Low-quality multi-view news data anchor graph regular division method based on diffusion completion

Abstract

A low-quality multi-view news data anchor graph regular division method based on diffusion completion belongs to the field of data division in low-quality multi-view news data, and comprises the following steps: firstly, inputting low-quality news data of each view and a corresponding similarity relation matrix into a heterogeneous relation convolutional network; to obtain a low-dimensional embedded representation of each view. Then, the low-dimensional embedded representation uses forward noise adding and reverse noise reduction processes of a conditional diffusion model to obtain predicted noise, and missing samples are complemented through the predicted noise; and then, an anchor point diagram is constructed by using the similarity between the complemented embedded sample and the anchor points, soft clustering distribution is obtained for the constructed anchor point diagram through an orthogonal normalization layer, and discriminative feature representation is obtained through anchor graph regularization constraint. And finally, soft clustering distribution is constrained by using a tensor Schatten p-norm so as to fully mine complementarity information and a sparse structure between the views.

Description

Technical Field

[0001] This invention belongs to the field of data partitioning in low-quality multi-view news data, specifically involving a regularized partitioning method for anchor graphs of low-quality multi-view news data based on diffusion completion. Background Technology

[0002] In the field of multimedia information processing, news data is not only widely available but also often collected from multiple different channels (images, text, sound, etc.) to describe the same sample. This type of news data, with multiple modes or collected from various sources, can be called multi-view news data. Because this data can comprehensively describe things from multiple perspectives, it is possible to extract potential category information such as news themes, emotional attitudes, and public opinion, thus enabling efficient and accurate segmentation of multi-view news data. However, multi-view news data is incomplete because information gaps or omissions from certain channels can occur at various stages of collection, transmission, and storage. Furthermore, the lack of views not only severely affects the accuracy of news data segmentation but also disrupts the consistency of the distribution of multi-view news data in its latent space. Based on this, the present invention proposes an anchor graph regularization partitioning method for low-quality multi-view news data based on diffusion completion. This method first maps the original features of the views to the latent embedding space through a heterogeneous relational convolutional network. Then, it completes the missing samples in the latent embedding space through a conditional diffusion model. Finally, it constrains the model through anchor graph regularization and tensor Schatten p-norm to reduce the model's running time and explore the complementary information between views, thereby achieving accurate partitioning of low-quality multi-view news data. Summary of the Invention

[0003] To address the aforementioned challenges, this invention provides a method for regularizing anchor graph partitioning of low-quality multi-view news data based on diffusion completion. The method includes the following steps:

[0004] The low-quality news data sample features and corresponding similarity matrix of each view are input into the heterogeneous relationship convolutional network of low-quality multi-view news data to obtain the low-dimensional embedding representation of each view.

[0005] In the embedding space, the conditional diffusion model is used to complete the missing samples. The conditional diffusion model first obtains the noise embedding representation by adding noise to the existing samples, then obtains the predicted noise through the noise prediction module, and finally completes the missing samples with the predicted noise.

[0006] An anchor point map is constructed by utilizing the similarity between the features of the completed embedded samples and the anchor points. The anchor point map is then input into an orthogonal normalization layer to obtain soft clustering assignments. Finally, anchor point map regularization constraints are applied to make the obtained soft clustering assignments more discriminative.

[0007] The soft clustering assignments of different views are constrained by the tensor Schatten p-norm to fully exploit the complementary information and sparse structure between views.

[0008] The overall loss function of the anchor graph regularization method for low-quality multi-view news data based on diffusion completion is optimized to convergence using the mini-batch stochastic gradient descent algorithm.

[0009] Furthermore, the low-quality news data sample features and corresponding similarity relation matrix of each view are input into the heterogeneous relation convolutional network of low-quality multi-view news data to obtain the low-dimensional embedding representation of each view. The proposed expression of the heterogeneous relation convolutional network is as follows:

[0010]

[0011] in, ReLU(·) represents the output of the heterogeneous relational convolutional network in the l-th layer of the v-th view, and ReLU(·) represents the activation function. This represents the output of the (l-1)th layer in the v-th view. When l = 1, This represents the original input data. D represents the learnable weight parameters of the heterogeneous relational convolutional network in the l-th layer of the v-th view. v Let I represent the degree matrix of the v-th view, and let I represent the identity matrix.

[0012] Furthermore, the conditional diffusion model is used to complete the missing samples in the embedding space. The conditional diffusion model first obtains the noise embedding representation by adding noise to the existing samples, then obtains the predicted noise through the noise prediction module, and finally completes the missing samples through the predicted noise.

[0013] The noise embedding representation is obtained by using the noise addition process of samples existing in the v-th view. The proposed expression formula is as follows:

[0014]

[0015] in, This represents the noise embedding representation in the v-th view after the t-th step of noise addition. Let represent the unnoised embedding representation in the v-th view, ∈ represent standard Gaussian noise, and t represent the number of steps to add noise. This represents the product of the control coefficients added to the noise at each step.

[0016] The noise embedding representation of each view is used to obtain the predicted noise through the noise prediction module's denoising process. The proposed expression formula is as follows:

[0017]

[0018] in, This represents the noise embedding representation in the v-th view after the t-th step of noise addition. Z represents the noise embedding representation of the v-th view after the (t-1)-th step of noise addition. j This represents a conditional view, where I represents the identity matrix. and σ t Let represent the mean and variance of the prediction noise at step t, respectively, and their definitions are as follows:

[0019]

[0020] in, Let θ represent the predicted noise for the v-th view, and let α represent the learnable parameters of the prediction noise network. t This represents the control coefficient for adding noise in step t. This represents the product of the control coefficients for adding noise at each step. It represents the product of all control coefficients with added noise before step t-1.

[0021] By constraining the predicted noise and standard Gaussian noise using a noise prediction loss function, the proposed expression is as follows:

[0022]

[0023] in, Let ∈ represent the noise prediction loss function, and let ∈ represent standard Gaussian noise. This represents the predicted noise for the v-th view. Z represents the noise embedding representation of the v-th view after the t-th step of noise addition. j ‖·‖2 represents the conditional view, and ‖·‖2 represents the L2 norm of the matrix.

[0024] Furthermore, the anchor point map is constructed by utilizing the similarity between the features of the completed embedded samples and the anchor points, and the anchor point map is input into the orthogonal normalization layer to obtain soft clustering assignments. Finally, the anchor map regularization constraint makes the obtained soft clustering assignments more discriminative.

[0025] An anchor point map is constructed using the similarity between the completed embedded sample features and common anchor points in each view. The proposed expression is:

[0026]

[0027] in, This represents the element in the i-th row and j-th column of the v-th anchor point graph. Let a represent the feature of the i-th embedded sample in the v-th view. jLet represent the j-th anchor point, k represent the number of nearest neighbors between the embedded sample feature and the anchor point, and S(·) represent the similarity between the completed embedded sample feature and the anchor point.

[0028] The obtained anchor point map is input into the orthogonal normalization layer to obtain the soft clustering assignment for each view. The proposed expression formula is as follows:

[0029]

[0030] Among them, Y v B represents the soft clustering assignment of the v-th view. v This represents the anchor point graph of the v-th view, and FC(·) represents a fully connected layer. This represents the orthogonal normalized layer weight of the v-th view.

[0031] The obtained soft clustering assignments are subjected to anchor graph regularization constraints to enhance the model's ability to extract discriminative features. The proposed expression formula is as follows:

[0032]

[0033] in, This represents the anchor plot regularization loss function, where V represents the number of views, and Y represents the number of views. v Let I represent the soft clustering assignment of the v-th view, and let I denote the identity matrix. Let represent the Laplacian regularization matrix of the v-th view, and Tr(·) represent the trace of the matrix.

[0034] Furthermore, the soft clustering assignment of different views is constrained by the tensor Schatten p-norm to fully exploit the complementary information and sparse structure between views.

[0035] The soft clustering assignments of different views are constrained by the tensor Schatten p-norm, and the proposed expression is as follows:

[0036]

[0037] Where V represents the number of views and K represents the number of clusters. A 3D tensor representing the soft clustering assignments of all views. express Fast Fourier Transform along the third dimension, express The k-th slice, σ v (·)express The vth singular value decomposition.

[0038] Furthermore, the overall loss function of the low-quality multi-view news data anchor graph regularization partitioning method based on diffusion completion is optimized to convergence using the mini-batch stochastic gradient descent algorithm.

[0039] The overall loss function of the anchor graph regularization partitioning method for low-quality multi-view news data based on diffusion completion is obtained by summing the noise prediction loss function and the anchor graph regularization loss function. The proposed expression formula is as follows:

[0040]

[0041] in, Represents the noise prediction loss function. This represents the anchor graph regularization loss function.

[0042] This invention provides a regularized partitioning method for anchor graphs in low-quality multi-view news data based on diffusion completion, which has the following advantages:

[0043] (1) The method uses the conditional diffusion model to complete the missing samples in the view, thereby effectively avoiding the problem of reduced segmentation accuracy caused by missing views.

[0044] (2) The method reduces the time complexity of the model by constructing anchor graphs and introducing orthogonal normalization layers, thereby effectively reducing the running time of the model.

[0045] (3) The method introduces the tensor Schatten p-norm to constrain the soft clustering assignment of each view, effectively capturing the complementary information between views and realizing the end-to-end data partitioning method, thereby avoiding the impact of partitioning accuracy caused by two-stage data partitioning. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0047] Figure 1 This is a flowchart of the low-quality multi-view news data anchor graph regularization method based on diffusion completion provided by the present invention.

[0048] Figure 2 This invention provides a comparison of the training time of the method provided in this invention with several common methods for segmenting incomplete multiview news data, using BBC-Sport multiview news data. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0050] Exemplary methods

[0051] like Figure 1 As shown, after the processing flow begins, step S110 is executed first.

[0052] Step S110: Input the low-quality news data sample features and corresponding similarity relation matrix of each view into the heterogeneous relation convolutional network of low-quality multi-view news data to obtain a low-dimensional embedding representation of each view. As an example, the expression of the heterogeneous relation convolutional network defined in step S110 is as follows:

[0053]

[0054] in, ReLU(·) represents the output of the heterogeneous relational convolutional network in the l-th layer of the v-th view, and ReLU(·) represents the activation function. This represents the output of the (l-1)th layer in the v-th view. When l = 1, This represents the original input data. D represents the learnable weight parameters of the heterogeneous relational convolutional network in the l-th layer of the v-th view. v Let I represent the degree matrix of the v-th view, and let I represent the identity matrix.

[0055] Step S120: Use the conditional diffusion model to complete the missing samples. The model first obtains the noise embedding representation through the noise addition process, then obtains the predicted noise through the noise prediction module, and finally completes the missing samples through the predicted noise.

[0056] The forward denoising process of the conditional diffusion model is used to progressively add noise to the embedding representation of each view to obtain the corresponding noisy embedding representation. As an example, the expression for the forward denoising process defined in step S120 is as follows:

[0057]

[0058] in, This represents the noise embedding representation in the v-th view after the t-th step of noise addition. Let represent the unnoised embedding representation in the v-th view, ∈ represent standard Gaussian noise, and t represent the number of steps to add noise. This represents the product of the control coefficients added to the noise at each step.

[0059] The reverse denoising process using a conditional diffusion model is employed to progressively predict the predicted noise for single-step denoising. As an example, the expression for the single-step reverse denoising process defined in step S120 is as follows:

[0060]

[0061] in, This represents the noise embedding representation in the v-th view after the t-th step of noise addition. Z represents the noise embedding representation of the v-th view after the (t-1)-th step of noise addition. j This represents a conditional view, where I represents the identity matrix. and σ t Let represent the mean and variance of the prediction noise at step t, respectively, and their definitions are as follows:

[0062]

[0063] in, Let θ represent the predicted noise for the v-th view, and let α represent the learnable parameters of the prediction noise network. t This represents the control coefficient for adding noise in step t. This represents the product of the control coefficients for adding noise at each step. It represents the product of all control coefficients with added noise before step t-1.

[0064] The predicted noise and standard Gaussian noise are constrained by a noise prediction loss function to keep their distributions aligned. As an example, the expression for the noise prediction loss function defined in step S120 is:

[0065]

[0066] in, Let ∈ represent the noise prediction loss function, and let ∈ represent standard Gaussian noise. This represents the predicted noise for the v-th view. Z represents the noise embedding representation of the v-th view after the t-th step of noise addition. j ‖·‖2 represents the conditional view, and ‖·‖2 represents the L2 norm of the matrix.

[0067] The missing samples in the target view are completed using the predicted noise obtained during the inverse denoising process of the noise prediction module. As an example, the view completion expression defined in step S120 is as follows:

[0068]

[0069] in, Let α represent the embedding representation after denoising at step t. t This represents the control coefficient for adding noise in step t. σ represents the sum of control coefficients for adding noise at each step. t Let denot be the variance of the noise added at step t, and ∈ denote standard Gaussian noise. This represents the predicted noise for the v-th view. H represents the denoised embedding representation of the v-th view. i This represents the missing samples that have been completed in the i-th view.

[0070] Step S130: Construct an anchor point map using the similarity between the features of the completed embedded samples and the anchor points, and input the anchor point map into the orthogonal normalization layer to obtain soft clustering assignments. Finally, make the obtained soft clustering assignments more discriminative through anchor map regularization constraints.

[0071] An anchor point map is constructed for each view by utilizing the similarity between the completed embedded sample features and the common anchor points in each view. As an example, the anchor point map expression defined in step S130 is as follows:

[0072]

[0073] in, This represents the element in the i-th row and j-th column of the v-th anchor point graph. Let a represent the feature of the i-th embedded sample in the v-th view. j Let represent the j-th anchor point, k represent the number of nearest neighbors between the embedded sample feature and the anchor point, and S(·) represent the similarity between the completed embedded sample feature and the anchor point.

[0074] The obtained anchor point maps are input into the orthogonal normalization layer of the corresponding views to obtain soft clustering assignments, thereby effectively reducing the time complexity of the model. As an example, the orthogonal normalization layer expression defined in step S130 is as follows:

[0075]

[0076] Among them, Y v B represents the soft clustering assignment of the v-th view. v This represents the anchor point graph of the v-th view, and FC(·) represents a fully connected layer. This represents the orthogonal normalized layer weight of the v-th view.

[0077] Anchor graph regularization constraints make the obtained soft cluster assignments more discriminative. As an example, the anchor graph regularization constraint expression defined in step S130 is as follows:

[0078]

[0079] in, This represents the anchor plot regularization loss function, where V represents the number of views, and Y represents the number of views. vLet I represent the soft clustering assignment of the v-th view, and let I denote the identity matrix. Let represent the Laplacian regularization matrix of the v-th view, and Tr(·) represent the trace of the matrix.

[0080] Step S140: The soft clustering assignments of different views are constrained by the tensor Schatten p-norm to fully exploit the complementary information and sparse structure between views.

[0081] The soft clustering assignments for different views are constrained by the tensor Schatten p-norm. As an example, the expression for the tensor Schatten p-norm defined in step S140 is as follows:

[0082]

[0083] Where V represents the number of views and K represents the number of clusters. A 3D tensor representing the soft clustering assignments of all views. express Fast Fourier Transform along the third dimension, express The k-th slice, σ v (·)express The vth singular value decomposition.

[0084] Step S150: Use the mini-batch stochastic gradient descent algorithm to optimize the overall loss function of the anchor graph regularization method for low-quality multi-view news data based on diffusion completion until convergence.

[0085] The noise prediction loss function and the anchor graph regularization loss function are summed to obtain the overall loss function of the anchor graph regularization partitioning method for low-quality multi-view news data based on diffusion completion. As an example, the expression for the overall loss function defined in step S150 is as follows:

[0086]

[0087] in, Represents the noise prediction loss function. This represents the anchor graph regularization loss function.

[0088] In this implementation, the low-quality news data sample features and corresponding similarity relation matrices of each view are first input into a convolutional network for heterogeneous relationships in low-quality multi-view news data to obtain a low-dimensional embedding representation for each view. Then, the low-dimensional embedding representations of different views are subjected to forward noise addition and backward noise reduction processes using a conditional diffusion model to obtain predicted noise, which is then used to complete the missing samples in each view. Next, an anchor point map is constructed using the similarity between the completed embedding sample features and the anchor points. This constructed anchor point map is input into an orthogonal normalization layer to obtain soft clustering assignments while reducing model runtime, and discriminative feature representations are obtained through anchor point map regularization constraints. The soft clustering assignments of different views are then constrained using the tensor Schatten p-norm to fully exploit the complementary information and sparse structure between views. Finally, a mini-batch stochastic gradient descent algorithm is used to optimize the overall loss function of the low-quality multi-view news data anchor point map regularization partitioning method to convergence.

[0089] Detailed Implementation Results

[0090] This implementation uses publicly available BBC-Sport sports news multi-view data and simulates low-quality multi-view news data according to different sample missing ratios. The BBC-Sport sports news multi-view data consists of sports news from five thematic areas: athletics, cricket, football, rugby, and tennis. This dataset has the following three views:

[0091] See Figure 1 It is a word-document co-occurrence feature. The characteristic of this view is that it measures the relevance of a word to a document by the number of times the word appears in the document.

[0092] See Figure 2 It is a word-word co-occurrence feature. The characteristic of this view is that if a word appears in the word window, the number of times the two words co-occur is incremented by one, thereby measuring the correlation between the two words.

[0093] View 3 is a TF-TDF feature, also known as term frequency-inverse document frequency. The characteristic of this view is that it measures the distinguishing ability of a word by the frequency of its occurrence in different documents.

[0094] To verify the superiority of the method of the present invention, it is compared with several common low-quality multi-view news data classification methods, including AIMC, ICMVC, GIMVC, and DVIMVC. The clustering accuracy (ACC) of these methods on the above-mentioned news multi-view data is compared, and the specific classification results are shown in Table 1.

[0095] Table 1. Accuracy of Category Classification for BBC-Sport Multiview Data

[0096]

[0097]

[0098] The comparison of the segmentation accuracy of different methods in the table above clearly shows that, for low-quality multi-view news data with different missing proportions, the method of this invention (Ours) achieves the best performance index and has a relatively high segmentation accuracy. This invention's method obtains the latent embedding representation of samples by using a convolutional network with heterogeneous relationships in the low-quality multi-view news data. Then, it completes the missing views using a conditional diffusion model and constrains the model using anchor graph regularization and the Schatten p-norm to reduce training time and obtain complementary information between views, thereby achieving the goal of accurately classifying low-quality view news data.

[0099] This implementation proposes an anchor graph regularization method for classifying low-quality multi-view news data based on diffusion completion, used to categorize common low-quality multi-view news data. This method obtains latent embedding features between different views through a heterogeneous relational convolutional network and completes missing views using a conditional diffusion model. Furthermore, the model reduces time complexity by constructing an anchor graph and introducing an orthogonal normalization layer, and constrains the model through anchor graph regularization and the Schatten p-norm, thereby achieving accurate classification of low-quality multi-view news data. Experimental results using five different proportions of missing samples from BBC-Sport sports news multi-view data demonstrate that this implementation achieves higher classification accuracy and superior performance compared to other methods.

[0100] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of the invention and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of the invention should be included within the protection scope of the invention. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A regularized partitioning method for anchor graphs in low-quality multi-view news data based on diffusion completion, characterized in that, The method includes the following steps: The low-quality news data sample features and the corresponding similarity matrix of each view are input into the heterogeneous relationship convolutional network of low-quality multi-view news data to obtain the low-dimensional embedding representation of each view. In the embedding space, the conditional diffusion model is used to complete the missing samples. The conditional diffusion model first obtains the noise embedding representation by adding noise to the existing samples, then obtains the predicted noise through the noise prediction module, and finally completes the missing samples with the predicted noise. An anchor point map is constructed by utilizing the similarity between the features of the completed embedded samples and the anchor points. The anchor point map is then input into an orthogonal normalization layer to obtain soft clustering assignments. Finally, anchor point map regularization constraints are applied to make the obtained soft clustering assignments more discriminative. The soft clustering assignments of different views are constrained by the tensor Schatten p-norm to fully explore the complementary information and sparse structure between views; The overall loss function of the anchor graph regularization method for low-quality multi-view news data based on diffusion completion is optimized to convergence using the mini-batch stochastic gradient descent algorithm.

2. The method for regularizing anchor graph partitioning of low-quality multi-view news data based on diffusion completion according to claim 1, characterized in that, The low-quality news data sample features and corresponding similarity matrix of each view are input into a convolutional network for heterogeneous relationships in low-quality multi-view news data, thereby obtaining a low-dimensional embedding representation for each view, expressed as follows: in, ReLU(·) represents the output of the heterogeneous relational convolutional network in the l-th layer of the v-th view, and ReLU(·) represents the activation function. This represents the output of the (l-1)th layer in the v-th view. When l = 1, This represents the original input data. D represents the learnable weight parameters of the heterogeneous relational convolutional network in the l-th layer of the v-th view. v Let I represent the degree matrix of the v-th view, and let I represent the identity matrix.

3. The method for regularizing anchor graph partitioning of low-quality multi-view news data based on diffusion completion according to claim 1, characterized in that, In the embedding space, the conditional diffusion model is used to complete the missing samples. The conditional diffusion model first obtains the noise embedding representation by adding noise to the existing samples, then obtains the predicted noise through the noise prediction module, and finally completes the missing samples with the predicted noise. The noise embedding representation is obtained by using the noise addition process of existing samples, and the expression formula is as follows: in, This represents the noise embedding representation in the v-th view after the t-th step of noise addition. Let represent the unnoised embedding representation in the v-th view, ∈ represent standard Gaussian noise, and t represent the number of steps to add noise. This represents the product of the control coefficients added to the noise at each step; The predicted noise is obtained through the noise prediction module, expressed by the following formula: in, This represents the noise embedding representation in the v-th view after the t-th step of noise addition. Z represents the noise embedding representation of the v-th view after the (t-1)-th step of noise addition. j This represents a conditional view, where I represents the identity matrix. and σ t Let represent the mean and variance of the prediction noise at step t, respectively, and their definitions are as follows: in, Let θ represent the predicted noise for the v-th view, and let α represent the learnable parameters of the prediction noise network. t This represents the control coefficient for adding noise in step t. This represents the product of the control coefficients for adding noise at each step. This represents the product of all control coefficients with added noise before step t-1; The predicted noise and standard Gaussian noise are constrained by a noise prediction loss function, expressed as follows: in, Let ∈ represent the noise prediction loss function, and let ∈ represent standard Gaussian noise. This represents the predicted noise for the v-th view. Z represents the noise embedding representation of the v-th view after the t-th step of noise addition. j This represents a conditional view, and ||·|2 represents the L2 norm of the matrix; The missing samples are filled with predicted noise obtained through the noise prediction module, expressed by the following formula: in, Let α represent the embedding representation after denoising at step t. t This represents the control coefficient for adding noise in step t. σ represents the sum of control coefficients for adding noise at each step. t Let denot be the variance of the noise added at step t, and ∈ denote standard Gaussian noise. This represents the predicted noise for the v-th view. H represents the denoised embedding representation of the v-th view. i This represents the missing samples that have been completed in the i-th view.

4. The method for regularizing anchor graph partitioning of low-quality multi-view news data based on diffusion completion according to claim 1, characterized in that, An anchor point map is constructed by utilizing the similarity between the features of the completed embedded samples and the anchor points. The anchor point map is then input into an orthogonal normalization layer to obtain soft clustering assignments. Finally, anchor point map regularization constraints are applied to make the obtained soft clustering assignments more discriminative. An anchor point map is constructed by utilizing the similarity between the features of the completed embedded samples and the anchor points. The formula is as follows: in, This represents the element in the i-th row and j-th column of the v-th anchor point graph. Let a represent the feature of the i-th embedded sample in the v-th view. j Let represent the j-th anchor point, k represent the number of nearest neighbors between the embedded sample feature and the anchor point, and S(·) represent the similarity between the completed embedded sample feature and the anchor point; The anchor point map is input into the orthogonal normalization layer to obtain soft clustering assignments, expressed by the following formula: Among them, Y v B represents the soft clustering assignment of the v-th view. v This represents the anchor point graph of the v-th view, and FC(·) represents a fully connected layer. This represents the orthogonal normalized layer weight of the v-th view; Anchor graph regularization constraints make the obtained soft cluster assignments more discriminative, as expressed by the following formula: in, This represents the anchor plot regularization loss function, where V represents the number of views, and Y represents the number of views. v Let I represent the soft clustering assignment of the v-th view, and let I denote the identity matrix. Let represent the Laplacian regularization matrix of the v-th view, and Tr(·) represent the trace of the matrix.

5. The method for regularizing anchor graph partitioning of low-quality multi-view news data based on diffusion completion according to claim 1, characterized in that, The soft clustering assignments of different views are constrained by the tensor Schatten p-norm to fully explore the complementary information and sparse structure between views; The soft clustering assignments for different views are constrained by the tensor Schatten p-norm, expressed as follows: Where V represents the number of views and K represents the number of clusters. A 3D tensor representing the soft clustering assignments of all views. express Fast Fourier Transform along the third dimension, express The k-th slice, σ v (·)express The vth singular value decomposition.

6. The overall loss function of the anchor graph regularization method for low-quality multi-view news data based on diffusion completion is optimized to convergence using the mini-batch stochastic gradient descent algorithm; The overall loss function of the anchor graph regularization partitioning method for low-quality multi-view news data based on diffusion completion is obtained by summing the noise prediction loss function and the anchor graph regularization loss function. The formula is as follows: in, Represents the noise prediction loss function. This represents the anchor graph regularization loss function.

Images