Paper recommendation method based on directed graph convolutional network

Abstract

The paper recommendation method based on the directed graph convolutional network comprises the following steps: constructing a literature node feature matrix; constructing a literature adjacency matrix and constructing a literature sparse adjacency matrix for storage; constructing a DAGCN network to generate literature nodes containing directional information, and performing category division on the literature nodes; performing category division on the new literature; a DAGRaphSAGE model is constructed to be used for link prediction of new literatures; generating a category prediction probability of the new literature node and each existing literature node; updating the quotation relationship; the similarity between the new literature node and the existing literature node is evaluated through the category prediction probability, the category prediction probability is larger than or equal to the boundary threshold value, it is indicated that the existing literature node corresponding to the category prediction probability is related to the academic content of the new literature node, and the academic literature database recommends the existing literature to a user. According to the method, literatures related to user research are accurately recommended based on the similarity between literature nodes and dynamic changes in a reference network, and it is ensured that the recommended literatures have high academic relevance, influence and timeliness.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular relates to a paper recommendation method based on a directed graph convolutional network. Background Art

[0002] Existing academic citation network analysis mainly relies on traditional graph algorithms and static graph structures, which are difficult to adapt to the dynamic changes of citation relationships. With the development of academic research, new documents are constantly added to the citation network, resulting in a large number of complex citation relationships, which puts higher requirements on the real-time update and accurate prediction of academic networks. Current citation network analysis is usually based on static graph models and cannot cope with the structural changes of citation networks over time.

[0003] In recent years, graph neural networks (GNNs) have shown superior performance in the representation and analysis of graph structured data, especially in the fields of social network analysis, recommendation systems, and knowledge graphs. Traditional GNN models are mostly based on undirected graphs and cannot effectively capture the directional characteristics of citation networks, which limits the performance of the model when processing academic citation networks. The academic citation network is not just a static graph structure, but also contains multi-dimensional information such as time and citation direction, which is crucial for accurately predicting citation relationships and making literature recommendations.

[0004] The academic citation network is a dynamic network that is constantly expanding and updating. With the continuous addition of new documents, existing GNN models usually require a lot of computing resources to process large-scale data, and it is difficult to effectively adjust model parameters in real-time data streams. Therefore, when facing new research fields, new documents and terms, how to accurately capture the characteristics of new documents and classify them is still an important problem faced by current methods.

[0005] The Journal of Northwest Minzu University, No. 3, 2023, published a study titled "Research on Academic Paper Recommendation Methods Based on Machine Learning". It is disclosed that existing recommendation methods face limitations of collaborative filtering and graph recommendation, especially when faced with data scarcity or cold start problems, the recommendation effect is not good. Traditional collaborative filtering methods rely on user historical behavior data to predict what users may be interested in. However, when faced with new research fields or scholars who have just joined the academic network, the lack of historical data makes it impossible for collaborative filtering methods to effectively recommend relevant papers. At the same time, although graph recommendation methods can use graph structures to capture the citation relationships between papers, they usually rely on static graph structures, which makes them unable to quickly adapt to structural changes in dynamic academic citation networks. Especially when research in new fields has not yet formed a relatively complete citation network, traditional recommendation methods cannot fully explore the potential correlation between papers. Therefore, how to make accurate paper recommendations in data scarcity and cold start situations is still a difficulty in academic recommendation systems. Summary of the invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method for paper recommendation based on a directed graph convolutional network that can fully utilize the directional relationship and similarity information between nodes when processing the dynamic changes of the academic citation network, so as to achieve efficient and accurate paper recommendation.

[0007] The technical solution adopted to solve the above technical problems is: a paper recommendation method based on directed graph convolutional network, including the following steps:

[0008] Step 1. Construct the document node feature matrix

[0009] The keyword features of each document are extracted from the academic literature database through the natural language processing method NLP, and the keyword feature vector is generated as the document node feature using the BERT model. The publication time of each document is recorded as the document node time label, and the document node feature matrix is ​​constructed through the document node features and document node time labels.

[0010] Step 2. Construct the document adjacency matrix and construct the document sparse adjacency matrix for storage

[0011] Based on the citation relationship of the literature in the academic literature database, the literature adjacency matrix A is constructed. The size of the literature adjacency matrix A is N×N, and the elements are A ij , N represents the number of documents, A ij Indicates whether document node i cites document node j. If document node i cites document node j, then A ij The value of is 1; otherwise it is 0;

[0012] Based on the document adjacency matrix A, the CSR method is used to construct the document sparse adjacency matrix A. * To store;

[0013] Step 3. Build a DAGCN network to generate document nodes containing directional information and classify the document nodes

[0014] The DAGCN network replaces the graph convolution module in the GCN network with a directed graph convolution module, which is used to perform weighted aggregation of first-order neighbors and second-order similarity neighbors on the input document node features to generate document node features containing directional information;

[0015] The document node feature matrix and document sparse adjacency matrix A *Input into the DAGCN network, perform first-order neighbor propagation and second-order similar neighbor propagation, the DAGCN network aggregates the first-order neighbor propagation results and the second-order similar neighbor propagation results, updates the document node feature matrix, and classifies the document nodes based on the similarity between the document node features and the logical relationship of the time sequence, and outputs the document category classification structure diagram;

[0016] Step 4. Categorize new documents

[0017] Step 4.1. Introduce new documents as new document nodes, construct keyword feature vectors and time labels for the new document nodes to generate new document node feature matrices;

[0018] Step 4.2. According to the similarity between the new document node and each document category in the category division structure diagram, the new document node is classified into the document category with the greatest similarity;

[0019] Step 5. Build DAGraphSAGE model for link prediction of new documents

[0020] The DAGraphSAGE model is composed of a link prediction directed graph convolution layer, a feature nonlinear activation module, a Dropout module, and a Softmax classification layer connected in sequence; the link prediction directed graph convolution layer is used to aggregate the features of the first-order neighbors and second-order similarity neighbors of each document node in link prediction to generate document node features containing directional information; the feature nonlinear activation module introduces nonlinearity through the ReLU activation function to enhance the expressiveness of the model; the Dropout module is used to randomly discard document node features to prevent model overfitting; the Softmax classification layer converts the feature vector of the document node into a category prediction probability distribution;

[0021] Step 6. Generate the category prediction probability of the new document node and each existing document node

[0022] Step 6.1. Connect the new document node with other existing document nodes in its category based on the time tag and construct a new document adjacency matrix with each existing document node and convert it into a new document sparse adjacency matrix;

[0023] Step 6.2. Input the sparse adjacency matrix of all new documents into the DAGraphSAGE model and output the category prediction probability of the new document node and each existing document node;

[0024] Step 7. Citation relationship update

[0025] Preset the demarcation threshold, compare each category prediction probability with the demarcation threshold, and if the category prediction probability is ≥ the demarcation threshold, generate a citation relationship between the new document node and the existing document node corresponding to the category prediction probability. The citation direction is determined according to the time label sequence, and the newly generated citation relationship is added to the citation network to update the document graph structure.

[0026] Step 8. Evaluate the similarity between the new document node and the existing document node through the category prediction probability. If the category prediction probability is ≥ the demarcation threshold, it means that the existing document node corresponding to the category prediction probability is related to the academic content of the new document node, and the academic document database recommends the existing document to the user.

[0027] As a preferred technical solution, the first-order neighbor propagation includes forward propagation from the source document node to the target document node and reverse propagation from the target document node to the source document node, and the result of the forward propagation is combined with the result of the reverse propagation to obtain a first-order neighbor propagation result representation;

[0028] The forward propagation passes through the sparse adjacency matrix A * The propagation is performed according to the following formula to capture the citation relationship from the source document node to the target document node:

[0029]

[0030] In the formula, represents the feature matrix from the source document node to the target document node after updating in the network layer l+1, σ(·) represents the activation function, represents the result of normalized forward propagation of the sparse adjacency matrix of the document, H (l) represents the feature matrix of the document node at the lth layer of the network, W (l) Represents the weight matrix of the network layer l;

[0031] The back propagation is done by transposing the sparse adjacency matrix A *T The propagation is performed according to the following formula to capture the citation relationship from the target document node to the source document node:

[0032]

[0033] In the formula, represents the feature matrix from the target document node to the source document node after updating at the l+1th layer of the network, Represents the result of the transposed back-propagation normalization of the sparse adjacency matrix of the document;

[0034] The first-order neighbor propagation result H (l+1) for:

[0035]

[0036] In the formula, α is a hyperparameter that controls the weight ratio of forward and backward propagation;

[0037] The second-order similarity neighbor propagation is used to capture the similarity between two document nodes that are not directly connected but have common in-degree neighbors or out-degree neighbors, including second-order in-degree similarity propagation and second-order out-degree similarity propagation;

[0038] The second-order in-degree similarity propagation is:

[0039]

[0040] In the formula, represents the weighted adjacency matrix, Norm(·) represents the normalization process, (1-δ) represents the removal of self-loops to prevent itself from becoming its own first-order neighbor, The document node feature matrix representing the l+1th layer of the network updates the document node features by combining the information of the second-order in-degree similar neighbor document nodes;

[0041] The second-order out-degree similarity propagation is:

[0042]

[0043] In the formula, The document node feature matrix representing the l+1th layer of the network updates the document node features by combining the information of the second-order out-degree similar neighbor document nodes;

[0044] The DAGCN network aggregates the first-order neighbor propagation results and the second-order similar neighbor propagation results as follows:

[0045]

[0046] In the formula, represents the node feature matrix of the l+1th layer after aggregation, β represents the weighting coefficient, α1 represents the contribution weight of the two-directional information sources of the first-order neighbors, and α2 represents the contribution weight of the two-directional information sources of the second-order similar neighbors.

[0047] As a preferred technical solution, the similarity between the new document node in step 4.2 and each document category in the category division structure diagram is the cosine similarity between the new node and the category center of each document category, and the category center of each document category is the average value of the feature vectors of all document nodes in that category.

[0048] As a preferred technical solution, the preset demarcation threshold is 60%.

[0049] The beneficial effects of the present invention are as follows:

[0050] The directed graph convolutional network (DAGCN) adopted in the present invention can efficiently capture the directional characteristics in the academic citation network, thereby adapting to the dynamic changes in the academic network and timely updating the characteristics and citation relationships of the document nodes.

[0051] Based on the similarity between document nodes and the dynamic changes in the citation network, the present invention can accurately recommend documents related to user research, ensure that the recommended documents have high academic relevance, influence and timeliness, and have the characteristics of high recommendation accuracy.

[0052] The present invention can efficiently process large-scale academic data and improve the accuracy of node classification and link prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flowchart of the paper recommendation method based on a directed graph convolutional network of the present invention.

[0054] Figure 2 It is a graph structure evolution diagram of the paper recommendation method based on directed graph convolutional network of the present invention.

[0055] Figure 3 This is a schematic diagram of the first-order neighbor and second-order similar neighbor analysis. DETAILED DESCRIPTION

[0056] The present invention will be further described in detail below in conjunction with the accompanying drawings and examples, but the present invention is not limited to the following embodiments.

[0057] exist Figure 1 In the embodiment, the paper recommendation method based on directed graph convolutional network includes the following steps:

[0058] Step 1. Construct node feature matrix

[0059] The keyword features of each document are extracted from the academic document database through the natural language processing method NLP, and the keyword feature vector is generated by the BERT model as the document node feature. The publication time of each document is recorded as the document node time label. The time label can ensure the time directionality of the citation relationship and prevent the "time loop" phenomenon, that is, the citation relationship that violates the time logic. The document node feature matrix is ​​constructed through the document node features and document node time labels;

[0060] Step 2. Construct the document adjacency matrix and construct the document sparse adjacency matrix for storage

[0061] Based on the citation relationship of the literature in the academic literature database, the literature adjacency matrix A is constructed. The size of the literature adjacency matrix A is N×N, and the elements are A ij , N represents the number of documents, A ijIndicates whether document node i cites document node j. If document node i cites document node j, then A ij The value of is 1; otherwise it is 0;

[0062] Based on the document adjacency matrix A, the CSR method is used to construct the document sparse adjacency matrix A. * To store;

[0063] Step 3. Build a DAGCN network to generate document nodes containing directional information and classify the document nodes

[0064] The DAGCN network replaces the graph convolution module in the GCN network with a directed graph convolution module. The directed graph convolution module is used to perform weighted aggregation of the first-order neighbors and second-order similarity neighbors of the input document node features to generate document node features containing directional information;

[0065] The document node feature matrix and document sparse adjacency matrix A * Input into the DAGCN network, perform first-order neighbor propagation and second-order similar neighbor propagation, the DAGCN network aggregates the first-order neighbor propagation results and the second-order similar neighbor propagation results, updates the document node feature matrix, and classifies the document nodes based on the similarity between the document node features and the logical relationship of the time sequence, and outputs the document category classification structure diagram;

[0066] Among them, the first-order neighbor propagation includes forward propagation from the source document node to the target document node and backward propagation from the target document node to the source document node, and the results of the forward propagation and the backward propagation are combined to obtain the first-order neighbor propagation result representation.

[0067] Forward propagation through the sparse adjacency matrix A of the document * The propagation is performed according to the following formula to capture the citation relationship from the source document node to the target document node:

[0068]

[0069] In the formula, represents the feature matrix from the source document node to the target document node after the update in the l+1th layer of the DAGCN network, σ(·) represents the activation function, represents the result of normalized forward propagation of the sparse adjacency matrix of the document, H (l) represents the feature matrix of the document node at the lth layer of the DAGCN network, W (l) Represents the weight matrix of the lth layer of the DAGCN network.

[0070] Back propagation through the document adjacency matrix transpose A *T The propagation is performed according to the following formula to capture the citation relationship from the target document node to the source document node:

[0071]

[0072] In the formula, represents the feature matrix from the target document node to the source document node after updating in the l+1th layer of the DAGCN network, Represents the result of the transposed back-propagation normalization of the sparse adjacency matrix of the document;

[0073] First-order neighbor propagation result H (l+1) for:

[0074]

[0075] In the formula, α is a hyperparameter that controls the weight ratio of forward and backward propagation.

[0076] Second-order similarity neighbor propagation, used to capture the similarity between two document nodes that are not directly connected but have common in-degree neighbors or out-degree neighbors, including second-order in-degree similarity propagation and second-order out-degree similarity propagation;

[0077] The second-order in-degree similarity propagation is:

[0078]

[0079] In the formula, represents the weighted adjacency matrix, Norm(·) represents the normalization process, (1-δ) represents the removal of self-loops to prevent itself from becoming its own first-order neighbor, Represents the feature matrix of the document node at the l+1th layer of the DAGCN network. The document node features are updated through the information of the second-order in-degree similar neighbor nodes;

[0080] The second-order out-degree similarity propagation is:

[0081]

[0082] In the formula, The document node feature matrix representing the l+1th layer of the network updates the document node features by combining the information of the second-order out-degree similar neighboring document nodes.

[0083] The DAGCN network aggregates the first-order neighbor propagation results and the second-order similar neighbor propagation results into the following formula:

[0084]

[0085] In the formula, represents the node feature matrix of the l+1th layer after aggregation, β represents the weighting coefficient, which is used to balance the contribution weights of the first-order neighbors and the second-order similar neighbors, where β∈[0,1]. α1 represents the contribution weights of the first-order neighbors in two directions, and α2 represents the contribution weights of the second-order similar neighbors in two directions, α1, α2∈[0,1].

[0086] Examples of first-order neighbor propagation and second-order similarity neighbor analysis, such as Figure 3 :

[0087] right Figure 3 Analyze the first-order neighbors of node 1 and node 2:

[0088] 1. The first-order out-degree neighbors of node 1: nodes 3, 4, 5, 9

[0089] 2. First-order in-degree neighbors of node 1: nodes 6, 7

[0090] 3. The first-order out-degree neighbors of node 2: nodes 3, 4, 5

[0091] 4. First-order in-degree neighbors of node 2: nodes 6, 7

[0092] Analyze the second-order similar neighbors of node 1 and node 2:

[0093] 1. The second-order similarity out-degree neighbors of node 1 and node 2: nodes 3, 4, and 5

[0094] 2. The second-order similar in-degree neighbors of node 1 and node 2: nodes 6 and 7.

[0095] Step 4. Categorize new documents

[0096] Step 4.1. Introduce new documents as new document nodes, construct keyword feature vectors and time labels for the new document nodes to generate new document node feature matrices;

[0097] Step 4.2. According to the similarity between the new document node and each document category in the category division structure diagram, the similarity is the cosine similarity between the new node and the category center of each document category, and the category center of each document category is the average value of the feature vectors of all document nodes in the category, the new document node is classified into the document category with the greatest similarity;

[0098] Step 5. Build DAGraphSAGE model for link prediction of new documents

[0099] The DAGraphSAGE model consists of a link prediction directed graph convolution layer, a feature nonlinear activation module, a Dropout module, and a Softmax classification layer connected in sequence; the link prediction directed graph convolution layer is used to aggregate the features of the first-order neighbors and second-order similarity neighbors of each document node in link prediction to generate document node features containing directional information; the feature nonlinear activation module introduces nonlinearity through the ReLU activation function to enhance the expressiveness of the model; the Dropout module is used to randomly discard document node features to prevent model overfitting; the Softmax classification layer converts the feature vector of the document node into a category prediction probability distribution.

[0100] Among them, the link prediction directed graph convolution layer aggregates the features of the first-order neighbors and second-order similarity neighbors of each document node in link prediction, and the method of generating document node features containing directional information is the same as the method of aggregating the first-order neighbor features and second-order similarity neighbor features of each document node in the DAGCN network.

[0101] Step 6. Generate the category prediction probability of the new document node and each existing document node

[0102] Step 6.1. Connect the new document node with other existing document nodes in its category based on the time tag and construct a new document adjacency matrix with each existing document node and convert it into a new document sparse adjacency matrix;

[0103] Step 6.2. Input the sparse adjacency matrix of all new documents into the DAGraphSAGE model and output the category prediction probability of the new document node and each existing document node;

[0104] Step 7. Citation relationship update

[0105] The preset demarcation threshold is 60%. The predicted probability of each category is compared with the demarcation threshold. If the predicted probability of a category is ≥ the demarcation threshold, a citation relationship between a new document node and the existing document node corresponding to the predicted probability of the category is generated. The citation direction is determined according to the time label sequence. The newly generated citation relationship is added to the citation network to update the document graph structure.

[0106] Step 8. Evaluate the similarity between the new document node and the existing document node through the category prediction probability. If the category prediction probability is ≥ the demarcation threshold, it means that the existing document node corresponding to the category prediction probability is related to the academic content of the new document node, and the academic document database recommends the existing document to the user.

[0107] Whenever new documents are added, the document graph structure is updated according to steps 4, 6, and 7 to continuously optimize the recommendation results, ensure the timeliness, accuracy, and relevance of the recommended documents, and ensure that scholars are provided with the latest and most relevant academic resources.

Claims

1. A paper recommendation method based on directed graph convolutional network, characterized by: The following steps are involved: Step 1. Construct the document node feature matrix The keyword features of each document are extracted from the academic literature database through the natural language processing method NLP, and the keyword feature vector is generated as the document node feature using the BERT model. The publication time of each document is recorded as the document node time label, and the document node feature matrix is ​​constructed through the document node features and document node time labels. Step 2. Construct the document adjacency matrix and construct the document sparse adjacency matrix for storage Based on the citation relationship of the documents in the academic literature database, the document adjacency matrix A is constructed. The size of the document adjacency matrix A is N×N, and the elements are A ij , N represents the number of documents, A ij Indicates whether document node i cites document node j. If document node i cites document node j, then A ij The value of is 1; otherwise it is 0; Based on the document adjacency matrix A, the CSR method is used to construct the document sparse adjacency matrix A. * To store; Step 3. Build a DAGCN network to generate document nodes containing directional information and classify the document nodes The DAGCN network replaces the graph convolution module in the GCN network with a directed graph convolution module, which is used to perform weighted aggregation of first-order neighbors and second-order similarity neighbors on the input document node features to generate document node features containing directional information; The document node feature matrix and document sparse adjacency matrix A * Input into the DAGCN network, perform first-order neighbor propagation and second-order similar neighbor propagation, the DAGCN network aggregates the first-order neighbor propagation results and the second-order similar neighbor propagation results, updates the document node feature matrix, and classifies the document nodes based on the similarity between the document node features and the logical relationship of the time sequence, and outputs the document category classification structure diagram; Step 4. Categorize new documents Step 4.

1. Introduce new documents as new document nodes, construct keyword feature vectors and time labels for the new document nodes to generate new document node feature matrices; Step 4.

2. According to the similarity between the new document node and each document category in the category division structure diagram, the new document node is classified into the document category with the greatest similarity; Step 5. Build DAGraphSAGE model for link prediction of new documents The DAGraphSAGE model is composed of a link prediction directed graph convolution layer, a feature nonlinear activation module, a Dropout module, and a Softmax classification layer connected in sequence; the link prediction directed graph convolution layer is used to aggregate the features of the first-order neighbors and second-order similarity neighbors of each document node in link prediction to generate document node features containing directional information; feature The nonlinear activation module introduces nonlinearity through the ReLU activation function to enhance the expressiveness of the model; the Dropout module is used to randomly discard the features of the document nodes to prevent the model from overfitting; the Softmax classification layer converts the feature vector of the document node into the category prediction probability distribution; Step 6. Generate the category prediction probability of the new document node and each existing document node Step 6.

1. Connect the new document node with other existing document nodes in its category based on the time tag and construct a new document adjacency matrix with each existing document node and convert it into a new document sparse adjacency matrix; Step 6.

2. Input the sparse adjacency matrix of all new documents into the DAGraphSAGE model and output the category prediction probability of the new document node and each existing document node; Step 7. Citation relationship update Preset the demarcation threshold, compare each category prediction probability with the demarcation threshold, and if the category prediction probability is ≥ the demarcation threshold, generate a citation relationship between the new document node and the existing document node corresponding to the category prediction probability. The citation direction is determined according to the time label sequence, and the newly generated citation relationship is added to the citation network to update the document graph structure. Step 8. Evaluate the similarity between the new document node and the existing document node through the category prediction probability. If the category prediction probability is ≥ the demarcation threshold, it means that the existing document node corresponding to the category prediction probability is related to the academic content of the new document node, and the academic document database recommends the existing document to the user.

2. The paper recommendation method based on directed graph convolutional network according to claim 1 is characterized in that: The first-order neighbor propagation includes forward propagation from the source document node to the target document node and reverse propagation from the target document node to the source document node, and the result of the forward propagation is combined with the result of the reverse propagation to obtain a first-order neighbor propagation result representation; The forward propagation passes through the sparse adjacency matrix A * The propagation is performed according to the following formula to capture the citation relationship from the source document node to the target document node: In the formula, represents the feature matrix from the source document node to the target document node after updating in the network layer l+1, σ(·) represents the activation function, represents the result of normalized forward propagation of the sparse adjacency matrix of the document, H (l) represents the feature matrix of the document node at layer l of the network, W (l) Represents the weight matrix of the network layer l; The back propagation is done by transposing the sparse adjacency matrix A *T The propagation is performed according to the following formula to capture the citation relationship from the target document node to the source document node: In the formula, represents the feature matrix from the target document node to the source document node after updating at the l+1th layer of the network, Represents the result of the transposed back-propagation normalization of the sparse adjacency matrix of the document; The first-order neighbor propagation result H (l+1) for: In the formula, α is a hyperparameter that controls the weight ratio of forward and backward propagation; The second-order similarity neighbor propagation is used to capture the similarity between two document nodes that are not directly connected but have common in-degree neighbors or out-degree neighbors, including second-order in-degree similarity propagation and second-order out-degree similarity propagation; The second-order in-degree similarity propagation is: In the formula, represents the weighted adjacency matrix, Norm(·) represents the normalization process, (1-δ) represents the removal of self-loops to prevent itself from becoming its own first-order neighbor, The document node feature matrix representing the l+1th layer of the network updates the document node features by combining the information of the second-order in-degree similar neighbor document nodes; The second-order out-degree similarity propagation is: In the formula, The document node feature matrix representing the l+1th layer of the network updates the document node features by combining the information of the second-order out-degree similar neighbor document nodes; The DAGCN network aggregates the first-order neighbor propagation results and the second-order similar neighbor propagation results as follows: In the formula, represents the node feature matrix of the l+1th layer after aggregation, β represents the weighting coefficient, α1 represents the contribution weight of the two-directional information sources of the first-order neighbors, and α2 represents the contribution weight of the two-directional information sources of the second-order similar neighbors.

3. The paper recommendation method based on directed graph convolutional network according to claim 1 is characterized by: The similarity between the new document node in step 4.2 and each document category in the category division structure diagram is the cosine similarity between the new node and the category center of each document category, and the category center of each document category is the average value of the feature vectors of all document nodes in that category.

4. The paper recommendation method based on directed graph convolutional network according to claim 1 is characterized by: The preset demarcation threshold is 60%.

Images