View clustering analysis method and device, computer equipment, readable storage medium and program product

Abstract

The invention relates to a view clustering analysis method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: acquiring an original view data set; according to the view type of each view in the original view data set, a target domain division algorithm is determined, each view is processed according to the domain number of each scale level through the target domain division algorithm, and a label indication matrix of each view under each scale level is obtained; based on the label indication matrix of each view under each scale level, determining each domain feature under each scale level; performing fuzzy clustering analysis on each domain feature to obtain a space-level membership matrix under each scale level; and determining a clustering result indication matrix of the original view data set based on the space-level membership matrix, the tensor low-rank item and the clustering indication learning item under each scale level. By adopting the method, the calculation efficiency can be improved while the clustering precision is improved, and the flexibility can also be improved.

Description

Technical Field

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

[0002] With the advent of the big data era, multi-view data analysis has become a research hotspot in the fields of machine learning and data mining. Multi-view data is widely present in practical applications; for example, images can be described by both color and texture features, and web pages contain both textual information and link structure information. Multi-view clustering, by fusing information from different views, can discover data structures that are difficult to identify using single-view methods, and has significant application value in fields such as image segmentation, social network analysis, and bioinformatics.

[0003] Current mainstream multi-view clustering methods mainly include: 1. Data clustering methods based on multi-feature fusion; 2. Multidimensional data clustering methods based on tensor data structures; and 3. Image segmentation clustering methods based on superpixel preprocessing. The core idea of ​​data clustering methods based on multi-feature fusion is to combine features from different data sources or feature extractors to form richer data representations. Its main advantages are its simplicity and intuitiveness, high degree of engineering feasibility, rapid deployment to production environments, and easy integration with various existing feature extraction modules. Multidimensional data clustering methods based on tensor data structures organize multidimensional data into tensor form, utilizing the unique multilinear structure of tensors to capture the correlations between data in different dimensions. Its technical advantages lie in preserving the original multidimensional structure of the data, avoiding information loss in traditional methods, and the natural dimensionality reduction effect of tensor decomposition, which can discover potential low-dimensional structures in the data. Image segmentation clustering methods based on superpixel preprocessing divide the original image into several superpixel regions. Pixels within each region have similar color, texture, or spatial location features, and subsequent clustering analysis is then performed based on these superpixels. Commonly used superpixel algorithms include Simple Linear Iterative Clustering (SLIC), Fast Shift Clustering, and graph cutting. These algorithms generate superpixels through different similarity measures and optimization strategies. The main advantage of this method is that it can significantly reduce the computational cost of data processing, transforming pixel-level processing into region-level processing, while preserving important boundary information of the image, thus providing a good foundation for subsequent advanced vision tasks.

[0004] However, traditional methods mainly analyze data at a single scale, which cannot simultaneously capture the global structure and local details of the data, resulting in low clustering accuracy. Summary of the Invention

[0005] Therefore, it is necessary to provide a view clustering analysis method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can improve the accuracy of data clustering in response to the above-mentioned technical problems.

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

[0007] Obtain the original view dataset;

[0008] Based on the view type of each view in the original view dataset, a target domain segmentation algorithm is determined. The target domain segmentation algorithm is then used to process each view according to the number of domains at each scale level to obtain a label indicator matrix for each view at each scale level. The label indicator matrix is ​​used to identify the domain in which each sample in each view is located.

[0009] Based on the label indication matrix of each view at each scale level, determine the domain features at each scale level;

[0010] Fuzzy clustering analysis is performed on each of the domain features to obtain the spatial membership matrix at each of the scale levels.

[0011] Based on the spatial membership matrix, tensor low-rank term, and clustering indicator learning term at each of the aforementioned scale levels, a clustering result indicator matrix for the original view dataset is determined; wherein, the tensor low-rank term is used to constrain the consistency of each of the aforementioned views at each of the aforementioned scale levels, and the clustering indicator learning term represents the error between the low-rank features of the view and the clustering structure.

[0012] In one embodiment, the step of processing each view according to the number of domains at each scale level using the target domain segmentation algorithm to obtain the label indication matrix of each view at each scale level includes:

[0013] When the view type is multi-view data, extract the features of each view of the multi-view data, and standardize each feature to obtain the standardized features of each view.

[0014] The number of cluster centers at each scale level is determined based on the number of domains at each scale level.

[0015] At each scale level, the clustering algorithm processes the standardized features of each view based on each cluster center to obtain a label indication matrix for each view at each scale level.

[0016] When the view type is image data, each view of the image data is acquired, and the view is segmented at each scale level using a superpixel segmentation algorithm to obtain superpixel regions at each scale level; wherein, the number of superpixel regions at each scale level is determined based on the number of domains at each scale level; the scale level is inversely proportional to the number of superpixel regions at the scale level; the expected size of the superpixels in each superpixel region is determined based on the total number of pixels in the view and the number of superpixel regions at the scale level;

[0017] The superpixel regions at each scale level are converted into the label indication matrices at each scale level.

[0018] In one embodiment, the method for determining the number of domains at each scale level includes:

[0019] Based on the total number of views in the acquired original view dataset and the expected number of neighbors of a domain at the minimum scale level, the number of domains at each scale level is determined, wherein the scale level is inversely proportional to the number of domains corresponding to the scale level.

[0020] In one embodiment, determining the domain features at each scale level based on the label indication matrix of each view at each scale level includes:

[0021] For each scale level, based on the label indication matrix of each view at that scale level, calculate the mean vector and covariance matrix of the domain at that scale level;

[0022] The mean vector and covariance matrix of the domain at each scale level are concatenated to determine the features of each domain at each scale level.

[0023] In one embodiment, performing fuzzy clustering analysis on each of the domain features to obtain the spatial membership matrix at each of the scale levels includes:

[0024] Perform fuzzy clustering analysis on each of the domain features to obtain the domain-level membership matrix at each of the scale levels;

[0025] Align the domain-level membership matrix corresponding to each scale level with the label indicator matrix at each scale level to obtain the spatial-level membership matrix at each scale level.

[0026] In one embodiment, the spatial membership matrix, tensor low-rank term, and clustering indicator learning term at each of the aforementioned scale levels determine the clustering result indicator matrix of the original view dataset, including:

[0027] Based on the spatial membership matrix at each of the aforementioned scale levels, a membership multidimensional tensor is obtained;

[0028] Based on the membership degree multidimensional tensor and the spatial-level membership degree matrix at each of the scale levels, structural constraints are determined.

[0029] Based on the structural constraints, normalization constraints, nonnegativity constraints, and clustering constraints, the constraint conditions are determined; wherein, the normalization constraints are used to transform the membership matrix into a probability distribution, and the membership includes a domain-level membership matrix and a spatial-level membership matrix; the nonnegativity constraints are used to determine the range of the domain-level membership matrix and the low-rank tensor; and the clustering constraints are used to enhance the clustering result indicator matrix.

[0030] The objective function is determined based on the tensor low-rank term, the data fitting term, the fuzzy clustering regularization term, and the clustering indicator learning term; wherein, the data fitting term represents the difference between the low-rank tensor and the membership multidimensional tensor; the fuzzy clustering regularization term is determined based on the cardinality vector of the domain at each scale level, the domain-level membership matrix at each scale level, and the distance from the domain features at each scale level to the cluster center;

[0031] Based on the constraints, the objective function is solved to obtain the clustering result indication matrix of the original view dataset.

[0032] Secondly, this application provides a view clustering analysis apparatus, the apparatus comprising:

[0033] The retrieval module is used to retrieve the original view dataset;

[0034] The first determining module is used to determine a target domain segmentation algorithm based on the view type of each view in the original view dataset, and to process each view according to the number of domains at each scale level using the target domain segmentation algorithm to obtain a label indication matrix for each view at each scale level; wherein, the label indication matrix is ​​used to identify the domain in which the sample in each view is located.

[0035] The second determining module is used to determine the domain features at each scale level based on the label indication matrix of each view at each scale level.

[0036] The fuzzy clustering module is used to perform fuzzy clustering analysis on each of the domain features to obtain the spatial membership matrix at each of the scale levels;

[0037] The result determination module is used to determine the clustering result indication matrix of the original view dataset based on the spatial membership matrix, tensor low-rank term, and clustering indicator learning term at each of the scale levels; wherein, the tensor low-rank term is used to constrain the consistency of each of the views at each of the scale levels, and the clustering indicator learning term represents the error between the low-rank features of the view and the clustering structure.

[0038] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0039] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0040] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method.

[0041] The aforementioned view clustering analysis method, apparatus, computer device, computer-readable storage medium, and computer program product first acquire the original view dataset; determine the target domain division algorithm based on the view type of each view in the original view dataset; and use a generalized preprocessing framework to uniformly process data of different view types, breaking through the scope of application and not being limited to the processing of multi-view or single-view data. Second, the target domain division algorithm processes each view according to the number of domains at each scale level, obtaining the label indicator matrix for each view at each scale level; based on the label indicator matrix for each view at each scale level, determine the features of each domain at each scale level; automatically extract hierarchical feature structures through domain decomposition, which can simultaneously capture the global structure and local details of the data, significantly improving clustering accuracy; finally, perform fuzzy clustering analysis on each domain feature to obtain the spatial membership matrix at each scale level; and determine the clustering result indicator matrix of the original view dataset based on the spatial membership matrix, tensor low-rank terms, and clustering indicator learning terms at each scale level. On the one hand, the tensor low-rank constraint and probabilistic membership mechanism ensure the consistency and stability of the clustering results, reduce the adverse effects of random initialization on the results, and achieve a comprehensive improvement in algorithm performance. On the other hand, deriving the clustering results directly from the tensor representation avoids the re-clustering steps required by traditional methods, effectively improving computational efficiency while ensuring clustering accuracy. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating a view clustering analysis method in one embodiment;

[0044] Figure 2 This is a schematic diagram illustrating the architecture and data processing of a view clustering analysis method in one embodiment;

[0045] Figure 3 This is a schematic diagram of the process for determining domain features in one embodiment;

[0046] Figure 4 This is a schematic diagram of the process for determining the label indication matrix in one embodiment;

[0047] Figure 5 This is a schematic diagram illustrating the process of determining the clustering result indicator matrix in one embodiment;

[0048] Figure 6 Here is a flowchart of a tensor representation learning and clustering algorithm in one embodiment;

[0049] Figure 7 This is a structural block diagram of a view clustering analysis device in one embodiment;

[0050] Figure 8 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] In one embodiment, such as Figure 1 As shown, a view clustering analysis method is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps S102 to S110. Wherein:

[0053] Step S102: Obtain the original view dataset.

[0054] For example, the terminal obtains the raw view dataset, which may include typical multi-view data or single-view data, i.e., image data.

[0055] Step S104: Based on the view type of each view in the original view dataset, determine the target domain division algorithm, and process each view according to the number of domains at each scale level using the target domain division algorithm to obtain the label indicator matrix of each view at each scale level.

[0056] The label indicator matrix, also known as the preprocessing matrix, is used to identify the domain where each sample in each view is located. ,in, An index represents the scale level. A domain is a set of samples or pixels with similar characteristics at a specific scale. Since directly processing each view in the original view dataset is computationally too complex and involves too much data, using domains can achieve dimensionality reduction, thus reducing the amount of data processing.

[0057] For example, based on the view type of each view in the original view dataset, the terminal can automatically select the most suitable target domain segmentation algorithm according to the type characteristics of the input data. The target domain segmentation algorithm preprocesses each view under its respective view type according to the number of domains at each scale level, obtaining the label indication matrix for each view at each scale level. For example, the scale levels include three levels: the first scale level represents the large-scale layer, denoted as k=1; the second scale level represents the medium-scale layer, denoted as k=2; and the third scale level represents the small-scale layer, denoted as k=3. The terminal obtains the view... Figure 1 The label indication matrix at the first scale level, view Figure 1 The label indication matrix at the second scale level, view Figure 1 Label indicator matrix at the third scale level.

[0058] Step S106: Based on the label indication matrix of each view at each scale level, determine the domain features at each scale level.

[0059] For example, the terminal models the domain features using a multivariate normal distribution. Based on the label indication matrix of each view at each scale level, that is, by performing feature processing on each domain using a normal distribution, the features of each domain at each scale level are obtained, denoted as... .

[0060] For example, such as Figure 2 As shown, different target domain segmentation algorithms are used for data of different view types, and the domains are divided at multiple scale levels to obtain the label indicator matrix for each view at each scale level. Based on the label indicator matrix of each view at each scale level and each view in the original view dataset Determine the features of each domain at each scale level. Where j represents the index of the view, which takes values ​​from 1 to N, where N is a positive integer greater than 1. i represents the index of the field.

[0061] Step S108: Perform fuzzy clustering analysis on the features of each domain to obtain the spatial membership matrix at each scale level.

[0062] For example, the terminal performs fuzzy clustering on the domain features at each scale level to obtain a domain-level membership matrix, and then scales and aligns the domain-level membership matrix to the space at the smallest scale level through domain scaling to obtain a spatial-level membership matrix.

[0063] Step S110: Based on the spatial membership matrix, tensor low-rank term, and clustering indicator learning term at each scale level, determine the clustering result indicator matrix of the original view dataset.

[0064] Among them, the tensor low-rank term is used to constrain the consistency of each view at each scale level, and the clustering indicator learning term represents the error between the low-rank features of the view and the clustering structure.

[0065] For example, the terminal concatenates the spatial membership matrices at each scale level into a tensor, learns the consistency of each view at each scale level by introducing a low-rank term of the tensor, and introduces a clustering indicator learning term to directly generate a clustering result indicator matrix of the original view dataset.

[0066] For example, the entire data processing process is as follows: Figure 3 As shown. The process consists of two stages. The first stage (multi-scale preprocessing) includes steps S102 to S106. Multi-scale preprocessing achieves hierarchical abstraction of the data, decomposing the original data at different scale levels and extracting multi-level feature representations from global to local. The second stage (tensor representation majority graph clustering algorithm) includes steps S108 to S110, concatenating the spatial membership matrices at each scale level into a three-dimensional tensor. The three-dimensional tensor is then processed by introducing a low-rank term, a clustering indicator learning term, and a multi-view fuzzy clustering algorithm. Through optimization and convergence judgment, a clustering result indicator matrix is ​​obtained.

[0067] In the aforementioned view clustering analysis method, firstly, the original view dataset is obtained; based on the view type of each view in the original view dataset, the target domain division algorithm is determined; a generalized preprocessing framework is used to uniformly process data of different view types, breaking through the scope of application and not being limited to the processing of multi-view or single-view data; secondly, the target domain division algorithm processes each view according to the number of domains at each scale level, obtaining the label indicator matrix of each view at each scale level; based on the label indicator matrix of each view at each scale level, the features of each domain at each scale level are determined; through domain decomposition, hierarchical feature structures are automatically extracted, which can simultaneously capture the global structure and local details of the data, significantly improving the clustering accuracy; finally, fuzzy clustering analysis is performed on the features of each domain to obtain the spatial membership matrix at each scale level; based on the spatial membership matrix, tensor low-rank term, and clustering indicator learning term at each scale level, the clustering result indicator matrix of the original view dataset is determined. On the one hand, the tensor low-rank constraint and probabilistic membership mechanism ensure the consistency and stability of the clustering results, reduce the adverse effects of random initialization on the results, and achieve a comprehensive improvement in algorithm performance. On the other hand, deriving the clustering results directly from the tensor representation avoids the re-clustering steps required by traditional methods, effectively improving computational efficiency while ensuring clustering accuracy.

[0068] In an exemplary embodiment, determining the target domain segmentation algorithm based on the view type of each view in the original view dataset includes: determining the target domain segmentation algorithm as a clustering algorithm when the view type is multi-view data; and determining the target domain segmentation algorithm as a superpixel segmentation algorithm when the view type is image data.

[0069] Traditional tensor clustering methods have significant limitations in practical applications: high computational complexity, especially for large-scale high-dimensional tensors, where the decomposition process requires substantial computational resources and time; poor algorithm stability, as the results of tensor decomposition are sensitive to initialization, and different random initializations can lead to completely different clustering results; lack of multi-scale processing capabilities, as existing methods typically can only capture data patterns at a single granularity, making it difficult to simultaneously process the global structure and local details of the data; and limited application scope, as most methods are specifically designed for native multidimensional tensor data and cannot be directly applied to data types that require preprocessing, such as images.

[0070] For example, when the view type is multi-view data, the terminal determines the target domain segmentation algorithm as a clustering algorithm, such as K-means. When the view type is image data, the target domain segmentation algorithm is determined to be a superpixel segmentation algorithm (Simple Linear Iterative Clustering, SLIC).

[0071] In this embodiment, by determining different target domain division algorithms for different view types, it is possible to achieve unified processing of typical multi-view data and image data, which greatly expands the scope of application.

[0072] In one exemplary embodiment, such as Figure 4 As shown, each view is processed according to the number of domains at each scale level using a target domain segmentation algorithm to obtain the label indication matrix for each view at each scale level, including the following steps S402 to S406. Wherein:

[0073] Step S402: When the view type is multi-view data, extract the features of each view of the multi-view data, and standardize each feature to obtain the standardized features of each view.

[0074] For example, for traditional multi-view data ,in This represents the feature matrix of the v-th view. N represents the feature dimension (rows), and N represents the view index (columns). Represents the real number space. The terminal uses an improved K-means clustering algorithm for domain segmentation. To eliminate the influence of dimensions between different views, the feature matrix of each view is standardized by Z-score, as shown in formula (1).

[0075] Formula (1)

[0076] In the formula, The vector representing the mean of the i-th view. The vector representing the standard deviation of the i-th view. The feature matrix representing the v-th view. The feature matrix represents the v-th view after standardization.

[0077] Step S404: Determine the number of cluster centers at each scale level based on the number of domains at each scale level.

[0078] For example, the terminal sets the number of cluster centers at each scale level: the number of domains at the k-th scale level. The number of cluster centers. If the number of domains at the first scale level is x, then the number of cluster centers at the first scale level is x. The same applies to other scale levels.

[0079] Step S406: At each scale level, the standardized features of each view are processed by a clustering algorithm based on each cluster center to obtain the label indication matrix of each view at each scale level.

[0080] For example, at each scale k, the terminal uses the K-means clustering algorithm on the standardized view features of each view i: the initial cluster centers are selected using the K-means++ initialization strategy, and the cluster center positions are iteratively optimized through the EM algorithm; the Euclidean distance from the features of each view to each cluster center is calculated; until all cluster centers equal to the number of domains are calculated, the domain to which each view belongs at each scale level is determined.

[0081] In this embodiment, a clustering algorithm is used to process the multi-view type to obtain the label indication matrix of each view at each scale level corresponding to the multi-view type.

[0082] In an exemplary embodiment, each view is processed according to the number of domains at each scale level using a target domain segmentation algorithm to obtain a label indication matrix for each view at each scale level. This includes: when the view type is image data, acquiring each view of the image data, segmenting the view at each scale level using a superpixel segmentation algorithm to obtain superpixel regions at each scale level; wherein, the number of superpixel regions at each scale level is determined based on the number of domains at each scale level; the scale level is inversely proportional to the number of superpixel regions at the scale level; the expected size of the superpixels in each superpixel region is determined based on the total number of pixels in the view and the number of superpixel regions at each scale level; and converting the superpixel regions at each scale level into label indication matrices at each scale level.

[0083] For example, for image data, this invention employs an improved SLIC superpixel segmentation algorithm for domain processing. The SLIC algorithm, by combining color similarity and spatial proximity, can generate compact and boundary-fitting superpixel regions. The superpixel regions in image data are also the domains in multi-view data.

[0084] For example, in the large-scale layer corresponding to the first scale level, fewer seed points are set to generate larger superpixel regions. In the medium-scale layer corresponding to the second scale level, a moderate seed point density balances the size and number of regions. In the small-scale layer corresponding to the small-scale layer, a dense seed point layout generates fine superpixel segmentation.

[0085] In each scale layer, each view is segmented by a similarity metric function that simultaneously considers color similarity and spatial proximity, as shown in Equations (2), (3) and (4).

[0086] Formula (2)

[0087] Formula (3)

[0088] Formula (4)

[0089] In the formula, and It is a pixel and pixels Location; and It is a pixel and pixels The color value; It is spatial distance. It's color distance. It is a balancing parameter that controls the relative importance of spatial and color constraints; Let S represent the similarity measure function, considering both color features and spatial location, and let S represent the expected size of the superpixels in the superpixel region. Here, S = N / K, where N represents the total number of pixels and K represents the number of superpixel regions.

[0090] For example, the terminal converts the superpixel regions at each scale level into label indication matrices at each scale level. Pixel coordinate (0,0) corresponds to the field with id 1, denoted as [(0,0),1]; for example, the SLIC segmentation result is [(0,0),1][(0,1),1][(0,2),1][(0,3),2][(0,4),2]... The label indication matrix is ​​a matrix composed of pixel index as columns and field index as rows. Pixel 1 (0,0) corresponds to the field with id 1; pixel 2 (0,1) corresponds to the field with id 1; pixel 3 (0,2) corresponds to the field with id 1; pixel 4 (0,3) corresponds to the field with id 2; pixel 5 (0,4) corresponds to the field with id 2.

[0091] In this embodiment, the label indication matrix at each scale level is obtained through multi-scale SLIC, which can achieve the label indication matrix of each view corresponding to the image type at each scale level, and the application scenarios are expanded through preprocessing.

[0092] In an exemplary embodiment, the method for determining the number of domains at each scale level includes: determining the number of domains at each scale level based on the total number of views in the acquired original view dataset and the expected number of neighbors of the domain at the minimum scale level.

[0093] The scale level is inversely proportional to the number of domains corresponding to that scale level. For example, at scale level 3, the large-scale layer corresponding to the first scale level has the fewest domains, primarily capturing the global clustering structure and main patterns of the data. The medium-scale layer corresponding to the second scale level has a moderate number of domains, extracting medium-granularity local structural information. In the small-scale layer corresponding to the small-scale layer, the number of domains is the largest, preserving fine local details and boundary information. This hierarchical design ensures complete information coverage from coarse to fine granular, providing a rich multi-scale feature foundation for subsequent tensor learning.

[0094] For example, the domain number calculation is based on hierarchical information theory, which balances the relationship between global structure capture and local detail preservation by controlling the information density at different scale levels. The calculation is shown in Equation (5). The total number of views in the acquired original view dataset, the expected number of neighbors of the domain at the current scale level and the minimum scale level are substituted into Equation (5) to obtain the domain number at each scale level.

[0095] Formula (5)

[0096] In the formula, The number of domains at the k-th scale level, ensuring that the number of domains decreases as the scale level increases; This represents the total number of views in the original dataset, reflecting the data scale. This represents the expected number of neighbors for each domain at the minimum scale, used to control the fineness of the local structure; This represents the current scale level index (r=1, 2, 3, ..., R), where R is the total number of scale levels; : Round down to the nearest integer to ensure the number of fields is a positive integer.

[0097] To adapt to the characteristics of different datasets, an adaptive adjustment mechanism is also included. When the calculated number of domains is too large (greater than N / 2) or too small (less than the cluster number m), the terminal will automatically adjust. The parameter or scale level R ensures the rationality and effectiveness of domain decomposition.

[0098] In this embodiment, the reasonable determination of the number of multi-scale domains is the foundation of the entire preprocessing process, and the optimal number of domains at each scale level can be automatically determined according to the data scale and feature complexity.

[0099] In an exemplary embodiment, determining the domain features at each scale level based on the label indication matrix of each view at each scale level includes: for each scale level, calculating the mean vector and covariance matrix of the domain at the scale level based on the label indication matrix of each view at the scale level; concatenating the mean vector and covariance matrix of the domain at each scale level to determine the domain features at each scale level.

[0100] For example, suppose there are samples in each view of the original view dataset. , No. The set of domains at each scale is A preprocessing (label indication) matrix can be generated using K-means (general data) or SLIC (image pixels). The label indicator matrix is ​​as shown in formula (6).

[0101] Formula (6)

[0102] In the formula, The representative label indicator matrix has a size of N*N. (k) ; The element in the label indicator matrix represents whether sample j belongs to domain i.

[0103] The terminal obtains the elements in the label indication matrix of each view at the scale level, the corresponding features of each view, and the domain cardinality vector, the domain mean vector, and the domain covariance matrix of each scale level, as shown in formula (7).

[0104] Formula (7)

[0105] In the formula, This represents the mean vector of the i-th domain across k scales; The covariance matrix represents the i-th domain across k scales; The feature vector representing the j-th view; T represents the cardinality of the i-th domain at the k-th scale, i.e., the number of views belonging to that domain; T is the transpose matrix.

[0106] The domain cardinality vectors at each scale level are determined as shown in formula (8). The terminal performs statistical analysis on the domains at each scale to obtain the domain cardinality vector at the k-th scale: p (k) =[ p 1 (k) , … , p N (k) (k) ] ⊤ .

[0107] Formula (8)

[0108] In the formula, The domain cardinality vector representing the k-th scale is derived from the cardinality of all domains. composition.

[0109] The terminal will generate the mean vector of the domain at each scale level. Covariance matrix of the sum domain By splicing, the features of each domain at each scale level are determined. .

[0110] In this embodiment, domain features, through mean and covariance matrices, can capture the central tendency and dispersion of data within the domain, providing richer statistical information and helping to improve the accuracy of subsequent clustering. They can also uniformly represent the fusion of information from different types of data at different scales, ensuring the accuracy of subsequent clustering.

[0111] In an exemplary embodiment, fuzzy clustering analysis is performed on each domain feature to obtain a spatial membership matrix at each scale level, including: performing fuzzy clustering analysis on each domain feature to obtain a domain-level membership matrix at each scale level; aligning the domain-level membership matrix at the corresponding scale level with the label indicator matrix at each scale level to obtain a spatial membership matrix at each scale level.

[0112] For example, for each domain feature Perform fuzzy clustering analysis to obtain the domain-level membership matrix at each scale level. This means The sum of each row is 1, meaning the sum of all cluster memberships for each data point is 1, ensuring the normalization of memberships. Where 1... M This represents a column vector of length M consisting entirely of 1s, used to represent the summation along the clustering dimension. This represents a column vector of length N(k) filled with all ones, used to represent the summation across the data points. Since the domain-level membership matrices generated by fuzzy clustering at various scales have different dimensions, it is necessary to align the membership matrices from different scales to the original scale. The terminal uses the label indicator matrix at each scale level. For the corresponding scale-level domain membership matrix Alignment is performed to obtain the spatial membership matrix at each scale level. As shown in formula (9).

[0113] Formula (9)

[0114] In the formula, Represents the spatial membership matrix at each scale level; A label indicator matrix representing each scale level; Domain-level membership matrices at each scale level.

[0115] In this embodiment, the goal of the entire process is to unify the domain-level membership matrices obtained at different scales to the original scale so as to facilitate subsequent analysis or fusion.

[0116] In one exemplary embodiment, such as Figure 5 As shown, the spatial membership matrix, tensor low-rank term, and clustering indicator learning term at each scale level are used to determine the clustering result indicator matrix of the original view dataset, including steps S502 to S510. Wherein:

[0117] Step S502: Based on the spatial membership matrix at each scale level, obtain the membership multidimensional tensor.

[0118] For example, the terminal, based on the spatial membership matrix at each scale level, i.e., the aligned membership matrix, through... The method constructs a three-dimensional tensor. ,in, It is the three-dimensional tensor corresponding to the spatial membership matrix.

[0119] Step S504: Determine structural constraints based on the membership degree multidimensional tensor and the spatial membership degree matrix at each scale level.

[0120] For example, the terminal establishes the relationship between the membership multidimensional tensor and the spatial membership matrix at each scale level. And the relationship between the spatial membership matrix and the domain membership matrix at each scale level. As a structural constraint. Among them, Ensure that the input tensor correctly reflects information from all views and scales, and connect the matrix representation with the tensor representation; Ensure the mapping relationship from the "domain space" to the "sample space" and ensure mathematical consistency between the two processing stages.

[0121] Step S506: Determine the constraint conditions based on structural constraints, normalization constraints, nonnegativity constraints, and clustering constraints.

[0122] Among them, the normalization constraint is used to transform the membership matrix into a probability distribution, and the membership includes the domain-level membership matrix and the spatial-level membership matrix; the nonnegativity constraint is used to determine the range of the domain-level membership matrix and the low-rank tensor; and the clustering constraint is used to enhance the clustering result indicator matrix.

[0123] Normalization constraints include and , This is used to ensure that the sum of the "probabilities" of each view belonging to each cluster is 1. This indicates that the sum of the "probabilities" of each domain belonging to each cluster is 1. This ensures mathematical plausibility and interpretability. Specifically, for... Each row represents the membership degree of a sample to each cluster in the low-rank space, and the sum of the rows being one indicates that these membership degrees constitute a probability distribution. Similarly, Each row represents the membership degree of a domain to each cluster, and the sum of the rows also constitutes a probability distribution.

[0124] Nonnegativity constraints include and .in, This means that all elements of a low-rank feature representation must be non-negative. All elements representing domain-level membership must be non-negative.

[0125] Clustering constraints include This is used to indicate that the clustering result indicator matrix must be a 0-1 matrix; This indicates that each row of the clustering indicator matrix contains exactly one 1.

[0126] Step S508: Determine the objective function based on the tensor low-rank term, data fitting term, fuzzy clustering regularization term, and clustering indicator learning term.

[0127] Among them, the data fitting term represents the difference between the low-rank tensor and the membership multidimensional tensor; the fuzzy clustering regularization term is determined based on the cardinality vector of the domain at each scale level, the domain-level membership matrix at each scale level, and the distance from the domain features at each scale level to the cluster center.

[0128] The low-rank term of a tensor is denoted as The matrix nuclear norm is a generalization of tensors, used to measure the "complexity" or "rank" of a tensor. It is used to discover shared factors in multi-view, multi-scale data, filtering out noise and redundant information specific to each view, and ensuring the quality of the final learned feature representation. It is consistent.

[0129] The data fitting term is denoted as Measure the learned low-rank tensor With the original input tensor Differences, ensuring consistent features And will not deviate from the original data Too far.

[0130] Fuzzy clustering regularization term, denoted as .in, It is the element in the j-th row and i-th column of the membership matrix at the k-th scale. It is the cardinality of the i-th field at the k-th scale. It is the i-th cluster center at the k-th scale. It represents the distance from the domain feature to the cluster center.

[0131] Clustering instruction learning item, denoted as .in, It is a low-rank tensor The k-th front slice It is the center matrix of the low-rank eigenvalue at the k-th scale. It is the final clustering result indicator matrix.

[0132] For example, the terminal adds these four parts together and uses the minimization of the sum as the objective function.

[0133] Step S510: Based on the constraints, solve the objective function to obtain the clustering result indicator matrix of the original view dataset.

[0134] The model consisting of the objective function and constraints is shown in formula (10).

[0135] Formula (10)

[0136] In the formula, The membership matrix is ​​obtained through The method constructs a three-dimensional tensor. For low-rank tensors, It is the element in the j-th row and i-th column of the membership matrix at the k-th scale. It is the cardinality of the i-th field at the k-th scale. It is the i-th cluster center at the k-th scale. It is a low-rank tensor The k-th front slice It is the center matrix of the low-rank eigenvalue at the k-th scale. It is the final clustering result indicator matrix. Represents the Frobenius norm. Represents the tensor nuclear norm (TNN).

[0137] The terminal uses an algorithm based on constraints to optimize and solve the objective function and determine convergence, thereby obtaining a clustering result indicator matrix.

[0138] In this embodiment, feature fusion and cluster analysis are achieved through tensor low-rank learning.

[0139] In an exemplary embodiment, the objective function is solved based on constraints to obtain a clustering result indication matrix of the original view dataset. This includes: solving the objective function based on constraints using the augmented Lagrange method and the alternating direction method to obtain a clustering result indication matrix of the original view dataset.

[0140] For example, the terminal employs the augmented Lagrangian method and the alternating direction method for efficient solution. An auxiliary tensor is introduced. , multiplier With penalty factor The augmented Lagrange is shown in equation (11).

[0141] Formula (11)

[0142] like Figure 6 As shown, by solving the subproblems as shown in formula (12) to... Update:

[0143] Formula (12)

[0144] By solving the subproblems shown in formula (13) Update:

[0145] Formula (13)

[0146] By solving the subproblems shown in formula (14) Update:

[0147] Formula (14)

[0148] By solving the subproblem shown in formula (15) Update:

[0149] Formula (15)

[0150] By solving the subproblem shown in formula (16) Update:

[0151] Formula (16)

[0152] Lagrange multipliers and penalty coefficient Update using the following formula (17):

[0153] Formula (17)

[0154] in, A coefficient of variation greater than 1 yes The maximum value.

[0155] Solve the subproblem pair shown in formula (18) Update:

[0156] Formula (18)

[0157] Termination condition is If the iteration reaches its limit, the final terminal outputs a clustering result indicator matrix.

[0158] In this embodiment, by solving the constraints and objective function using the augmented Lagrange method and the alternating direction method, the final clustering result indicator matrix can be obtained.

[0159] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0160] Based on the same inventive concept, this application also provides a view clustering analysis apparatus for implementing the view clustering analysis method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more view clustering analysis apparatus embodiments provided below can be found in the limitations of the view clustering analysis method described above, and will not be repeated here.

[0161] In one exemplary embodiment, such as Figure 7 As shown, a view clustering analysis device is provided, including: an acquisition module 701, a first determination module 702, a second determination module 703, a fuzzy clustering module 704, and a result determination module 705, wherein:

[0162] Module 701 is used to obtain the original view dataset.

[0163] The first determining module 702 is used to determine the target domain division algorithm based on the view type of each view in the original view dataset, and to process each view according to the number of domains at each scale level by the target domain division algorithm to obtain the label indicator matrix of each view at each scale level; wherein, the label indicator matrix is ​​used to identify the domain in which each sample in each view is located.

[0164] The second determining module 703 is used to determine the domain features at each scale level based on the label indication matrix of each view at each scale level.

[0165] The fuzzy clustering module 704 is used to perform fuzzy clustering analysis on the features of each domain to obtain the spatial membership matrix at each scale level.

[0166] The result determination module 705 is used to determine the clustering result indicator matrix of the original view dataset based on the spatial membership matrix, tensor low-rank term and clustering indicator learning term at each scale level; wherein, the tensor low-rank term is used to constrain the consistency of each view at each scale level, and the clustering indicator learning term represents the error between the low-rank features of the view and the clustering structure.

[0167] In an exemplary embodiment, the first determining module 702 is further configured to extract features of each view of the multi-view data when the view type is multi-view data, and standardize each feature to obtain each standardized view feature; determine the number of cluster centers at each scale level according to the number of domains at each scale level; and at each scale level, process each standardized view feature based on each cluster center using a clustering algorithm to obtain a label indication matrix for each view at each scale level.

[0168] In an exemplary embodiment, the first determining module 702 is further configured to, when the view type is image data, acquire each view of the image data, segment the view at each scale level using a superpixel segmentation algorithm, and obtain superpixel regions at each scale level; wherein, the number of superpixel regions at each scale level is determined based on the number of domains at each scale level; the scale level is inversely proportional to the number of superpixel regions at the scale level; the expected size of the superpixels in each superpixel region is determined based on the total number of pixels in the view and the number of superpixel regions at the scale level; and convert the superpixel regions at each scale level into label indication matrices at each scale level.

[0169] In one exemplary embodiment, a view clustering analysis apparatus further includes a domain number determination module, used to determine the number of domains at each scale level based on the total number of views in the acquired original view dataset and the expected number of neighbors of a domain at the minimum scale level, wherein the scale level is inversely proportional to the number of domains corresponding to the scale level.

[0170] In an exemplary embodiment, the second determining module 703 is further configured to, for each scale level, calculate the mean vector and covariance matrix of the domain at the scale level based on the label indication matrix of each view at the scale level; and concatenate the mean vector and covariance matrix of the domain at each scale level to determine the features of each domain at each scale level.

[0171] In an exemplary embodiment, the fuzzy clustering module 704 is further configured to perform fuzzy clustering analysis on each domain feature to obtain a domain-level membership matrix at each scale level; and align the domain-level membership matrix at the corresponding scale level with the label indicator matrix at each scale level to obtain a spatial-level membership matrix at each scale level.

[0172] In an exemplary embodiment, the result determination module 705 is further configured to: obtain a membership degree multidimensional tensor at each scale level based on the spatial membership degree matrix at each scale level; determine structural constraints based on the membership degree multidimensional tensor at each scale level and the spatial membership degree matrix at each scale level; and determine constraint conditions based on structural constraints, normalization constraints, nonnegativity constraints, and clustering constraints; wherein, the normalization constraint is used to transform the membership degree matrix into a probability distribution, and the membership degree includes a domain-level membership degree matrix and a spatial-level membership degree matrix; the nonnegativity constraint is used to determine domain-level membership... The range of the degree matrix and low-rank tensor is used to enhance the clustering result indicator matrix. The objective function is determined based on the tensor low-rank term, data fitting term, fuzzy clustering regularization term, and clustering indicator learning term. Among them, the data fitting term represents the difference between the low-rank tensor and the membership multidimensional tensor; the fuzzy clustering regularization term is determined based on the cardinality vector of the domain at each scale level, the domain-level membership matrix at each scale level, and the distance from the domain features at each scale level to the cluster center. Based on the constraints, the objective function is solved to obtain the clustering result indicator matrix of the original view dataset.

[0173] Each module in the aforementioned view clustering analysis device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0174] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 8 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores clustering data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a view clustering analysis method.

[0175] Those skilled in the art will understand that Figure 8The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0176] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0177] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0178] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0179] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0180] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0181] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A view clustering analysis method, characterized in that, The method includes: Obtain the original view dataset; Based on the view type of each view in the original view dataset, a target domain segmentation algorithm is determined. The target domain segmentation algorithm is then used to process each view according to the number of domains at each scale level to obtain a label indicator matrix for each view at each scale level. The label indicator matrix is ​​used to identify the domain in which each sample in each view is located. Based on the label indication matrix of each view at each scale level, determine the domain features at each scale level; Fuzzy clustering analysis is performed on each of the domain features to obtain the spatial membership matrix at each of the scale levels. Based on the spatial membership matrix, tensor low-rank term, and clustering indicator learning term at each of the aforementioned scale levels, a clustering result indicator matrix for the original view dataset is determined; wherein, the tensor low-rank term is used to constrain the consistency of each of the aforementioned views at each of the aforementioned scale levels, and the clustering indicator learning term represents the error between the low-rank features of the view and the clustering structure.

2. The method according to claim 1, characterized in that, The step of processing each view according to the number of domains at each scale level using the target domain segmentation algorithm to obtain the label indication matrix of each view at each scale level includes: When the view type is multi-view data, extract the features of each view of the multi-view data, and standardize each feature to obtain the standardized features of each view. The number of cluster centers at each scale level is determined based on the number of domains at each scale level. At each scale level, the clustering algorithm processes the standardized features of each view based on each cluster center to obtain a label indication matrix for each view at each scale level. When the view type is image data, each view of the image data is acquired, and the view is segmented at each scale level using a superpixel segmentation algorithm to obtain superpixel regions at each scale level; wherein, the number of superpixel regions at each scale level is determined based on the number of domains at each scale level; the scale level is inversely proportional to the number of superpixel regions at the scale level; the expected size of the superpixels in each superpixel region is determined based on the total number of pixels in the view and the number of superpixel regions at the scale level; The superpixel regions at each scale level are converted into the label indication matrices at each scale level.

3. The method according to claim 1, characterized in that, The methods for determining the number of domains at each scale level include: Based on the total number of views in the acquired original view dataset and the expected number of neighbors of a domain at the minimum scale level, the number of domains at each scale level is determined, wherein the scale level is inversely proportional to the number of domains corresponding to the scale level.

4. The method according to claim 1, characterized in that, The determination of domain features at each scale level based on the label indication matrix of each view at each scale level includes: For each scale level, based on the label indication matrix of each view at that scale level, calculate the mean vector and covariance matrix of the domain at that scale level; The mean vector and covariance matrix of the domain at each scale level are concatenated to determine the features of each domain at each scale level.

5. The method according to claim 1, characterized in that, The step of performing fuzzy clustering analysis on each of the domain features to obtain the spatial membership matrix at each of the scale levels includes: Perform fuzzy clustering analysis on each of the domain features to obtain the domain-level membership matrix at each of the scale levels; Align the domain-level membership matrix corresponding to each scale level with the label indicator matrix at each scale level to obtain the spatial-level membership matrix at each scale level.

6. The method according to claim 1, characterized in that, The spatial membership matrix, tensor low-rank term, and clustering indicator learning term at each of the aforementioned scale levels determine the clustering result indicator matrix of the original view dataset, including: Based on the spatial membership matrix at each of the aforementioned scale levels, a membership multidimensional tensor is obtained; Based on the membership degree multidimensional tensor and the spatial-level membership degree matrix at each of the scale levels, structural constraints are determined. Based on the structural constraints, normalization constraints, nonnegativity constraints, and clustering constraints, the constraint conditions are determined; wherein, the normalization constraints are used to transform the membership matrix into a probability distribution, and the membership includes a domain-level membership matrix and a spatial-level membership matrix; the nonnegativity constraints are used to determine the range of the domain-level membership matrix and the low-rank tensor; and the clustering constraints are used to enhance the clustering result indicator matrix. The objective function is determined based on the tensor low-rank term, the data fitting term, the fuzzy clustering regularization term, and the clustering indicator learning term; wherein, the data fitting term represents the difference between the low-rank tensor and the membership multidimensional tensor; the fuzzy clustering regularization term is determined based on the cardinality vector of the domain at each scale level, the domain-level membership matrix at each scale level, and the distance from the domain features at each scale level to the cluster center; Based on the constraints, the objective function is solved to obtain the clustering result indication matrix of the original view dataset.

7. A view clustering analysis device, characterized in that, The device includes: The retrieval module is used to retrieve the original view dataset; The first determining module is used to determine a target domain segmentation algorithm based on the view type of each view in the original view dataset, and to process each view according to the number of domains at each scale level using the target domain segmentation algorithm to obtain a label indication matrix for each view at each scale level; wherein, the label indication matrix is ​​used to record the domain in which each view is located. The second determining module is used to determine the domain features at each scale level based on the label indication matrix of each view at each scale level. The fuzzy clustering module is used to perform fuzzy clustering analysis on each of the domain features to obtain the spatial membership matrix at each of the scale levels; The result determination module is used to determine the clustering result indication matrix of the original view dataset based on the spatial membership matrix, tensor low-rank term, and clustering indicator learning term at each of the scale levels; wherein, the tensor low-rank term is used to constrain the consistency of each of the views at each of the scale levels, and the clustering indicator learning term represents the error between the low-rank features of the view and the clustering structure.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images