Heterogeneous relation convolution spectral clustering method for incomplete multi-view streaming news data

Abstract

The invention discloses a heterogeneous relation convolution spectral clustering method for incomplete multi-view streaming news data, and belongs to the field of data clustering processing in the incomplete multi-view streaming news data. And constructing a heterogeneous relation matrix and a high-dimensional manifold distribution matrix for each view. And then inputting the sample features of each view and the corresponding heterogeneous relationship matrix into an incomplete multi-view streaming news data heterogeneous relationship convolutional spectral clustering network to obtain a low-dimensional embedded representation of each view, and obtaining a common low-dimensional embedded representation of all views through a feature fusion layer. Then, the common low-dimensional embedding representation is input into an orthogonal normalization layer to obtain approximate spectrum embedding, and then the approximate spectrum embedding is processed through a self-supervision module to obtain soft distribution of single input sample features. And finally, optimizing the heterogeneous relation convolution spectral clustering total loss function of the incomplete multi-view streaming news data to convergence by using a small-batch stochastic gradient descent algorithm. And when the model reaches convergence, clustering single incomplete multi-view streaming news data samples continuously generated after the t0 moment by using the model to obtain a clustering result of single streaming news data, thereby realizing the purpose of clustering the incomplete multi-view streaming news data.

Description

Technical Field

[0001] The present invention belongs to the field of data clustering processing in incomplete multi-view streaming news data, and specifically relates to a heterogeneous relational convolutional spectral clustering method for incomplete multi-view streaming news data. Background Art

[0002] In the field of multimedia information processing, news data is not only a widely available data, but also often collected from multiple different channels (pictures, text, sound, etc.) to describe the same sample. This kind of news data with multiple modes or collected from multiple different sources is called multi-view news data. However, since multi-view news data may have missing information or some channels may not report it at various stages such as collection, transmission, and storage, the integrity of the multi-view news data will be destroyed. In addition, news data in real applications not only presents multiple views, but also shows the characteristics of data streams, that is, news data will be continuously generated in the form of streams over time. This kind of news data that is continuously generated in the form of data streams is called streaming news data. For the clustering problem of incomplete multi-view streaming news data, the traditional incomplete multi-view clustering method needs to use the relationship information between sample data to extract sample features for obtaining clustering results. However, for streaming data with only one sample data input model, the relationship information between sample data cannot be provided. Therefore, when clustering streaming data, the traditional incomplete multi-view clustering method needs to merge the current single sample data with the previous multiple sample data and recalculate the relationship information to obtain the cluster division of the current sample. Therefore, the traditional incomplete multi-view clustering method cannot be directly used for clustering single sample data. In addition, since the traditional method needs to repeatedly calculate the relationship information between samples when clustering streaming data, it will cause a lot of computational costs. Based on this, the present invention proposes a clustering method for incomplete multi-view streaming news data, which can achieve the purpose of clustering a single incomplete streaming news data in an end-to-end manner, and also avoids the problem of excessive computational cost caused by repeated calculation of relationship information between samples in the traditional incomplete multi-view clustering method. Summary of the invention

[0003] In order to solve the above problems, the present invention provides a heterogeneous relational convolutional spectral clustering method for incomplete multi-view streaming news data, the method comprising the steps of:

[0004] By obtaining t 0 Multiple incomplete multi-view streaming news data samples within a period of time before the moment are normalized, and then a heterogeneous relationship matrix and a high-dimensional manifold distribution matrix are constructed for each view.

[0005] The sample features of each view and the corresponding heterogeneous relation matrix are input into the heterogeneous relational convolutional spectral clustering network to obtain a low-dimensional embedding representation of each view, which is then passed through a feature fusion layer to obtain a common low-dimensional embedding representation for all views. At the same time, the representation learned by the network is constrained by the embedding loss function.

[0006] The public low-dimensional embedding representation is input into the orthogonal normalization layer, and the approximate spectral embedding is obtained through the multi-view spectral clustering loss function.

[0007] The approximate spectral embedding is passed through the KL divergence constraint in the self-supervised module to obtain the cluster assignment of the input sample features.

[0008] The overall loss function of heterogeneous relational convolutional spectral clustering on incomplete multi-view streaming news data is optimized to convergence using a mini-batch stochastic gradient descent algorithm.

[0009] When the model reaches convergence, use this model to 0 The single incomplete multi-view streaming news data that is continuously generated after the time is clustered to obtain the clustering result of a single streaming news data.

[0010] Furthermore, the method is for 0 Multiple incomplete multi-view streaming news data samples within a period of time before the moment are normalized, and the heterogeneous relationship matrix and high-dimensional manifold distribution matrix are constructed respectively.

[0011] For t 0 For multiple incomplete multi-view streaming news data samples obtained before time, each component of its feature vector is normalized to a standard format with a mean of 0 and a standard deviation of 1. The proposed normalization expression is as follows:

[0012]

[0013] in, represents the feature vector of the i-th sample in the v-th view, and Respectively The mean and standard deviation of .

[0014] For each view, a heterogeneous relationship matrix and a high-dimensional manifold distribution matrix are constructed. The heterogeneous relationship matrix A of the proposed v-th view is v In the element The expression is:

[0015]

[0016] in, is a heterogeneous relationship matrix The element of the i-th row and j-th column of , i and j represent the i-th and j-th nodes on the heterogeneous relationship matrix, respectively. v (i) is the neighborhood of the ith node. If an edge exists between the ith node and the jth node, then otherwise

[0017] The proposed high-dimensional manifold distribution matrix P of the v-th view v In the element The expression is:

[0018]

[0019] in, is the manifold distribution matrix P v The element in the i-th row and j-th column in , which represents the distance between the i-th sample and the j-th sample in the high-dimensional space. are the feature vectors of the i-th, j-th, l-th, and k-th sample data respectively. is the Gaussian kernel width of the ith sample, determined by the perplexity parameter ppl. ‖·‖ F Represents the F-norm of the matrix.

[0020] Furthermore, a low-dimensional embedding representation of each view is obtained through a heterogeneous relational convolutional spectral clustering network, and a common low-dimensional embedding representation of all views is obtained through a feature fusion layer, and the representation learned by the network is constrained by an embedding loss function.

[0021] The proposed low-dimensional embedding representation H of the v-th view v The expression formula is:

[0022] Z v =ReLU(BN(Conv(X v ,A v )))

[0023] H v =BN(Conv(Z v ,A v ))

[0024] Among them, ReLU is a nonlinear activation function layer, BN is a batch normalization layer, and Conv is a heterogeneous relational convolution layer. v is the sample feature matrix, A v is the corresponding heterogeneous relationship matrix, Z v is the output of the first heterogeneous relation convolutional module, H v is the low-dimensional embedding representation of the vth view. The final output of the multi-view heterogeneous relationship convolution layer is the low-dimensional embedding representation H of the m views. v ,v=1,2,...,m.

[0025] The low-dimensional embedding representation H of all views is obtained through the feature fusion layer v Perform weighted fusion to obtain the common low-dimensional embedding representation H of all views. The expression formula of the proposed feature fusion layer is:

[0026]

[0027]

[0028] Among them, h ij ∈H, Indicates the use of missing indicator matrix M :,v The vth column represents the low-dimensional embedding H v The latent representation of the v-th view after the missing processing. v is the weight of the vth view, and m is the number of views.

[0029] The representation learned by the network is constrained by the embedding loss function. The proposed network embedding loss function expression formula is:

[0030]

[0031] Among them, P v is the high-dimensional manifold distribution matrix of the v-th view, Q v is the low-dimensional embedding representation matrix of the v-th view. In the low-dimensional embedding representation matrix, the elements The expression formula is:

[0032]

[0033] in, is the matrix Q v The element in the i-th row and j-th column represents the probability that the i-th sample is a neighbor of the j-th sample in the low-dimensional space. They are low-dimensional embedding representations H v The corresponding feature vectors of the i, j, l, k sample data in the v-th view, ‖·‖ F Represents the F-norm of the matrix.

[0034] Furthermore, the public low-dimensional embedding representation is input into the orthogonal normalization layer, and the approximate spectral embedding is obtained through the multi-view spectral clustering loss function. The expression of the proposed orthogonal normalization layer is:

[0035]

[0036] Among them, linear is a fully connected layer with weights and biases, and orthnorm is an orthogonal normalization layer; is the approximate spectral embedding obtained by the network through learning, is the spectral embedding that is not orthogonally normalized, and H is the public low-dimensional embedding representation matrix.

[0037] The function of the orthogonal normalization layer is to orthogonalize the common low-dimensional embedding representation matrix H so that (I c is a c-order unit matrix), the purpose of which is to make the features of the input matrix independent of each other and make it closer to the spectral embedding matrix.

[0038] The proposed multi-view spectral clustering loss function is expressed as:

[0039]

[0040] Y T Y=I

[0041] Among them, L v is the Laplace regularization matrix of the vth view, that is W v is the affinity matrix of view v, is a diagonal matrix defined as Represents the trace of the matrix.

[0042] Furthermore, the approximate spectrum is embedded in the self-supervisory module to constrain the soft assignment Q and the target distribution P to obtain the cluster distribution probability of the input sample. The proposed KL divergence expression is:

[0043]

[0044] Among them, Q is the soft assignment of the embedded sample, which represents the similarity between each embedded sample point and the cluster center. In the soft assignment Q, the element q ij The expression is:

[0045]

[0046] Among them, q ij ∈Q represents the yth i The probability that an embedded sample point belongs to the jth cluster, is the cluster center, which is obtained by k-means clustering of the embedded sample features in the approximate spectral embedding Y, and c represents the number of clusters. The target distribution P can be obtained by soft assignment Q. In the target distribution P, the element p ij The expression is:

[0047]

[0048] Among them, p ij ∈P, is the soft class frequency.

[0049] Furthermore, the mini-batch stochastic gradient descent algorithm is used to optimize the overall loss function of heterogeneous relational convolutional spectral clustering of incomplete multi-view streaming news data to convergence. The embedding loss function, the multi-view spectral clustering loss function and the KL divergence in the self-supervision module are combined to obtain the overall loss function of heterogeneous relational convolutional spectral clustering of incomplete multi-view streaming news data. The overall loss function of heterogeneous relational convolutional spectral clustering is:

[0050]

[0051] Furthermore, after the model reaches convergence, t 0 The single incomplete multi-view streaming news data that is continuously input into the model after the moment is generated into a certain number of copy data using a mask matrix, and the heterogeneous relationship matrix and high-dimensional manifold distribution matrix of the copy data in each view are obtained. Then, the common low-dimensional embedding representation of all views is obtained through the spectral clustering network and feature fusion layer. Finally, the self-supervision module is used to obtain the clustering result of a single streaming news data sample.

[0052] The present invention provides a heterogeneous relational convolutional spectral clustering method for incomplete multi-view streaming news data, which has the following advantages:

[0053] (1) The method adopts an incomplete multi-view clustering framework, which makes full use of the consistency and complementary information between views in incomplete multi-view streaming news data, and can effectively mine the potential cluster structure of the data, thereby achieving accurate clustering.

[0054] (2) The method adopts a technology that combines heterogeneous relational convolutional networks and spectral clustering. The heterogeneous relational convolutional network is used to construct a low-dimensional embedding representation of the view, and then an orthogonal normalization layer is used on this basis to obtain an approximate spectral embedding, thereby effectively reducing the training time cost of the model.

[0055] (3) In order to achieve the purpose of clustering incomplete multi-view streaming news data, the method collects t 0 All streaming news data samples from a period of time before time t are input into the model for training. When the model reaches convergence, 0 The streaming news data samples after the time point are input into the model one by one and continuously to obtain the clustering results, thereby achieving the purpose of clustering the incomplete multi-view streaming news data. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the drawings without paying creative labor.

[0057] Figure 1 It is a flow chart of the heterogeneous relational convolutional spectral clustering method for incomplete multi-view streaming news data provided by the present invention.

[0058] Figure 2 The figure shows a comparison of the model training time of the method provided by the present invention with that of several common incomplete multi-view news data clustering methods under BBC-Sport multi-view streaming news data. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present invention.

[0060] Exemplary Methods

[0061] like Figure 1 As shown, after the processing flow starts, step S110 is first executed.

[0062] Step S110: normalize the stream data and construct a heterogeneous relationship matrix and a high-dimensional manifold distribution matrix.

[0063] For t 0 The multiple incomplete multi-view streaming news data samples obtained before the time are normalized to each component of their feature vectors. As an example, the normalization defined in step S110 is expressed as follows:

[0064]

[0065] in, represents the feature vector of the i-th sample in the v-th view, and Respectively The mean and standard deviation of .

[0066] A heterogeneous relationship matrix and a high-dimensional manifold distribution matrix are constructed for each view. As an example, the heterogeneous relationship matrix A of the vth view defined in step S110 is v In the element The expression is as follows:

[0067]

[0068] in, is a heterogeneous relationship matrix The element of the i-th row and j-th column of , i and j represent the i-th and j-th nodes on the heterogeneous relationship matrix respectively. is the neighborhood of the ith node. If an edge exists between the ith node and the jth node, then otherwise

[0069] As an example, the high-dimensional manifold distribution matrix P of the v-th view defined in step S110 is v In the element The expression is as follows:

[0070]

[0071] in, is the manifold distribution matrix P v The element in the i-th row and j-th column in , which represents the distance between the i-th sample and the j-th sample in the high-dimensional space. are the feature vectors of the i-th, j-th, l-th, and k-th sample data respectively. is the Gaussian kernel width of the ith sample, determined by the perplexity parameter ppl. ‖·‖ F Represents the F-norm of the matrix.

[0072] Step S120: Obtain a low-dimensional embedding representation of each view, and pass it through a feature fusion layer to obtain a common low-dimensional embedding representation.

[0073] The sample features and the corresponding heterogeneous relation matrix of each view are input into the heterogeneous relational convolutional spectral clustering network. As an example, the low-dimensional embedding representation H of the v-th view defined in step S120 is v The expression is as follows:

[0074]

[0075] Among them, ReLU is a nonlinear activation function layer, BN is a batch normalization layer, Conv is a heterogeneous relational convolution layer, and X v is the sample feature matrix, A v is the corresponding heterogeneous relationship matrix, Z v is the output of the first heterogeneous relation convolutional module, H v is the low-dimensional embedding representation of the vth view. The final output of the multi-view heterogeneous relationship convolution layer is the low-dimensional embedding representation H of the m views. v ,v=1,2,...,m.

[0076] The low-dimensional embedding representation H of all views is obtained through the feature fusion layer v Perform weighted fusion to obtain a common low-dimensional embedding representation H of all views. As an example, the expression of the feature fusion layer defined in step S120 is as follows:

[0077]

[0078] Among them, h ij ∈H, Indicates the use of missing indicator matrix M :,v The vth column represents the low-dimensional embedding H v The latent representation of the v-th view after the missing processing. v is the weight of the vth view, and m is the number of views.

[0079] The representation learned by the network is constrained by the embedding loss function. As an example, the network embedding loss function expression is defined as follows:

[0080]

[0081] Among them, P v is the high-dimensional manifold distribution matrix of the v-th view, Q v is the low-dimensional embedding representation matrix of the v-th view, in which the elements The expression formula is:

[0082]

[0083] in, is the matrix Q v The element in the i-th row and j-th column represents the probability that the i-th sample is a neighbor of the j-th sample in the low-dimensional space. They are low-dimensional embedding representations H v The corresponding feature vectors of the i, j, l, k sample data in the v-th view, ‖·‖ F Represents the F-norm of the matrix.

[0084] Step S130: input the public low-dimensional embedding representation into an orthogonal normalization layer, and obtain an approximate spectral embedding through a multi-view spectral clustering loss function.

[0085] As an example, the orthogonal normalization layer expression defined in step S130 is as follows:

[0086]

[0087] Among them, linear is a fully connected layer with weights and biases, and orthnorm is an orthogonal normalization layer. is the approximate spectral embedding obtained by the network through learning, is the spectral embedding that is not orthogonally normalized, and H is the public low-dimensional embedding representation matrix.

[0088] As an example, the multi-view spectral clustering loss function defined in step S130 is expressed as follows:

[0089]

[0090] Among them, L v is the Laplace regularization matrix of the vth view, that is W v is the affinity matrix of the vth view, is a diagonal matrix defined as Tr(·) represents the trace of the matrix.

[0091] Step S140: embedding the approximate spectrum into the KL divergence in the self-supervision module to constrain the soft assignment Q and the target distribution P to obtain the cluster distribution probability of the input sample.

[0092] As an example, the KL divergence expression defined in step S140 is as follows:

[0093]

[0094] Among them, Q is the soft assignment of the embedded sample, which represents the similarity between each embedded sample point and the cluster center. In the soft assignment Q, the element q ij The expression is:

[0095]

[0096] Among them, q ij ∈Q represents the yth i The probability that an embedded sample point belongs to the jth cluster, is the cluster center, which is obtained by k-means clustering of the embedded sample features in the approximate spectral embedding Y, and c represents the number of clusters. The target distribution P can be obtained by soft assignment Q. In the target distribution P, the element p ij The expression is:

[0097]

[0098] Among them, p ij ∈P, is the soft class frequency.

[0099] Step S150: Use a mini-batch stochastic gradient descent algorithm to optimize the overall loss function of heterogeneous relational convolutional spectral clustering of incomplete multi-view streaming news data to convergence.

[0100] The embedding loss function, the multi-view spectral clustering loss function and the KL divergence in the self-supervision module are summed to obtain the overall loss function of heterogeneous relational convolutional spectral clustering of incomplete multi-view streaming news data. As an example, the overall loss function of heterogeneous relational convolutional spectral clustering defined in step S150 is expressed as follows:

[0101]

[0102] Step S160: When the model reaches convergence, t 0 After the time instant, the single incomplete multi-view streaming news data continuously input into the model is used to generate a certain number of replica data using a mask matrix, and the heterogeneous relationship matrix and high-dimensional manifold distribution matrix of the replica data in each view are obtained through S110, and then the public low-dimensional embedding representation of all views is obtained through the spectral clustering network and feature fusion layer in S120, and finally the clustering result of a single streaming news data sample is obtained using S130 and S140. Therefore, based on this clustering characteristic of the model, the proposed method can achieve the purpose of clustering incomplete multi-view streaming news data.

[0103] In this embodiment, firstly, by obtaining t 0 Multiple incomplete multi-view streaming news data samples before time t are taken, and a heterogeneous relationship matrix and a high-dimensional manifold distribution matrix are constructed for each view. Then the sample features of each view and the corresponding heterogeneous relationship matrix are input into the heterogeneous relational convolutional spectral clustering network of the incomplete multi-view streaming news data to obtain a low-dimensional embedding representation for each view, and then it is passed through the feature fusion layer to obtain a common low-dimensional embedding representation for all views. Next, the common low-dimensional embedding representation is input into the orthogonal normalization layer to obtain an approximate spectral embedding, which is then passed through the self-supervision module to obtain a soft assignment of a single input sample feature. Finally, the overall loss function of the heterogeneous relational convolutional spectral clustering of the incomplete multi-view streaming news data is optimized to convergence by using a mini-batch stochastic gradient descent algorithm. When the model reaches convergence, the model is used to train the spectral clustering model on t 0 The single incomplete multi-view streaming news data samples that are continuously generated after the time are clustered to obtain the clustering results of the single streaming news data, thereby achieving the purpose of clustering the incomplete multi-view streaming news data.

[0104] Detailed description of the invention Results

[0105] This implementation uses BBC-Sport sports news multi-view data to simulate streaming news data, and simulates incomplete multi-view streaming news data according to different sample missing ratios. BBC-Sport sports news multi-view data consists of sports news in five subject areas, namely athletics, cricket, football, rugby, and tennis. This dataset has the following three views:

[0106] See Figure 1 It is a word-document co-occurrence feature. The characteristic of this view is to measure the relevance of a word to a document by the number of times word i appears in document j.

[0107] See Figure 2 It is a word-word co-occurrence feature. The characteristic of this view is that if word j appears in the window of word i, the number of co-occurrences between the two is increased by one, thereby measuring the correlation between the two words.

[0108] View 3 is the TF-TDF feature, also known as term frequency-inverse document frequency. The feature of this view is to measure the ability of a word to distinguish documents by the frequency of its appearance in different articles.

[0109] In order to verify the superiority of the method of the present invention (Ours), the method of the present invention is compared with several common incomplete multi-view news data clustering methods, including HCLS-CGL, SGC-IMC, UMIC, IMVC-CBG and other methods, and the clustering accuracy (ACC) of these methods on the above streaming news data is compared. The specific clustering results are shown in Table 1. It should be noted that since streaming news data is obtained by inputting a single incomplete multi-view news data into the model in the form of a data stream for clustering to obtain the clustering result, and for a single sample, it is impossible to obtain the clustering accuracy of the model through the obtained clustering result. Therefore, in order to obtain the model clustering accuracy of the streaming news data for evaluating the model, starting from a certain moment, the clustering results of multiple incomplete multi-view news data within a period of time are continuously obtained to detect the clustering accuracy of the model and then evaluate the performance of the model.

[0110] Table 1 Clustering accuracy of BBC-Sport multi-view streaming news data (%)

[0111]

[0112] The comparison of clustering accuracy of different methods in the above table clearly shows that for incomplete multi-view streaming news data with different missing ratios, the method of the present invention (Ours) achieves the best performance index and has a relatively high clustering accuracy. 0 A certain amount of incomplete BBC-Sport multi-view streaming news data before time t is fed into the model in sequence for model training. 0 The single incomplete BBC-Sport multi-view streaming news data after time is clustered to obtain the clustering result, thereby achieving the purpose of clustering the incomplete multi-view streaming news data.

[0113] This embodiment proposes a heterogeneous relational convolutional spectral clustering method for incomplete multi-view streaming news data, which is used to perform clustering analysis on incomplete multi-view streaming news data commonly seen in life. 0 A certain amount of incomplete multi-view streaming news data within a period of time before time t is input into the model in sequence for model training. When the model reaches convergence, 0 The single incomplete multi-view streaming news data after the time is clustered to obtain the clustering result, thereby achieving the purpose of clustering the incomplete multi-view streaming news data. The experimental results on the five incomplete sample ratios of the simulated incomplete multi-view streaming news data BBC-Sport show that this implementation method has a higher clustering accuracy and better superiority than other methods.

[0114] It should be understood that the above specific embodiments of the present invention are only used to illustrate or explain the principles of the present invention, and do not constitute a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc. made without departing from the spirit and scope of the present invention should be included in the protection scope of the present invention. In addition, the appended claims of the present invention are intended to cover all changes and modifications that fall within the scope and boundaries of the appended claims, or the equivalent forms of such scope and boundaries.

Claims

1. A heterogeneous relational convolutional spectral clustering method for incomplete multi-view streaming news data, characterized by: The method comprises the steps of: S110. normalize multiple incomplete multi-view streaming news data samples within a period of time before time t0, and construct a heterogeneous relationship matrix and a high-dimensional manifold distribution matrix respectively; S120. Obtain a low-dimensional embedding representation of each view through a heterogeneous relational convolutional spectral clustering network, and obtain a common low-dimensional embedding representation of all views through a feature fusion layer; S130. Input the public low-dimensional embedding representation into an orthogonal normalization layer, and obtain an approximate spectral embedding through a multi-view spectral clustering loss function; S140. Embed the approximate spectrum through the KL divergence constraint in the self-supervision module to obtain the cluster assignment of the input sample features; S150. Use mini-batch stochastic gradient descent algorithm to optimize the overall loss function of heterogeneous relational convolutional spectral clustering of incomplete multi-view streaming news data to convergence; S160. After the model converges, the model is used to cluster the single incomplete multi-view streaming news data continuously generated after time t0 to obtain the clustering result of the single streaming news data; In S110, for the multiple incomplete multi-view streaming news data samples acquired before time t0, each component of the feature vector is normalized, and the normalized expression is as follows: in, represents the feature vector of the i-th sample in the v-th view, and Respectively The mean and standard deviation of A heterogeneous relationship matrix and a high-dimensional manifold distribution matrix are constructed for each view; the heterogeneous relationship matrix A of the vth view v In the element The expression is: in, is a heterogeneous relationship matrix The element of the i-th row and j-th column of , i and j represent the i-th and j-th nodes on the heterogeneous relationship matrix respectively; is the neighborhood of the ith node; if an edge between the ith node and the jth node exists, then otherwise The high-dimensional manifold distribution matrix P of the v-th view v In the element The expression is: in, is the manifold distribution matrix P v The element in the i-th row and j-th column of , which represents the distance between the i-th sample and the j-th sample in the high-dimensional space; are the feature vectors of the i-th, j-th, l-th, and k-th sample data respectively; is the Gaussian kernel width of the i-th sample, determined by the perplexity parameter ppl; ‖·‖ F represents the F norm of the matrix; In S120, the sample features of each view and the corresponding heterogeneous relationship matrix are input into the heterogeneous relationship convolutional spectral clustering network to obtain a low-dimensional embedding representation of each view; and the low-dimensional embedding representation of all views is obtained through a feature fusion layer; and the representation learned by the network is constrained by an embedding loss function; The low-dimensional embedding representation H of the v-th view v The expression is: Z v =ReLU(BN(Conv(X v ,A v ))) H v =BN(Conv(Z v ,A v )) Among them, ReLU is a nonlinear activation function layer, BN is a batch normalization layer, and Conv is a heterogeneous relational convolution layer; X v is the sample feature matrix, A v is the corresponding heterogeneous relationship matrix, Z v is the output of the first heterogeneous relation convolutional module, H v is the low-dimensional embedding representation of the vth view; the final output of the multi-view heterogeneous relationship convolution layer is the low-dimensional embedding representation H of the m views v ,v=1,2,...,m; The low-dimensional embedding representation H of all views is obtained through the feature fusion layer v Fusion is performed to obtain a common low-dimensional embedding representation H of all views; the expression of the feature fusion layer is: s.t.h ij ≥0,1 T h i =1 Among them, h ij ∈H, Indicates the use of missing indicator matrix M :,v The vth column represents the low-dimensional embedding H v The potential representation of the v-th view obtained after the missing processing; w v is the weight of the vth view, and m is the number of views. The representation learned by the network is constrained by the embedding loss function; the embedding loss function expression of the network is: Among them, P v is the high-dimensional manifold distribution matrix of the v-th view, which is obtained by S110; Q v is the low-dimensional embedding representation matrix of the v-th view. In the low-dimensional embedding representation matrix, the elements The expression formula is: in, is the matrix Q v The element in the i-th row and j-th column represents the probability that the i-th sample is a neighbor of the j-th sample in the low-dimensional space; They are low-dimensional embedding representations H v The corresponding feature vectors of the i, j, l, k sample data in the v-th view, ‖·‖ F represents the F norm of the matrix; In S130, the public low-dimensional embedding representation is input into an orthogonal normalization layer, and an approximate spectral embedding is obtained through a multi-view spectral clustering loss function constraint. The expression of the orthogonal normalization layer is: Among them, linear is a fully connected layer with weights and biases, and orthnorm is an orthogonal normalization layer; is the approximate spectral embedding obtained by the network through learning, is the spectral embedding that is not orthogonally normalized, and H is the public low-dimensional embedding representation matrix; The multi-view spectral clustering loss function is expressed as: s.t.Y T Y=I Among them, L v is the Laplace regularization matrix of the vth view, that is W v is the affinity matrix of the vth view, is a diagonal matrix defined as Tr(·) represents the trace of the matrix; In S140, the soft assignment Q and the target distribution P are constrained by the KL divergence in the self-supervision module to obtain the cluster distribution probability of the input sample. The expression of the KL divergence is: Among them, Q is the soft assignment of the embedded sample, which represents the similarity between each embedded sample point and the cluster center. In the soft assignment Q, the element q ij The expression is: Among them, q ij ∈Q represents the i-th embedded sample point y i The probability of belonging to the jth cluster; is the cluster center, which is obtained by k-means clustering of the embedded sample features in the approximate spectral embedding Y, and c represents the number of clusters; the target distribution P can be obtained by soft assignment Q; in the target distribution P, the element p ij The expression is: Among them, p ij ∈P, is the soft category frequency; In S150, a mini-batch stochastic gradient descent algorithm is used to optimize the overall loss function of heterogeneous relational convolutional spectral clustering of incomplete multi-view streaming news data to convergence; the embedding loss function, the multi-view spectral clustering loss function and the KL divergence in the self-supervision module are combined to obtain the overall loss function of heterogeneous relational convolutional spectral clustering of incomplete multi-view streaming news data, and the overall loss function of heterogeneous relational convolutional spectral clustering is expressed as: In S160, after the model converges, a certain number of copy data are generated for the single incomplete multi-view streaming news data that is continuously input into the model after time t0 using a mask matrix, and the heterogeneous relationship matrix and high-dimensional manifold distribution matrix of the copy data in each view are obtained through S110, and then the common low-dimensional embedding representation of all views is obtained through the spectral clustering network and feature fusion layer in S120, and finally, the clustering result of a single streaming news data sample is obtained using S130 and S140.

Images