A deep multi-view clustering method and system based on co-training

Abstract

The application provides a deep multi-view clustering method and system based on collaborative training, relates to the fields of data mining and machine learning, and specifically includes the following steps: obtaining multi-views to be clustered, and pre-training a deep self-encoding model for each view; calculating an affinity matrix of each view, setting an initial weight of each view and an initial dynamic learning factor between the views; performing iterative formal training on the deep self-encoding model of each view, updating the weight of each view and the dynamic learning factor between the views until a preset iteration stopping condition is met, and outputting final hidden layer features and clustering centers; and obtaining a clustering result based on the final hidden layer features and the clustering centers; the application designs a deep multi-view weighted graph embedding clustering algorithm based on dynamic collaborative training, adopts a dynamic collaborative training idea for multi-view data, deeply mines complementary information between the multi-views, and improves the clustering effect for multi-view data sets.

Description

Technical Field

[0001] This invention belongs to the field of data mining and machine learning, and particularly relates to a deep multi-view clustering method and system based on collaborative training. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Clustering analysis, as an efficient unsupervised processing method, plays an important role in data mining and machine learning, and is widely used in fields such as machine learning, computer vision, data mining, and pattern recognition. Traditional clustering algorithms perform well on small datasets, but are only suitable for low-dimensional data. With the rapid development of science and technology and the rapid expansion of data scale, traditional clustering algorithms have encountered bottlenecks when dealing with high-dimensional data scenarios. How to effectively handle high-dimensional data has become a major challenge in clustering problems.

[0004] In response to this scenario, the emerging deep learning framework has come into the researchers' view with its powerful feature representation and nonlinear fitting capabilities. Deep neural networks can effectively connect the feature representation and the clustering target. The combination of deep learning and clustering algorithms is essentially a clustering method that improves performance by leveraging the powerful representation capabilities of deep learning. This approach effectively solves the problems of convergence difficulty and low accuracy of traditional clustering in high-dimensional, massive data scenarios.

[0005] For example: information data with different language translations, rich text information in different forms, and human motion capture data collected by cameras from different angles.

[0006] Furthermore, with the rapid development of science and technology, multi-faceted information about the same object has become an increasingly common phenomenon in real life. Figure 1 This is an example diagram of a multi-view dataset, such as... Figure 1 As shown, multi-view data, compared to single-view data, possesses complementarity and consistency, providing richer and more comprehensive information, thus becoming a popular research topic. Multi-view clustering, as an analytical method, can merge data from different views for cluster analysis to reveal correlations between different views and the inherent structure of the data. Among numerous multi-view clustering algorithms, deep multi-view clustering algorithms based on co-training have become a research hotspot in this subfield due to their concise and effective approach and excellent algorithmic performance.

[0007] Current mainstream deep multi-view algorithms tend to focus more on the information interaction between different views when processing multi-view data, while ignoring the unique information of each view itself. In this case, the performance of the algorithm often fails to reach its optimal level. At the same time, since the learning strategies between views are fixed, the algorithm cannot make full use of the advantages of collaborative training, which limits the algorithm's performance. Summary of the Invention

[0008] To overcome the shortcomings of the existing technologies, and addressing the current mainstream deep multi-view clustering algorithms based on collaborative training's difficulty in focusing on the unique information of the view itself, as well as the overly coarse-grained control of the learning process in collaborative training, this invention provides a deep multi-view clustering method and system based on collaborative training. It designs a deep multi-view weighted graph embedding clustering algorithm based on dynamic collaborative training. For multi-view data, it adopts the idea of ​​dynamic collaborative training to deeply mine the complementary information between multiple views, thereby improving the clustering effect for multi-view datasets.

[0009] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0010] The first aspect of this invention provides a deep multi-view clustering method based on collaborative training.

[0011] A deep multi-view clustering method based on collaborative training includes:

[0012] Obtain multiple views to be clustered, and pre-train your own deep autoencoder model for each view;

[0013] Calculate the affinity matrix for each view, and set the initial weights for each view and the initial dynamic learning factors between views;

[0014] Based on the affinity matrix, initial weights, and initial dynamic learning factors between views for each view, the deep autoencoder model for each view is iteratively trained, updating the weights of each view and the dynamic learning factors between views until the preset iteration stopping condition is met, and the final hidden features and cluster centers are output.

[0015] Based on the final hidden layer features and cluster centers, the clustering results are obtained;

[0016] The deep autoencoder model takes samples from the view as input and hidden features and cluster centers as output. During pre-training, the target function is MSE reconstruction loss. During formal training, the target function is constructed based on the affinity matrix, weights, and dynamic learning factors between views for each view.

[0017] Furthermore, the deep autoencoder model includes an encoder, a decoder, and a clustering layer;

[0018] The encoder performs dimensionality reduction processing on the input view samples to obtain the hidden layer features of the view;

[0019] The decoder reconstructs new view samples based on the hidden layer features of the view;

[0020] The clustering layer uses k-means to obtain cluster centers based on the hidden features of the view.

[0021] Furthermore, the pre-trained objective function, used to minimize the error between the original samples and the reconstructed samples in each view, specifically includes:

[0022]

[0023] in, Let x represent the reconstruction loss term for the v-th view, where N represents the number of samples in view v, and x represents the number of samples in view v. i f represents the original sample. Θ Indicates encoder, g Ω Indicates decoder, g Ω (f Θ (x i ) represents the reconstructed sample.

[0024] Furthermore, the affinity matrix, based on the distance of the radial basis functions, quantifies the similarity between samples, and the specific formula is as follows:

[0025]

[0026] Where σ represents the bandwidth parameter of the radial basis function, NB i Indicates belonging to sample x i The nearest neighbor x j A set of.

[0027] Furthermore, the objective function for formal training consists of a reconstruction loss function, a clustering loss function, a graph embedding constraint loss function, and a weight entropy loss function, with the specific formula as follows:

[0028] L = L rec +γL clustering +ηL graph +ρL w

[0029]

[0030]

[0031]

[0032]

[0033] Among them, Lrec L clustering L graph L w These represent the reconstruction loss function, clustering loss function, graph embedding constraint loss function, and weight entropy loss function, respectively. Let represent the reconstruction loss term, clustering loss term, graph embedding constraint loss term, and weight entropy loss term of view v, respectively, and let γ, η, and ρ represent the hyperparameters of the clustering loss function, graph embedding constraint loss function, and weight entropy loss function, respectively.

[0034] Furthermore, the clustering term, used to maximize the similarity between the original student t-distribution and the sharpened auxiliary distribution, is specifically formulated as follows:

[0035]

[0036] Among them, Q v P represents the student t-distribution of view v. r Q represents view r v The auxiliary distribution after sharpening the distribution, ξ vr D represents the dynamic learning factor that view v learns from view r. KL This represents the KL divergence formula.

[0037] Furthermore, the graph embedding constraint term is used to constrain the proximity relationship between samples, and the specific formula is as follows:

[0038]

[0039] in, Let N represent the similarity between the i-th sample and the j-th sample in view v, where N represents the number of samples in view v, K is the number of cluster centers in view v, and L represents the number of neighbors of each sample. Let i and j represent the membership degrees of the i-th and j-th samples in view v, respectively.

[0040] Furthermore, the weight of each view is given by the following formula:

[0041]

[0042] Where, ω v This indicates the weight of view v. represents the clustering term and graph embedding constraint term of view v, respectively, and γ and η represent the hyperparameters of the clustering and graph embedding constraints, respectively.

[0043] Furthermore, the dynamic learning factor between the views is specifically formulated as follows:

[0044]

[0045]

[0046]

[0047] Where θ is a parameter greater than 0, used to control the speed of learning factor updates, β represents the measure of the value of the two views, and Q... v Q is the student distribution t of the current view v. r The student t-distribution of the target view r. This represents the global student t-distribution, used as a global consensus for all views.

[0048] A second aspect of the present invention provides a deep multi-view clustering system based on collaborative training.

[0049] A deep multi-view clustering system based on collaborative training includes a pre-training module, an initialization module, a training module, and a clustering module.

[0050] The pre-training module is configured to: acquire multiple views to be clustered, and pre-train its own deep autoencoder model for each view;

[0051] The initialization module is configured to: calculate the affinity matrix for each view, set the initial weights for each view, and set the initial dynamic learning factors between views;

[0052] The training module is configured to: perform iterative formal training on the deep autoencoder model of each view based on the affinity matrix of each view, the initial weights and the initial dynamic learning factors between views, update the weights of each view and the dynamic learning factors between views until the preset iteration stopping condition is met, and output the final hidden features and cluster centers.

[0053] The clustering module is configured to obtain clustering results based on the final hidden layer features and cluster centers;

[0054] The deep autoencoder model takes samples from the view as input and hidden features and cluster centers as output. During pre-training, the target function is MSE reconstruction loss. During formal training, the target function is constructed based on the affinity matrix, weights, and dynamic learning factors between views for each view.

[0055] The above one or more technical solutions have the following beneficial effects:

[0056] This invention provides prior knowledge of the nearest neighbor affinity for each view, ensuring that the algorithm focuses on global information while also utilizing local information; it introduces a dynamic learning factor, enabling the algorithm to fully satisfy the complementarity principle between views during the clustering task; to focus on the true value of each view, this invention assigns a weight to each view to reflect the differences in the view data itself; these optimization measures further improve the clustering effect for multi-view datasets.

[0057] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0058] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0059] Figure 1 Example diagram for a multi-view dataset.

[0060] Figure 2 This is a flowchart of the method in the first embodiment.

[0061] Figure 3 This is a schematic diagram illustrating the construction of the objective function in the first embodiment.

[0062] Figure 4 This is a system structure diagram of the second embodiment. Detailed Implementation

[0063] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0064] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0065] This invention relates to a deep multi-view clustering method based on collaborative training. Considering the current state of mainstream research, the technical problems encountered and their solutions are as follows:

[0066] (1) The core idea of ​​the collaborative training approach is to promote mutual learning among the views to achieve the optimization of the final result. However, this approach focuses too much on the differences between views, only considering the distance between the sample and the cluster center, while ignoring the nearest neighbor relationship between samples, i.e., the information unique to each view itself. In this case, the partitioning of samples at the cluster boundary will cause incorrect results, thus affecting the clustering performance. To solve this problem, this invention provides an additional affinity matrix based on the similarity between samples as the model input while performing clustering assignment. This allows the algorithm to focus on the overall sample clustering assignment task while also paying attention to the nearest neighbor information between samples.

[0067] (2) Considering that the descriptions of different views of the same object in a multi-view dataset contain their own unique information, but in most current algorithms, the degree to which a single view learns from other views is fixed, that is, the learning strategy between views is fixed and cannot be changed during the learning process. Therefore, it is impossible to fully leverage the advantages of collaborative training in practice. In addition, considering that it is difficult to take into account all the information of all views by fixing a global learning factor, a small value will lead to insufficient learning of the current source view, while a large value will lead to an excessive proportion of knowledge from other views, causing the target of the view itself to be ignored. This invention believes that in deep clustering, the collaborative learning between multiple views should intelligently adjust the learning rate according to the deepening of the learning process. To this end, a dynamic learning factor is designed to adjust the degree of learning of each view from other views. Specifically, a global consensus is designed as a benchmark for measuring the value of knowledge. The size of the learning factor in the current state is determined by comparing the source view and the target view with the global consensus, thereby realizing dynamic learning.

[0068] (3) Considering that although the dynamic learning strategy optimizes the learning method, it is not reliable in evaluating the value of the view. The algorithm needs to pay more attention to the data of each view itself and understand the differences between them. This can more accurately show the importance and value of each view. Therefore, this invention assigns a weight to each view. In the process of global collaborative training, the importance of each view to the overall clustering can be monitored at all times, and the promoting effect of reliable views on global clustering can be emphasized in the iteration process. In this process, by analyzing the contribution of each view, its characteristics and importance are determined, and the degree of tilt towards these important views is increased in the subsequent training process.

[0069] Example 1

[0070] In one or more embodiments, a deep multi-view clustering method based on collaborative training is disclosed, which has the characteristics of wide applicability and excellent performance; for multi-view data, a dynamic collaborative training approach is adopted to deeply mine the complementary information between multiple views. Figure 2It is the process of the method, such as Figure 2 As shown, it includes the following steps:

[0071] Step S1: Obtain multiple views to be clustered, and pre-train a deep autoencoder model for each view.

[0072] Read in a multi-view dataset X = {X_{V}} containing V views that needs to be clustered. 1 ,X 2 ,…,X V Each view is represented as} Where i represents the number of samples in view v.

[0073] During the pre-training phase, for each view in the multi-view dataset, a custom deep autoencoder model is constructed. The objective function, based on the MSE reconstruction loss, minimizes the error between the original samples and the reconstructed samples in each view during pre-training. The specific objective function is as follows:

[0074]

[0075] in, Let x represent the reconstruction loss term for the v-th view, where N represents the number of samples in view v, and x represents the number of samples in view v. i f represents the original sample. Θ Indicates encoder, g Ω Indicates decoder, g Ω (f Θ (x i ) represents the reconstructed sample.

[0076] Deep autoencoder models, including encoder f Θ decoder g Ω The deep autoencoder model takes view samples as input and outputs hidden features and clustering layers; the encoder performs dimensionality reduction on the input view samples to obtain the hidden features of the view; the decoder reconstructs new view samples based on the hidden features of the view; the clustering layer obtains cluster centers using k-means based on the hidden features of the view; the deep autoencoder model takes view samples as input and hidden features and cluster centers as output.

[0077] Specifically, the number of epochs in the pre-training phase is set to 300. During the pre-training process, the encoder f is used... Θ Dimensionality reduction of the entire multi-view dataset yields the hidden layer features Z = {Z 1 Z 2 ,…,Z V}, then use decoder g Ω Reconstruct new view samples, and finally, randomly initialize K cluster centers C for each view. V ={c1,c2,...,c kThen, k-means is used to process the hidden features of all views to obtain the initial cluster centers C of all views. K ={C 1 C 2 ,...,C V}, and initialize the clustering layer in the pre-trained deep autoencoder model with the cluster centers of all views.

[0078] Step S2: Calculate the affinity matrix for each view, and set the initial weights for each view and the initial dynamic learning factors between views.

[0079] Set the initial weight ω for each view. v =1 / V, set an appropriate batch size (batch_size), set the number of iterations for formal training (epochs) to 300, and set the initial dynamic learning factor ξ between views. vr The value is 1.

[0080] Calculate the affinity matrix for each view in the multi-view dataset, and quantify the similarity between samples based on the distance of the radial basis functions. The specific formula is as follows:

[0081]

[0082] Where σ represents the bandwidth parameter of the radial basis function, set to 1, and NB i Indicates belonging to sample x i The nearest neighbor x j A set of.

[0083] Step S3: Based on the affinity matrix of each view, the initial weights, and the initial dynamic learning factors between views, perform iterative formal training on the deep autoencoder model of each view until the preset iteration stopping condition is met, and output the final hidden layer features and cluster centers.

[0084] During formal training, a for loop is used to continuously learn the proximity relationships and sample distribution between samples of each view, update the weights of each view and the dynamic learning factors between views, and construct the objective function based on the affinity matrix, weights, and dynamic learning factors between views. The specific steps are as follows:

[0085] Step S301: Increment the number of iteration epochs by one.

[0086] Step S302: Using the same iterative training method as pre-training, the pre-trained deep autoencoder model is used to reconstruct the samples of each view using the objective function of formula (1), and outputs the hidden layer features Z and cluster centers C. K .

[0087] Reconstruction loss term for each view Based on this, the reconstruction loss function L is obtained. rec Specifically:

[0088]

[0089] Step S303: Based on hidden layer features Z and cluster centers C K The clustering term for each view is calculated to maximize the similarity between the original student t-distribution and the sharpened auxiliary distribution, specifically:

[0090] Through formula Calculate the membership degree of the samples, construct the student t-distribution, and then use the membership degree q ij Through formula Construct an auxiliary distribution p with higher reliability than the original student t-scores. ij By using these two distributions and the KL divergence formula to maximize the similarity between q and p, and considering that multi-view data itself has multiple view formats, combined with the weights assigned to each view and the dynamic learning factors for mutual learning between views, the clustering term formula for each view is obtained:

[0091]

[0092] Among them, P r Q represents the sharpened auxiliary distribution of view r. v Let ξ represent the student t-distribution of view v. vr D represents the dynamic learning factor that view v learns from view r. KL This represents the KL divergence formula.

[0093] Clustering items in each view Based on this, the clustering loss function L is obtained. clustering Specifically:

[0094]

[0095] Where, ω v It is the weight of view v.

[0096] Step S304: Calculate the graph embedding constraint term between the original sample and its nearest neighbor data for each view v, which is used to constrain the proximity relationship between samples. The specific formula is as follows:

[0097]

[0098] in, Let N represent the similarity between the i-th sample and the j-th sample in view v, where N represents the number of samples in view v, K is the number of cluster centers in view v, and L represents the number of neighbors of each sample. Let i and j represent the membership degrees of the i-th and j-th samples in view v, respectively.

[0099] Embed constraints in the graph of each view. Based on this, the graph embedding constraint loss function L is obtained. graph Specifically:

[0100]

[0101] Where, ω v It is the weight of view v.

[0102] Step S305: After each epoch, based on the reconstruction loss function L obtained in steps S302-304... rec Clustering loss function L clustering And graph embedding constraint loss function L graph And by adding the weighted entropy loss function L w We construct the overall objective function and use stochastic gradient descent to complete one iteration of the overall algorithm.

[0103] Weighted entropy loss function L w The weight entropy term, used to limit the weights, is the balancing weight, and its formula is:

[0104]

[0105] Where, ω v This indicates the weight of view v.

[0106] Figure 3 This is a schematic diagram of the construction of the objective function, such as... Figure 3 As shown, the overall objective function is:

[0107] L = L rec +γL clustering +ηL graph +ρL w (9)

[0108] Among them, L rec L clustering L graph L w Let represent the reconstruction loss function, clustering loss function, graph embedding constraint loss function, and weight entropy loss function, respectively, and let γ, η, and ρ represent the hyperparameters of the clustering loss function, graph embedding constraint loss function, and weight entropy loss function, respectively.

[0109] Step S306: After each epoch, update the dynamic learning factor between views. The specific formula is as follows:

[0110]

[0111]

[0112]

[0113] Where θ is a parameter greater than 0, used to control the speed of learning factor updates, β represents the measure of the value of the two views, and Q... v Q is the student distribution t of the current view v. r The student t-distribution of the target view r. This represents the global student t-distribution, used as a global consensus for all views.

[0114] Step S307: After each epoch, update the weight of each view using the following formula:

[0115]

[0116] Where, ω v This indicates the weight of view v. represents the clustering term and graph embedding constraint term of view v, respectively, and γ and η represent the hyperparameters of the clustering and graph embedding constraints, respectively.

[0117] Step S308: After each epoch, calculate the clustering accuracy for the current epoch. The specific calculation method is as follows:

[0118] First, calculate the label for each sample:

[0119]

[0120] in, Let l be the global student t-distribution. i The label is calculated by the algorithm for the i-th sample.

[0121] Then, based on the sample labels, the clustering accuracy is calculated. Where N is the number of samples, y i Let be the true label of sample i, map be the redistribution function, and δ be the indicator function, as follows:

[0122]

[0123] Step S309: If the current iteration number epoch reaches the set threshold, terminate the iteration; if not, repeat steps S301-308. After the iteration stops, take the hidden features and cluster centers from the iteration with the highest clustering accuracy during the formal training process as the final hidden features and cluster centers.

[0124] Step S4: Based on the final hidden layer features and cluster centers, the sample labels calculated by formula (12) are used as the final clustering results.

[0125] Example 2

[0126] In one or more embodiments, a deep multi-view clustering system based on collaborative training is disclosed, such as Figure 4 As shown, it includes a pre-training module, an initialization module, a training module, and a clustering module:

[0127] The pre-training module is configured to: acquire multiple views to be clustered, and pre-train its own deep autoencoder model for each view;

[0128] The initialization module is configured to: calculate the affinity matrix for each view, set the initial weights for each view, and set the initial dynamic learning factors between views;

[0129] The training module is configured to: perform iterative formal training on the deep autoencoder model of each view based on the affinity matrix of each view, the initial weights and the initial dynamic learning factors between views, update the weights of each view and the dynamic learning factors between views until the preset iteration stopping condition is met, and output the final hidden features and cluster centers.

[0130] The clustering module is configured to obtain clustering results based on the final hidden layer features and cluster centers;

[0131] The deep autoencoder model takes samples from the view as input and hidden features and cluster centers as output. During pre-training, the target function is MSE reconstruction loss. During formal training, the target function is constructed based on the affinity matrix, weights, and dynamic learning factors between views for each view.

[0132] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep multi-view clustering method based on collaborative training, characterized in that, include: A multi-view clustering process is used to obtain multiple views, and each view is pre-trained with its own deep autoencoder model. The multiple views include information data with different language translations, rich text information with different forms, and human motion capture data collected by cameras at different angles. The deep autoencoder model includes an encoder, a decoder, and a clustering layer. The encoder performs dimensionality reduction on the input view samples to obtain the hidden features of the view. The decoder reconstructs new view samples based on the hidden layer features of the view; The clustering layer obtains cluster centers using k-means based on the hidden features of the view; Calculate the affinity matrix for each view, which quantifies the similarity between samples based on the distance of the radial basis functions; Set the initial weights for each view and the initial dynamic learning factors between views; Based on the affinity matrix, initial weights, and initial dynamic learning factors between views for each view, the deep autoencoder model for each view is iteratively trained, updating the weights of each view and the dynamic learning factors between views until a preset iteration stopping condition is met, outputting the final hidden features and cluster centers; the weights of each view are specifically calculated using the following formula: in, Represents a view The weight, Representing views respectively Clustering terms and graph embedding constraints These represent the hyperparameters of clustering and graph embedding constraints, respectively. Based on the final hidden layer features and cluster centers, the clustering results are obtained; The deep autoencoder model takes view samples as input and hidden features and cluster centers as output. During pre-training, the MSE reconstruction loss is used as the objective function. During formal training, the objective function is constructed based on the affinity matrix, weights, and dynamic learning factors between views for each view. The formally trained objective function consists of a reconstruction loss function, a clustering loss function, a graph embedding constraint loss function, and a weight entropy loss function. One iteration of the overall algorithm is completed using stochastic gradient descent. The dynamic learning factors between views are specifically formulated as follows: in, This is a parameter greater than 0, used to control the speed at which the learning factor is updated. A measure representing the value of two views. This is the current view. The student t-distribution It is the target view The student t-distribution This represents the global student t-distribution, used as a global consensus for all views.

2. The deep multi-view clustering method based on collaborative training as described in claim 1, characterized in that, The pre-trained objective function, used to minimize the error between the original samples and the reconstructed samples in each view, is as follows: in, Indicates the first The reconstruction loss term for each view. Represents a view The number of samples, Represents the original sample. Indicates encoder, Indicates decoder, This represents the reconstructed sample.

3. The deep multi-view clustering method based on collaborative training as described in claim 1, characterized in that, The affinity matrix is ​​specifically formulated as follows: in, The bandwidth parameter represents the radial basis function. Indicates belonging to the sample Neighbors A set of.

4. The deep multi-view clustering method based on collaborative training as described in claim 1, characterized in that, The objective function for the formal training is specifically formulated as follows: in, , , , These represent the reconstruction loss function, clustering loss function, graph embedding constraint loss function, and weight entropy loss function, respectively. , , , Representing views respectively The loss terms include reconstruction loss, clustering loss, graph embedding constraint loss, and weight entropy loss. represents the hyperparameters of the clustering loss function, the graph embedding constraint loss function, and the weight entropy loss function, respectively.

5. The deep multi-view clustering method based on collaborative training as described in claim 4, characterized in that, The clustering term is used to maximize the similarity between the original student t-distribution and the sharpened auxiliary distribution, and the specific formula is as follows: in, Represents a view Student t-distribution Represents a view of The auxiliary distribution after sharpening the distribution. Represents a view To view Dynamic learning factors This represents the KL divergence formula.

6. The deep multi-view clustering method based on collaborative training as described in claim 4, characterized in that, The graph embedding constraint term is used to constrain the proximity relationship between samples, and the specific formula is as follows: in, Represents a view The Middle Sample and the first Similarity between samples Represents a view The number of samples, K is the view The number of cluster centers, where L represents the number of neighbors of each sample. Representing views respectively The membership degrees of the i-th and j-th samples.

7. A deep multi-view clustering system based on collaborative training, characterized in that, It includes a pre-training module, an initialization module, a training module, and a clustering module: The pre-training module is configured to: acquire multiple views to be clustered, and pre-train its own deep autoencoder model for each view; the multiple views include information data with different language translations, rich text information with different forms, and human motion capture data collected by cameras at different angles; the deep autoencoder model includes an encoder, a decoder, and a clustering layer; the encoder performs dimensionality reduction processing on the input view samples to obtain the hidden features of the view; The decoder reconstructs new view samples based on the hidden layer features of the view; The clustering layer obtains cluster centers using k-means based on the hidden features of the view; The initialization module is configured to: calculate the affinity matrix for each view, the affinity matrix being based on the distance of the radial basis functions, quantifying the similarity between samples; Set the initial weights for each view and the initial dynamic learning factors between views; The training module is configured to: iteratively train the deep autoencoder model for each view based on the affinity matrix, initial weights, and initial dynamic learning factors between views; update the weights of each view and the dynamic learning factors between views until a preset iteration stopping condition is met; and output the final hidden features and cluster centers. The weights of each view are specifically calculated using the following formula: in, Represents a view The weight, Representing views respectively Clustering terms and graph embedding constraints These represent the hyperparameters of clustering and graph embedding constraints, respectively. The clustering module is configured to obtain clustering results based on the final hidden layer features and cluster centers; The deep autoencoder model takes view samples as input and hidden features and cluster centers as output. During pre-training, the MSE reconstruction loss is used as the objective function. During formal training, the objective function is constructed based on the affinity matrix, weights, and dynamic learning factors between views for each view. The formally trained objective function consists of a reconstruction loss function, a clustering loss function, a graph embedding constraint loss function, and a weight entropy loss function. One iteration of the overall algorithm is completed using stochastic gradient descent. The dynamic learning factors between views are specifically formulated as follows: in, This is a parameter greater than 0, used to control the speed at which the learning factor is updated. A measure representing the value of two views. This is the current view. The student t-distribution It is the target view The student t-distribution This represents the global student t-distribution, used as a global consensus for all views.

Images