A semi-supervised multi-view clustering ensemble method and system based on width learning

Abstract

The present invention relates to the field of machine learning, and discloses a semi-supervised multi-view clustering integration method and system based on width learning, the method comprising the following steps: constructing an autoencoder based on width learning that can perform feature processing based on multi-view face image data, for extracting features from input face images, and calculating the weights of different node features based on width learning; constructing a clustering integration model by combining the autoencoder based on width learning and bidirectional constraint propagation; and realizing clustering integration of multi-view face images using the clustering integration model. The present invention utilizes the performance advantages of the width learning system to perform clustering processing on multi-view data, obtains a model network with both efficiency and performance, introduces a flow structure and pairwise constraints, and adopts constraint propagation in the consensus process to enrich sample information, effectively improving the robustness and accuracy of the clustering integration network model based on width learning, and thus is more applicable in actual scenarios.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a semi-supervised multi-view clustering integration method and system based on width learning. Background Art

[0002] Clustering integration of multi-view face images refers to clustering analysis of face image data under multiple views (or feature subspaces), and then integrating the clustering results of these views through a certain integration strategy to obtain a more comprehensive and accurate clustering result. Multi-view clustering integrates heterogeneous data of multiple views of face images. Different views can provide different information, so that face images can be analyzed and understood more comprehensively and objectively. However, clustering of multi-view data of face images needs to comprehensively consider the information between different views, weigh the importance of different views and solve the processing of high-dimensional data. Taking a dataset containing 1,000 face images as an example, the data may be distributed in 50 different categories and contain different lighting, expressions and angle changes (i.e. different views). This leads to large computational complexity and insufficient clustering quality of existing methods, which limits the application of multi-view clustering methods in reality. Summary of the invention

[0003] The purpose of the present invention is to solve the problems in the prior art.

[0004] The technical solution adopted by the present invention to solve the technical problem is: to provide a semi-supervised multi-view clustering integration method based on width learning, comprising the following steps:

[0005] Construct a width-learning-based autoencoder for feature processing based on multi-view face image data, which is used to extract features from input face images and calculate the weights of different node features based on width learning;

[0006] Combining width-learning-based autoencoders and bidirectional constraint propagation to build a clustering ensemble model;

[0007] Use clustering ensemble model to realize clustering ensemble of multi-view face images;

[0008] The process of extracting features of the width-learning-based autoencoder includes the following steps:

[0009] Design the overall objective function and solve the objective function to obtain the weight matrix;

[0010] The subspace representation is calculated based on the weight matrix and is expressed as:

[0011]

[0012] Among them, F vrepresents the subspace representation of the v-th view; X v W represents the image set that needs to be clustered; v represents the weight matrix;

[0013] Running k-means on the subspace representation of each view generates a set of basic partitions, denoted as P v ={π1,π2,...,π V}, where π1,π2,...,π V represents a series of basic partitions generated by the subspace of the v-th view, and the set of basic partitions is denoted by P v express.

[0014] Preferably, the designing of the overall objective function and solving the objective function to obtain a weight matrix comprises the following steps:

[0015] Design the overall objective function, expressed as:

[0016]

[0017]

[0018] Where n represents the number of samples; is the regularization component; Representing flow pattern constraints, ensuring that local and global affinities in the projected subspace are consistent; represents the F-norm, represents the l-2 norm; A v represents the hidden layer; α represents the trade-off parameter for the regularization term; β represents the trade-off parameter for the flow constraint term; r ij represents the (i, j)th value of the constraint affinity matrix R, which contains pairwise relationships from global affinity and constraints; M represents must-connect constraints, C represents do-not-connect constraints; V represents the number of views; N v (x i ,x j ) is the neighbor parameter, if x i (x j ) is x j (x i )’s K nearest neighbors, then N v (x i ,x j )=1, otherwise N v (x i ,x j )=0.

[0019] In order to minimize the objective function, the objective function is rewritten as:

[0020]

[0021] Among them, L v is the Laplace matrix of R, Tr represents the trace;

[0022] Solve the objective function to get the weight matrix W v , expressed as:

[0023]

[0024] Where I represents the identity matrix and l represents A v Dimension.

[0025] Preferably, the method of combining the width learning-based autoencoder and the bidirectional constraint propagation to construct a clustering integration model network for width learning comprises the following steps:

[0026] The cluster structure information contained in P is encapsulated through an association matrix C, which is expressed as:

[0027]

[0028] Among them, π v (x i ) is π v The cluster labels of

[0029] Adopt bidirectional constraint propagation to construct undirected graphs;

[0030] The graph cut algorithm is used to partition the undirected network to obtain a consensus partition.

[0031] Preferably, the method of constructing an undirected graph by using bidirectional constraint propagation comprises the following steps:

[0032] Construct the matrix G, and use the matrix G to calculate the sample unified similarity matrix S, which is expressed as:

[0033]

[0034] Among them, S ij , G ij and C ij Respectively represent the data of the i-th row and j-th column in the sample unified similarity matrix S, the data of the i-th row and j-th column in the matrix G, and the data of the i-th row and j-th column in the matrix C;

[0035] Finally, the samples are regarded as nodes, and S is used as the node connection matrix to construct an undirected graph.

[0036] Preferably, the constructing matrix G comprises the following steps:

[0037] In order to absorb the supervision information, bidirectional constraint propagation is used to propagate constraints in the horizontal and vertical directions. The expression of bidirectional constraint propagation is expressed as:

[0038]

[0039] Where L is the Laplacian matrix of C;

[0040] The closed-form solution to the expression for bidirectional constraint propagation is given by:

[0041]

[0042] in L is the Laplacian matrix of C, and λ represents the regularization coefficient.

[0043] Preferably, the method of realizing cluster integration of multi-view face images by using a cluster integration model comprises the following steps:

[0044] Input multi-view face images into the clustering ensemble model;

[0045] The clustering ensemble model uses a width-based learning autoencoder to learn a subspace representation for each view, and then applies K-means to the subspace to obtain a series of partitions;

[0046] The clustering ensemble model adopts bidirectional constraint propagation, constructs an undirected graph based on a series of partitions, and uses a graph cut algorithm to obtain consensus partitions.

[0047] The present invention also provides a semi-supervised multi-view clustering integration system based on width learning, comprising the following steps:

[0048] The encoder construction module constructs an autoencoder based on width learning for feature processing based on multi-view face image data, which is used to extract features from the input face image and calculate the weights of different node features based on width learning;

[0049] Model building module, combining width-learning-based autoencoders and bidirectional constraint propagation to build a clustering ensemble model;

[0050] Clustering integration module, which uses clustering integration model to realize clustering integration of multi-view face images;

[0051] The process of extracting features of the width-learning-based autoencoder includes the following steps:

[0052] Design the overall objective function and solve the objective function to obtain the weight matrix;

[0053] The subspace representation is calculated based on the weight matrix and is expressed as:

[0054]

[0055] Among them, F v represents the subspace representation of the vth view; Xv W represents the image set that needs to be clustered; v represents the weight matrix;

[0056] Running k-means on the subspace representation of each view generates a set of basic partitions, denoted as P v ={π1,π2,...,π V}, where π1,π2,...,π V represents a series of basic partitions generated by the subspace of the v-th view, and the set of basic partitions is denoted by P v express.

[0057] The present invention has the following beneficial effects: the present invention utilizes the performance advantage of the width learning system to perform clustering processing of multi-view data, and obtains a model network with both efficiency and performance. This method does not require large-scale computing resources and can be easily run even on an ordinary computer; the data processing introduces manifold structure and pairwise constraints, constructs a constrained multi-view width autoencoder network, takes label information into account, and adopts constraint propagation in the consensus process to enrich sample information, effectively improving the robustness and accuracy of the clustering integration network model based on width learning, and is therefore more applicable in actual scenarios.

[0058] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A method step diagram of an embodiment of the present invention;

[0060] Figure 2 A width learning network structure diagram of an embodiment of the present invention;

[0061] Figure 3 This is a network structure diagram of a width learning autoencoder according to an embodiment of the present invention.

[0062] Figure 4 A structural diagram of a clustering integration model according to an embodiment of the present invention;

[0063] Figure 5 It is a schematic diagram of experimental results of an embodiment of the present invention;

[0064] Figure 6 4 is a system structure diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0065] See also Figure 1 FIG. 1 is a diagram showing steps of a method according to an embodiment of the present invention, comprising the following steps:

[0066] S101, constructing a width learning autoencoder that can perform feature processing based on multi-view face image data, for extracting features from an input face image, and calculating weights of different node features based on width learning;

[0067] S102, building a clustering ensemble model by combining width-learning-based autoencoders and bidirectional constraint propagation;

[0068] S103, using a clustering integration model to implement clustering integration of multi-view face images.

[0069] Specifically, the width learning autoencoder constructed in S101 can integrate cross-attempt affinity and pairwise constraints. The width learning autoencoder includes a width learning network, which efficiently determines the weights between feature nodes and enhancement nodes by directly calculating the pseudo-inverse, avoiding the repeated adjustment of feature extractor parameters in the complex back-propagation process in traditional deep neural networks. This method significantly improves the efficiency and flexibility of training, and at the same time shows obvious advantages in processing speed and effect compared to existing deep neural network structures. Specifically, the width learning network is as follows Figure 2 As shown in the figure, the original input is first converted into random features to form a feature layer. Then, the overall rough structure is expanded in the form of an enhancement layer. In particular, the input is converted into some random features through feature mapping technology. Multiple feature node parts are connected and enhanced nodes are generated through different transformation functions. All these nodes are gathered together and connected to the output layer, and the weights of the output layer are calculated by a fast pseudo-inverse calculation strategy. In the figure, Mapped Feature is a mapping node and Enhancement Feature is an enhancement node.

[0070] join Figure 3 As shown, this is a network structure diagram of the width learning autoencoder of an embodiment of the present invention. First, the original feature distribution is mapped to a random feature space; then, the feature space is further mapped to an enhanced feature space to study more abstract feature information; finally, the ridge regression technique is used to accelerate the reconstruction of the original input.

[0071] Specifically, the width learning autoencoder extracts the compressed feature subspace and derives the basic partition. The process is as follows:

[0072] On the basis of the width learning autoencoder, in order to combine pairwise constraints and cross-attempt affinity information, the designed width learning-based autoencoder aims to preserve the popular structure of the data in the subspace. The overall objective function is as follows:

[0073]

[0074] in, represents the data set to be solved, n represents the number of samples, dv Represents the sample feature dimension; is the regularization component; the third term is the flow constraint, which ensures that the local and global affinities in the projected subspace are consistent. v represents the hidden layer; α represents the trade-off parameter for the regularization term; β represents the trade-off parameter for the flow constraint term; r ij corresponds to the (i,j)th value of the constraint affinity matrix R, which contains pairwise relationships from global affinities and constraints and is calculated as follows:

[0075]

[0076] In order to minimize the objective function, the objective function is rewritten as:

[0077]

[0078] Where L v is the Laplace matrix of R. Solving the objective function yields the weight matrix W v :

[0079]

[0080] The subspace representation F of the vth view is obtained using the following formula: v :

[0081]

[0082] To improve efficiency, k-means is run on each subspace to generate a series of basic partitions, which are denoted as P v ={π1,π2,...,π V}.

[0083] Specifically, S102 absorbs information from pairwise constraints and multiple partitions and uses bidirectional constraint propagation to refine the unified similarity matrix to obtain better consensus partitions. The specific process is as follows:

[0084] P v The cluster structure information contained in can be encapsulated by an association matrix C:

[0085]

[0086] where π v (x i ) = j is π v In order to absorb the supervision information, two-way constraint propagation is adopted:

[0087]

[0088] Where L is the Laplace matrix of C and the closed-form solution is calculated as follows:

[0089]

[0090] in Using G, we can get the sample unified similarity matrix S:

[0091]

[0092] Finally, the samples are regarded as nodes, S is used as the node connection matrix, an undirected graph is constructed, and the graph is partitioned using the graph cut algorithm to obtain the consensus partition.

[0093] Specifically, the clustering ensemble model network is shown in Figure 4 As shown in the figure, in the generation process, a constrained multi-view broad self-encoding network is proposed to learn the subspace representation of each view, which combines information from cross-view affinity, intra-view projection and pairwise constraints, and then applies K-means to the subspace to obtain a series of partitions; in the consensus process, bidirectional constraint propagation is used to construct an undirected graph, and the graph cut algorithm is used to obtain the consensus partition. In the figure, constrained affinity matrix R is the pairwise constraint matrix R, must-link M is the must-link constraint M, cannot-link C is the cannot-link constraint C, Generation: constrained multi-view broad self-encoding network is the generation process: constrained multi-view broad self-encoding network, Consensus: bidirectional constraint propagation and graph cut is the consensus process: bidirectional constraint propagation and graph cut, and consensus partition is the consensus partition.

[0094] The embodiments of the present invention are experimentally verified. ORL is a face image dataset consisting of 400 face images of 40 people. The shooting time, lighting, angle, facial expression and other facial expressions of these images are different. Experimental verification is carried out on the public face image dataset, and the two widely used evaluation indicators, Normalized Mutual Information (NMI) and Adjusted Rand Index (ARI), are used to comprehensively evaluate the clustering results. The value range of NMI is [0,1], and the value range of ARI is [-1,1]; the higher the values ​​of NMI and ARI, the better the clustering performance. The experimental results are shown in Table 1.

[0095] Table 1:

[0096]

[0097] In addition, two main parameters, i.e., α and β, are analyzed by selecting different values ​​of α and β from the set {0.01, 0.1, 1, 10, 100} to evaluate the performance. Figure 5 As shown, the x-axis represents α, the y-axis represents β, and the z-axis represents the NMI score. In this 3D graph, the red area indicates that the embodiment of the present invention performs better, while the blue area indicates the opposite. It should be emphasized that when the value of α is in a smaller range {0.01, 0.1}, the performance of the embodiment of the present invention remains stable without being affected by the change of β, and shows better stability and higher overall performance. This suggests that setting α within this range may be a better choice for these data sets.

[0098] Experiments show that when clustering is performed using the multi-view clustering integration method of the embodiment of the present invention, the NMI value reaches 0.862 and the ARI value reaches 0.813, which is significantly improved compared with single-view clustering. This shows that multi-view clustering integration can effectively utilize the complementarity of different feature views, thereby improving the clustering effect.

[0099] See also Figure 6 FIG. 1 is a system structure diagram of an embodiment of the present invention, including:

[0100] The encoder construction module 601 constructs an autoencoder based on width learning that can perform feature processing based on multi-view face image data, and is used to extract features from the input face image and calculate weights of different node features based on width learning;

[0101] A model building module 602, combining the width-based learning autoencoder and bidirectional constraint propagation to build a clustering ensemble model;

[0102] The clustering integration module 603 implements clustering integration of multi-view face images using a clustering integration model.

[0103] The present invention discloses a semi-supervised multi-view clustering integration method for face image data based on width learning. The method utilizes the performance advantages of the width learning system to perform clustering processing on multi-view data, and obtains a model network with both efficiency and performance. This method does not require large-scale computing resources, and can be easily run even on an ordinary computer. The data processing introduces flow structure and pairwise constraints, constructs a constrained multi-view width autoencoder network, takes label information into account, and adopts constraint propagation in the consensus process to enrich sample information, effectively improving the robustness and accuracy of the clustering integration network model based on width learning, so it is more applicable in actual scenarios.

[0104] It can be seen that the present invention utilizes the performance advantages of the width learning system to perform clustering processing on multi-view data, and obtains a model network with both efficiency and performance. It introduces flow structure and pairwise constraints, and uses constraint propagation in the consensus process to enrich sample information, effectively improving the robustness and accuracy of the clustering ensemble network model based on width learning, and is therefore more applicable in actual scenarios.

[0105] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A semi-supervised multi-view clustering ensemble method based on width learning, characterized in that: The following steps are involved: Construct a width-learning-based autoencoder for feature processing based on multi-view face image data, which is used to extract features from input face images and calculate the weights of different node features based on width learning; Combining width-learning-based autoencoders and bidirectional constraint propagation to build a clustering ensemble model; Use clustering ensemble model to realize clustering ensemble of multi-view face images; The process of extracting features of the width-learning-based autoencoder includes the following steps: Design the overall objective function and solve the objective function to obtain the weight matrix; The subspace representation is calculated based on the weight matrix and is expressed as: Among them, F v represents the subspace representation of the vth view; X v W represents the image set that needs to be clustered; v represents the weight matrix; Running k-means on the subspace representation of each view generates a set of basic partitions, denoted as P v ={π1,π2,...,π V }, where π1,π2,...,π V represents a series of basic partitions generated by the subspace of the v-th view, and the set of basic partitions is denoted by P v express; The design of the overall objective function and solving the objective function to obtain a weight matrix comprises the following steps: Design the overall objective function, expressed as: Where n represents the number of samples; is the regularization component; Representing flow pattern constraints, ensuring that local and global affinities in the projected subspace are consistent; represents the F-norm, represents the l-2 norm; A v represents the hidden layer; α represents the trade-off parameter for the regularization term; β represents the trade-off parameter for the flow constraint term; r ij represents the (i,j)th value of the constraint affinity matrix R, which contains pairwise relationships from global affinities and constraints; represents a must-connect constraint, C represents a do-not-connect constraint; V represents the number of views; N v (x i ,x j ) is the neighbor parameter, if x i (x j ) is x j (x i )’s K nearest neighbors, then N v (x i ,x j )=1, otherwise N v (x i ,x j )=0; In order to minimize the objective function, the objective function is rewritten as: Among them, L v is the Laplace matrix of R, Tr represents the trace; Solve the objective function to get the weight matrix W v , expressed as: Among them, I represents the identity matrix, l represents A v Dimension.

2. The semi-supervised multi-view clustering ensemble method based on width learning according to claim 1, characterized in that: The method of combining the width learning-based autoencoder and bidirectional constraint propagation to construct a clustering integration model network for width learning includes the following steps: The cluster structure information contained in P is encapsulated through an association matrix C, which is expressed as: Among them, π v (x i ) is π v The cluster labels of Adopt bidirectional constraint propagation to construct undirected graphs; The graph cut algorithm is used to partition the undirected network to obtain a consensus partition.

3. The semi-supervised multi-view clustering ensemble method based on width learning according to claim 2, characterized in that: The method of constructing an undirected graph by using bidirectional constraint propagation includes the following steps: Construct the matrix G and use the matrix G to calculate the sample unified similarity matrix S, which is expressed as: Among them, S ij , G ij and C ij Respectively represent the data of the i-th row and j-th column in the sample unified similarity matrix S, the data of the i-th row and j-th column in the matrix G, and the data of the i-th row and j-th column in the matrix C; Finally, the samples are regarded as nodes, and S is used as the node connection matrix to construct an undirected graph.

4. The semi-supervised multi-view clustering ensemble method based on width learning according to claim 3, characterized in that: The construction matrix G comprises the following steps: In order to absorb the supervision information, bidirectional constraint propagation is used to propagate constraints in the horizontal and vertical directions. The expression of bidirectional constraint propagation is expressed as: Where L is the Laplacian matrix of C; The closed-form solution to the expression for bidirectional constraint propagation is given by: in L is the Laplacian matrix of C, and λ represents the regularization coefficient.

5. The semi-supervised multi-view clustering ensemble method based on width learning according to claim 1, characterized in that: The method of realizing cluster integration of multi-view face images by using a cluster integration model comprises the following steps: Input multi-view face images into the clustering ensemble model; The clustering ensemble model uses a width-based learning autoencoder to learn a subspace representation for each view, and then applies K-means to the subspace to obtain a series of partitions; The clustering ensemble model adopts bidirectional constraint propagation, constructs an undirected graph based on a series of partitions, and uses a graph cut algorithm to obtain consensus partitions.

6. A semi-supervised multi-view clustering ensemble system based on width learning, characterized in that: The following steps are involved: The encoder construction module constructs an autoencoder based on width learning for feature processing based on multi-view face image data, which is used to extract features from the input face image and calculate the weights of different node features based on width learning; Model building module, combining width-learning-based autoencoders and bidirectional constraint propagation to build a clustering ensemble model; Clustering integration module, which uses clustering integration model to realize clustering integration of multi-view face images; The process of extracting features of the width-learning-based autoencoder includes the following steps: Design the overall objective function and solve the objective function to obtain the weight matrix; The subspace representation is calculated based on the weight matrix and is expressed as: Among them, F v represents the subspace representation of the vth view; X v W represents the image set that needs to be clustered; v represents the weight matrix; Running k-means on the subspace representation of each view generates a set of basic partitions, denoted as P v ={π1,π2,...,π V }, where π1,π2,...,π V represents a series of basic partitions generated by the subspace of the v-th view, and the set of basic partitions is denoted by P v express; The design of the overall objective function and solving the objective function to obtain a weight matrix comprises the following steps: Design the overall objective function, expressed as: Where n represents the number of samples; is the regularization component; Representing flow pattern constraints, ensuring the consistency of local and global affinities in the projected subspace; represents the F-norm, represents the l-2 norm; A v represents the hidden layer; α represents the trade-off parameter for the regularization term; β represents the trade-off parameter for the flow constraint term; r ij represents the (i, j)th value of the constraint affinity matrix R, which contains pairwise relationships from global affinity and constraints; M represents must-connect constraints, C represents do-not-connect constraints; V represents the number of views; N v (x i ,x j ) is the neighbor parameter, if x i (x j ) is x j (x i )’s K nearest neighbors, then N v (x i ,x j )=1, otherwise N v (x i ,x j )=0; In order to minimize the objective function, the objective function is rewritten as: Among them, L v is the Laplace matrix of R, Tr represents the trace; Solve the objective function to get the weight matrix W v , expressed as: Among them, I represents the identity matrix, l represents A v Dimension.

Images