501, memory 502, and communication component 503 are connected via a bus 504.

[0226] The memory 502 stores computer-executed instructions;

[0227] The processor 501 executes computer execution instructions stored in the memory 502 to implement the method as described in any of the preceding descriptions.

[0228] At least one processor 501 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0229] Optionally, in specific implementations, the processor 501 and memory 502 are implemented independently. In this case, the processor 501 and memory 502 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc., but this does not imply that there is only one bus or one type of bus.

[0230] Optionally, in a specific implementation, if the processor 501 and the memory 502 are integrated on a single chip, the processor 501 and the memory 502 can communicate through an internal interface.

[0231] This application also provides a computer storage medium storing computer execution instructions, which, when executed by a processor, implement the aforementioned method for handling damaged components.

[0232] The aforementioned computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The computer-readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0233] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Alternatively, the readable storage medium can be an integral part of the processor. Both the processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the control device of a garment handling apparatus.

[0234] The division of units described herein is merely a logical functional division. In actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0235] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0236] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0237] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0238] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0239] The technical solutions of this application have been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it is readily understood by those skilled in the art that the scope of protection of this application is obviously not limited to these specific embodiments. The above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A multi-view clustering method, characterized in that, The method includes: Obtain the view dataset corresponding to each view, where each view dataset contains feature elements of multiple samples under the corresponding view; Based on the similarity between feature elements within each view dataset, a similarity matrix is ​​constructed for each view dataset; The similarity matrices of multiple view datasets are multiplied to obtain a consistency matrix, and the similarity matrix of each view dataset is reconstructed based on the consistency matrix to obtain a reconstructed similarity matrix; Multi-view clustering is performed based on multiple reconstructed similarity matrices to obtain clustering results, wherein the clustering results are used to indicate the clustering label corresponding to each sample.

2. The method according to claim 1, characterized in that, The process of reconstructing the similarity matrix of each view dataset based on the consistency matrix to obtain the reconstructed similarity matrix includes: Based on the consistency matrix, the similarity matrix of each view dataset is reconstructed to obtain the reconstructed similarity matrix; Based on the multiple reconstruction similarity matrices, the consistency matrices of the multiple view datasets are obtained again, and based on the newly obtained consistency matrices and the one-time matrix before this reconstruction, it is determined whether the reconstruction termination condition is met. If the condition is not met, the reconstruction similarity matrix of each view dataset is reconstructed based on the re-obtained consistency matrix until the reconstruction termination condition is met.

3. The method according to claim 2, characterized in that, The process of reconstructing the similarity matrix of each view dataset based on the consistency matrix to obtain the reconstructed similarity matrix includes: For any column of similarity elements in each similarity matrix, the feature weight of the corresponding column of similarity elements is determined based on the difference between the similarity element and the reference element at the same position in the consistency matrix, wherein the feature weight is inversely proportional to the difference. Based on all feature weights corresponding to each similarity matrix, determine the weight matrix of each similarity matrix; Based on the weight matrix and corresponding row constraints of each similar matrix, a reconstruction operation of the similar matrix is ​​performed to obtain the reconstructed similar matrix.

4. The method according to claim 2, characterized in that, The determination of whether the reconstruction termination condition is met based on the re-acquired consistency matrix and the one-time matrix before this reconstruction includes: Summing the newly acquired consistency matrix and the consistency matrix before this reconstruction yields the position information value corresponding to the newly acquired consistency matrix, and the position information value corresponding to the consistency matrix before this reconstruction. If the position information value corresponding to the consistency matrix obtained again is less than the position information value corresponding to the consistency matrix before this reconstruction, then the reconstruction end condition is met. If the position information value corresponding to the consistency matrix obtained again is greater than or equal to the position information value corresponding to the consistency matrix before this reconstruction, then it is determined that the reconstruction termination condition is not met.

5. The method according to claim 1, characterized in that, The multi-view clustering based on multiple reconstructed similarity matrices yields clustering results, including: The unified graph matrix before iteration is determined based on multiple reconstructed similarity matrices; Based on the unified graph matrix before iteration, determine the view weights of each of the reconstructed similarity matrices after iteration, and determine the unified graph matrix after iteration based on the multiple reconstructed similarity matrices and the multiple view weights after iteration; Based on the unified graph matrix after this iteration and the unified graph matrix before this iteration, determine whether the iteration termination condition is met; If not satisfied, the view weights of each reconstructed similarity matrix are iterated again based on the unified graph matrix after the iteration to obtain the view weights after the kth iteration and the corresponding unified graph matrix after the kth iteration, until the iteration termination condition is met, where k is the iteration number; If satisfied, then apply rank constraints to the unified graph matrix after this iteration to obtain the clustering results.

6. The method according to claim 5, characterized in that, The step of determining whether the iteration termination condition is met based on the unified graph matrix after this iteration and the unified graph matrix before this iteration includes: Determine the difference matrix based on the unified graph matrix before and after this iteration; The F-norm is calculated based on the difference matrix, and the relationship between the F-norm and the convergence threshold is determined. If the F-norm is less than the convergence threshold, then the iteration termination condition is determined to be met. If the F-norm is greater than or equal to the convergence threshold, then the iteration termination condition is determined not to be met.

7. The method according to claim 5, characterized in that, The rank constraint is applied to the unified graph matrix after this iteration to obtain the clustering results, including: Determine the diagonal matrix of the iterated unified graph matrix, and determine the Laplace matrix based on the diagonal matrix and the iterated unified graph matrix; Given a preset number of clusters, apply a rank constraint to the Laplacian matrix; If the Laplacian matrix satisfies the rank constraint, the iterated unified graph matrix is ​​split into multiple sub-matrices based on the number of clusters, and the clustering label of each sub-matrix is ​​determined. Based on each submatrix and its corresponding clustering label, determine the clustering label for each sample, and determine the clustering result based on the clustering label for each sample.

8. A multi-view clustering device, characterized in that, include: The module includes a retrieval module, a construction module, and a clustering module, among which: The acquisition module is used to acquire the view dataset corresponding to each view, wherein each view dataset contains feature elements of multiple samples under the corresponding view; The construction module is used to construct a similarity matrix for each view dataset based on the similarity between feature elements within each view dataset; The construction module is further configured to multiply the similarity matrices of multiple view datasets to obtain a consistency matrix, and reconstruct the similarity matrix of each view dataset based on the consistency matrix to obtain a reconstructed similarity matrix; The clustering module is used to perform multi-view clustering based on multiple reconstructed similarity matrices to obtain clustering results, wherein the clustering results are used to indicate the clustering label corresponding to each sample.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 7.

11. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images