A hybrid diffusion-based graph convolutional network document classification method and system

Abstract

The present invention discloses a document classification method and system based on a hybrid diffusion graph convolutional network, comprising the following steps: obtaining a training dataset containing document data; performing information diffusion between neighboring nodes in the feature space of a node using a feature similarity diffusion model to obtain a dataset that has undergone feature space information diffusion; performing information diffusion between adjacent nodes using an adjacency matrix in a graph convolutional model for the dataset that has undergone feature space information diffusion; training the dataset after hybrid diffusion using a semi-supervised graph convolutional network, and constraining the final predicted labels using the diffusion distance between nodes; and performing document classification using the trained hybrid diffusion-based graph convolutional network. The present invention utilizes hybrid diffusion of node features to diffuse information about each node to both nearby nodes and adjacent nodes in the feature space, thereby ensuring that the node features contain more information and significantly improving document classification accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of document classification, and in particular relates to a document classification method and system based on a hybrid diffusion graph convolutional network. Background Art

[0002] The classification of scientific literature has always been a key issue in the field of data mining. With the continuous maturity of machine learning and the emergence of deep learning, scientific literature classification has become increasingly efficient and accurate. Traditional convolutional neural networks have a wide range of applications in image processing. However, the vast majority of scientific literature datasets exist in the form of graph structures, which are non-Euclidean and cannot be processed using convolutional neural networks used in the image field. Graph convolutional neural networks effectively introduce the concept of convolution to graph-structured data, making them suitable for scientific literature classification with high accuracy. Summary of the Invention

[0003] In view of the above research status and problems, the purpose of this invention is to provide a graph convolutional network document classification method based on hybrid diffusion, which enhances the expressive ability of node features and reduces the noise component in node information, thereby improving network performance.

[0004] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0005] A document classification method based on hybrid diffusion graph convolutional network includes the following steps:

[0006] S100: Access to public scientific paper datasets;

[0007] S200: Using a node feature similarity diffusion model, the node information of the imported public scientific paper dataset is diffused and mapped, and the diffusion distance between nodes is calculated. The diffusion process is used to diffuse information between neighboring nodes in the feature space of the node, thereby obtaining new feature information of each node and obtaining a dataset that has undergone node feature space information diffusion.

[0008] S300: Through the semi-supervised graph convolution model, the data set after S200 is subjected to graph convolution operation to diffuse information between adjacent nodes of the data set;

[0009] S400: Using the graph convolution model in S300, a semi-supervised graph convolutional network is trained on the dataset after the mixed diffusion of S200 and S300. During the training process, the diffusion distance between nodes obtained in S200 is used to constrain the final predicted label, making the final classification surface smoother.

[0010] S500: Using the graph convolutional neural network trained in step S400 to perform document classification.

[0011] Furthermore, the training dataset is one or more of a Cora dataset, a Pubmed dataset, and a Citeseer dataset.

[0012] Furthermore, step S200 includes the following sub-steps:

[0013] Step S201: The data set is recorded as G(X,A), X∈R n×d ,A∈R n×n , X represents the node information of the data set, A represents the adjacency matrix of the data set, n represents the number of samples in the data set, D represents the characteristic dimension of the sample, and the similarity matrix L constructed by the data set X is constructed;

[0014] Step S202: normalize the similarity matrix L to obtain the Markov state transition matrix M on the data set X;

[0015] Step S203: Using the obtained Markov state transition matrix M, define the diffusion distance between any two points on the adjacency graph G;

[0016] Step S204: Obtain a diffusion mapping from the original data to the k-dimensional space embedded in the original space through the node feature similarity diffusion model.

[0017] Furthermore, the diffusion distance between any two points on the adjacency graph G is defined as:

[0018]

[0019] in, m i =∑ j p ij , where p ij Represents node x i and node x j The probability of a random walk between i Represents all other nodes to node x i The sum of the probabilities of the random walks of is node x i The static distribution of N is the number of nodes, g∈N, and t represents the number of diffusion times.

[0020] Furthermore, the node feature similarity diffusion model is:

[0021]

[0022] Among them, (t)(x) is the node information after the node feature space information diffusion model is used to diffuse the information between neighboring nodes in the feature space of the dataset node information X, where X represents the node information of the introduced public scientific paper dataset; λ k is the k-dimensional eigenvalue of the node feature x, t represents the number of diffusion times, ψ k (x) is the k-dimensional feature vector of node feature x.

[0023] Furthermore, in step S300, the semi-supervised graph convolution model is as follows:

[0024]

[0025] In the above formula: σ(·) represents the activation function; H ( l +1) H represents the (l+1)th layer output of the semi-supervised graph convolutional network; ( l ) Represents the lth layer output of the semi-supervised graph convolutional network, and the 0th layer output H of the semi-supervised graph convolutional network ( 0 ) That is, the initial input is the current training data set Ψ (t) (x), we get: H ( 0 ) =Ψ (t) (x); represents the first adjacency matrix with self-connection added; represents the degree matrix of the matrix; W ( l ) Represents the weight matrix corresponding to the lth layer of the semi-supervised graph convolutional network.

[0026] Furthermore, in step S400, the diffusion distance is used to regularize the loss by node feature similarity, where

[0027] The loss function is as follows:

[0028]

[0029] In the above formula: Represents the loss function of the network; F i* represents the i-th row in the label matrix F; F j* represents the jth row in the label matrix F, y ik represents the true label distribution of the i-th sample; L is the number of labeled samples; C is the characteristic dimension of the label matrix F; N is the number of nodes; α is the balance parameter that balances these two terms; A ij is the value between node i and node j in the adjacency matrix A of the dataset; x i and x j is the node feature, x i ,xj ∈X, X represents the node information of the data set, D ( t ) (x i ,x j ) 2 is node x i and node x j Diffusion distance obtained by diffusion mapping, t represents diffusion time; log represents logarithmic function.

[0030] The present invention also provides a document classification system based on a hybrid diffusion graph convolutional network, comprising:

[0031] The data set acquisition module is used to obtain the data set needed for training;

[0032] The feature space neighborhood node information diffusion module uses the node feature similarity diffusion model and the diffusion process of the diffusion mapping to diffuse the information between the feature space neighborhood nodes of the dataset nodes;

[0033] The graph convolution model's adjacent node information diffusion module uses a dataset that has undergone node feature similarity diffusion and uses the adjacency matrix through the graph convolution model to diffuse information between adjacent nodes.

[0034] The graph convolution training module performs semi-supervised graph convolution network training on the mixed diffusion dataset and uses the diffusion distance between nodes to constrain the final predicted labels during training.

[0035] The classification module uses the network model trained by the graph convolution training module to perform document classification.

[0036] Furthermore, the feature space neighborhood node information diffusion module performs diffusion mapping on the node information of the introduced public scientific paper dataset through the node feature similarity diffusion model. The node feature similarity diffusion model is:

[0037]

[0038] Among them, (t) (x) is the node information after the node feature space information diffusion model is used to diffuse the information between neighboring nodes in the feature space of the dataset node information X, where X represents the node information of the introduced public scientific paper dataset; λ k is the k-dimensional eigenvalue of node feature x, t is the diffusion time, ψ k (x) is the k-dimensional feature vector of node feature x.

[0039] Furthermore, the semi-supervised graph convolution model in the adjacent node information diffusion module of the graph convolution model is as follows:

[0040]

[0041] In the above formula: σ(·) represents the activation function; H ( l +1) H represents the (l+1)th layer output of the semi-supervised graph convolutional network; ( l ) Represents the lth layer output of the semi-supervised graph convolutional network, and the 0th layer output H of the semi-supervised graph convolutional network ( 0 ) That is, the initial input is the current training data set Ψ (t) (x), we get: H ( 0 ) =Ψ (t) (x); represents the first adjacency matrix with self-connection added; represents the degree matrix of the matrix; W ( l ) Represents the weight matrix corresponding to the lth layer of the semi-supervised graph convolutional network.

[0042] Compared with the prior art, this application has the following beneficial effects:

[0043] The present invention takes into account that diffusion mapping diffuses information on data while preserving the spatial structure of the data manifold, and proposes a graph convolutional network document classification method and system based on hybrid diffusion. This method uses a diffusion mapping method through a node feature similarity diffusion model to diffuse information between neighboring nodes in the feature space of the nodes of the data set, so that the features of the nodes absorb the information of the neighboring nodes in the feature space. Then, through the graph convolution model, information diffusion is performed between adjacent nodes through the node adjacency matrix. After such hybrid diffusion of node information, the node feature information is enriched, the sample noise is reduced, the model classification surface is smoother, and the possibility of overfitting is reduced, thereby improving the prediction accuracy of the graph convolutional neural network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. 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 these drawings without creative work.

[0045] Figure 1 It is a flow chart adopted by the present invention. DETAILED DESCRIPTION

[0046] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0048] The present invention will be described in detail below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0049] Example: Figure 1 As shown, the embodiment provides a hybrid diffusion-based graph convolutional network document classification method, including:

[0050] S100: Access to public scientific paper datasets;

[0051] Before conducting document classification, model training is required, which requires collecting a dataset of document-related sample data. In this embodiment, the training dataset collected is the Cora dataset (i.e., a public scientific paper dataset), which includes a small amount of labeled sample data and a large amount of unlabeled sample data. The labeled sample data includes at least sample feature information, category information, and edge information. Edge information is used to represent the connection relationship between samples; the unlabeled sample data includes at least sample feature information and edge information.

[0052] S200: The node information of the introduced public scientific paper dataset is diffused and mapped using the node feature similarity diffusion model, and the diffusion distance between nodes is calculated. The diffusion process is used to diffuse information between neighboring nodes in the feature space, thereby obtaining new feature information of each node and obtaining a dataset that has undergone node feature space information diffusion. The node feature space information diffusion model mainly uses a diffusion mapping method to diffuse node information. The specific implementation method is as follows:

[0053] Let the dataset be G(X,A), X∈R n×d ,A∈R n×n, X represents the node information of the dataset, A represents the adjacency matrix of the dataset, n represents the number of samples in the dataset, and d represents the characteristic dimension of the samples. First, the similarity matrix L is constructed using the node information X of the given dataset. Diffusion mapping exploits the relationship between heat kernel diffusion and random walk Markov chains. The connectivity between two data points can be defined as the probability of a random walk from one point to another. Typically, this probability is specified using the kernel function of the two data points: k:X×X→R. This kernel function generally meets the conditions of symmetry, non-negativeness, and positive semi-definiteness, namely:

[0054] Symmetry: k(x,y)=k(y,x),

[0055] Non-negativity: k(x,y)≥0.

[0056] To some extent, this kernel function represents the connectivity between data points in the dataset X. The kernel function constitutes a priori definition of the local geometric structure of the dataset. Since a given kernel function will capture the specific characteristics of the dataset, the kernel function will be selected based on the characteristics of the dataset. Generally, a Gaussian kernel function is selected to define the similarity between data points, that is:

[0057]

[0058] Among them, σ is the variance of the Gaussian kernel function. The larger the variance σ, the worse the connectivity between data points. With the given kernel function k(x,y), we can get the similarity matrix L constructed for the data set X:

[0059] L ij =k(x i ,x j )x i ,x j ∈X

[0060] Then normalize the similarity matrix L so that the sum of each row is 1, so that we can get a normalized matrix The elements are: This normalized matrix P can be interpreted as the Markov state transition matrix M on the data set X, which represents the probability of random walks between each data point. In other words, this Markov state transition matrix M represents the probability of a data point randomly walking to another data point. Therefore, M represents the probability of two data points moving from one point to another after one step of random walk. If there are t steps of random walk, the corresponding transition probability M t =(M) t .

[0061] By using the obtained Markov state transfer matrix M, the diffusion distance between any two points on the adjacency graph G can be defined as:

[0062]

[0063] in, m i =∑ j p ij , where p ij Represents node x i and node x j The probability of a random walk between i Represents all other nodes to node x i The sum of the probabilities of the random walks of is node x i The static distribution of N is the number of nodes, g∈N, and t represents the number of diffusion times.

[0064] The diffusion distance formula shows that the diffusion distance is directly proportional to the two data points x. i and x j This formula also shows that the diffusion distance is extremely robust, because the distance between two points depends on the length of all possible paths between the points.

[0065] Calculate the Markov state transition matrix M corresponding to the t-step random walk t The k largest eigenvalues ​​and corresponding eigenvectors of M can be obtained: t ψ(x)=ψ(x). Then the diffusion distance can be calculated using the eigenvector as:

[0066] D (t) (x i , x j ) 2 =∑ k λ k 2t (ψ k (xi)-ψ k (x j )) 2 x i , x j eX

[0067] From this we can know that the corresponding eigenvector can represent a new set of coordinates of the data set in the feature space, and we get the diffusion mapping from the original data to the k-dimensional space embedded in the original space as:

[0068]

[0069] Then the node information X of the data set is transformed into Ψ after diffusion mapping. (t) (x)=∑ k λ k t ψ k (x).

[0070] That is Ψ (t) (x) is the node information of the dataset node X after the node feature space information diffusion model performs information diffusion between neighboring nodes in the feature space. X represents the node information of the introduced public scientific paper dataset; λ k is the k-dimensional eigenvalue of the node feature x, t represents the number of diffusion times, ψ k (x) is the k-dimensional feature vector of node feature x.

[0071] S300: Using the dataset node information Ψ after the node feature space information is diffused (t) (x), and using the dataset's adjacency matrix A, a graph convolutional model is used to diffuse information between adjacent nodes in the dataset. In graph theory and computer science, an adjacency matrix is ​​a square matrix used to represent a finite graph, where each element indicates whether there is an edge connecting the nodes.

[0072] In this embodiment, the semi-supervised graph convolutional network used is as follows:

[0073]

[0074] In the above formula:

[0075] σ(·) represents the activation function, and this embodiment adopts the RELU function;

[0076] H (l+1) Represents the (l+1)th layer output of the semi-supervised graph convolutional network;

[0077] H (l) Represents the lth layer output of the semi-supervised graph convolutional network, and the 0th layer output H of the semi-supervised graph convolutional network (0) (ie initial input) is the current training data set Ψ (t) (x), we get: H (0) =Ψ (t) (x);

[0078] represents the first adjacency matrix with self-connections added. The so-called "adding self-connections" means adding the matrix to the identity matrix of the same size as the matrix. The resulting matrix is ​​the matrix with self-connections added.

[0079] represents the degree matrix of the matrix;

[0080] W(l) Represents the weight matrix corresponding to the lth layer of the semi-supervised graph convolutional network.

[0081] After the two-layer graph convolution model performs information diffusion between adjacent nodes on the adjacency matrix on the dataset, the node information obtained is:

[0082] S400: Use the graph convolution model in S300 to perform semi-supervised graph convolution network training on the dataset after mixed diffusion of S200 and S300, and use the diffusion distance D between nodes obtained in S200 during training. () Constrain the final predicted labels to make the final classification surface smoother.

[0083] Cross-entropy measures the degree of dissimilarity between multiple probability distributions for the same random variable. In machine learning, it is expressed as the difference between the true and predicted probability distributions. A smaller cross-entropy value indicates a better model prediction. Diffusion distance preserves the underlying manifold structure of the dataset in feature space. This distance is robust to noise because the distance between two points depends on all possible paths of length t between them. Graph convolutional models ignore the neighborhood relationships of nodes in feature space. Some nodes may not have adjacencies in the adjacency matrix but may be neighbors in feature space. Our diffusion mapping method, however, diffuses nodes in feature space based on node feature similarity, preserving neighbor node information in feature space. Furthermore, after a dataset undergoes a graph convolution operation, its original distribution in feature space is not retained, and the original neighborhood relationships are destroyed. Therefore, the diffusion distance obtained from diffusion mapping can be used to constrain the predicted label category relationships of nodes. We use the diffusion distance to regularize the loss based on node feature similarity, ensuring that neighboring nodes in the feature space produce predicted labels that are as similar as possible.

[0084] The loss function in this embodiment is as follows:

[0085]

[0086] In the above formula: Represents the loss function of the network; F i* represents the i-th row in the label matrix F; F j* represents the jth row in the label matrix F, y ik represents the true label distribution of the i-th sample; L is the number of labeled samples; C is the characteristic dimension of the label matrix F; N is the number of nodes; α is the balance parameter that balances these two terms; A ij is the value between node i and node j in the adjacency matrix A of the dataset; x i and x j is the node feature, x i ,xj ∈X, X represents the node information of the data set, D ( t ) (x i ,x j ) 2 is node x i and node x j The diffusion distance obtained by diffusion mapping, t represents the diffusion time; log represents the logarithmic function. In specific implementation, the final node information H (2) Through a Softmax function operation, the final prediction category of each node can be obtained, and then the above loss function is used to train the network.

[0087] S500: Using the graph convolutional neural network trained in step S400 to perform document classification.

[0088] When classifying documents, the feature information of the node is first input into the trained graph convolutional data network, and then the corresponding prediction value is output.

[0089] This example uses three datasets, Cora, Pubmed, and Citeseer, for validation. Using the method in step S100, the Cora, Pubmed, and Citeseer datasets are introduced into the system. First, in step S200, a diffusion mapping operation is performed on the dataset using the node feature space information diffusion model. Node information can be diffused from the perspective of similar points in the feature space, allowing node information to diffuse between neighboring nodes in the feature space. Then, using the graph convolution model in step S300, information diffusion is performed between adjacent nodes in the dataset. Then, in step S400, the node information of the dataset, which has been mixed and diffused through steps S100 and S200, is trained. Finally, in step S500, the trained graph convolutional neural network is used to perform classification predictions on the test dataset.

[0090] In this embodiment, other existing methods are also used to first build a model, and then use the built model to perform classification prediction on the test data set. Other methods include back propagation neural network method (BP), support vector machine (SVM), simplified graph convolutional neural network (SGC), and graph convolutional neural network (GCN). In this embodiment, when training SGC and GCN, the model we used is the benchmark model in PyG, and the model parameters are not changed. The correct results in the prediction results are compared with the total number of test data to obtain the accuracy data in Table 1. From Table 1, the accuracy data of the comparative test

[0091] method Cora Pubmed Citeseer BP 0.555 0.714 0.584 Support Vector Machine 0.527 0.699 0.499 SGC 0.711 0.736 0.675 GCN 81.8 79.9 70.1 Method of the present invention 83.0 80.0 74.5

[0092] It can be seen that the method of the present invention has a better classification accuracy.

[0093] In some embodiments, the training dataset is one or more of a Cora dataset, a Pubmed dataset, and a Citeseer dataset.

[0094] In some embodiments, in step S200, information diffusion processing is performed on each node in the graph dataset using a node feature similarity diffusion model.

[0095] In some embodiments, in step 400, a cross entropy loss function is used to train the graph convolutional neural network, and a regularization term related to the diffusion distance between nodes is added to the loss to constrain the predicted labels of the nodes.

[0096] The above is a specific embodiment of the technical solution of the present invention and the technical principles used. Any modification, supplement or equivalent replacement made by those skilled in the art based on the embodiment is within the scope of protection required by the claims of the present invention.

Claims

1. A document classification method based on hybrid diffusion graph convolutional network, characterized by: The following steps are involved: S100: Access to public scientific paper datasets; S200: Using a node feature similarity diffusion model, the node information of the imported public scientific paper dataset is diffused and mapped, and the diffusion distance between nodes is calculated. The diffusion process is used to diffuse information between neighboring nodes in the feature space of the node, thereby obtaining new feature information of each node and obtaining a dataset that has undergone node feature space information diffusion. The node feature similarity diffusion model is: Among them, (t) (x) is the node information after the node feature space information diffusion model is used to diffuse the information between neighboring nodes in the feature space of the dataset node X, and Z represents the node information of the introduced public scientific paper dataset; λ k is the k-dimensional eigenvalue of the node feature x, t represents the number of diffusion times, ψ k (x) is the k-dimensional feature vector of node feature x; S300: Through the semi-supervised graph convolution model, the data set after S200 is subjected to information diffusion between adjacent nodes through graph convolution operation; the semi-supervised graph convolution model is as follows: In the above formula: σ(·) represents the activation function; H (l+1) H represents the (l+1)th layer output of the semi-supervised graph convolutional network; (l) Represents the lth layer output of the semi-supervised graph convolutional network, and the 0th layer output H of the semi-supervised graph convolutional network (0) That is, the initial input is the current training data set Ψ (t) (x), we get: H (0) =Ψ (t) (x); represents the first adjacency matrix with self-connection added; represents the degree matrix of the matrix; W (l) Represents the weight matrix corresponding to the lth layer of the semi-supervised graph convolutional network; S400: Using the graph convolution model in S300, a semi-supervised graph convolutional network is trained on the dataset after the mixed diffusion of S200 and S300. During the training process, the diffusion distance between nodes obtained in S200 is used to constrain the final predicted label, making the final classification surface smoother. S500: Using the graph convolutional neural network trained in step S400 to perform document classification.

2. The document classification method based on hybrid diffusion graph convolutional network according to claim 1, characterized in that: The training dataset is one or more of the Cora dataset, the Pubmed dataset, and the Citeseer dataset.

3. The document classification method based on hybrid diffusion graph convolutional network according to claim 1, characterized in that: Step S200 includes the following sub-steps: Step S201: The data set is recorded as G(X,A), X∈R n×d ,A∈R n×n , X represents the node information of the data set, A represents the adjacency matrix of the data set, n represents the number of samples in the data set, d represents the characteristic dimension of the sample, and the similarity matrix L constructed by the data set X is constructed; Step S202: normalize the similarity matrix L to obtain the Markov state transition matrix M on the data set X; Step S203: Using the obtained Markov state transition matrix M, define the diffusion distance between any two points on the adjacency graph G; Step S204: Obtain a diffusion mapping from the original data to the k-dimensional space embedded in the original space through the node feature similarity diffusion model.

4. The document classification method based on hybrid diffusion graph convolutional network according to claim 3, characterized in that: The diffusion distance between any two points on the adjacency graph G is defined as: in, m i =∑ j p ij , where p ij Represents node x i and node x j The probability of a random walk between i Represents all other nodes to node x i The sum of the probabilities of the random walks of is node x i The static distribution of N is the number of nodes, g∈N, and t represents the number of diffusion times.

5. The document classification method based on hybrid diffusion graph convolutional network according to claim 1, characterized in that: In step S400, the diffusion distance is used to regularize the loss by node feature similarity, where The loss function is as follows: In the above formula: Represents the loss function of the network; F i* represents the i-th row in the label matrix F; F j* represents the jth row in the label matrix F, y ik represents the true label distribution of the i-th sample; L is the number of labeled samples; C is the characteristic dimension of the label matrix F; N is the number of nodes; α is the balance parameter that balances these two terms; A ij is the value between node i and node j in the adjacency matrix A of the dataset; x i and x j is the node feature, x i , x j ∈X, X represents the node information of the data set, D (t) (x i ,x j ) 2 is node x i and node x j Diffusion distance obtained by diffusion mapping, t represents diffusion time; log represents logarithmic function.

6. A document classification system based on hybrid diffusion graph convolutional network, characterized by: include: The data set acquisition module is used to obtain the data set needed for training; The feature space neighborhood node information diffusion module diffuses information between the dataset nodes in the feature space neighborhood through the diffusion process of the diffusion mapping using the node feature similarity diffusion model. The feature space neighborhood node information diffusion module diffuses the node information of the introduced public scientific paper dataset using the node feature similarity diffusion model. The node feature similarity diffusion model is: Among them, (t) (x) is the node information after the node feature space information diffusion model is used to diffuse the information between neighboring nodes in the feature space of the dataset node information X, where X represents the node information of the introduced public scientific paper dataset; λ k is the k-dimensional eigenvalue of node feature x, t is the diffusion time, ψ k (x) is the k-dimensional feature vector of node feature x; The graph convolution model adjacent node information diffusion module uses a dataset that has undergone node feature similarity diffusion and uses the adjacency matrix through the graph convolution model to diffuse information between adjacent nodes. The semi-supervised graph convolution model in the graph convolution model adjacent node information diffusion module is as follows: In the above formula: σ(·) represents the activation function; H (l+1) H represents the (l+1)th layer output of the semi-supervised graph convolutional network; (l) Represents the lth layer output of the semi-supervised graph convolutional network, and the 0th layer output H of the semi-supervised graph convolutional network (0) That is, the initial input is the current training data set Ψ (t) (x), we get: H (0) =Ψ (t) (x); represents the first adjacency matrix with self-connection added; represents the degree matrix of the matrix; W (l) Represents the weight matrix corresponding to the lth layer of the semi-supervised graph convolutional network; The graph convolution training module performs semi-supervised graph convolution network training on the mixed diffusion dataset and uses the diffusion distance between nodes to constrain the final predicted labels during training. The classification module uses the network model trained by the graph convolution training module to perform document classification.

Images