Rapid depth multi-view subspace clustering method based on dynamic anchor points

Abstract

The invention discloses a rapid depth multi-view subspace clustering method based on dynamic anchor points, which comprises the following steps: obtaining multi-view sample data, inputting the multi-view sample data into a pre-training network model for reconstruction and processing, preliminarily obtaining the characteristics of each view anchor point, and updating the network parameters of an auto-encoder and the dynamic anchor points; based on a pre-trained network model, constructing a fine tuning network, learning a sparse representation relation matrix of each view, adaptively allocating a weight to the sparse representation relation matrix of each view by using a self-weight network, obtaining a weighted sparse representation relation matrix of each view, and fusing the weighted sparse representation relation matrixes of each view to obtain a sparse representation relation matrix of each view; a multi-view consensus sparse representation relation matrix is obtained, and a final clustering result is obtained through a fast spectral clustering method. According to the method, quadratic and cubic of space-time complexity of an existing method about the number of samples are reduced to linear overhead, and efficient clustering of multi-view data is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-view clustering, and in particular to a fast deep multi-view subspace clustering method based on dynamic anchor points. Background Art

[0002] In the "data oil era", mining important information from massive data and making data valuable can promote the organic integration of artificial intelligence, big data, economic and social development. Cluster analysis, as a typical data mining method, is a key technology for unlabeled data processing and provides high-quality data analysis support for human social activities. With the rapid development of information technology, the explosively growing data presents characteristics such as high dimensionality, nonlinearity, and high complexity. In addition, with the diversification of information collection methods and data types, the data of the same object can usually be obtained by different sensors, media systems, cameras, feature extractors, etc., forming multi-view data with characteristics such as polymorphism, multi-source, multi-feature, and high-dimensional nonlinearity. High-dimensional nonlinear and complex multi-view data make traditional clustering methods unable to meet the needs of practical applications, and also pose new challenges to clustering methods based on linear mapping functions or shallow models. With the rapid development of artificial intelligence technology, deep learning has achieved good application results in clustering tasks with its powerful nonlinear representation ability and effective feature extraction ability. However, it has been criticized for its "black box" operation. The subspace clustering, an effective high-dimensional data processing technology with a solid theoretical foundation in linear algebra and signal processing, can make up for its shortcomings of insufficient theoretical support to a certain extent.

[0003] The deep multi-view subspace clustering method uses a deep neural network to extract its effective feature representation, and to a certain extent removes redundant information and noise in the data; then, based on the feature space, a self-representation learning network is constructed to learn to obtain a multi-view consensus self-representation coefficient matrix with consistency and complementarity, revealing the representation relationship between data samples; finally, the similarity matrix is ​​calculated based on the consensus self-representation coefficient matrix, and spectral clustering is used to complete the multi-view data clustering. Figure 1 Some solutions are provided based on the consistency and complementarity of the proposed method, and excellent clustering performance is achieved. However, these methods are based on the self-representation property, use the feature representation of all data samples to learn the multi-view consensus self-representation coefficient matrix, construct the similarity matrix through the self-representation coefficient matrix, and perform spectral clustering to complete the final data partition. The space and time complexities are respectively and This makes the application of existing methods to large-scale data analysis very limited.

[0004] In the field of traditional multi-view subspace clustering, an anchor-based method is proposed. It applies bipartite graph learning and sparse representation ideas in graph theory to construct a sparse representation relationship between data anchors and data samples, avoiding the high memory requirements caused by learning data self-representation relationships. However, this method constructs multi-view sparse representation relationships based on linear mapping functions, which is not very applicable to nonlinear data and has the following defects: (1) The data anchors are obtained from multi-view data by random sampling, uniform sampling, etc., which are not representative of the source data and will cause the loss of data structure information, making the sparse representation relationship matrix based on anchors unable to fully reflect the similarity relationship between multi-view data; (2) The k-m Although the anchor samples obtained by the eans clustering algorithm can effectively represent multi-view data, the anchor samples selected by k-means are easily mixed with noise based on the local information of the sample data space, and will also increase the time complexity of the algorithm; (3) Moreover, in most traditional anchor-based multi-view subspace clustering methods, the anchor points are selected and determined before the clustering process, and remain fixed in the subsequent clustering process, which is separated from the construction of the sparse representation coefficient matrix, and will have an adverse effect on the clustering performance; (4) In addition, after obtaining the sparse representation relationship matrix, most traditional anchor-based multi-view subspace clustering methods still use general spectral clustering to realize data sample clustering, which still has a high time complexity. Summary of the invention

[0005] To solve the above technical problems, the present invention proposes a fast deep multi-view subspace clustering method based on dynamic anchor points, with the aim of improving the efficiency of the deep multi-view subspace clustering method and expanding its application to large-scale data, breaking through the limitations of the self-representation characteristics of the data subspace, avoiding the shortcomings of the sampling anchor points being unrepresentative of the data and the additional time overhead brought by the k-means anchor points, and overcoming the defect of separating the selection of fixed anchor points from the clustering process.

[0006] To achieve the above object, the present invention provides a fast deep multi-view subspace clustering method based on dynamic anchor points, comprising:

[0007] Obtain multi-view sample data, input the multi-view sample data into a pre-trained network model for reconstruction and processing, preliminarily obtain the anchor features of each view, and update the autoencoder and dynamic anchor network parameters, wherein the pre-trained network model is composed of a deep autoencoder and a dynamic anchor learning network;

[0008] Based on the pre-trained network model, a fine-tuning network is constructed to learn the sparse representation relationship matrix of each view, and a self-weight network is used to adaptively assign weights to the sparse representation relationship matrix of each view to obtain a weighted sparse representation relationship matrix of each view, and the weighted sparse representation relationship matrix of each view is fused to obtain a multi-view consensus sparse representation relationship matrix;

[0009] The final clustering result is obtained by using a fast spectral clustering method on the multi-view consensus sparse representation relationship matrix.

[0010] Preferably, obtaining the multi-view sample data comprises:

[0011] A multi-view public data set is obtained, and multi-view data with diverse sample numbers, diverse categories and rich details are selected. Then, the matrix type data in the multi-view data is normalized, and all the multi-view data are processed into a batch of data to obtain the multi-view sample data.

[0012] Preferably, initially acquiring the anchor features of each view and updating the autoencoder and dynamic anchor network parameters include:

[0013] Reconstructing the sample data by the deep autoencoder to obtain an effective potential feature representation of each view;

[0014] Based on the predefined anchor features, the dynamic anchor learning network is used to reversely map and reconstruct the effective potential feature representation, and after end-to-end training, the anchor features of each view are obtained, and the parameters of the autoencoder and the dynamic anchor network are updated.

[0015] Preferably, constructing the fine-tuning network to learn the sparse representation relationship matrix of each view includes:

[0016] Designing a fully connected sparse representation learning layer based on the anchor point features and the effective potential feature representation;

[0017] Embedding the fully connected sparse representation learning layer between the deep autoencoder and the dynamic anchor learning network to construct the fine-tuning network;

[0018] The view anchor features are dynamically updated through the fine-tuning network and the sparse representation relationship matrix of each view is learned.

[0019] Preferably, obtaining the weighted sparse representation relationship matrix of each view includes:

[0020] A sparse representation tensor is constructed based on the sparse representation relationship matrix of each view, and the sparse representation tensor is compressed along the channel dimension using global average pooling to obtain a global description factor of each channel, i.e., each sparse representation relationship matrix;

[0021] For the global description factor of each sparse representation relationship matrix, a two-layer unbiased fully connected network is used to capture the relationship between the view self-representation coefficient matrices and stimulate them to a weight between 0 and 1;

[0022] Re-scaling the sparse representation tensor using the weight to obtain a weighted sparse representation tensor;

[0023] The self-weighted sparse representation tensor is split along the channel dimension to obtain the weighted sparse representation relationship matrix of each view.

[0024] Preferably, the weighted sparse representation relationship matrices of the views are fused using a splicing mechanism, specifically:

[0025]

[0026] In the formula, C s is the multi-view consensus sparse representation relation matrix, is the weighted sparse representation relationship matrix of each view, and V is the number of views.

[0027] Preferably, obtaining the final clustering result includes:

[0028] The multi-view consensus sparse representation relationship matrix is ​​subjected to truncated singular value decomposition, a left singular vector matrix is ​​calculated, and k-means is used to complete multi-view data clustering based on the left singular vector matrix to obtain the final clustering result.

[0029] Preferably, the method for calculating the left singular vector matrix is:

[0030] make

[0031] In the formula, C s is the multi-view consensus sparse representation relationship matrix, and the diagonal elements of Λ are C s The sum of each column of ;

[0032] right Perform SVD decomposition:

[0033] In the formula, is the left singular vector matrix; is the singular value matrix; is the right singular vector matrix.

[0034] Compared with the prior art, the present invention has the following advantages and technical effects:

[0035] The present invention breaks through the limitation of high time and space overhead of deep multi-view subspace methods based on self-representation characteristics, reduces the time and space complexity of existing methods with respect to the quadratic and cubic powers of the number of samples to a linear overhead, effectively realizes efficient clustering of multi-view data, and provides a simple, easy-to-implement, and scalable network model for the application of deep multi-view subspace clustering methods on large-scale data. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0037] Figure 1 This is a pre-training network framework diagram of an embodiment of the present invention;

[0038] Figure 2 A fine-tuning network framework diagram of an embodiment of the present invention;

[0039] Figure 3 4 is a block diagram of a self-weighted network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0041] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0042] The present invention specifically implements the fast deep multi-view subspace clustering method based on dynamic anchor points, which consists of three stages: pre-training, fine-tuning and fast spectral clustering. First, the pre-training stage preliminarily updates the network parameters to preliminarily obtain the potential features and anchor features of each view; then, the fine-tuning network is initialized using the network parameters in the pre-training stage, and the sparse representation relationship matrix of each view is obtained through end-to-end training. The self-weighting network is used to assign discriminant weights to it and perform splicing and fusion; finally, fast spectral clustering is used to obtain the final clustering result.

[0043] A fast deep multi-view subspace clustering method based on dynamic anchors, such as Figure 1 ,include:

[0044] Obtain multi-view sample data, input the multi-view sample data into a pre-trained network model for reconstruction and processing, preliminarily obtain the anchor features of each view, and update the autoencoder and dynamic anchor network parameters, wherein the pre-trained network model is composed of a deep autoencoder and a dynamic anchor learning network;

[0045] Based on the pre-trained network model, a fine-tuning network is constructed to learn the sparse representation relationship matrix of each view, and a self-weight network is used to adaptively assign weights to the sparse representation relationship matrix of each view to obtain a weighted sparse representation relationship matrix of each view, and the weighted sparse representation relationship matrix of each view is fused to obtain a multi-view consensus sparse representation relationship matrix;

[0046] The final clustering result is obtained by using a fast spectral clustering method on the multi-view consensus sparse representation relationship matrix.

[0047] Furthermore, obtaining multi-view sample data includes:

[0048] A multi-view public data set is obtained, and multi-view data with diverse sample numbers, diverse categories and rich details are selected. Then, the matrix type data in the multi-view data is normalized, and all the multi-view data are processed into a batch of data to obtain multi-view sample data.

[0049] Initially obtain the anchor features of each view and update the autoencoder and dynamic anchor network parameters, including:

[0050] Reconstruct the sample data through deep autoencoders to obtain effective latent feature representations for each view;

[0051] Based on the predefined anchor features, the dynamic anchor learning network is used to reversely map and reconstruct the effective potential feature representation. After end-to-end training, the anchor features of each view are obtained, and the parameters of the autoencoder and the dynamic anchor network are updated.

[0052] Furthermore, a fine-tuning network is constructed to learn the sparse representation relationship matrix of each view, including:

[0053] Design a fully connected sparse representation layer based on anchor features and effective potential feature representation; embed the fully connected sparse representation learning layer between the deep autoencoder and the dynamic anchor learning network to construct a fine-tuning network;

[0054] The view anchor features are dynamically updated by fine-tuning the network and learning the sparse representation relationship matrix of each view.

[0055] Obtain the weighted sparse representation relationship matrix of each view, including:

[0056] A sparse representation tensor is constructed based on the sparse representation relationship matrix of each view, and the sparse representation tensor is compressed along the channel dimension using global average pooling to obtain the global description factor of each channel, i.e., each sparse representation relationship matrix;

[0057] For the global description factor of each sparse representation relationship matrix, a two-layer unbiased fully connected network is used to capture the relationship between the view self-representation coefficient matrix and stimulate it to a weight between 0 and 1;

[0058] Re-scaling the sparse representation tensor using the weights to obtain a weighted sparse representation tensor;

[0059] The self-weighted sparse representation tensor is split along the channel dimension to obtain a weighted sparse representation relationship matrix of each view.

[0060] The weighted sparse representation relationship matrix of each view is fused using the splicing mechanism, specifically:

[0061]

[0062] In the formula, C s is the multi-view consensus sparse representation relation matrix, is the weighted sparse representation relationship matrix of each view, and V is the number of views.

[0063] Furthermore, the truncated singular value decomposition is performed on the multi-view consensus sparse representation relationship matrix, the left singular vector matrix is ​​calculated, and the multi-view data clustering is completed based on the left singular vector matrix using k-means to obtain the final clustering result;

[0064] Among them, the method for calculating the left singular vector matrix is:

[0065] make

[0066] In the formula, C s is the multi-view consensus sparse representation relationship matrix, and the diagonal elements of Λ are C s The sum of each column of ;

[0067] right Perform SVD decomposition:

[0068] In the formula, is the left singular vector matrix; is the singular value matrix; is the right singular vector matrix.

[0069] In order to more clearly express the technical solution of the present invention, the following specific embodiments are provided to introduce the solution:

[0070] A fast deep multi-view subspace clustering method based on dynamic anchors, which mainly includes five parts: the first part is to preprocess the data set; the second part is to pre-train the deep autoencoder and the dynamic anchor network to extract multi-view data features and anchor features; the third part is to build a fine-tuning network based on the second part to learn the sparse representation relationship between multi-view data; the fourth part uses the self-weight network to adaptively assign discriminant weights to the sparse representation relationship matrix of each view, and perform splicing and fusion; the fifth part is to use fast spectral clustering to perform data clustering.

[0071] Specifically:

[0072] Part 1: Preprocessing the dataset:

[0073] Step 1: Download the multi-view public dataset, select multi-view data with diverse sample numbers, diverse categories and rich details, then normalize the matrix type data and process all the multi-view data into a batch data. Assume that the multi-view data There are V views and N samples, where d v is the data dimension of the vth view.

[0074] Part 2: Pre-training deep autoencoders and dynamic anchor networks to extract multi-view data features and anchor features:

[0075] Step 2: Build an arbitrary view encoder E v (·) and decoder D v (·) network, minimizing the encoder input and decoder output The reconstruction loss between them is used to obtain the latent feature representation Z of each view. v =E v (X v );

[0076] Step 3: For all views, a dynamic anchor learning network f(·) is constructed using a fully connected neural network. The predefined anchor features of each view are transformed into Reverse mapping encoding into latent feature representation Where K is the number of anchor features, and K<<N, Θ a Learning network parameters for dynamic anchors, is the anchor feature corresponding to any view, T is the transposition function, and multiple view anchor features are updated by minimizing the potential feature reconstruction loss.

[0077] Step 4: Input the multi-view data into the pre-trained network ( Figure 1 As shown in the figure), by minimizing the dual reconstruction loss of the deep autoencoder and the dynamic anchor learning network, as shown in formula (1), the potential features and anchor features of each view are preliminarily obtained, the pre-trained network parameters are updated, and the pre-trained model is saved;

[0078]

[0079] In the formula, is the pre-trained network loss function, are the encoder network parameters, are the decoder network parameters.

[0080] Part 3: Build a fine-tuning network based on Part 2 to learn the sparse representation relationship between multi-view data:

[0081] Figure 2To fine-tune the network framework of the present invention, this stage is carried out according to the following steps in this embodiment:

[0082] Step 5: Fine-tune the network ( Figure 2 As shown in the figure, based on the pre-trained network model, for each view, the potential feature Z output by its encoder is v With its anchor feature A v A fully connected sparse representation relationship learning layer is embedded between them to learn the sparse representation relationship matrix between sample data.

[0083] Part 4: Use the self-weight network to adaptively assign discriminant weights to the sparse representation relationship matrix of each view, and perform splicing and fusion:

[0084] Step 6: Based on the discriminant contribution of each view, use the self-weight network to adaptively assign weights to the sparse representation relationship matrix of each view. Figure 3 The network diagram is shown below. The specific implementation details are as follows:

[0085] Step 6-1: Sparsely represent the coefficient matrix of multiple views Stacked along the channel dimension into a sparse representation tensor

[0086] Step 6-2: Use global average pooling to sparsely represent the tensor along the channel dimension Perform compression operation to obtain the global description factor of each channel, i.e., each sparse representation relationship matrix;

[0087] Step 6-3: For each global description factor, a two-layer unbiased fully connected network is used to capture the relationship between the view self-representation coefficient matrix and stimulate it to a weight between 0 and 1

[0088] Step 6-4: Rescale the sparse representation tensor using weight U Get the sparse representation tensor after self-weighting

[0089] Step 6-5: Split the sparse representation tensor after self-weighting along the channel dimension Get the weighted sparse representation relationship matrix of each view

[0090] Step 7: Use the pre-trained model to initialize the fine-tuning network, and input the multi-view data into the fine-tuning network. Train the fine-tuning network end-to-end through the loss function shown in formula (2) to obtain the weighted sparse representation relationship matrix of multiple views

[0091]

[0092] In the formula, is the loss function of the fine-tuned network, λ1 and λ2 are the balancing parameters;

[0093] Step 8: Use the splicing mechanism to Perform fusion, see formula (3), and obtain the multi-view consensus sparse representation relationship matrix;

[0094]

[0095] Part V: Data clustering using fast spectral clustering:

[0096] Step 9: Calculation for The truncated singular value decomposition is used to obtain its left singular value vector matrix, where the diagonal elements of Λ are C s The sum of each column of .

[0097] Step 10: Based on the left singular value vector matrix, use the k-means algorithm to obtain the pseudo labels corresponding to the multi-view samples and obtain the final clustering results.

[0098] In order to solve the problem that the existing deep multi-view subspace clustering is limited in application on large-scale data, the present invention breaks through the limitation of the self-representation characteristics of the data subspace, avoids the shortcomings of the sampling anchor point being unrepresentative of the data and the additional time overhead caused by the k-means anchor point, and proposes a scalable deep multi-view subspace clustering method based on dynamic anchor point learning to overcome the defect of separating the selection of fixed anchor points from the clustering process. This method is a new attempt to study and develop efficient deep multi-view subspace clustering, which expands the application on large-scale data. It is simple to implement and easy to adjust, has strong operability, and has wide applicability. It can be further expanded based on this general framework in the future, combined with the actual application background, by adding different constraints and fusion strategies to meet the actual production and life needs.

[0099] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A fast deep multi-view subspace clustering method based on dynamic anchor points, characterized in that: include: Obtain multi-view sample data, input the multi-view sample data into a pre-trained network model for reconstruction and processing, preliminarily obtain the anchor features of each view, and update the autoencoder and dynamic anchor network parameters, wherein the pre-trained network model is composed of a deep autoencoder and a dynamic anchor learning network; Based on the pre-trained network model, a fine-tuning network is constructed to learn the sparse representation relationship matrix of each view, and a self-weight network is used to adaptively assign weights to the sparse representation relationship matrix of each view to obtain a weighted sparse representation relationship matrix of each view, and the weighted sparse representation relationship matrix of each view is fused to obtain a multi-view consensus sparse representation relationship matrix; The final clustering result is obtained by using a fast spectral clustering method on the multi-view consensus sparse representation relationship matrix.

2. The fast deep multi-view subspace clustering method based on dynamic anchor points according to claim 1, characterized in that: Obtaining the multi-view sample data includes: A multi-view public data set is obtained, and multi-view data with diverse sample numbers, diverse categories and rich details are selected. Then, the matrix type data in the multi-view data is normalized, and all the multi-view data are processed into a batch of data to obtain the multi-view sample data.

3. The fast deep multi-view subspace clustering method based on dynamic anchor points according to claim 1, characterized in that: Initially obtain the anchor features of each view and update the autoencoder and dynamic anchor network parameters, including: Reconstructing the sample data by the deep autoencoder to obtain an effective potential feature representation of each view; Based on the predefined anchor features, the dynamic anchor learning network is used to reversely map and reconstruct the effective potential feature representation, and after end-to-end training, the anchor features of each view are obtained, and the parameters of the autoencoder and the dynamic anchor network are updated.

4. The fast deep multi-view subspace clustering method based on dynamic anchor points according to claim 1 or 3, characterized in that: Construct the fine-tuning network and learn the sparse representation relationship matrix of each view, including: Designing a fully connected sparse representation learning layer based on the anchor point features and the effective potential feature representation; Embedding the fully connected sparse representation learning layer between the deep autoencoder and the dynamic anchor learning network to construct the fine-tuning network; The view anchor features are dynamically updated through the fine-tuning network and the sparse representation relationship matrix of each view is learned.

5. The fast deep multi-view subspace clustering method based on dynamic anchor points according to claim 1, characterized in that: Obtaining the weighted sparse representation relationship matrix of each view includes: A sparse representation tensor is constructed based on the sparse representation relationship matrix of each view, and the sparse representation tensor is compressed along the channel dimension using global average pooling to obtain a global description factor of each channel, i.e., each sparse representation relationship matrix; For the global description factor of each sparse representation relationship matrix, a two-layer unbiased fully connected network is used to capture the relationship between the view self-representation coefficient matrices and stimulate them to a weight between 0 and 1; Re-scaling the sparse representation tensor using the weight to obtain a weighted sparse representation tensor; The self-weighted sparse representation tensor is split along the channel dimension to obtain the weighted sparse representation relationship matrix of each view.

6. The fast deep multi-view subspace clustering method based on dynamic anchor points according to claim 5, characterized in that: The weighted sparse representation relationship matrices of the views are fused using a splicing mechanism, specifically: In the formula, C s is the multi-view consensus sparse representation relation matrix, is the weighted sparse representation relationship matrix of each view, and V is the number of views.

7. The fast deep multi-view subspace clustering method based on dynamic anchor points according to claim 1, characterized in that: Obtaining the final clustering result includes: The multi-view consensus sparse representation relationship matrix is ​​subjected to truncated singular value decomposition, a left singular vector matrix is ​​calculated, and k-means is used to complete multi-view data clustering based on the left singular vector matrix to obtain the final clustering result.

8. The fast deep multi-view subspace clustering method based on dynamic anchor points according to claim 7, characterized in that: The method for calculating the left singular vector matrix is: make In the formula, C s is the multi-view consensus sparse representation relationship matrix, and the diagonal elements of Λ are C s The sum of each column of ; right Perform SVD decomposition: In the formula, is the left singular vector matrix; is the singular value matrix; is the right singular vector matrix.

Images