A citation network skeleton construction method based on multi-view graph learning

Abstract

The present invention discloses a method for constructing a citation network skeleton based on multi-view graph learning. The multi-view graph learning method is used to effectively combine the external attribute information of nodes in the citation network, the network structure, and the various relationships between nodes, and an attention mechanism is introduced to assign different weights to different nodes and attributes, and feature learning is performed on important nodes to achieve accurate node representation. The idea of random walk is further used to measure the mutual influence of paper nodes to reflect the dissemination and application of knowledge. The experimental results of the present invention also verify that the citation network skeleton construction helps to display the knowledge transfer relationship between nodes in the citation network in a fine-grained manner, and more objectively reflects the dissemination, utilization and innovation of knowledge in the citation network. At the same time, the experimental results can also well reflect the composition and structure of the network. The present invention provides a new method for constructing a skeleton network in an academic citation network, which provides a new solution for studying citation networks.

Description

Technical Field

[0001] The present invention relates to a method for constructing a skeleton of a citation network based on papers in the academic field, and in particular to a method for constructing a skeleton of a citation network based on a combination of important node identification based on multi-view learning and influence strength between nodes. Background Art

[0002] A skeleton network is a subnetwork composed of key nodes and the edges between them in the original network, and can largely maintain the topological structure of the original network. Important elements extracted during the construction of the skeleton network have a greater impact on the structure and function of the network and are relatively important. In recent years, skeleton construction has been a hot topic, with related research focusing primarily on human brain networks and power systems. Citation networks can trace the history and development of science. In the era of big data, where the number of papers is growing exponentially, applying complex network theory and methods can identify important paper nodes in the paper citation network and organize these nodes into a skeleton rich in information about the entire citation network, revealing the deep relationships between papers. This not only helps scholars understand the characteristics of citation networks and the context of scientific development, and grasp scientific frontiers and dynamics, but also helps scholars extract effective information and improve information utilization in the face of information overload.

[0003] Existing research has achieved some success in skeleton construction, but applying these methods to citation network skeleton construction still faces the following limitations: Identifying important nodes is a key step in citation network skeleton construction. With the rapid development of information and communication technologies and science and technology, scientific research activities are generating an ever-increasing amount of data, and citation relationships have formed a massive, complex network system. Accurately identifying important nodes in citation networks with numerous nodes and complex relationships is one of the challenges in citation network skeleton construction. Current important node identification methods typically measure node importance within a single-layer network, failing to fully capture the characteristics of important nodes through the complex and numerous relationships, resulting in inaccurate identification results. Furthermore, compared to other complex networks, the relationships between paper nodes are complex and mutually influential. The citation relationships between nodes reflect the dissemination and application of knowledge, and the papers that inspire important papers are equally important. Therefore, measuring the mutual influence between nodes is particularly important for citation network skeleton construction, but existing methods do not consider this mutual influence between nodes. Therefore, how to effectively integrate multi-dimensional features, identify important nodes in citation networks, and calculate the mutual influence between papers in a fine-grained manner are key issues in citation network skeleton construction. Summary of the Invention

[0004] The purpose of the present invention is mainly to address the problems that existing traditional algorithms cannot effectively integrate multi-dimensional attributes, have low recognition accuracy, and do not consider the mutual influence between nodes. A citation network skeleton construction method is proposed, which strives to accurately establish the hierarchical structure of the citation network and show the development process of related research. First, this method is based on multi-perspective graph learning, while considering network structure, external information and various relationships between nodes, to construct an important node identification algorithm that effectively integrates multi-dimensional attribute features, and learns important node features from multiple perspectives. Secondly, based on the idea of random walk, the mutual influence between paper nodes is calculated to construct a skeleton network for the citation network. The present invention uses the method of combining multi-perspective graph learning with random walk for the first time to provide a new method for the problem of citation network skeleton construction.

[0005] The technical solution of the present invention:

[0006] A method for constructing a citation network skeleton based on multi-view graph learning includes the following steps:

[0007] (1) Construct a multi-layer relationship network based on the citation relationship, co-citation relationship, and co-keyword relationship between papers;

[0008] (2) Identify important nodes in the multi-layer relationship network constructed based on step (1)

[0009] 2.1) Attribute View Feature Learning

[0010] 2.1.1) Obtaining external attribute information of nodes based on the constructed multi-layer relationship network, including influencing papers Factors of importance - impact factor of the journal in which the paper is published , average citations by all authors of the paper , the popularity of the topic of the paper and the average number of citations in the paper's references The formula for calculating the average citation volume of all authors of a paper is:

[0011]

[0012] in, is the set of all authors of the paper, For the author The number of citations, For the data set papers;

[0013] Papers based on popular topics tend to attract more attention from scholars, and the topic popularity TP index is proposed. The calculation formula is:

[0014]

[0015] in, For the paper A keyword in For keywords The number of occurrences, It is a paper Keywords collection, keywords is the set of all keywords in the paper collection, is a keyword Number of occurrences; average number of citations in references It is an intuitive indicator to measure the quality of references, and the calculation formula is:

[0016]

[0017] in, For the paper The number of references in is the number of citations of the reference;

[0018] Use the impact factor of the journal in which the paper is published , average citations by all authors of the paper , the popularity of the topic of the paper and the average number of citations in the paper's references Constructing attribute feature matrix ;

[0019] 2.1.2) Use the attribute feature matrix obtained in step 2.1.1) to calculate the node representation under the attribute view; the importance of a node is not only related to its own attributes, but also to the contribution of other nodes in the network, especially the neighboring nodes to it; in order to effectively integrate the contribution of neighboring nodes, a graph convolutional neural network is used to learn the function Implement attribute feature aggregation, where It is the citation network in the constructed multi-layer relational network. is the parameter for attribute view feature learning, in the In the convolutional layer, the output of attribute feature aggregation is expressed as:

[0020]

[0021] in, , is the adjacency matrix of the paper citation network, is the identity matrix; yes The angle matrix of is calculated as follows: , Representation matrix Middle Rank Column elements, Representation matrix Middle Rank Column elements; It is The training parameters of the convolutional layer, is the activation function; when When input is the attribute feature matrix, that is ; In order to fully measure the importance difference of each neighbor node, the attention mechanism is used to learn different weights for each neighbor node; in the Layer convolution layer, the correlation matrix between nodes is expressed as:

[0022]

[0023] in, are the training parameters of the node itself; is the training parameter of the neighbor node; It is the product operation of the adjacency matrix and the parameter, which represents the propagation ability; is the Sigmod activation function, that is ; Finally, the correlation coefficient matrix of the entire network By calculating the correlation matrix between nodes, we can get:

[0024]

[0025] in, Representative paper nodes The set of all neighbor nodes; The original network correlation coefficient matrix Replacement, to assign different weights to different neighbor nodes during the convolution process;

[0026] 2.2) Structural View Feature Learning

[0027] 2.2.1) Based on the constructed multi-layer relationship network, we obtain the structural characteristics of the network, including the degree, betweenness, closeness centrality and clustering coefficient of the node, and calculate the structural characteristics of the reference network. ;

[0028] 2.2.2) Use the structural feature matrix obtained in step 2.2.1) to calculate the node representation under the structural view, and use the convolution operation to calculate the final representation of the structural feature view. The convolution process is as follows:

[0029]

[0030] in, ,In structured view feature learning, the first The correlation matrix between nodes in the convolutional layer is expressed as:

[0031]

[0032] Finally, the correlation coefficient matrix of the entire graph is ;

[0033] 2.3) Relational View Feature Learning

[0034] 2.3.1) Obtaining a multi-layer network adjacency matrix set based on the constructed multi-layer relationship network , Save in Relationship View The connection relationship between the nodes, Indicates the number of layers of the constructed multi-layer relationship network;

[0035] 2.3.2) Use the relationship feature matrix obtained in step 2.3.1) to calculate the node representation under the relationship view; adopt a collaborative training method to maximize the mutual consistency between different relationship views, reduce the computational complexity, and obtain the representation of nodes under multiple relationship views; express the Laplacian matrix of the aggregated multi-layer network as ,in is the weight of each relational view, and the Laplace matrix under each relational view is ,in yes The degree matrix of is calculated as follows: ; In each relationship view Next, add node self-connections to preserve the node's own structural characteristics:

[0036]

[0037] in, , , the Laplace matrix under each relational view is formalized as ; Dynamically adjust the Laplace matrix weights under each relationship to achieve joint learning, define ; Weight parameter The calculation formula is:

[0038]

[0039] Represented in the relationship view The representation of the next node, The calculation formula for the layer is:

[0040]

[0041] Based on the idea of collaborative training, the final representation of the nodes in the relational view is calculated: ;

[0042] 2.4) Multi-view aggregation: When aggregating the final representation of the node, in order to effectively integrate the attribute feature representation, structural feature representation, and relationship feature representation, an attention mechanism is used to learn the importance of different views:

[0043]

[0044] in Presentation View The final output of feature learning is Represents the view when aggregating the final representation The importance of is calculated as follows:

[0045]

[0046] in, Is a view The eigenvector of

[0047] 2.5) Important node label prediction; the representation of the final node As the input of the SoftMax classifier, the SoftMax classifier is used to identify important nodes. The classification results are compared with the node importance ranking results obtained from the SIR simulation experiment to optimize the negative log-likelihood loss function; the final classification-based cross entropy loss function is:

[0048]

[0049]

[0050] in, is a node The true label, The predicted node label is The probability value of , The models are classified into and Classification parameters of

[0051] 2.6) Important node labeling: Use the node labels predicted in step 2.5) to label the important nodes in the network;

[0052] (3) Build a skeleton network based on the important nodes identified in step (2)

[0053] 3.1) For each paper The references of the node constitute the parent node set of the node , cited The papers constitute the descendant set of this node ,paper The parent node right The influence of express, Indicates the paper and Similarity between The papers in companion, if the paper The higher the similarity between the papers in its peer set, the higher the parent node right The greater the impact;

[0054] 3.2) Use the idea of random walk to measure the similarity between two papers; To Node The two-step random walk of can go through their common reference, in which case The contributions are:

[0055]

[0056] 3.3) Compute slave nodes To Node The two-step random walk probability of metric is the similarity between two papers:

[0057]

[0058] 3.4) Linear Aggregation and , calculate the parent node to its child nodes Impact:

[0059]

[0060] 3.5) For each central paper ,according to The size of the value, keep the node The parent node with the greatest influence is added to the skeleton network;

[0061] 3.6) Repeat steps (1) to (5) for the nodes in the skeleton network to construct the citation network skeleton.

[0062] Beneficial effects of the present invention: The present invention only utilizes the external attribute information of nodes in the citation network, the network structure and the various relationships between nodes, combines the multi-view graph learning method with the random walk method, breaks through the shortcomings of traditional methods in the problem of citation network skeleton construction, and introduces the attention mechanism to better characterize nodes. The experimental results show that the method proposed by the present invention is more suitable for processing complex academic network structures, and has higher accuracy and reliability than other methods. The experimental results of the present invention also verify that the citation network skeleton construction helps to display the knowledge transfer relationship between nodes in the citation network in a fine-grained manner, and more objectively reflects the dissemination, utilization and innovation of knowledge in the citation network. At the same time, the experimental results can also well reflect the composition and structure of the network. The present invention provides a new method for constructing a skeleton network in an academic citation network, which provides a new solution for studying citation networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 For the experimental process of this invention, a citation network skeleton construction method is proposed based on the paper data in Cord-19 and MAG, combining multi-view graph learning and random walk.

[0064] Figure 2(a) shows the original citation network in the medical field, and Figure 2(b) shows the constructed skeleton network in the medical field. Figure 2(c) shows the original citation network in the biological field, and Figure 2(d) shows the constructed skeleton network in the biological field. DETAILED DESCRIPTION

[0065] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0066] A method for constructing a citation network skeleton based on multi-view graph learning includes the following steps:

[0067] Step 1: Build an academic multi-layer network based on real paper publications

[0068] This paper constructs a citation network based on the open research dataset CORD-19. CORD-19 contains not only relevant papers but also other research findings. Metadata for these papers is collected from PubMed Central (PMC), BioRXiv, and MedRXiv. Data preprocessing, including data filtering, was first performed. During the experiment, papers published after 2020 whose titles or abstracts contained the keyword "2019-NCoV" were selected from CORD-19. To categorize papers by research field, the paper collections were matched in Microsoft Academic Graph (MAG), and a citation network was constructed based on citation relationships between papers. Metadata from MAG was also used to calculate paper attribute features and construct an academic multi-layer network based on citation relationships, co-citation relationships, and shared keyword relationships between papers. Finally, this paper conducted experiments on citation networks constructed from papers in the fields of biology and medicine.

[0069] Step 2: Identify important nodes in the network based on the multi-layer network constructed in step 1, including the following steps:

[0070] (2.1) Based on the constructed multi-layer network, the external attribute information of the nodes is obtained and the following indicators are calculated according to the formula, including factors that affect the importance of the paper, such as the impact factor of the journal in which the paper is published. , average citations by all authors of the paper , the popularity of the topic of the paper and the average number of citations in the paper's references , the calculation formulas are:

[0071]

[0072]

[0073]

[0074] The above features are used to construct the feature matrix as the input of attribute feature learning, that is, .

[0075] (2.2) Construct an attribute feature learning view and use the attribute feature matrix to obtain node representation. Based on the graph convolutional neural network, the contributions of neighboring nodes are effectively integrated and the importance differences of neighboring nodes are distinguished. First, the attention mechanism is used to learn different weights for each neighboring node:

[0076]

[0077] Finally, the correlation coefficient matrix of the entire network By calculating the correlation matrix between nodes, we can get:

[0078]

[0079] Neural Network Output of attribute feature learning of convolutional layer:

[0080] .

[0081] (2.3) Based on the constructed multi-layer network, we obtain the structural characteristics of the network, including the degree of the node, the betweenness of the node, the closeness centrality, and the clustering coefficient of the node, and calculate the structural characteristics of the reference network. .

[0082] (2.4) Construct the structural feature view and calculate the The correlation matrix between nodes in the convolutional layer:

[0083]

[0084] Finally, the correlation coefficient matrix of the entire graph is ;

[0085] in The convolution operation is used to calculate the final representation of the structural feature view. The convolution process is as follows:

[0086]

[0087] (2.5) Obtaining a multi-layer network adjacency matrix set based on the constructed multi-layer network , Save in Relationship View The connection relationship between the nodes, Indicates the number of layers in the constructed multi-layer relationship network.

[0088] (2.6) Construct relational views and adopt collaborative training to obtain the representation of nodes under multiple relational views. Next, add node self-connections to preserve the node's own structural characteristics:

[0089]

[0090] in, , , the Laplace matrix under each relational view is formalized as ; Dynamically adjust the Laplace matrix weights under each relationship to achieve joint learning, define .

[0091] (2.7) In the relationship view Next, learn the feature representation of the node:

[0092]

[0093] (2.8) The Laplacian matrix of the aggregated multilayer network can be expressed as ,in is the weight of each relational view, and the Laplace matrix under each relational view is ,in . Using attention mechanism to learn weight parameters :

[0094]

[0095] (2.9) Based on the idea of collaborative training, the final representation of the nodes in the relational view is calculated:

[0096]

[0097] (2.10) Multi-view aggregation to obtain the final representation of the node:

[0098]

[0099] in, Represents the view when aggregating the final representation The importance of different views is learned by using the attention mechanism. The calculation formula is:

[0100]

[0101] in, Is a view The eigenvector of .

[0102] (2.11) Identify important nodes based on the obtained node representation. As the input of the classifier, a SoftMax classifier is used to compare the classification results with the node importance ranking results obtained from the SIR simulation experiment to optimize the negative log-likelihood loss function. The node labels and final representations are obtained by minimizing the cross entropy loss function:

[0103]

[0104]

[0105] (2.12) Train the model parameters based on the training set, apply the adjusted model to the entire network, and obtain and mark important nodes in the network.

[0106] Step 3: Build a skeleton network based on the important nodes identified in step 2, including the following steps:

[0107] (3.1) Add the identified important nodes to the skeleton network;

[0108] (3.2) Perform the following operations for the nodes in the skeleton network in step (3.1): The references of the node constitute the parent node set of the node , cited The papers constitute the descendant set of this node ,paper The parent node right The influence of express, Indicates the paper and Similarity between The papers in companion, if the paper The higher the similarity between the papers in its peer set, the higher the parent node right The greater the impact;

[0109] (3.3) The idea of random walk is used to measure the similarity between two papers. Calculate the similarity from node To Node The two-step random walk probability of is used to measure the similarity between nodes:

[0110]

[0111] (3.3) Calculation slave nodes To Node The two-step random walk probability of is used to measure the similarity between nodes:

[0112]

[0113] (3.4) Linear aggregation and , calculate the parent node To its child nodes Impact:

[0114]

[0115] (3.5) For each important paper ,according to The size of the value, keep the node The parent node with the greatest influence is added to the skeleton network.

[0116] (3.6) Repeat steps (3.1) to (3.5) for the nodes in the skeleton network to construct the citation network skeleton. The skeleton construction results for the medical field are shown in Figure 2 (b), and the skeleton construction results for the biological field are shown in Figure 2 (d).

[0117] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any non-substantial changes and replacements made by those skilled in the art on the basis of the present invention shall fall within the scope of protection required by the present invention. The scope of protection of the present invention shall be based on the scope of protection of the claims.

Claims

1. A method for constructing a citation network skeleton based on multi-view graph learning, characterized in that: The following steps are involved: (1) Construct a multi-layer relationship network based on the citation relationship, co-citation relationship, and co-keyword relationship between papers; (2) Identify important nodes in the multi-layer relationship network constructed based on step (1) 2.1) Attribute View Feature Learning 2.1.1) Obtaining external attribute information of nodes based on the constructed multi-layer relationship network, including influencing papers Factors of importance - impact factor of the journal in which the paper is published , average citations by all authors of the paper , the popularity of the topic of the paper and the average number of citations in the paper's references The formula for calculating the average citation volume of all authors of a paper is: ， in, is the set of all authors of the paper, For the author The number of citations, For the data set papers; Papers based on popular topics tend to attract more attention from scholars, and the topic popularity TP index is proposed. The calculation formula is: ， in, For the paper A keyword in For keywords The number of occurrences, It is a paper Keywords collection, keywords is the set of all keywords in the paper collection, is a keyword Number of occurrences; average number of citations in references It is an intuitive indicator to measure the quality of references, and the calculation formula is: ， in, For the paper The number of references in is the number of citations of the reference; Use the impact factor of the journal in which the paper is published , average citations by all authors of the paper , the popularity of the topic of the paper and the average number of citations in the paper's references Constructing attribute feature matrix ; 2.1.2) Use the attribute feature matrix obtained in step 2.1.1) to calculate the node representation under the attribute view; the importance of a node is not only related to its own attributes, but also to the contribution of other nodes and neighboring nodes in the network to it; in order to effectively integrate the contribution of neighboring nodes, a graph convolutional neural network is used to learn the function Implement attribute feature aggregation, where It is the citation network in the constructed multi-layer relational network. is the parameter for attribute view feature learning, in the In the convolutional layer, the output of attribute feature aggregation is expressed as: ， in, , is the adjacency matrix of the paper citation network, is the identity matrix; yes The angle matrix of is calculated as follows: , Representation matrix Middle Rank Column elements, Representation matrix Middle Rank Column elements; It is The training parameters of the convolutional layer, is the activation function; when When input is the attribute feature matrix, that is ; In order to fully measure the importance difference of each neighbor node, the attention mechanism is used to learn different weights for each neighbor node; in the Layer convolution layer, the correlation matrix between nodes is expressed as: ， in, are the training parameters of the node itself; is the training parameter of the neighbor node; It is the product operation of the adjacency matrix and the parameter, which represents the propagation ability; is the Sigmod activation function, that is ; Finally, the correlation coefficient matrix of the entire network By calculating the correlation matrix between nodes, we can get: ， in, Representative paper nodes The set of all neighbor nodes; The original network correlation coefficient matrix Replacement, to assign different weights to different neighbor nodes during the convolution process; 2.2) Structural View Feature Learning 2.2.1) Based on the constructed multi-layer relationship network, we obtain the structural characteristics of the network, including the degree, betweenness, closeness centrality and clustering coefficient of the node, and calculate the structural characteristics of the reference network. ; 2.2.2) Use the structural feature matrix obtained in step 2.2.1) to calculate the node representation under the structural view, and use the convolution operation to calculate the final representation of the structural feature view. The convolution process is as follows: ， in, ,In structured view feature learning, the first The correlation matrix between nodes in the convolutional layer is expressed as: ， Finally, the correlation coefficient matrix of the entire graph is ; 2.3) Relational View Feature Learning 2.3.1) Obtaining a multi-layer network adjacency matrix set based on the constructed multi-layer relationship network , Save in Relationship View The connection relationship between the nodes, Indicates the number of layers of the constructed multi-layer relationship network; 2.3.2) Use the relationship feature matrix obtained in step 2.3.1) to calculate the node representation under the relationship view; adopt a collaborative training method to maximize the mutual consistency between different relationship views, reduce the computational complexity, and obtain the representation of nodes under multiple relationship views; express the Laplacian matrix of the aggregated multi-layer network as ,in is the weight of each relational view, and the Laplace matrix under each relational view is ,in yes The degree matrix of is calculated as follows: ; In each relationship view Next, add node self-connections to preserve the node's own structural characteristics: ， in, , , the Laplace matrix under each relational view is formalized as ; Dynamically adjust the Laplace matrix weights under each relationship to achieve joint learning, define ; Weight parameter The calculation formula is: ， Represented in the relationship view The representation of the next node, The calculation formula for the layer is: ， Based on the idea of collaborative training, the final representation of the nodes in the relational view is calculated: ; 2.4) Multi-view aggregation: When aggregating the final representation of the node, in order to effectively integrate the attribute feature representation, structural feature representation, and relationship feature representation, an attention mechanism is used to learn the importance of different views: ， in Presentation View The final output of feature learning is Represents the view when aggregating the final representation The importance of is calculated as follows: ， in, Is a view The eigenvector of 2.5) Important node label prediction; the representation of the final node As the input of the SoftMax classifier, the SoftMax classifier is used to identify important nodes. The classification results are compared with the node importance ranking results obtained from the SIR simulation experiment to optimize the negative log-likelihood loss function; the final classification-based cross entropy loss function is: ， ， in, is a node The true label, The predicted node label is The probability value of , The models are classified into and Classification parameters of 2.6) Important node labeling: Use the node labels predicted in step 2.5) to label the important nodes in the network; (3) Build a skeleton network based on the important nodes identified in step (2) 3.1) For each paper The references of the node constitute the parent node set of the node , cited The papers constitute the descendant set of this node ,paper The parent node right The influence of express, Indicates the paper and Similarity between The papers in companion, if the paper The higher the similarity between the papers in its peer set, the higher the parent node right The greater the impact; 3.2) Use the idea of random walk to measure the similarity between two papers; To Node The two-step random walk of goes through their common reference, in this case The contributions are: ， 3.3) Compute slave nodes To Node The two-step random walk probability of metric is the similarity between two papers: ， 3.4) Linear Aggregation and , calculate the parent node to its child nodes Impact: ， 3.5) For each central paper ,according to The size of the value, keep the node The parent node with the greatest influence is added to the skeleton network; 3.6) Repeat steps 3.1) to 3.5) for the nodes in the skeleton network to construct the citation network skeleton.

Images