Rapid classification method and system for large-scale multi-view data and medium

Abstract

The invention discloses a rapid classification method and system for large-scale multi-view data and a medium, and the method comprises the steps: obtaining a large-scale multi-view data set, and forming a multi-view data matrix; based on the multi-view data matrix, obtaining a bipartite graph of each view data, and forming a first clustering model; optimizing the first clustering model by introducing a graph learning item and a double-anchor learning mechanism to obtain a second clustering model; the second clustering model is a multi-view clustering model formed by joint learning of a consistent bipartite graph and a complementary bipartite graph; performing optimization solution on the second clustering model by adopting an iterative optimization algorithm to obtain a solved second clustering model; and inputting the large-scale multi-view data set into the solved second clustering model for clustering to obtain a clustering result. The method aims at improving the performance of an existing bipartite graph-based multi-view clustering method in processing a large-scale data set, and further improving the clustering efficiency of the large-scale data set.

Description

Technical Field

[0001] The present invention relates to the technical field of data classification, and in particular to a fast classification method, system and medium for large-scale multi-view data. Background Art

[0002] With the development of information technology, efficient processing and in-depth analysis of large-scale data have become the key to modern scientific research, business decision-making and public services. However, in practical applications, traditional classification techniques face problems such as classification bias, low efficiency and neglect of the intrinsic connection of data. To meet these challenges, cluster analysis has received widespread attention as an unsupervised learning method. It forms groups by identifying the similarities between data, which can improve the accuracy and efficiency of subsequent classification. In particular, when processing large-scale multi-view datasets, multi-view clustering methods based on bipartite graphs have shown unique advantages. This method overcomes the limitations of traditional algorithms in processing large-scale multi-view data by selecting a small number of representative data points (i.e., anchor points) and measuring their similarities with other data points.

[0003] However, existing multi-view clustering methods based on bipartite graphs tend to rely on the global structural characteristics of the data to construct bipartite graphs, while ignoring the local structural characteristics of the data; secondly, the method of selecting anchor points through k-means or random sampling introduces a certain degree of randomness, which may cause the selected anchor points to not represent the overall distribution of the data well. In addition, directly applying consistent bipartite graphs or fused bipartite graphs for clustering may limit the information exchange between different views, thereby affecting the clustering effect.

[0004] In view of this, this application is hereby filed. Summary of the invention

[0005] The technical problem to be solved by the present invention is that the existing multi-view clustering method based on bipartite graph tends to rely on the global structural characteristics of the data to construct the bipartite graph, while ignoring the local structural characteristics of the data; secondly, the method of selecting anchor points by k-means or random sampling introduces a certain randomness, which may cause the selected anchor points to not represent the overall distribution of the data well. In addition, directly applying consistent bipartite graphs or fused bipartite graphs for clustering may limit the information exchange between different views, thereby affecting the clustering effect. The purpose of the present invention is to provide a fast classification method, system and medium for large-scale multi-view data. On the basis of the existing anchor-based multi-view data clustering method, a new joint consistent and complementary bipartite graph multi-view clustering model is proposed, which not only captures the global and local structural characteristics of the data at the same time, but also integrates consistent bipartite graphs and complementary bipartite graphs for joint learning, thereby realizing information exchange between different views. The present invention aims to improve the performance of the existing multi-view clustering method based on bipartite graphs when processing large-scale data sets, and further improve the clustering efficiency and accuracy of large-scale data sets.

[0006] The present invention is achieved through the following technical solutions:

[0007] In a first aspect, the present invention provides a fast classification method for large-scale multi-view data, the method comprising:

[0008] Acquire a large-scale multi-view dataset to form a multi-view data matrix;

[0009] Based on the multi-view data matrix, obtain a bipartite graph of each view data and form a first clustering model;

[0010] The first clustering model is optimized by introducing a graph learning item and a dual anchor learning mechanism to obtain a second clustering model; the second clustering model is a multi-view clustering model formed by jointly learning a consistent bipartite graph and a complementary bipartite graph;

[0011] An iterative optimization algorithm is used to optimize and solve the second clustering model to obtain a solved second clustering model;

[0012] The large-scale multi-view data set is input into the solved second clustering model for clustering to obtain a clustering result.

[0013] Furthermore, the large-scale multi-view dataset refers to a dataset containing a large number of samples, which are represented by a variety of different feature sets or description methods; each feature set or description method is regarded as an independent perspective or view in the dataset;

[0014] Large-scale multi-view datasets include indoor scene image datasets, animal image datasets, or face image datasets.

[0015] Furthermore, the first clustering model is optimized by introducing graph learning items and dual anchor learning mechanism to obtain the second clustering model, including:

[0016] Using subspace learning and graph learning to capture global structural features and local structural features of the first clustering model data to form a first intermediate clustering model;

[0017] By introducing a dual-anchor learning mechanism, different types of anchors are learned in high-dimensional space and low-dimensional space respectively, so as to obtain more representative anchors and more comprehensively capture the relationship and features between data points, thus forming a second intermediate clustering model.

[0018] Based on the second intermediate clustering model, the consistent bipartite graph and the complementary bipartite graph are jointly learned to form a second clustering model, thereby realizing information exchange between different views.

[0019] Furthermore, the expression of the second clustering model is:

[0020]

[0021] stα T 1=1,(W v ) T W v =I c , A T A=I m , Z ≥ 0, Z T 1=1,G v ≥0, (G v ) T 1=1.

[0022] Where min represents the objective function. The first term of the objective function measures the reconstruction error, that is, the degree to which the original data in the low-dimensional space is approximated by the relationship between the consensus anchor matrix A and the consistent bipartite graph Z; the second term is a regularization term used to obtain a better bipartite graph; the third term is measured by the anchor matrix O v The Euclidean distance from the original data point is to obtain a bipartite graph with local structural information;

[0023] V represents the total number of views; α ν is the weight of the v-th view, which is used to measure the importance of the v-th view; W ν is the dimension reduction matrix of the vth view, which is used to map the original features to a new feature space; () T represents the transpose of the matrix; X νis the data matrix of the vth view, which contains the feature information of n samples; A is the consensus anchor matrix, which represents the representative samples shared by all views, i.e., anchor points; Z is the consistent bipartite graph, which describes the consistency information between different views; is the square of the F norm; λ and β are regularization parameters used to control model complexity and avoid overfitting; G ν represents the bipartite graph of the vth view obtained in the high-dimensional space, that is, the complementary bipartite graph; m is the number of anchor points, n is the number of samples; O v is the anchor matrix obtained by the vth view in the high-dimensional space, which reflects the distribution of anchor points under each view; X v is the multi-view data matrix; For X ν The i-th row vector in ; O v The j-th column vector in ; is the square of the 2-norm; G v The element in row i and column j of ;

[0024] st is the constraint condition; I is the identity matrix, I m represents the identity matrix of dimension m×m; I c represents the identity matrix of dimension c×c, where c is the dimension of the reduced matrix; α T 1=1 means the sum of the weights of each view is 1; Z T 1=1,(G v ) T 1=1 constrains Z and G respectively v The sum of the elements in each column is 1.

[0025] Furthermore, an iterative optimization algorithm is used to optimize and solve the second clustering model, including:

[0026] Using iterative optimization algorithm, fix A, Z, G in the second clustering model v , α v , and delete the v The irrelevant terms of the reduced dimension matrix W v ;

[0027] Fix W in the second clustering model v , Z, G v , α v , and delete the irrelevant items related to A to solve the consensus anchor matrix A;

[0028] Fix W in the second clustering model v , Z, A, α v , and delete the G v Irrelevant terms, solve the complementary bipartite graph G v;

[0029] Fix W in the second clustering model v , G v , A, α v , and delete the irrelevant items of Z to solve the consistent bipartite graph Z;

[0030] Fix W in the second clustering model v , G v , A, Z, and delete the irrelevant items with respect to α, and solve the matrix α;

[0031] Based on the obtained dimension reduction matrix W v , consensus anchor matrix A, complementary bipartite graph G v , consistent bipartite graph Z and matrix α (i.e., the iterative expression of the variables to be optimized) are used to perform optimization solution.

[0032] Furthermore, the consistent bipartite graph preserves the consistency information between different views.

[0033] In a second aspect, the present invention provides a fast classification system for large-scale multi-view data, the system comprising:

[0034] An acquisition unit, used for acquiring a large-scale multi-view data set to form a multi-view data matrix;

[0035] A first model building unit, configured to obtain a bipartite graph of each view data based on the multi-view data matrix and form a first clustering model;

[0036] The second model building unit is used to optimize the first clustering model by introducing a graph learning item and a dual anchor learning mechanism to obtain a second clustering model; the second clustering model is a multi-view clustering model formed by jointly learning a consistent bipartite graph and a complementary bipartite graph;

[0037] A model solving unit, used for optimizing and solving the second clustering model by using an iterative optimization algorithm to obtain a solved second clustering model;

[0038] The clustering unit is used to input the large-scale multi-view data set into the solved second clustering model for clustering to obtain a clustering result.

[0039] Furthermore, the second model building unit includes:

[0040] A first optimization subunit is used to enhance the learning ability of the first clustering model for local features by introducing a graph learning item to obtain a first intermediate clustering model;

[0041] The second optimization subunit is used to capture the relationship and features between data points in the first intermediate clustering model by introducing a dual anchor learning mechanism, and learn the set anchor matrix and the corresponding complementary bipartite graph of each view in the high-dimensional space; introduce a dimensionality reduction matrix to map the data in the high-dimensional space to the low-dimensional space, and learn the consensus anchor matrix in the low-dimensional space; construct a consistent bipartite graph across views based on the consensus anchor matrix to obtain the second intermediate clustering model;

[0042] The joint learning subunit is used to perform joint learning on the consistent bipartite graph and the complementary bipartite graph based on the second intermediate clustering model to form a second clustering model.

[0043] In a third aspect, the present invention provides a fast classification device for large-scale multi-view data, comprising a memory, a processor, and computer executable instructions stored in the memory and executable on the processor. When the processor executes the computer executable instructions, it implements the above-mentioned fast classification method for large-scale multi-view data.

[0044] In a fourth aspect, the present invention further provides a computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed by a processor, the above-mentioned method for rapid classification of large-scale multi-view data is implemented.

[0045] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0046] The present invention provides a fast classification method, system and medium for large-scale multi-view data. The new joint consistent bipartite graph and complementary bipartite graph multi-view clustering model designed by the present invention first uses subspace learning and graph learning to capture the global and local structural features of the data; then, by introducing a dual anchor learning mechanism, different types of anchors are learned in high-dimensional space and low-dimensional space respectively, so as to obtain more representative anchors; finally, by constructing a consistent bipartite graph and a complementary bipartite graph, and performing joint learning, information exchange between different views is realized. In the above process, the present invention not only directly uses graph learning to capture local structural features in the data, but more importantly, two technologies are designed for high-dimensional space (i.e., original space) and low-dimensional space to generate different anchors. This is done to ensure that the generated anchors can cover a wider range of data information, improve their representativeness, and then enhance the capture effect of subspace learning and graph learning on the global and local information of the data. At the same time, the present invention constructs a consistent bipartite graph and a series of complementary bipartite graphs, and performs joint learning on this basis. This process allows information from different perspectives to complement and optimize each other, thereby improving the performance of the overall model.

[0047] The present invention improves the performance of existing bipartite graph-based multi-view clustering methods when processing large-scale data sets, and further improves the clustering efficiency and accuracy of large-scale data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings:

[0049] Figure 1 This is a flow chart of a fast classification method for large-scale multi-view data of the present invention;

[0050] Figure 2 This is a structural block diagram of a fast classification system for large-scale multi-view data according to the present invention. DETAILED DESCRIPTION

[0051] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0052] Existing multi-view clustering methods based on bipartite graphs tend to rely on the global structural characteristics of the data to construct bipartite graphs, while ignoring the local structural characteristics of the data; in addition, the method of selecting anchor points through k-means or random sampling introduces a certain degree of randomness, which may cause the selected anchor points to not represent the overall distribution of the data well. Even more, directly applying consistent bipartite graphs or fused bipartite graphs for clustering may limit the information exchange between different views, thereby affecting the clustering effect. Therefore, the present invention designs a fast classification method, system and medium for large-scale multi-view data. Based on the existing anchor-based multi-view data clustering method, a new joint consistent bipartite graph and complementary bipartite graph multi-view clustering (JC 2BG, Joint Consistent and Complementary Bipartite Graph Multi-ViewClustering) model, not only captures the global and local structural features of the data at the same time, but also obtains more representative anchor points. In addition, it combines consistent bipartite graphs with complementary bipartite graphs for joint learning, which effectively promotes information exchange between different perspectives. Specifically, the new joint consistent bipartite graph and complementary bipartite graph multi-view clustering model designed by the present invention first uses subspace learning and graph learning to capture the global and local structural features of the data; then, by introducing a dual anchor learning mechanism, different types of anchor points are learned in high-dimensional space and low-dimensional space respectively, so as to obtain more representative anchor points and capture the relationship and features between data points more comprehensively. Finally, by constructing consistent bipartite graphs and complementary bipartite graphs and performing joint learning, information exchange between different views is achieved. The present invention aims to improve the performance of the existing multi-view clustering method based on bipartite graphs when processing large-scale data sets, and further improve the clustering efficiency and accuracy of large-scale data sets.

[0053] The new joint consistent and complementary bipartite graph multi-view clustering (JC 2 BG, Joint Consistent and Complementary Bipartite Graph Multi-View Clustering) model (that is, the final JC 2 The design process of the BG model is as follows:

[0054] Existing multi-view data clustering methods based on bipartite graphs can efficiently handle clustering tasks of large-scale data sets by using the similarity between anchor points and data points to construct a smaller bipartite graph (also known as a similarity matrix). Mathematically, the first clustering model is expressed as follows:

[0055]

[0056] Z v ≥0,Z v 1=1.

[0057] In the formula, X is the data matrix, which is the matrix obtained by digitizing a picture; is the feature submatrix containing n samples corresponding to the vth view, d is the feature dimension of the view, V represents the total number of views, A v ∈R d×m is the anchor matrix corresponding to the vth view, Z v ∈R m×n Represents the bipartite graph of each view. I is the identity matrix, ZT is the transpose of Z; I m represents the unit matrix with dimension m×m, where m is the number of anchor points and n is the number of samples; (Z v ) T 1=1 constrains the sum of each column element of Zv to be 1. v ≥0 means that every element in the bipartite graph (similarity matrix) is non-negative; Represents the square derivative of the F norm: Example: Tr represents the trace of a matrix. For example, the sum of the elements on the main diagonal of a matrix B (the diagonal from the upper left to the lower right) is called the trace (or trace number) of the matrix B; A is the anchor matrix, which specifically refers to selecting some representative samples (called "anchor points") from the original data matrix to construct a smaller matrix.

[0058] Formula (1) is used to find Z v The minimum, self-expression To measure the similarity between the data points and the anchor points in each view, to construct the similarity matrix (also known as the bipartite graph) of the corresponding views.

[0059] However, this model mainly uses the global structural characteristics of the data to build a bipartite graph, while ignoring the importance of local structural features in improving clustering effects. To this end, the present invention introduces graph learning items to enhance the model's learning ability for local features to improve the overall clustering performance. Then formula (1) becomes formula (2), that is, the first intermediate clustering model is:

[0060]

[0061] In the formula, Represents the square of the 2-norm, for example: For X v The i-th row vector in ; is the j-th column vector in Av;

[0062] In the model of formula (2), we retain Add graph learning items Formula (2) is based on the constraint Z v ≥0,(Z v ) T 1=1. So that Z v The smallest.

[0063] Usually, the above model uses k-means algorithm or random sampling to obtain anchor points. Since these two methods have a certain degree of randomness, the obtained anchor points cannot completely simulate the original data. In order to solve this problem, the present invention proposes a dual anchor point learning mechanism, which learns two different anchor points in high-dimensional space and low-dimensional space respectively. First, in high-dimensional space, the present invention transforms X v Concatenate along the feature dimension to get Then calculate the variance of each feature of each data point i , and perform standardization on it. Finally, the present invention selects the data point with the largest variance as the initial anchor point, and uses the formula (Υ i-1 Represents the variance of the anchor point selected in the previous round, i ≥ 2) and continuously updates to obtain the anchor point matrix O v ; Secondly, by introducing a dimensionality reduction matrix, the present invention maps the data in the high-dimensional space to the low-dimensional space to remove redundant information and highlight the potential structure of the data. Then, in this low-dimensional space, by combining the learning of anchor points and the construction of bipartite graphs into one framework, the consensus anchor matrix A is continuously and alternately updated through the bipartite graph. Combined with the model of formula (2), the second intermediate clustering model can be obtained as follows:

[0064]

[0065] In the formula, α v is the weight coefficient for balancing each view; β is the balance coefficient; I c represents the identity matrix of dimension c×c, where c is the dimension of the reduced matrix; For X v The i-th row vector in ; O v The jth column vector in W v represents the dimension reduction matrix of the v-th view; A represents the consensus anchor matrix in the low-dimensional space, O v represents the anchor matrix of the vth view in the original space; (Z v ) T 1=1 constrains Z v The sum of the elements in each column of is 1;

[0066] Formula (3) under the constraint condition α T 1=1,(W v ) T W v =I c , A T A=I m , Z ≥ 0, Z T 1=1,G v ≥0, (G v )T 1=1. Set α, W v , Z, A, G v Minimum.

[0067] In order to effectively solve the problem of information exchange between different views, the present invention first learns the complementary bipartite graph G unique to each view. v And a consistent bipartite graph across views to capture the unique information and consistent information between different views. Then, by jointly learning the consistent bipartite graph Z and the complementary bipartite graph G v , the present invention can effectively promote the exchange and integration of shared information and unique information between different views, thereby further improving the efficiency of clustering. Therefore, the final JC 2 The BG model, that is, the second clustering model is as follows:

[0068]

[0069] Where min represents the objective function. The first term of the objective function measures the reconstruction error, that is, the degree to which the original data in the low-dimensional space is approximated by the relationship between the consensus anchor matrix A and the consistent bipartite graph Z; the second term is a regularization term used to obtain a better bipartite graph; the third term is measured by the anchor matrix O v The Euclidean distance from the original data point is to obtain a bipartite graph with local structural information;

[0070] α v is the weight of the v-th view, which is used to measure the importance of the v-th view; λ and β are regularization parameters, which are used to control the complexity of the model and avoid overfitting; I c represents the identity matrix of dimension c×c, where c is the dimension of the reduced matrix; Represents the square derivative of the F norm: Example: Tr represents the trace of a matrix. For example, the sum of the elements on the main diagonal of a matrix A (the diagonal from the upper left to the lower right) is called the trace (or trace number) of the matrix A. Represents the square of the 2-norm: Example: For X v The i-th row vector in G v represents the bipartite graph of the vth view obtained in the high-dimensional space, that is, the complementary bipartite graph; Z represents the consistent bipartite graph, which describes the consistency information between different views; O v The jth column vector in Z T 1=1,(G v ) T 1=1 constrains Z and G respectively v The sum of the elements in each column is 1.

[0071] Example 1

[0072] like Figure 1 As shown, the present invention provides a fast classification method for large-scale multi-view data, the method comprising:

[0073] Step 1, obtain a large-scale multi-view dataset to form a multi-view data matrix;

[0074] Step 2: based on the multi-view data matrix, obtain a bipartite graph of each view data and form a first clustering model;

[0075] Step 3, optimizing the first clustering model by introducing a graph learning item and a dual anchor learning mechanism to obtain a second clustering model; the second clustering model is a multi-view clustering model formed by jointly learning a consistent bipartite graph and a complementary bipartite graph;

[0076] In step 3, the first clustering model is optimized by introducing graph learning items and dual anchor learning mechanism to obtain the second clustering model, which specifically includes:

[0077] Step 31, using subspace learning and graph learning to capture the global structural features and local structural features of the first clustering model data to form a first intermediate clustering model;

[0078] Step 32, by introducing a dual anchor point learning mechanism, different types of anchor points are learned in high-dimensional space and low-dimensional space respectively, so as to obtain more representative anchor points and more comprehensively capture the relationship and features between data points, thereby forming a second intermediate clustering model;

[0079] Step 33: Based on the second intermediate clustering model, the consistent bipartite graph and the complementary bipartite graph are jointly learned to form a second clustering model, thereby realizing information exchange between different views.

[0080] Among them, in step 31, the learning ability of the first clustering model for local structural features is enhanced by introducing graph learning items; in step 32, the relationship and features between data points in the first intermediate clustering model are captured by introducing a dual anchor learning mechanism, and the set anchor matrix and the corresponding complementary bipartite graph of each view are learned on each view in the high-dimensional space; a dimensionality reduction matrix is ​​introduced to map the data in the high-dimensional space to the low-dimensional space, and a consensus anchor matrix is ​​learned in the low-dimensional space; and a consistent bipartite graph across views is constructed based on the consensus anchor matrix.

[0081] It should be noted that the above steps 2 and 3 refer to the final JC 2 The design process of the BG model and the final JC 2 The BG model is the second clustering model, which will not be described in detail here.

[0082] Step 4, using an iterative optimization algorithm to optimize and solve the second clustering model to obtain a solved second clustering model;

[0083] In this embodiment, after obtaining the final JC 2 After the BG model (second clustering model), because there are multiple variables and constraints in the model, it is very difficult to directly solve all the variables. Therefore, the present invention solves it by adopting an efficient iterative optimization algorithm. The solution process of step 4 is:

[0084] (1) Update W v

[0085] Fix A, Z, G in the second clustering model v , α v , and delete the v The irrelevant terms of the reduced dimension matrix W v ; W v It can be solved by the following formula:

[0086]

[0087] Because W v is learned independently in each view, so it remains independent between different views. Based on this feature, we can transform formula (5) into the following equivalent problem:

[0088]

[0089] Where D v =X v Z T A T ,W v The optimal solution of can be obtained by singular value decomposition (SVD), that is, UV T Among them, U and V are D v The left and right singular vector matrices obtained after singular value decomposition.

[0090] (2) Update A

[0091] Fix W in the second clustering model v , Z, G v , α v , and delete the irrelevant items of A to solve the consensus anchor matrix A; A can be solved by the following formula:

[0092]

[0093] With the matrix W v The solution is similar, and the final solution of matrix A is as follows:

[0094]

[0095] in Similarly, the optimal solution of A is also the product of the left singular value and the right singular value of M.

[0096] (3) Update G v

[0097] Fix W in the second clustering model v , Z, A, α v , and delete the G v Irrelevant terms, solve the complementary bipartite graph G v ; Matrix G v It can be solved by the following formula:

[0098]

[0099] In order to facilitate the solution, the present invention further represents a vector whose jth entry Therefore, the above problem in formula (9) can be reformulated for each i of G separately as:

[0100]

[0101] By taking the derivative of the above formula, we can get the following formula:

[0102]

[0103] It is worth noting that the above formula is a Euclidean projection problem on a simple space, and its Lagrangian function is:

[0104]

[0105] Among them, φ, are the scalar and vector of the Lagrange coefficients respectively. The optimal solution based on KKT (Karush-Kuhn-Tucker) is as follows:

[0106]

[0107] Note: For the matrix O (v) , the j-th column vector and the ij-th element are respectively and Express to update.

[0108] (4) Update Z

[0109] Fix W in the second clustering model v , G v , A, αv , and delete the irrelevant items of Z to solve the consistent bipartite graph Z; Z can be solved by the following formula:

[0110]

[0111] In order to obtain the optimal solution for Z, we transform the above formula into a quadratic programming (QP) problem, namely:

[0112]

[0113] in, Therefore, each column of Z is optimized using the QP problem.

[0114] (5) Update α

[0115] Fix W in the second clustering model v , G v , A, Z, and delete the irrelevant items with respect to α, and solve the matrix α; α can be solved by the following formula:

[0116]

[0117] This problem can be solved by the Cauchy-Bunyakovsky-Schwarz inequality, namely

[0118]

[0119] Among them, r v =||(W v ) T X v -AZ|| F . Similar to this r v Symbols represent intermediate variables, and their purpose is to simplify the solution process.

[0120] After obtaining the iterative expression of the variable to be optimized through the iterative optimization algorithm (that is, when optimizing one variable, the other variables are fixed), the optimization solution is performed through the specific algorithm flow in Table 1 below.

[0121] Table 1 Specific algorithm flow

[0122]

[0123]

[0124] The specific steps of optimizing the solution of the present invention are:

[0125] (1) Input multi-view data matrix X v , parameters λ, β, number of anchor points m, anchor point matrix O v ;

[0126] (2) Initialization: consistent bipartite graph (Z), consensus anchor matrix (A), dimension reduction matrix (W v ), each view weight vector (α v ), complementary bipartite graph (G v ) and the number of iterations t are initialized to 0; the maximum number of iterations maxIter = 20;

[0127] (3) When the algorithm does not converge and t<maxItler, perform the following steps;

[0128] 3.1) Assumption A t , Z t , (G v ) t , (α v ) t Given, find (W v ) t+1 ;

[0129] 3.2) Assumption (W v ) t , Z t , (G v ) t , (α v ) t Known, find A t+1 ;

[0130] 3.3) Assumption (W v ) t , A t , (G v ) t , (α v ) t Given, find Z t+1 ;

[0131] 3.4) Assumption (W v ) t , A t , Z t , (α v ) t Given, find (G v ) t+1 ;

[0132] 3.5) Assumption (W v ) t , A t , Z t , (G v ) t Given, find (α v ) t+1 ;

[0133] 3.6) Update t=t+1;

[0134] (4) The algorithm ends when the algorithm converges or t>maxIter.

[0135] Step 5: Input the large-scale multi-view dataset into the solved second clustering model for clustering to obtain a clustering result.

[0136] In the specific implementation, the present invention carried out relevant experimental verification on three different types of public large-scale datasets, including the SUNRGBD dataset, the AWA dataset and the YoutubeFace dataset. The detailed information of the datasets is shown in Table 2.

[0137] Table 2 Dataset

[0138]

[0139]

[0140] In order to be fair, all the comparison methods in this paper are tested according to the following rules: 1) Since all clustering algorithms involve the k-means algorithm, all the comparison algorithms are run 50 times of k-means to eliminate the randomness of k-means initialization; 2) The clustering performance of all comparison methods is evaluated by three commonly used clustering indicators (ACC, NMI and Purity). 2 The BG model was compared with several advanced models and achieved the best results in ACC, NMI and Purity, which effectively proved the rationality and superiority of the model.

[0141] The large-scale multi-view dataset refers to a data set containing a large number of samples, which are represented by a variety of different feature sets or description methods; each feature set or description method is regarded as an independent perspective or view in the dataset; the large-scale multi-view dataset includes an indoor scene image dataset, an animal image dataset or a face image dataset.

[0142] Such views can include various types of data, such as numerical data, text, or images, which can be obtained from different sources or generated through a variety of feature extraction techniques. Taking the SUNRGBD dataset as an example, it consists of two parts: one is the RGB image and the other is the corresponding depth information; there are 10,335 samples in the entire dataset, and each sample has both types of views.

[0143] Table 3 Comparison of large-scale datasets (%)

[0144]

[0145]

[0146] It can be seen from Table 3 that the final JC proposed by the present invention 2 The BG model (the second clustering model) is combined with Multi-view Low-rank Sparse Subspace Clustering (MLRSSC), Large-scale Multi-view Subspace Clustering in Linear Time (LMVSC), Scalable multi-view subspace clustering with unified anchors (SMVSC), Fast Parameter-Free Multi-View Subspace Clustering With Consensus AnchorGuidance (FPMVS-CAG), Anchor-based multi-view subspace clustering with hierarchical feature descent (An ... The proposed model has achieved good results in the three indicators of ACC, NMI and Purity. We obtained 8 best values ​​in 9 indicators of three large-scale datasets, SUNRGBD, AWA and YoutubeFace. (In the bold part of Table 3, 'OM' means that the algorithm has an out of memory error on the corresponding dataset.)

[0147] (1) Comparison with the conventional multi-view clustering MLRSSC method: When the data scale continues to grow, it becomes increasingly difficult for the MLRSSC method to perform cluster analysis under limited computing resources. This is because MLRSSC needs to calculate the similarity between all data points. When the size of the data set increases, this global similarity measurement process becomes very computationally intensive and time-consuming. Therefore, this method cannot efficiently process large-scale data sets. In contrast, the anchor-based method and the JC 2 BG models are not only able to effectively process large-scale datasets, but also achieve better results in terms of clustering performance. This is because they only need to calculate the similarity between anchor points and data points, which significantly reduces computing resources and time.

[0148] (2) Comparison of four anchor-based multi-view methods, LMVSC, SMVSC, FPMVS-CAG, and MVSC-HFD: In general, we have achieved almost the best clustering results on all datasets. Specifically, on the SUNRGBD dataset, we have improved ACC, NMI, and Purity by 2.52%, 1.66%, and 2.07%, respectively; on the AWA dataset, we have improved ACC, NMI, and Purity by 1.13%, 1.58%, and 1.27%, respectively; on the YoutubeFace dataset, although our improvement on ACC is not significant, we have improved the normalized mutual information (NMI) and purity by 2.1% and 0.7%, respectively. In general, the JC 2 The BG model can be flexibly applied to data sets of different sizes and can improve clustering performance to a great extent. In summary, it is proved that the method of the present invention is effective.

[0149] The new joint consistent bipartite graph and complementary bipartite graph multi-view clustering model designed by the present invention first uses subspace learning and graph learning to capture the global and local structural features of the data; then, by introducing a dual-anchor learning mechanism, different types of anchors are learned in high-dimensional space and low-dimensional space respectively, so as to obtain more representative anchors and more comprehensively capture the relationship and features between data points. Finally, by constructing a consistent bipartite graph and a complementary bipartite graph and performing joint learning, information exchange between different views is achieved. The present invention aims to improve the performance of existing multi-view clustering methods based on bipartite graphs when processing large-scale data sets, and further improve the clustering efficiency and accuracy of large-scale data sets.

[0150] The new joint consistent bipartite graph and complementary bipartite graph multi-view clustering (JC 2 The BG) model achieved the best results on the SUNRGBD, AWA and YoutubeFace datasets, compared with the most typical multi-view subspace clustering method (MLRSSC) and other latest anchor-based methods (LMVSC, SMVSC, FPMVS-CAG and MVSC-HFD). The best results were achieved on the SUNRGBD dataset, 9.98% on the AWA dataset, and 25.88% on the YoutubeFace dataset.

[0151] Example 2

[0152] like Figure 2As shown, the difference between this embodiment and embodiment 1 is that this embodiment provides a fast classification system for large-scale multi-view data, which corresponds to the fast classification method for large-scale multi-view data in embodiment 1 in a one-to-one manner; the system includes:

[0153] An acquisition unit, used for acquiring a large-scale multi-view data set to form a multi-view data matrix;

[0154] A first model building unit, configured to obtain a bipartite graph of each view data based on the multi-view data matrix and form a first clustering model;

[0155] The second model building unit is used to optimize the first clustering model by introducing a graph learning item and a dual anchor learning mechanism to obtain a second clustering model; the second clustering model is a multi-view clustering model formed by jointly learning a consistent bipartite graph and a complementary bipartite graph;

[0156] A model solving unit, used for optimizing and solving the second clustering model by using an iterative optimization algorithm to obtain a solved second clustering model;

[0157] The clustering unit is used to input the large-scale multi-view data set into the solved second clustering model for clustering to obtain a clustering result.

[0158] As a further implementation, the second model building unit includes:

[0159] A first optimization subunit is used to enhance the learning ability of the first clustering model for local features by introducing a graph learning item to obtain a first intermediate clustering model;

[0160] The second optimization subunit is used to capture the relationship and features between data points in the first intermediate clustering model by introducing a dual anchor learning mechanism, and learn the set anchor matrix and the corresponding complementary bipartite graph of each view in the high-dimensional space; introduce a dimensionality reduction matrix to map the data in the high-dimensional space to the low-dimensional space, and learn the consensus anchor matrix in the low-dimensional space; construct a consistent bipartite graph across views based on the consensus anchor matrix to obtain the second intermediate clustering model;

[0161] The joint learning subunit is used to perform joint learning on the consistent bipartite graph and the complementary bipartite graph based on the second intermediate clustering model to form a second clustering model.

[0162] The execution process of each unit may be performed according to the process steps of a fast classification method for large-scale multi-view data in Example 1, and will not be described in detail in this embodiment.

[0163] At the same time, the present invention also provides a fast classification device for large-scale multi-view data, including a memory, a processor, and computer executable instructions stored in the memory and executable on the processor. When the processor executes the computer executable instructions, the fast classification method for large-scale multi-view data is implemented.

[0164] At the same time, the present invention also provides a computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed by a processor, the above-mentioned fast classification method for large-scale multi-view data is implemented.

[0165] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer executable instruction products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer executable instruction product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program codes.

[0166] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer-executable instruction products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer-executable instructions. These computer-executable instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0167] These computer executable instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0168] These computer executable instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0169] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A fast classification method for large-scale multi-view data, characterized in that: The method includes: Acquire a large-scale multi-view dataset to form a multi-view data matrix; Based on the multi-view data matrix, obtaining a bipartite graph of each view data and forming a first clustering model; The first clustering model is optimized by introducing a graph learning item and a dual anchor learning mechanism to obtain a second clustering model; the second clustering model is a multi-view clustering model formed by jointly learning a consistent bipartite graph and a complementary bipartite graph; Using an iterative optimization algorithm to optimize and solve the second clustering model to obtain a solved second clustering model; The large-scale multi-view data set is input into the solved second clustering model for clustering to obtain a clustering result.

2. The fast classification method for large-scale multi-view data according to claim 1, characterized in that: The large-scale multi-view dataset refers to a dataset containing a large number of samples, which are represented by a variety of different feature sets or description methods; each feature set or description method is regarded as an independent perspective or view in the dataset; The large-scale multi-view dataset includes an indoor scene image dataset, an animal image dataset or a human face image dataset.

3. The fast classification method for large-scale multi-view data according to claim 1, characterized in that: The first clustering model is optimized by introducing a graph learning item and a dual anchor point learning mechanism to obtain a second clustering model, including: Using subspace learning and graph learning to capture global structural features and local structural features of the first clustering model data to form a first intermediate clustering model; By introducing a dual-anchor learning mechanism, different types of anchors are learned in high-dimensional space and low-dimensional space respectively, so as to obtain more representative anchors and more comprehensively capture the relationship and features between data points, thus forming a second intermediate clustering model. Based on the second intermediate clustering model, the consistent bipartite graph and the complementary bipartite graph are jointly learned to form a second clustering model.

4. A fast classification method for large-scale multi-view data according to claim 1 or 3, characterized in that: The expression of the second clustering model is: s.t.α T 1=1,(W v ) T W v =I c ,A T A=I m ,z≥0,z T 1=1,G v ≥0,(G v ) T 1=1. Where min represents the objective function. The first term of the objective function measures the reconstruction error, that is, the degree to which the original data in the low-dimensional space is approximated by the relationship between the consensus anchor matrix A and the consistent bipartite graph z; the second term is a regularization term used to obtain a better bipartite graph; the third term is measured by the anchor matrix Ov 和 The Euclidean distance of the original data points aims to obtain a bipartite graph with local structural information; V represents the total number of views; α v is the weight of the v-th view, which is used to measure the importance of the v-th view; W v It is the dimension reduction matrix of the v-th view, which is used to map the original features to a new feature space; () I represents the transpose of the matrix; X v is the data matrix of the vth view, which contains the feature information of n samples; A is the consensus anchor matrix, which represents the representative samples shared by all views, i.e., anchor points; z is a consistent bipartite graph, which describes the consistency information between different views; is the square of the F norm; λ and β are regularization parameters used to control model complexity and avoid overfitting; Gv represents the bipartite graph of the vth view obtained in the high-dimensional space, that is, the complementary bipartite graph; m is the number of anchor points, n is the number of samples; O v is the anchor matrix obtained by the vth view in the high-dimensional space, which reflects the distribution of anchor points under each view; X v is the multi-view data matrix; For X v The i-th row vector in ; O v The j-th column vector in ; is the square of the 2-norm; G v The element in row i and column j of ; st is the constraint condition; I is the identity matrix, c is the dimension of the reduced matrix; α T 1=1 means the sum of the weights of each view is 1; z T 1=1,(G v ) T 1=1 constrains Z and G respectively v The sum of the elements in each column is 1.

5. The fast classification method for large-scale multi-view data according to claim 4, characterized in that: The second clustering model is optimized and solved by using an iterative optimization algorithm, including: Using an iterative optimization algorithm, A, Z, G in the second clustering model are fixed v , α v , and delete the v The irrelevant terms of the reduced dimension matrix W v ; Fix W in the second clustering model v , Z, G v , α v , and delete the irrelevant items related to A to solve the consensus anchor matrix A; Fix W in the second clustering model v , Z, A, α v , and delete the G v Irrelevant terms, solve the complementary bipartite graph G v ; Fix W in the second clustering model v , G v , A, α v , and delete the irrelevant items of Z to solve the consistent bipartite graph Z; Fix W in the second clustering model v , G v , A, z, and delete the irrelevant items with α, and solve the matrix α; Based on the obtained dimension reduction matrix W v , consensus anchor matrix A, complementary bipartite graph G v , consistent bipartite graph Z and matrix α, and perform optimization solution.

6. The fast classification method for large-scale multi-view data according to claim 3, characterized in that: The consistent bipartite graph retains the consistency information between different views; the complementary bipartite graph retains the difference information between different views.

7. A fast classification system for large-scale multi-view data, characterized in that: The system includes: An acquisition unit, used for acquiring a large-scale multi-view data set to form a multi-view data matrix; A first model building unit, configured to obtain a bipartite graph of each view data based on the multi-view data matrix, and form a first clustering model; A second model building unit, configured to optimize the first clustering model by introducing a graph learning item and a dual anchor learning mechanism to obtain a second clustering model; the second clustering model is a multi-view clustering model formed by jointly learning a consistent bipartite graph and a complementary bipartite graph; A model solving unit, used for optimizing and solving the second clustering model by using an iterative optimization algorithm to obtain a solved second clustering model; The clustering unit is used to input the large-scale multi-view data set into the solved second clustering model for clustering to obtain a clustering result.

8. The fast classification system for large-scale multi-view data according to claim 8, characterized in that: The second model building unit comprises: A first optimization subunit, configured to enhance the learning ability of the first clustering model for local features by introducing a graph learning item, so as to obtain a first intermediate clustering model; The second optimization subunit is used to capture the relationship and features between data points in the first intermediate clustering model by introducing a dual anchor learning mechanism, and learn the set anchor matrix and the complementary bipartite graph corresponding to each view in the high-dimensional space; introduce a dimensionality reduction matrix to map the data in the high-dimensional space to the low-dimensional space, and learn the consensus anchor matrix in the low-dimensional space; construct a consistent bipartite graph across views based on the consensus anchor matrix to obtain the second intermediate clustering model; The joint learning subunit is used to perform joint learning on the consistent bipartite graph and the complementary bipartite graph based on the second intermediate clustering model to form a second clustering model.

9. A fast classification device for large-scale multi-view data, comprising a memory and a processor, wherein the processor is used to implement various instructions, and the memory is used to store multiple computer executable instructions, characterized in that: The computer executable instructions enable the processor to execute the fast classification method for large-scale multi-view data as described in any one of claims 1 to 6.

10. A computer-readable storage medium, wherein a plurality of computer-executable instructions are stored in the storage medium, characterized in that: When the computer executable instructions are executed by a processor, a fast classification method for large-scale multi-view data is implemented as described in any one of claims 1 to 6.

Images