A Fast Depth Multi-View Subspace Clustering Method Based on Dynamic Anchors

Abstract

The present invention discloses a fast deep multi-view subspace clustering method based on dynamic anchors, including: obtaining multi-view sample data, inputting the multi-view sample data into a pre-trained network model for reconstruction and processing, initially obtaining anchor point features of each view, and updating the parameters of the autoencoder and the dynamic anchor network; based on the pre-trained network model, constructing a fine-tuning network, learning the sparse representation relationship matrix of each view, using a self-weight network to adaptively assign weights to the sparse representation relationship matrix of each view, obtaining the weighted sparse representation relationship matrix of each view, fusing the weighted sparse representation relationship matrix of each view, obtaining a multi-view consensus sparse representation relationship matrix, and obtaining a final clustering result through a fast spectral clustering method. The present invention reduces the quadratic and cubic space-time complexities of the existing method with respect to the number of samples to linear overhead, realizing efficient clustering of multi-view data.

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 anchors. Background Art

[0002] In the "data oil era", mining important information from massive data to make data valuable can promote the organic integration of artificial intelligence, big data, economic and social development. As a typical data mining method, clustering analysis is a key technology for processing unlabeled data 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, non-linearity, and high complexity. And with the diversification of information collection means and data types, data of the same object can usually be obtained by different sensors, media systems, cameras, feature extractors, etc., constituting multi-view data with characteristics such as polymorphism, multi-source, multi-feature, and high-dimensional non-linearity. The high-dimensional non-linear complex multi-view data makes traditional clustering methods no longer meet the actual application requirements and also poses new challenges to clustering methods based on linear mapping functions or shallow models. With the rapid development of artificial intelligence technology, by leveraging the powerful non-linear representation ability and effective feature extraction ability of deep learning, applying deep learning to clustering tasks has achieved good application effects. However, due to its "black box" operation, it has been widely criticized. Combining subspace clustering, an effective high-dimensional data processing technology with a solid theoretical foundation such as linear algebra and signal processing, can make up for the lack of its 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, removing redundant information and noise in the data to a certain extent; then based on the feature space, by constructing a self-representation learning network, a multi-view consensus self-representation coefficient matrix with consistency and complementarity is learned to reveal the representation relationship between data samples; finally, based on the consensus self-representation coefficient matrix, a similarity matrix is calculated, and spectral clustering is used to complete the clustering of multi-view data. Although existing deep multi-view subspace clustering algorithms have provided some solutions for mining multi-view Figure 1 consistency and complementarity and achieved excellent clustering performance, these methods are based on the self-representation property, use the feature representations of all data samples to learn the multi-view consensus self-representation coefficient matrix, construct a similarity matrix through the self-representation coefficient matrix, and perform spectral clustering to complete the final data division. Their space and time complexities are respectively and which 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. By applying bipartite graph learning and sparse representation ideas in graph theory, a sparse representation relationship between data anchors and data samples is constructed, avoiding the high memory requirements brought by learning the self-representation relationship of data. However, such methods construct the multi-view sparse representation relationship based on a linear mapping function, and are not very applicable to non-linear data, and there are also the following defects: (1) Using random sampling, uniform sampling, etc. to obtain data anchors from multi-view data is not representative of the source data, which will cause the loss of data structure information, making the anchor-based sparse representation relationship matrix unable to fully reflect the similarity relationship between multi-view data; (2) The anchor samples obtained by using the k-means clustering algorithm can effectively represent multi-view data, but the anchors selected by k-means are easily mixed with noise based on the local information of the sample data space, and at the same time will additionally increase the time complexity of the algorithm; (3) Moreover, in most traditional multi-view subspace clustering methods based on anchors, the anchors 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 impact on the clustering performance; (4) In addition, after obtaining the sparse representation relationship matrix, most traditional multi-view subspace clustering methods based on anchors still use general spectral clustering to implement the clustering of data sample clusters, and still have 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 anchors, aiming to improve the efficiency of the deep multi-view subspace clustering method and expand its application to large-scale data, break through the limitations of the self-representation characteristics of data subspaces, avoid the disadvantages of non-representative sampling anchors and the additional time overhead brought by k-means anchors, and overcome the defect of the separation of the selection of fixed anchors 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 anchors, including:

[0007] Obtain multi-view sample data, input the multi-view sample data into a pre-trained network model for reconstruction and processing, initially obtain the anchor features of each view, and update the autoencoder and dynamic anchor network parameters, where 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, construct a fine-tuning network, learn the sparse representation relationship matrix of each view, use the self-weight network to adaptively assign weights to the sparse representation relationship matrix of each view, obtain the weighted sparse representation relationship matrix of each view, and fuse the weighted sparse representation relationship matrix of each view to obtain a multi-view consensus sparse representation relationship matrix;

[0009] The final clustering result is obtained for the multi-view consensus sparse representation relationship matrix through the fast spectral clustering method.

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

[0011] Obtain a multi-view public dataset, select multi-view data with diverse sample quantities, diverse categories, and rich details, then perform normalization processing on the matrix-type data in the multi-view data, and process all the multi-view data into a batch of data to obtain the multi-view sample data.

[0012] Preferably, initially obtaining the anchor point features of each view and updating the autoencoder and dynamic anchor point network parameters includes:

[0013] Reconstruct the sample data through the deep autoencoder to obtain the effective latent feature representation of each view;

[0014] Based on the predefined anchor point features, use the dynamic anchor point learning network to inversely map and reconstruct the effective latent feature representation, and through end-to-end training, obtain the anchor point features of each view and update the autoencoder and dynamic anchor point network parameters.

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

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

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

[0018] Dynamically update the view anchor point features through the fine-tuning network and learn the sparse representation relationship matrix of each view.

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

[0020] Construct a sparse representation tensor based on the sparse representation relationship matrix of each view, and perform a compression operation on the sparse representation tensor along the channel dimension using global average pooling to obtain the global description factor of each channel, that is, the sparse representation relationship matrix;

[0021] For the global description factors of the sparse representation relationship matrices of each view, use a two-layer unbiased fully connected network to capture the relationship between the view self-representation coefficient matrices and stimulate their weights to be between 0 and 1;

[0022] Use the weights to re-measure the sparse representation tensor to obtain the self-weighted sparse representation tensor;

[0023] Split the sparse representation tensor after self-weight along the channel dimension to obtain the weighted sparse representation relationship matrix of each view.

[0024] Preferably, use the splicing mechanism to fuse the weighted sparse representation relationship matrices of each view, specifically:

[0025]

[0026] In the formula, C s is the multi-view consensus sparse representation relationship 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] Perform truncated singular value decomposition on the multi-view consensus sparse representation relationship matrix, calculate the left singular vector matrix, and complete the multi-view data cluster division 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] Let

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

[0032] Perform SVD decomposition on :

[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 the large time and space overhead of the deep multi-view subspace method based on self-representation characteristics, reduces the time and space complexity of the existing method from quadratic and cubic powers of the sample number to linear overhead, effectively realizes the efficient clustering of multi-view data, and provides a simple and easy-to-implement scalable network model for the application of the deep multi-view subspace clustering method on large-scale data. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings, which form a part of this application, are used to provide a further understanding of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation of this application. In the drawings:

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

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

[0039] Figure 3 It is a self-weight network block diagram of an embodiment of the present invention. Detailed implementation manners

[0040] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments may be combined with each other. The following will refer to the accompanying drawings and combine the embodiments to detail this application.

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

[0042] The specific implementation of the fast deep multi-view subspace clustering method based on dynamic anchors of the present invention consists of three stages: pre-training, fine-tuning, and fast spectral clustering. First, in the pre-training stage, the network parameters are initially updated, and the potential features and anchor features of each view are initially obtained; then, the fine-tuning network is initialized with the network parameters of the pre-training stage, and the sparse representation relationship matrix of each view is obtained through end-to-end training. The self-weight network assigns discriminant weights to it and performs splicing and fusion; finally, the fast spectral clustering is used to obtain the final clustering result.

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

[0044] Obtain multi-view sample data, input the multi-view sample data into the pre-training network model for reconstruction and processing, initially obtain the anchor features of each view, and update the autoencoder and dynamic anchor network parameters, where the pre-training network model consists of a deep autoencoder and a dynamic anchor learning network;

[0045] Based on the pre-training network model, construct a fine-tuning network, learn the sparse representation relationship matrix of each view, use the self-weight network to adaptively assign weights to the sparse representation relationship matrix of each view, obtain the weighted sparse representation relationship matrix of each view, and fuse the weighted sparse representation relationship matrices of each view to obtain a multi-view consensus sparse representation relationship matrix;

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

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

[0048] Obtain a multi-view public dataset, select multi-view data with diverse sample quantities, diverse categories, and rich details, then normalize the matrix-type data in the multi-view data, and process all the multi-view data into a batch of data to obtain multi-view sample data.

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

[0050] Reconstruct the sample data through a deep autoencoder to obtain the effective latent feature representation of each view;

[0051] Based on the predefined anchor point features, use the dynamic anchor point learning network to inversely map and reconstruct the effective latent feature representation, and through end-to-end training, obtain the anchor point features of each view and update the parameters of the autoencoder and the dynamic anchor point network.

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

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

[0054] Dynamically update the view anchor point features through the fine-tuning network and learn the sparse representation relationship matrix of each view.

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

[0056] Construct a sparse representation tensor based on the sparse representation relationship matrix of each view, and use global average pooling to perform a compression operation on the sparse representation tensor along the channel dimension to obtain the global description factor of each channel, that is, the global description factor of each sparse representation relationship matrix;

[0057] For the global description factors of the sparse representation relationship matrices of each view, use a two-layer fully connected network without bias to capture the relationship between the view self-representation coefficient matrices and stimulate their weights to be between 0 and 1;

[0058] Re-measure the sparse representation tensor using the weights to obtain the self-weighted sparse representation tensor;

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

[0060] Fuse the weighted sparse representation relationship matrices of each view by using the splicing mechanism, specifically as follows:

[0061]

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

[0063] Furthermore, perform truncated singular value decomposition on the multi-view consensus sparse representation relationship matrix, calculate the left singular vector matrix, and complete the multi-view data cluster division based on the left singular vector matrix by using k-means to obtain the final clustering result;

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

[0065] Let

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

[0067] Perform SVD decomposition on :

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

[0069] To more clearly express the technical solution of the present invention, the following provides specific embodiments for scheme introduction:

[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 a deep autoencoder and a dynamic anchor network to extract multi-view data features and anchor features; the third part is to construct a fine-tuning network based on the second part to learn the sparse representation relationship between multi-view data; the fourth part uses a self-weight network to adaptively assign discriminant weights to the sparse representation relationship matrices of each view and perform splicing fusion; the fifth part is to use fast spectral clustering for data cluster division.

[0071] Specifically:

[0072] The first part, preprocess the data set:

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

[0074] Part Two: Pre-train the deep autoencoder and the dynamic anchor network to extract multi-view data features and anchor features:

[0075] Step 2: Construct an arbitrary view encoder E v (·) and decoder D v (·) networks, minimize the reconstruction loss between the encoder input and the decoder output to obtain the latent feature representation Z of each view v = E v (X v );

[0076] Step 3: For all views, construct a dynamic anchor learning network f(·) using a fully connected neural network; map the predefined anchor features of each view backward through the dynamic anchor learning network to encode them as latent feature representations where K is the number of anchor features, and K << N, Θ a is the parameter of the dynamic anchor learning network, is the anchor feature corresponding to an arbitrary view, T is the transpose function, and update the anchor features of multiple views by minimizing the latent feature reconstruction loss.

[0077] Step 4: Input the multi-view data into the pre-trained network ( Figure 1 as shown), and initially obtain the latent features and anchor features of each view by minimizing the double reconstruction losses of the deep autoencoder and the dynamic anchor learning network, as shown in formula (1). Update the parameters of the pre-trained network and save the pre-trained model;

[0078]

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

[0080] Part Three: Based on Part Two, construct a fine-tuning network to learn the sparse representation relationship between multi-view data:

[0081] Figure 2This is the fine-tuning network framework of the present invention. In this implementation, it is carried out according to the following steps:

[0082] Step 5: The fine-tuning network ( Figure 2 as shown) is based on the pre-trained network model. For each view, between the latent feature Z output by its encoder v and its anchor feature A v embed a fully connected sparse representation relationship learning layer to learn the sparse representation relationship matrix between sample data

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

[0084] Step 6: According to the discriminant contributions of each view, use the self-weight network to adaptively assign weights to the sparse representation relationship matrices of each view. Figure 3 For its network block diagram, the specific implementation details are as follows:

[0085] Step 6-1: Stack multiple view sparse representation coefficient matrices along the channel dimension into a sparse representation tensor

[0086] Step 6-2: Use global average pooling to compress the sparse representation tensor along the channel dimension to obtain the global description factor of each channel, that is, each sparse representation relationship matrix;

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

[0088] Step 6-4: Use the weight U to re-measure the sparse representation tensor to obtain the self-weighted sparse representation tensor

[0089] Step 6-5: Split the self-weighted sparse representation tensor along the channel dimension to obtain the weighted sparse representation relationship matrices of each view

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

[0091]

[0092] In the formula, is the loss function for fine-tuning the network, and λ1 and λ2 are balance parameters;

[0093] Step 8: Use the splicing mechanism to perform fusion, as shown in formula (3), to obtain the multi-view consensus sparse representation relationship matrix;

[0094]

[0095] Part Five: Use fast spectral clustering to perform data cluster division:

[0096] Step 9: Calculate For perform truncated singular value decomposition to obtain its left singular value vector matrix, where the diagonal elements of Λ are the sum of each column of C s of C.

[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 result.

[0098] The present invention aims to solve the problem that the existing deep multi-view subspace clustering is limited in application to large-scale data, breaks through the limitation of the self-representation characteristics of the data subspace, avoids the disadvantages such as the lack of data representativeness of the sampling anchor points and the additional time overhead brought by the k-means anchor points, and in order to overcome the defect that the selection of fixed anchor points is separated from the clustering process, a scalable deep multi-view subspace clustering method based on dynamic anchor point learning is proposed. This method is a brand-new attempt to study and develop efficient deep multi-view subspace clustering, expands the application to large-scale data, is simple to implement and easy to adjust, has strong operability, has wide applicability, and can be further extended based on this general framework in the future. Combining the actual application background, by adding different constraints and fusion strategies, it can meet the needs of actual production and life.

[0099] The above is only a preferred specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in this application should be covered by the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

Claims

1. A fast depth multi-view subspace clustering method based on dynamic anchors, characterized in that Including: 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 point features of each view, and update the parameters of the autoencoder and the dynamic anchor point network, where the pre-trained network model consists of a deep autoencoder and a dynamic anchor point learning network; Based on the pre-trained network model, construct a fine-tuning network to learn the sparse representation relationship matrix of each view, use a self-weight network to adaptively assign weights to the sparse representation relationship matrix of each view, obtain the weighted sparse representation relationship matrix of each view, and fuse the weighted sparse representation relationship matrix of each view to obtain a multi-view consensus sparse representation relationship matrix; Obtain the final clustering result for the multi-view consensus sparse representation relationship matrix through a fast spectral clustering method; Among them, preliminarily obtaining the anchor point features of each view and updating the parameters of the autoencoder and the dynamic anchor point network includes: Reconstruct the sample data through the deep autoencoder to obtain an effective latent feature representation of each view; Based on predefined anchor point features, use the dynamic anchor point learning network to reverse-map and reconstruct the effective latent feature representation, and through end-to-end training, obtain the anchor point features of each view and update the parameters of the autoencoder and the dynamic anchor point network.

2. The fast depth multi-view subspace clustering method based on dynamic anchors according to claim 1, wherein Obtaining the multi-view sample data includes: Obtain a multi-view public dataset, select multi-view data with diverse sample quantities, diverse categories, and rich details, then normalize the matrix type data in the multi-view data, and process all the multi-view data into a batch of data to obtain the multi-view sample data.

3. The fast depth multi-view subspace clustering method based on dynamic anchors according to claim 1, wherein, Constructing the fine-tuning network to learn the sparse representation relationship matrix of each view includes: Design a fully connected sparse representation learning layer based on the anchor point features and the effective latent feature representation; Embed the fully connected sparse representation learning layer between the deep autoencoder and the dynamic anchor point learning network to construct the fine-tuning network; Dynamically update the view anchor point features through the fine-tuning network and learn the sparse representation relationship matrix of each view.

4. The fast depth multi-view subspace clustering method based on dynamic anchors according to claim 1, characterized in that Obtaining the weighted sparse representation relationship matrix of each view includes: Construct a sparse representation tensor based on the sparse representation relationship matrix of each view, use global average pooling to perform a compression operation on the sparse representation tensor along the channel dimension to obtain the global description factor of each channel, that is, the sparse representation relationship matrix; For the global description factors of the sparse representation relationship matrices of each view, use a two-layer unbiased fully connected network to capture the relationship between the view self-representation coefficient matrices and activate their weights to between 0 and 1; Use the weights to re-measure the sparse representation tensor to obtain a self-weighted sparse representation tensor; Split the self-weighted sparse representation tensor along the channel dimension to obtain the weighted sparse representation relationship matrix of each view.

5. The fast depth multi-view subspace clustering method based on dynamic anchors according to claim 4, wherein Fuse the weighted sparse representation relationship matrix of each view using a splicing mechanism, specifically: In the formula, is the multi-view consensus sparse representation relationship matrix, is the weighted sparse representation relationship matrix of each view, and N is the number of views.

6. The fast depth multi-view subspace clustering method based on dynamic anchors according to claim 1, wherein, Obtaining the final clustering result includes: Perform truncated singular value decomposition on the multi-view consensus sparse representation relationship matrix, calculate the left singular vector matrix, and complete multi-view data clustering based on the left singular vector matrix using k-means to obtain the final clustering result.

7. The fast depth multi-view subspace clustering method based on dynamic anchors according to claim 6, characterized in that, The method for calculating the left singular vector matrix is as follows: Let ; In the formula, is the multi-view consensus sparse representation relationship matrix, the diagonal elements of which are the sum of each column of Pair Perform SVD decomposition on: ; In the formula, is the left singular vector matrix; is the singular value matrix; is the right singular vector matrix.

Images