A citation network article classification and recommendation method based on graph integrated neural network

Abstract

The application discloses a citation network article classification and recommendation method based on a graph integrated neural network, which comprises the following steps: performing structure enhancement on an original citation network, performing feature enhancement on initial features of the original citation network, and obtaining sub-citation networks and new features respectively; training base classifiers based on the sub-citation networks and the new features and performing deep integration, performing integration when data flow passes through each layer of base classifiers, taking an integration result as an input of a next layer, adjusting the base classifiers in training based on neighbor information, and obtaining a trained graph integrated neural network model; inputting an article to be classified of a citation network into the trained graph integrated neural network model, obtaining a probability that the article is predicted to correspond to a category, and classifying and sorting the article based on the obtained probability. The base classifiers are trained based on the sub-citation networks and the new features and deep integration is performed, so that the close neighbor information between the base classifiers influences each other, and the classification performance of the article and the effectiveness of recommendation are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of citation network article classification and recommendation, and particularly relates to a citation network article classification and recommendation method based on a graph integrated neural network. BACKGROUND

[0002] Graph neural networks are currently the most effective method for processing graph data related tasks, and downstream tasks of graphs include node classification, link prediction and graph classification. The node classification task aims to predict and classify entities (nodes) that make up the graph, for example, predicting which field a node (article) in the citation network belongs to, which can save a lot of manpower and reduce human classification errors caused by a large amount of work. According to the number of nodes available for training, node classification can be divided into unsupervised, semi-supervised and fully supervised node classification. Learning, predicting and classifying entities (nodes) in the citation network (graph) is a basic and important part of graph learning, which can effectively improve the analysis and processing efficiency of the citation network in real life, such as article classification and recommendation.

[0003] For citation networks, the integrated method can improve the classification performance of the graph neural network (GNNs) of the node (article) classification task. The existing integrated method in the citation network combines a set of trained base classifiers to form a strong classifier, that is, the final output of the combined base classifier is used for prediction. However, such an integrated method cannot effectively promote many popular GNNs, such as graph convolutional networks, graph attention networks, and graph isomorphism networks, to perform well under the heteromorphism condition, which limits their applicability. In addition, such a method only combines the final output of the base classifier, ignoring the hierarchical nature of the layer-by-layer aggregation of information by GNNs, which makes the base classifiers unable to interact when aggregating neighbors during training, that is, during training, the articles of one base classifier cannot interact with the articles of other base classifiers. Therefore, before the final output is formed, important close-range information cannot positively affect other base classifiers. These two problems are called low applicability and shallow integration. Effectively solving these two problems can improve the classification performance of GNNs in the citation network and achieve better article classification and article recommendation. SUMMARY

[0004] In order to overcome the defects and deficiencies existing in the prior art, the present application provides a citation network article classification and recommendation method based on a graph integrated neural network. The present application designs a structure enhancement method to generate a set of graphs for training different base classifiers, and designs a feature enhancement method to weaken the errors caused by the initial features. The present application introduces an integrated operation into the neighborhood aggregation period in the training process, so that important close-range information between base classifiers can influence each other, thereby improving the article classification performance and further improving the effectiveness of article recommendation.

[0005] The second object of the present application is to provide a citation network article classification and recommendation system based on graph integrated neural network.

[0006] The third object of the present application is to provide a computer readable storage medium.

[0007] The fourth object of the present application is to provide a computer device.

[0008] In order to achieve the above objects, the present application adopts the following technical solutions:

[0009] The present application provides a citation network article classification and recommendation method based on graph integrated neural network, comprising the following steps:

[0010] The original citation network is structurally enhanced, and the initial features thereof are enhanced, to obtain a sub-citation network and new features, respectively;

[0011] The base classifier is trained based on the sub-citation network and the new features and deep integration is performed, the integration is performed when the data stream passes through each layer of base classifier, the integration result is taken as the input of the next layer, the base classifier is adjusted in the training based on the neighbor information, and a trained graph integrated neural network model is obtained;

[0012] The citation network article to be classified is input into the trained graph integrated neural network model, the probability that the article is predicted to be the corresponding category is obtained, and the article is classified and sorted based on the obtained probability.

[0013] As a preferred technical solution, the structural enhancement of the original citation network specifically comprises:

[0014] On the basis of retaining all articles in the original citation network, Bernoulli sampling is performed on the citation relationship between the articles to obtain a sub-citation network of the original citation network.

[0015] As a preferred technical solution, the feature enhancement of the initial features specifically comprises:

[0016] The feature mean value of each labeled node is taken as a category vector, and the category vector is added to the articles of the respective category according to a set probability.

[0017] As a preferred technical solution, a single-layer graph convolution network is adopted as the base classifier of the graph integrated neural network, and is represented as:

[0018]

[0019] Wherein, σ is an activation function, W l is a learnable parameter of the lth layer, is a formal adjacency matrix, and D ii =∑u A iu ,A is the adjacency matrix of the citation network, I is the identity matrix, D is the degree matrix, D ii represents the value of the ith row and the ith column of the matrix D.

[0020] As a preferred technical solution, the residual connection form of the graph convolution network is as follows:

[0021]

[0022] Wherein, H l-1 is the feature matrix of the l-1 layer.

[0023] As a preferred technical solution, the base classifier is trained based on the sub-citation network and the new features and deep integration is performed, specifically including:

[0024] The number of generated sub-citation networks is equal to the number of base classifiers of each layer of the graph integrated neural network;

[0025] Set the integrated mode to integrate C base classifiers of each layer, denoted as:

[0026]

[0027] Wherein, σ is the activation function, is the output of the lth layer and the cth base classifier, α is the residual coefficient, W l,c is the lth layer and the cth learnable parameter, is the formalized adjacency matrix of the cth sub-citation network, H l-1 is the feature matrix of the l-1 layer.

[0028] As a preferred technical solution, the base classifier is trained based on the sub-citation network and the new features and deep integration is performed, and the loss function is represented as:

[0029] L GEN =L cro +γL con

[0030]

[0031] Wherein, L GEN represents the loss function of GEN, L cro represents the supervised cross-entropy loss, L con represents the unsupervised consistency regularization loss, γ represents the coefficient of balancing L cro and L con , Y L represents the article set in the labeled training set, represents the transpose of the label vector of node r, and n is the number of articles in the citation network. is changed by a sharpening trick i .

[0032] As a preferred technical solution, the citation network article to be classified is input into the trained graph integrated neural network model to obtain the probability that the article is predicted to correspond to the category, which is specifically represented as:

[0033]

[0034] wherein T is temperature, a parameter for controlling the degree of category distribution, Q is the number of categories of articles in the citation network, Z ij is the value of the i-th row and the j-th column of the model output, that is, the probability that the article i is predicted to be of the j category.

[0035] The application also provides a citation network article classification and recommendation system based on a graph integrated neural network, which is used to implement the above-mentioned citation network article classification and recommendation method based on a graph integrated neural network, and the system comprises a structure enhancement module, a feature enhancement module, a training module and a prediction module.

[0036] The structure enhancement module is used to perform structure enhancement on the original citation network.

[0037] The feature enhancement module is used to perform feature enhancement on the initial features of the original citation network to obtain a sub-citation network and new features respectively.

[0038] The training module is used to train base classifiers based on the sub-citation network and the new features and perform deep integration, the integration is performed when the data stream passes through each layer of base classifiers, the integration result is taken as the input of the next layer, the base classifiers are adjusted in the training based on the near neighbor information, and a trained graph integrated neural network model is obtained.

[0039] The prediction module is used to input the citation network article to be classified into the trained graph integrated neural network model to obtain the probability that the article is predicted to correspond to the category, and perform article classification and sorting based on the obtained probability.

[0040] The application also provides a computer readable storage medium which stores a program, and the program is executed by a processor to implement the above-mentioned citation network article classification and recommendation method based on a graph integrated neural network.

[0041] Compared with the prior art, the application has the following advantages and beneficial effects:

[0042] (1) The application is based on a new structure enhancement technique, and multiple Bernoulli sampling is performed on the edges (citation relationship between articles) in the citation network according to a given probability. Since the noise citations (edges) contained in the original citation network are few, the sub-citation network after sampling contains fewer noise edges. The group of citation networks obtained by structure enhancement is reasonable and different, and is used to train a group of base classifiers, instead of training different base classifiers with the same citation network, which ensures the effectiveness of the base classifier and improves the diversity of the base classifier, and is more in line with the idea of ensemble learning.

[0043] (2) The application is based on a new feature enhancement technique, and the error caused by the initial features of the citation network is reduced. The initial features are usually represented by the keywords of the articles in the word bag. In the real world citation network, the initial features of the same type of articles are very different or completely different, and the initial features of different types of articles are very similar or completely the same, which is contradictory to the cognition of the classification problem and will interfere with the model training. The application calculates the class features of each type of article, concatenates the class features to the initial features of each article according to a set probability, obtains new features, and then forms new initial features of the citation network. Feature enhancement focuses on the possible problem that the initial features may be contrary to the cognition of the classification problem, and then alleviates this problem through class features and class matrices.

[0044] (3) The application extends the previous graph ensemble that only focuses on the output of the base classifier to each layer in the training, fully utilizes important neighbor information based on deep ensemble, and then adjusts the base classifier in the training to obtain better and suitable base classifiers for the ensemble, and finally improves the article classification performance and the effectiveness of article recommendation within an acceptable training time. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 The flowchart of the citation network article classification and recommendation method based on the graph ensemble neural network of the application;

[0046] Figure 2 The architecture diagram of the graph ensemble neural network model of the application;

[0047] Figure 3 The flowchart of the feature enhancement algorithm AugmFeat of the application algorithm 1;

[0048] Figure 4 The flowchart of the structure enhancement algorithm AugmStru of the application algorithm 2;

[0049] Figure 5 The flowchart of the graph ensemble neural network training process of the application algorithm 3;

[0050] Figure 6This is a diagram showing the comparison of article classification accuracy across three publicly available citation networks. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0052] Example 1

[0053] like Figure 1 As shown, this embodiment provides a citation network article classification and recommendation method based on graph ensemble neural networks, including the following steps:

[0054] S1: The original citation network is structurally enhanced, and its initial features are feature-enhanced to obtain a set of sub-citation networks and new features respectively;

[0055] like Figure 3 As shown, in Algorithm 1, feature enhancement specifically involves: first, calculating the category vector, that is, taking the average feature value of each labeled node as the category vector, and then adding the category vector to the articles of each category according to the set probability, so that the features between the same category are more similar, that is, their keywords are more similar.

[0056] like Figure 4 As shown, in Algorithm 2, structural enhancement involves Bernoulli sampling of the citation relationships between articles while retaining all articles in the original citation network. That is, some original citation relationships are retained while some citation relationships are deleted. The resulting new network must be a subnetwork of the original citation network.

[0057] S2: Train a set of base classifiers with a set of subcitation networks and new features, and perform deep ensemble so that the base classifiers can interact during training;

[0058] Previous shallow ensembles that only focused on the outputs of base classifiers have been extended to deep ensembles that integrate during training. Feature augmentation mitigates potential errors from initial features, and structural augmentation generates graphs for training different base classifiers. The entire model for article classification and recommendation consists of a certain number of layers, each containing a certain number of base classifiers, such as... Figure 2As shown, integration is performed when data flows through each layer of base classifier, the integration result is used as the input of the next layer, during training, the final loss is obtained according to a series of integration results, so when adjusting the parameters by back propagation, the base classifiers can affect each other. In addition to the cross-entropy loss function, the graph integration neural network for article classification and recommendation in the application adds a consistency regularization loss to assist the training of the model. After the model is trained, the network and its initial features to be classified and recommended are input into the model to obtain a feature matrix (Z in formula (1)), and according to the matrix, the articles in the network can be classified and recommended. Figure 2

[0059] As shown in Figure 2 , the graph integration neural network for article classification and recommendation in the citation network is divided into two main stages, namely the graph data enhancement stage and the graph integration training stage, wherein the former includes graph feature enhancement and graph structure enhancement, and the latter expands the shallow integration of only focusing on the output of the base classifier to deep integration in training.

[0060] Specifically, the graph integration neural network proposed in the embodiment is abbreviated as GEN, the articles in the citation network are referred to as nodes, the representation of the keywords of the articles in the word bag is referred to as a feature vector, and the matrix composed of the feature vectors of all articles is referred to as a feature matrix. The input of GEN is the original citation network and the initial feature matrix thereof, and the output is a matrix that can be used to predict the classification of nodes, and the graph feature enhancement makes the features of nodes of the same class more similar through the class feature vector and the class feature matrix. The graph structure enhancement of GEN generates a group of different sub-citation networks from the input citation network for training base classifiers, ensuring that the base classifiers are good and different, and GEN expands the shallow integration to deep integration in training, and the deep integration is spread to each message passing layer in the graph neural network, so that the base classifiers can affect and correct each other through important neighbor information during the training process, and generate better models;

[0061] In the embodiment, the edge homomorphism ratio h is the proportion of intra-class edges in the total edges in the graph, which is often used to judge whether the graph is homomorphic and help to judge the applicable range of certain models, and is used in the application to calculate the proportion of citation relationships in the citation network, and then help to judge whether the network is homomorphic, and the calculation method is as follows (formula 1):

[0062]

[0063] Wherein, u and v represent articles, Y u is the label vector of article u, and E is the edge set of the graph, that is, the set of citation relationships. The application uses a single-layer graph convolutional network as the base classifier of GEN, which has the following form (formula 2):

[0064]

[0065] ​​

[0066] where σ is an activation function, W l is the learnable parameter of the l-th layer, H l-1 is the feature matrix of the l-1-th layer, is the formalized adjacency matrix, which is calculated as follows (Equation 3):

[0067]

[0068] where, and D ii =∑ u A iu , A is the adjacency matrix of the citation network, I is the identity matrix, D is the degree matrix, D ii denotes the value of the i-th row and i-th column of the matrix D. DropEdge is a simple variant of the graph convolution network, which slows down the convergence speed by discarding some edges, thereby alleviating the over-smoothing problem, and its form is as follows (Equation 4):

[0069]

[0070] where, is the formalized adjacency matrix after discarding some edges, and the structure of GEN enhances and has similarities with it. This embodiment also imitates the residual connection in the graph convolution network and designs a similar residual connection for GEN to improve its applicability and depth and increase the influence of the initial features and feature enhancement. The form of the residual connection in the graph convolution network is as follows (Equation 5):

[0071]

[0072] In this embodiment, the graph data enhancement stage of GEN consists of graph feature enhancement and graph structure enhancement, as in Algorithms 1 AugmFeat and 2 AugmStru in Figure 3 and Figure 4 . Algorithm 1 consists of two large loops with a small loop embedded in each, the first large loop is used to calculate the class feature vector, i.e., lines 3-11 of the algorithm, and the second large loop is used to enhance the features according to the probability, i.e., lines 13-22 of the algorithm. Algorithm 2 consists of two large loops, the first large loop has a small loop embedded in it, which is used to generate a set of graphs, i.e., lines 1-8 of the algorithm, and the second large loop adds all known intra-class edges to the generated graph, i.e., lines 10-17 of the algorithm. Algorithm 3 GEN is the end-to-end data processing, model training and output process of GEN, where the training process of lines 9-16 mentions three formulas, which are important parts of the form and training of GEN;

[0073] When the original citation network and its initial features are output to the GEN, the structure enhancement and feature enhancement stages of the GEN's graph data augmentation phase generate a set of sub-citation networks and a new feature matrix, respectively, for training the base classifiers. This stage is the data processing procedure of the graph ensemble training stage. The number of sub-citation networks generated should be equal to the number of base classifiers in each layer of the GEN, and the number of ensemble operations should be equal to the depth (number of layers) of the GEN. Therefore, for an L-layer GEN, there are a total of L×C base classifiers, that is, C base classifiers in each layer. Thus, the graph data augmentation stage also needs to generate C additional, different citation networks. This invention chooses an averaging method, similar to a voting method, as the ensemble method to integrate the C base classifiers in each layer, as shown in the following form (Formula 6):

[0074]

[0075] Where σ is the activation function, This is the output of the c-th base classifier in the l-th layer. This embodiment incorporates residual connections to extract information from previous layers to maintain the effect of feature enhancement. The calculation method is as follows (Formula 7):

[0076]

[0077] Where α is the residual coefficient, W l,c It is the c-th learnable parameter in the l-th layer. It is the formalized adjacency matrix of the c-th sub-citation network. Besides Equations 6 and 7, such as... Figure 5 As shown in Algorithm 3, Formula 8, which is the calculation of GEN's loss, consists of two parts: supervised cross-entropy loss and unsupervised consistency regularization loss. The calculation method is as follows (Formula 8):

[0078] L GEN =L cro +γL con

[0079] Where L GEN Let L represent the loss function of GEN. cro L represents the supervised cross-entropy loss. con This represents the unsupervised consistency regularization loss, where γ is used to balance L. cro and L con The coefficient of L. cro The calculation is based on labeled nodes, hence it is called supervised, and its calculation method is as follows (Equation 9):

[0080]

[0081] Where Y LY represents a set of articles in the labeled training set r ⊥ represents the transpose of the label vector of node r, the cross-entropy loss function is often used as the loss function of the classification problem. For the graph neural network model that performs data augmentation to generate multiple graphs or features, the consistency regularization loss function L con is commonly used to balance the effectiveness of multiple graphs or features. Therefore, the present application uses L con to the classification and recommendation problem of GEN on the citation network, the present application adds it to the cross-entropy loss function by the coefficient γ, and obtains the loss function of GEN, L con The calculation method of L

[0082]

[0083] where n is the number of articles in the citation network, is obtained by sharpening the technique of Z i , is the probability that node i is predicted to be class j, it can be found that the calculation of L con involves all articles, including unlabeled articles, The calculation method of L

[0084]

[0085] where T is the temperature, a parameter for controlling the degree of class distribution, Q is the number of classes of articles in the citation network, Z ij is the value of the i-th row and the j-th column of the model output, i.e. the probability that article i is predicted to be class j, is the processed Z ij .

[0086] The present embodiment compares with fifteen published benchmark models on three public citation networks to verify the advancement of GEN proposed by the present application, the present embodiment takes the classification accuracy of the model for the articles in the citation network as the comparison standard, and the comparison result is shown in Figure 6 . Figure 6 It is shown that GEN achieves the best article classification result in the three public citation networks, exceeding the fifteen benchmark models compared by the present application, including early graph neural network models based on space and based on spectrum, recent graph neural network models based on space and based on polynomial spectrum, etc. The comparison test result verifies the advancement of GEN in article classification, and the present embodiment can be used for article recommendation by simply transforming GEN, i.e. the same category of articles should be recommended to the reader.

[0087] In summary, the present application extends the shallow ensemble that only focuses on the base classifier output to a deep ensemble in training, so that the base classifiers can interact in training, and important neighbor information can affect other classifiers. In order to train good and different base classifiers, the present application uses a single-layer graph convolution network as a base classifier, designs a structure enhancement technique to generate a set of different graphs from a source graph, which are used to train the base classifier. In order to weaken the error that the initial features may bring (the initial features of the nodes of the same class are not similar, even completely different), the present application designs a feature enhancement, which improves the similarity of the features of the nodes of the same class through a class vector and a matrix, weakens the error caused by the initial features, and designs a residual connection to maintain this effect when the model is deep.

[0088] Embodiment 2

[0089] The embodiment provides a citation network article classification and recommendation system based on a graph ensemble neural network, which is used for the citation network article classification and recommendation method based on the graph ensemble neural network in Embodiment 1, and the system comprises a structure enhancement module, a feature enhancement module, a training module and a prediction module.

[0090] In the embodiment, the structure enhancement module is used for performing structure enhancement on an original citation network.

[0091] In the embodiment, the feature enhancement module is used for performing feature enhancement on initial features of the original citation network, and obtaining a sub-citation network and new features.

[0092] In the embodiment, the training module is used for training base classifiers based on the sub-citation network and the new features and performing deep integration, performing integration when data flows through each layer of the base classifiers, taking an integration result as an input of a next layer, adjusting the base classifiers in training based on neighbor information, and obtaining a trained graph ensemble neural network model.

[0093] In the embodiment, the prediction module is used for inputting a citation network article to be classified into the trained graph ensemble neural network model, obtaining a probability that the article is predicted to be a corresponding category, and performing article classification and sorting based on the obtained probability.

[0094] Embodiment 3

[0095] The embodiment provides a storage medium, which can be a ROM, a RAM, a magnetic disk, an optical disk or the like storage medium. The storage medium stores one or more programs, and the programs are executed by a processor to implement the citation network article classification and recommendation method based on the graph ensemble neural network in Embodiment 1.

[0096] The above embodiments are the preferred embodiments of the present application, but the embodiments of the present application are not limited to the above embodiments, and any changes, modifications, substitutions, combinations, simplifications, etc. made without departing from the spirit and principles of the present application should be equivalent replacement manners and should be included in the protection scope of the present application.

Claims

1. A citation network article classification and recommendation method based on graph integrated neural network, characterized in that, Comprise the following steps: The original citation network is structurally enhanced, and the initial features thereof are feature-enhanced, to obtain a sub-citation network and new features respectively; The base classifiers are trained based on the sub-citation network and the new features and deep integration is performed, the integration is performed when the data stream passes through each layer of base classifier, the integration result is taken as the input of the next layer, the base classifier is adjusted in the training based on the neighbor information, to obtain the trained graph integration neural network model; The single-layer graph convolution network is adopted as the base classifier of the graph integration neural network, and is represented as: where σ is an activation function, W l is a learnable parameter of the lth layer, is a formalized adjacency matrix, and D ii =∑ u A iu D is the adjacency matrix of the citation network, I is the identity matrix, D is the degree matrix, D ii denotes the value of the i-th row and i-th column of the matrix D; The residual connection form of the graph convolution network is as follows: wherein is a characteristic matrix of the l-1 layer; The base classifiers are trained based on the sub-citation network and the new features and deep integration is performed, specifically comprising: The number of generated sub-citation networks is equal to the number of base classifiers of each layer of graph integration neural network; The C base classifiers of each layer are integrated in the set integration mode, and is represented as: where σ is an activation function, is the output of the lth layer cth base classifier, a is a residual coefficient, W l,c is the lth layer cth learnable parameter, is the formalized adjacency matrix of the cth sub-citation network, H l-1 is the feature matrix of the l-1th layer; The citation network article to be classified is input into the trained graph integration neural network model, to obtain the probability that the article is predicted to be the corresponding category, and the article is classified and sorted based on the obtained probability. 2.The graph-based integrated neural network-based citation network article classification and recommendation method according to claim 1, characterized in that, The original citation network is structurally enhanced, specifically comprising: The citation relationship between the articles is Bernoulli sampled based on the reservation of all articles in the original citation network, to obtain the sub-citation network of the original citation network. 3.The graph-based integrated neural network based citation network article classification and recommendation method according to claim 1, characterized in that, The initial features of the original citation network are feature-enhanced, specifically comprising: The feature mean of each labeled node is taken as a category vector, and the category vector is added to the article of the respective category according to the set probability. 4.The graph-based integrated neural network based citation network article classification and recommendation method according to claim 1, characterized in that, The base classifiers are trained based on the sub-citation network and the new features and deep integration is performed, and the loss function is represented as: L GEN = L cro + γL con where L GEN represents the loss function of GEN, L cro represents the supervised cross-entropy loss, L con represents the unsupervised consistency regularization loss, and γ represents the coefficient balancing L cro and L con , Y L represents the set of articles in the labeled training set, represents the transpose of the label vector of node r, and n is the number of articles in the citation network, is obtained by changing Z i through a sharpening trick. 5.The graph-based integrated neural network based citation network article classification and recommendation method according to claim 4, characterized in that, The citation network article to be classified is input into the trained graph integration neural network model, to obtain the probability that the article is predicted to be the corresponding category, and the article is classified and sorted based on the obtained probability. where T is the temperature, a parameter used to control the extent of the class distribution, Q is the number of classes of articles in the citation network, Z ij is the value of the i-th row and j-th column of the model output, i.e., the probability that article i is predicted to be of class j.

6. A citation network article classification and recommendation system based on graph integrated neural networks, characterized in that, The system is used for implementing the citation network article classification and recommendation method based on the graph integration neural network of any one of claims 1-5, and comprises a structure enhancement module, a feature enhancement module, a training module and a prediction module. The structure enhancement module is used for structurally enhancing the original citation network. The feature enhancement module is used for feature-enhancing the initial features of the original citation network, to obtain a sub-citation network and new features respectively. The training module is used for training the base classifiers based on the sub-citation network and the new features and performing deep integration, the integration is performed when the data stream passes through each layer of base classifier, the integration result is taken as the input of the next layer, the base classifier is adjusted in the training based on the neighbor information, to obtain the trained graph integration neural network model. The prediction module is used for inputting the citation network article to be classified into the trained graph integration neural network model, to obtain the probability that the article is predicted to be the corresponding category, and the article is classified and sorted based on the obtained probability.

7. A computer-readable storage medium storing a program, characterized in that, The program is executed by the processor to implement the citation network article classification and recommendation method based on the graph integration neural network of any one of claims 1-5.

Images