A method, device and medium for classifying papers of a citation network

Abstract

The application relates to the technical field of natural language processing, and particularly discloses a paper classification method for a citation network, a device and a medium, the method comprising the following steps: graph data preprocessing of the citation network, node centroid topology enhancement for the majority class and the minority class, and multi-period centroid node information fusion for the minority class. The method combines the node centroid topology enhancement and the multi-period centroid node information fusion mechanism, improves the classification performance of the graph neural network for the minority class nodes on the unbalanced graph, and improves the classification accuracy of different types of papers in the citation network.

Description

Technical Field

[0001] This invention relates to the field of graph neural network technology, specifically to a paper classification method, device, and medium using citation networks. Background Technology

[0002] In academic citation networks, there is often an imbalance in the number of citations among different types of literature (such as journal articles, conference papers, and technical reports). Categories with fewer citations (such as research in certain emerging disciplines) are often overlooked during model training, resulting in insufficient learning of their features by the model. A minority of literature categories also carry important academic information, such as cutting-edge research findings or interdisciplinary contributions; neglecting these categories may lead to the omission of crucial information in academic network analysis.

[0003] In conclusion, there is an urgent need for a citation network-based paper classification method, device, and medium to address the problems in existing technologies. Summary of the Invention

[0004] The purpose of this invention is to provide a method, device, and medium for classifying papers using citation networks. The specific technical solution is as follows:

[0005] A paper classification method based on citation networks includes the following steps:

[0006] S1: Obtain the data from the citation network, classify it into majority and minority classes according to the existing labels, and obtain the first classification result;

[0007] S2: The original graph neural network is used to pass messages to obtain the first classification result, and the first classification result map is obtained. The centroid node of each category is calculated using the centroid node formula. At the same time, dynamic majority class and dynamic minority class classification are performed.

[0008] S3: Perform node centroid topology enhancement for dynamic majority and dynamic minority classes;

[0009] S4: Perform minority class training on the original graph neural network, save the information of the minority class centroid nodes in different training cycles of the graph neural network, and construct a similarity matrix between historical centroid nodes using similarity metrics.

[0010] S5: For the similarity matrix, when the similarity between two historical centroid nodes exceeds a preset threshold, an additional connection between the two historical centroid nodes is established in the topology to complete the minority class training of the original graph neural network and obtain the paper classification model.

[0011] S6: Input the data from the citation network into the paper classification model to complete the paper classification using the citation network.

[0012] Optionally, in S1, the minority class represents the class with the number of nodes in the existing labels less than the average total number of each class, and the majority class represents the class with the number of nodes in the existing labels greater than or equal to the average total number of each class.

[0013] Optionally, in S2, the formula for calculating the centroid node is as follows:

[0014] ;

[0015] in, Indicates training period Belongs to the category The centroid node, Indicates training period Belongs to the category The set of nodes, Represents the first node in the set of nodes. 1 node Represents a node The embedding vector.

[0016] Optionally, in S3, node centroid topology enhancement includes the construction of dynamic majority class nodes and dynamic minority class nodes.

[0017] Optionally, in S3, the construction process of dynamic majority class nodes includes:

[0018] Obtain the node probability predictions output during the training cycle. The node probability is regarded as an indicator of the connection strength between the node and the centroid node of the class. The expression for the node probability is as follows:

[0019] ;

[0020] in, Represents a node Category The probability, Represents a node Category The label prediction results Indicates the total number of categories. Indicate category , Represents a node Category The label prediction results;

[0021] The classification results of the original graph neural network are rebalanced based on the connection strength index, targeting nodes of the majority class. and any category centroid node Construct the connection weights, as shown in the following expression:

[0022] ;

[0023] in, Represents a node and any category centroid node Connection weights between them This indicates that the result obtained after message passing is The distribution percentage of each category;

[0024] Construct the adjacency matrix of the majority class nodes, and define the connection matrix between the majority class nodes and the centroid nodes of each class as follows: ,in, The number of majority class nodes is represented by the following elements in the connection matrix:

[0025] ;

[0026] in, The set threshold is used to filter low-confidence connections.

[0027] Optionally, in S3, the construction process of dynamic minority class nodes is as follows:

[0028] For minority nodes ,satisfy ,in, This indicates the label prediction result. This indicates a minority category label, directly related to the centroid node of that category. Alternatively, connections can be made using stable self-clustering centroids;

[0029] The connectivity metrics are as follows:

[0030] ;

[0031] in, Represents a node Category The connectivity metrics, based on the connectivity metrics Construct a subgraph connection for each category centroid node, and connect all category centroid nodes. , forming a subgraph ,in, Represents a set of nodes. Represents the set of edges;

[0032] ;

[0033] in, The threshold for similarity determination. For temperature parameters, dynamic minority class nodes are established based on similarity. and centroid node The connection.

[0034] Optionally, in step S4, at the first In each training cycle, the category Centroid node representation The expression is as follows:

[0035] ;

[0036] in, Indicates training period Belongs to the category The set of nodes, The feature vector representing a node;

[0037] Introducing a sliding window mechanism to save the most recent The historical centroid nodes of each training cycle are represented by the following expression:

[0038] ;

[0039] in, Indicates recent The historical centroid nodes of each training cycle are represented;

[0040] For the representation of historical centroid nodes, a similarity matrix is ​​constructed. Similarity matrix elements in The definition is as follows:

[0041] ;

[0042] in, This represents the similarity metric, using cosine similarity or similarity values ​​converted from Euclidean distance.

[0043] Optionally, in step S5, additional connections are established for the historical centroid nodes, as follows:

[0044] Define index function , ;

[0045] Construct cumulative weighted connections , ;

[0046] Additional connections to historical centroid nodes across multiple training cycles are achieved through cumulative weighted connections.

[0047] In addition, the present invention also provides a computer device, including a memory and a processor;

[0048] The memory is used to store computer programs that can run on the processor;

[0049] The processor is used to implement the paper classification method of the citation network as described above when executing the computer program.

[0050] In addition, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described citation network paper classification method.

[0051] The application of the technical solution of the present invention has the following beneficial effects:

[0052] This invention addresses the imbalanced classification problem in citation networks by embedding balanced centroid nodes over multiple periods. It utilizes topology enhancement with dynamic majority and minority nodes to preserve the original minority node centroids, thus constructing a subgraph of the original minority centroid nodes. By establishing additional connections in the topology to represent nodes, the training process of the original graph neural network model is optimized, improving its classification performance for minority nodes on imbalanced graphs. This, in turn, enhances the classification accuracy for different categories of papers in citation networks.

[0053] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0054] To more clearly illustrate the technical solutions of the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart of the paper classification method of citation network in a preferred embodiment of the present invention. Detailed Implementation

[0056] To enable those skilled in the art to better understand the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Example:

[0058] like Figure 1 As shown, this embodiment provides a paper classification method based on citation networks, including the following steps (S1 to S6):

[0059] S1: Obtain the data from the citation network, classify it into majority and minority classes according to the existing labels, and obtain the first classification result.

[0060] In this embodiment, minority classes refer to classes where the number of nodes in existing labels is less than the average total number of each class, and majority classes refer to classes where the number of nodes in existing labels is greater than or equal to the average total number of each class.

[0061] S2: The original graph neural network is used to pass messages to obtain the first classification result, and the first classification result map is obtained. The centroid node of each category is calculated using the centroid node formula. At the same time, dynamic majority and dynamic minority classifications are performed.

[0062] In this embodiment, the calculation expression for the centroid node formula is as follows:

[0063] ;

[0064] in, Indicates training period Belongs to the category The centroid node, Indicates training period Belongs to the category The set of nodes, Represents the first node in the set of nodes. 1 node Represents a node The embedding vector.

[0065] S3: Perform node centroid topology enhancement for dynamic majority and dynamic minority classes. Node centroid topology enhancement includes the construction of dynamic majority class nodes and dynamic minority class nodes.

[0066] In this embodiment, the construction process of dynamic majority class nodes includes:

[0067] Obtain the node probability predictions output during the training cycle. The node probability is regarded as an indicator of the connection strength between the node and the centroid node of the class. The expression for the node probability is as follows:

[0068] ;

[0069] in, Represents a node Category The probability, Represents a node Category The label prediction results Indicates the total number of categories. Indicate category , Represents a node Category The label prediction results;

[0070] The classification results of the original graph neural network are rebalanced based on the connection strength index, targeting nodes of the majority class. and any category centroid node Construct the connection weights, as shown in the following expression:

[0071] ;

[0072] in, Represents a node and any category centroid node Connection weights between them This indicates that the result obtained after message passing is The distribution percentage of each category;

[0073] Construct the adjacency matrix of the majority class nodes, and define the connection matrix between the majority class nodes and the centroid nodes of each class as follows: ,in, The number of majority class nodes is represented by the following elements in the connection matrix:

[0074] ;

[0075] in, The set threshold is used to filter low-confidence connections.

[0076] Furthermore, the construction process of dynamic minority class nodes is as follows:

[0077] For minority nodes ,satisfy ,in, This indicates the label prediction result. This indicates a minority category label, directly related to the centroid node of that category. Alternatively, connections can be made using stable self-clustering centroids;

[0078] The connectivity metrics are as follows:

[0079] ;

[0080] in, Represents a node Category The connectivity metrics, based on the connectivity metrics Construct a subgraph connection for each category centroid node, and connect all category centroid nodes. , forming a subgraph ,in, Represents a set of nodes. Represents the set of edges;

[0081] ;

[0082] in, The threshold for similarity determination. For temperature parameters, dynamic minority class nodes are established based on similarity. and centroid node The connection.

[0083] S4: Perform minority class training on the original graph neural network, save the information of the minority class centroid nodes in different training cycles of the graph neural network, and construct a similarity matrix between historical centroid nodes using similarity metrics.

[0084] In this embodiment, in the first In each training cycle, the category Centroid node representation The expression is as follows:

[0085] ;

[0086] in, Indicates training period Belongs to the category The set of nodes, The feature vector representing a node;

[0087] Introducing a sliding window mechanism to save the most recent The historical centroid nodes of each training cycle are represented by the following expression:

[0088] ;

[0089] in, Indicates recent The historical centroid nodes of each training cycle are represented;

[0090] For the representation of historical centroid nodes, a similarity matrix is ​​constructed. Similarity matrix elements in The definition is as follows:

[0091] ;

[0092] in, The similarity metric is represented by the similarity value transformed from cosine similarity or Euclidean distance.

[0093] S5: For the similarity matrix, when the similarity between two historical centroid nodes exceeds a preset threshold, an additional connection between the two historical centroid nodes is established in the topology to complete the minority class training of the original graph neural network and obtain the paper classification model.

[0094] In this embodiment, the additional connection to the historical centroid node is established as follows:

[0095] Define index function , ;

[0096] Construct cumulative weighted connections , ;

[0097] Additional connections to historical centroid nodes across multiple training cycles are achieved through cumulative weighted connections.

[0098] S6: Input the data from the citation network into the paper classification model to complete the paper classification using the citation network.

[0099] This embodiment provides a paper classification method for citation networks, which has significant comprehensive advantages compared to existing technologies, maintaining high classification accuracy even in highly imbalanced citation networks. Firstly, at the feature understanding level, this embodiment overcomes the limitations of traditional data-level processing by addressing the imbalance problem at the fundamental level of feature representation through a centroid vector embedding mechanism. Unlike node synthesis methods such as GraphSMOTE and GAN-based generation methods such as DRGCN, this embodiment calculates class centroid vectors to accurately characterize the feature centers of each class, maintaining the authenticity of the original node features while avoiding graph topology destruction and feature distortion problems that may arise from synthesized nodes.

[0100] In addition, this embodiment also provides a computer device, including a memory and a processor;

[0101] The memory is used to store computer programs that can run on the processor;

[0102] The processor is used to implement the paper classification method of the citation network as described above when executing the computer program.

[0103] For example, the computer device may be a mobile phone, desktop computer, laptop, handheld computer, or cloud server, etc. The computer device may include, but is not limited to, a processor and memory. For instance, the computer device may also include input / output devices, network access devices, buses, etc.

[0104] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the computer device, connecting various parts of the computer device via various interfaces and lines.

[0105] The memory can be used to store the computer program and / or modules. The processor implements the computer program by running or executing the computer program and / or modules stored in the memory, and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0106] Wherein, if the modules / units integrated into the computer device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by a processor, it can implement the steps of the various method embodiments described above. Wherein, the computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0107] In addition, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described citation network paper classification method.

[0108] This embodiment of the method preserves the original minority class centroids by using topology enhancement of dynamic majority and minority class nodes, thereby constructing a subgraph of original minority class centroid nodes. By establishing additional connections for node representations in the topology, the training process of the original graph neural network model is optimized, improving the classification performance of the graph neural network model for minority class nodes on unbalanced graphs, thereby improving the classification accuracy for different categories of papers in the citation network.

[0109] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0110] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method of classifying papers in a citation network, characterized by, The method comprises the following steps: S1: obtaining data of a citation network, classifying majority classes and minority classes according to existing labels to obtain a first classification result; the minority classes represent classes in which the number of nodes in the existing labels is less than the average total number of each class, and the majority classes represent classes in which the number of nodes in the existing labels is greater than or equal to the average total number of each class; S2: performing message passing using an original graph neural network to obtain the first classification result, obtaining a first classification result graph, and calculating the centroid nodes of each class using a centroid node formula, while classifying dynamic majority classes and dynamic minority classes; S3: performing node centroid topology enhancement on the dynamic majority classes and the dynamic minority classes; S4: performing minority class training on the original graph neural network, saving the centroid node information of the minority classes in different training periods of the graph neural network, and constructing a similarity matrix between historical centroid nodes using similarity measurement; S5: for the similarity matrix, when the similarity between two historical centroid nodes exceeds a preset threshold, establishing an additional connection between the two historical centroid nodes in the topology, completing the minority class training of the original graph neural network, and obtaining a paper classification model; S6: inputting data of the citation network into the paper classification model to complete paper classification of the citation network.

2. The method of classifying papers of a citation network according to claim 1, wherein, In S2, the calculation expression of the centroid node formula is as follows: ； wherein, denotes a training period denotes a centroid node belonging to the class in the training period, denotes a training period denotes a set of nodes belonging to the class in the training period, denotes the i-th node in the set of nodes, denotes the i-th node in the set of nodes, denotes the embedding vector of the node .

3. The method of claim 2, wherein, In S3, the node centroid topology enhancement includes construction of dynamic majority class nodes and construction of dynamic minority class nodes.

4. The method of claim 3, wherein, In S3, the construction process of the dynamic majority class nodes includes: The output node probability in the training period is obtained, and the node probability is regarded as a connection strength index of the node and the centroid node of the class, and the expression of the node probability is as follows: ； wherein, representing a node belongs to a class with a probability, representing a node belongs to a class with a label prediction, representing a total number of classes, representing a class , representing a node belongs to a class with a label prediction; Based on the connection strength index, the classification results of the original graph neural network are re-balanced in categories, and the nodes of the majority class are re-balanced and the centroid nodes of any category are re-balanced , and the connection weight is constructed, and the expression is as follows: ； wherein, representing nodes and arbitrary classes of centroid nodes between the connection weights, representing the distribution proportion of the class after message passing the result is An adjacency matrix of the majority class node is constructed, and a connection matrix between the majority class node and each category centroid node is defined as wherein, represents the number of majority class nodes, and the elements in the connection matrix are defined as follows: ； wherein is a set threshold value for filtering low confidence connections.

5. The method of claim 4, wherein, In S3, the construction process of the dynamic minority class nodes is as follows: For minority class nodes , satisfy Where, Indicates the label prediction result, Indicates the minority class label, directly connected with the class centroid node Or stable self-clustering centroid; The connection index is as follows: ； in, Represents a node Category The connectivity metrics, based on the connectivity metrics Construct a subgraph connection for each category centroid node, and connect all category centroid nodes. , forming a subgraph ,in, Represents a set of nodes. Represents the set of edges; ； wherein, is a threshold value for the similarity determination, is a temperature parameter, the dynamic minority class node is established according to the similarity and the connection of the centroid node .

6. The method of classifying papers of a citation network according to claim 5, wherein, In step S4, in the first training cycle, the class centroid node representation is expressed as follows: ； wherein, denotes a training period belonging to the class , a set of nodes, denotes a feature vector of a node; A sliding window mechanism is introduced to keep the history of the last few training epochs, expressed as follows: ； wherein, represents the most recent history centroid node representation of the last training cycle; For the historical centroid node representation, a similarity matrix is built , the elements of the similarity matrix are defined as follows: ​ ； wherein, represents a similarity measure, using cosine similarity or a similarity value converted from Euclidean distance.

7. The method of claim 6, wherein, In step S5, the additional connection of the historical centroid nodes is established, and the process is as follows: Defining the index function , ; Constructing cumulative weighted connections , ; The additional connection of the historical centroid nodes in multiple training periods is realized by cumulative weighted connection.

8. A computer device, comprising: The computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the steps of the paper classification method of the citation network according to any one of claims 1 to 7. The computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the steps of the paper classification method of the citation network according to any one of claims 1 to 7. ​ 9. A computer-readable storage medium, characterized in that, ​

Images