Community discovery method and system based on heterogeneous academic network
Through the use of meta -path embedded and label dissemination in the heterogeneous academic network, and the comparative learning community discovery algorithm guided by the pseudo -label guidance, the problem of insufficient node information in the study of the heterogeneous diagram is solved. Community discovery.
Patent Information
- Application Number
- CN202411424443.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-12
- Publication Date
- 2025-01-28
AI Technical Summary
Existing heterogeneous graph representation learning methods have shortcomings in capturing node structure and semantic information, resulting in poor results in community discovery tasks.
The study algorithm based on the element path -based embedded and label dissemination is used to represent the learning algorithm, combined with the comparative learning community discovery algorithm guided by the multi -layer convolution network and the pseudo -label guidance, and enhance the semantic information and distinction of the node representation.
By fully capturing the structure and semantic information of the node, the accuracy and effect of community discovery have been significantly improved, and the scholars' community structure can be more effectively discovered.
Smart Images
Figure CN119377495A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of information retrieval technology, and more specifically, to a community discovery method and system based on heterogeneous academic networks. Background Art
[0002] The number of scholars in the world is growing explosively. A large number of scholars and the relationships between them contain a large amount of information that can be mined and used by researchers. Scholars and their achievements constitute a huge academic network diagram. The network contains rich information waiting for researchers to extract. The huge number of scholars and achievements also brings severe challenges to traditional data mining methods. For example, how to measure the similarity between multiple scholars; how to divide scholar communities based on the similarity of scholars.
[0003] Heterogeneous networks (also called heterogeneous networks) are graph networks composed of various types of nodes and edges. Compared with homogeneous networks, heterogeneous networks contain richer structural and semantic information and are more in line with real-world application scenarios. Therefore, heterogeneous graph representation learning has been proposed and applied to heterogeneous graph analysis tasks. However, most of the current heterogeneous graph representation learning methods have the problem of insufficient node representation semantic information, which leads to the fact that they cannot achieve good results in many downstream tasks. Therefore, how to design a model based on the characteristics of rich semantics and complex structure of heterogeneous graphs so that it can fully capture the node-related structural and semantic information has become an important issue in heterogeneous graph representation learning.
[0004] Community discovery is an important task in the field of graph network analysis. A community is a collection of nodes with similar attributes, and the nodes in the same community are closely connected, while the nodes in different communities are loosely connected. By analyzing the community structure in academic networks, researchers can discover the rules of cooperation between different scholars, which brings great application value. However, the node representations learned by existing community discovery methods are not discriminative enough and do not consider the high-order neighbor information of the graph. Therefore, how to design effective self-supervision signals and reasonable model structures to make the learned node representations more discriminative is a hot issue in current community discovery.
[0005] In order to solve the above problems, a technical solution is now provided. Summary of the invention
[0006] In order to overcome the above-mentioned defects of the prior art, the present invention provides a community discovery method and system based on heterogeneous academic networks, which solves the problem of insufficient semantic information of node representation through a heterogeneous graph representation learning algorithm based on meta-path embedding and label propagation, and solves the problem of insufficient node representation discrimination and the problem of insufficient node representation structural information of mainstream models through a comparative learning community discovery algorithm guided by pseudo-labels, so as to solve the problems raised in the above-mentioned background technology.
[0007] To achieve the above object, the present invention provides the following technical solutions:
[0008] A community discovery method based on heterogeneous academic networks includes the following steps:
[0009] Step 1: By constructing a heterogeneous graph representation learning algorithm based on deep learning, the semantic information of the nodes is supplemented by learning meta-path embedding using a multi-layer graph convolutional network;
[0010] Step 2: Based on the pseudo-label-guided contrastive learning community discovery model, the learning process of node representation is guided by contrastive learning, so that the node representations meet the requirements that the node representations of the same category are close and the node representations of different categories are far away, thereby improving the distinguishability of node representations; the low-order and high-order structural information of the graph is learned by message passing on the low-order neighbor graph and the high-order neighbor graph, thus supplementing the structural features of the node representation;
[0011] Step 3: Calculate scholar representations and discover scholar communities through heterogeneous graph representation learning algorithms and community discovery algorithms, and show users the community structure of the academic network and the importance of scholars in the community.
[0012] As a further solution of the present invention, in step 1, the heterogeneous graph representation learning algorithm based on deep learning specifically includes data preprocessing and model training: for a given heterogeneous academic network, the label information of the training set nodes is counted, a label set is constructed, label nodes are added to the heterogeneous academic network, and edges are added between the training set nodes and their corresponding label nodes; after the preprocessing is completed, the transformed heterogeneous academic network is used as the input of the heterogeneous graph representation learning model to train the model.
[0013] As a further solution of the present invention, a heterogeneous graph representation learning model includes an input feature conversion module, a meta-path embedding module, a meta-path information exchange module and a meta-path fusion module; the input feature conversion module is used to map the input features of different types of nodes to the same implicit space; the meta-path embedding module learns the meta-path embedding associated with the target node through a multi-layer graph neural network; the meta-path information exchange module uses a language model to complete the information exchange between meta-path embeddings and supplement the information required for meta-path embeddings; the meta-path fusion module uses a neural network to fuse all meta-path embeddings to form an embedded representation of the target node.
[0014] As a further solution of the present invention, the overall structure of the pseudo-label-guided contrastive learning community discovery model includes contrastive view generation, view feature extraction, and cross-view contrastive learning;
[0015] For the input adjacency matrix A and node feature matrix X, adjacency matrix perturbation and feature perturbation are performed to obtain two views. The same graph convolutional network and autoencoder are used to learn the structural features and attribute features of the two views respectively, and the graph structure embedding and attribute feature embedding under the two views are obtained, which are combined into a unified graph structure embedding and attribute feature embedding; the graph structure embedding and attribute feature embedding are fused to obtain the initial node representation, and then the high-order structural information of the graph is integrated to obtain the final node representation; cross-view comparative learning is used to guide the learning process of node embedding, so that the similarity between nodes with the same pseudo-label is high and the similarity between nodes with different pseudo-labels is low, while ensuring that the representation of nodes in different views is consistent. After the pseudo-label-guided comparative learning community discovery model completes the learning of node embedding, the KMeans algorithm is used to cluster the nodes to obtain the community division result.
[0016] A community discovery system based on heterogeneous academic networks, used to implement the above-mentioned community discovery method based on heterogeneous academic networks, including a scholar community visualization module, a scholar information display module, a paper information display module and an important scholar discovery module in the field;
[0017] The scholar community visualization module is used to reduce the dimension of the scholar embedding representation learned by the heterogeneous graph representation learning model and draw a scatter plot to display it. According to the structure of the community discovery model, each point in the scatter plot is marked with different colors to show the community to which the scholar belongs;
[0018] The scholar information display module includes the scholar search function and the scholar basic information display function. The scholar search function performs fuzzy query on the scholar's name based on keywords and displays the name, affiliated institution and total number of citations of each scholar in the search results in the form of a list; the scholar basic information display function mainly displays the scholar's name, affiliated institution, number of published papers, total number of citations and other inherent attributes, and also displays the main research fields calculated by the algorithm, and displays the papers published by the scholar in the form of a list;
[0019] The paper information display module includes the paper search function and the paper basic information display function. The paper search function performs fuzzy query on the paper title based on keywords and displays the title, number of citations, journal or conference, and abstract information of each paper in the result in the form of a list; the paper basic information display function displays the basic attributes of the paper in detail, including the title of the paper, the journal or conference, the author, the publication time, the number of cited papers, the number of citations, the abstract, and the research field;
[0020] The module for discovering important scholars in a field uses the embedded representation of scholars in the academic network calculated by intelligence and the embedded representation of the community center to calculate the distance from the scholar to the community center, and classifies the scholars according to their importance according to the pre-set distance range, into three categories: important, general and weak. The display order of scholars in the field is displayed in reverse order according to the importance of the scholars.
[0021] The technical effects and advantages of the community discovery method and system based on heterogeneous academic networks of the present invention are as follows: the present invention uses a heterogeneous graph representation learning algorithm based on meta-path embedding and label propagation, uses a multi-layer graph neural network to learn all meta-path embeddings related to nodes, and then fuses the meta-path embeddings to obtain node representations with sufficient semantic information, thereby solving the problem of insufficient semantic information of node representation. A pseudo-label-guided comparative learning community discovery algorithm is proposed, which solves the problem of insufficient node representation discrimination and the problem of insufficient node representation structural information of mainstream models; a scholar community discovery system in an academic network is designed and implemented, and the heterogeneous graph representation learning model and community discovery model proposed in this article are used to learn the representation of scholar nodes in an academic network composed of open source academic data, and to discover scholar communities, and finally the scholar community is displayed through visualization technology. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 A flowchart of a community discovery method based on a heterogeneous academic network provided in Example 1 of the present invention;
[0023] Figure 2 It is a schematic diagram of the structure of the heterogeneous graph representation learning algorithm based on meta-path embedding and label propagation;
[0024] Figure 3 Schematic diagram of the overall structure of the PSCD model for contrastive learning based on pseudo-label guidance;
[0025] Figure 4 This is a functional structure diagram of a community discovery system based on heterogeneous academic networks;
[0026] Figure 5 A schematic diagram of the overall architecture of a community discovery system based on heterogeneous academic networks. DETAILED DESCRIPTION
[0027] The following will be combined with the accompanying drawings in the present invention to clearly and completely describe the technical solution in the present invention. Obviously, the described technical solution is only a part of the present invention, not all of it. Based on the technical solution in the present invention, all other technical solutions obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0028] Example 1
[0029] Figure 1 FIG. 1 is a flow chart of a community discovery method based on a heterogeneous academic network provided by the first embodiment of the present invention. Figure 1 As shown, a community discovery method based on a heterogeneous academic network in this embodiment includes the following steps:
[0030] Step 1: By constructing a heterogeneous graph representation learning algorithm based on deep learning, the semantic information of the nodes is supplemented by learning meta-path embedding using a multi-layer graph convolutional network;
[0031] Step 2: Based on the pseudo-label-guided contrastive learning community discovery model, the learning process of node representation is guided by contrastive learning, so that the node representations meet the requirements that the node representations of the same category are close and the node representations of different categories are far away, thereby improving the distinguishability of node representations; the low-order and high-order structural information of the graph is learned by message passing on the low-order neighbor graph and the high-order neighbor graph, thus supplementing the structural features of the node representation;
[0032] Step 3: Calculate scholar representations and discover scholar communities through heterogeneous graph representation learning algorithms and community discovery algorithms, and show users the community structure of the academic network and the importance of scholars in the community.
[0033] Specifically, in step 1, the heterogeneous graph representation learning algorithm based on deep learning is as follows: Figure 2 The structural diagram of the heterogeneous graph representation learning algorithm based on meta-path embedding and label propagation shown in the figure specifically includes data preprocessing and model training: for a given heterogeneous academic network, the label information of the training set nodes is counted, a label set is constructed, label nodes are added to the heterogeneous academic network, and edges are added between the training set nodes and their corresponding label nodes; after the preprocessing is completed, the transformed heterogeneous academic network is used as the input of the heterogeneous graph representation learning model for model training.
[0034] The heterogeneous graph representation learning model includes an input feature conversion module, a meta-path embedding module, a meta-path information exchange module and a meta-path fusion module; the input feature conversion module is used to map the input features of different types of nodes to the same implicit space; the meta-path embedding module learns the meta-path embedding associated with the target node through a multi-layer graph neural network; the meta-path information exchange module uses a language model to complete the information exchange between meta-path embeddings and supplement the information required for meta-path embeddings; the meta-path fusion module uses a neural network to fuse all meta-path embeddings to form an embedded representation of the target node.
[0035] Specifically, Figure 3 The overall structure diagram of the PSCD community discovery model based on pseudo-label guidance is shown in FIG. The overall structure of the PSCD community discovery model based on pseudo-label guidance includes contrastive view generation, view feature extraction, and cross-view contrastive learning.
[0036] For the input adjacency matrix A and node feature matrix X, two views are obtained by adjacency matrix perturbation and feature perturbation. The adjacency matrix A is perturbed to generate an adjacency view (structural view), that is, a small modification or noise is added to the structure of the original graph. The node feature matrix X is perturbed to generate a feature view (attribute view), that is, the attributes of the nodes are changed to reflect the differences in different features between nodes. After the adjacency matrix perturbation is performed on the adjacency matrix A and the node feature matrix X, a structural view A is obtained. k and a feature view X k , after perturbing the adjacency matrix A and the node feature matrix X, we obtain: a structural view A p and a feature view X p , use the same graph convolutional network (GCN) and autoencoder (AE) to learn the structural features and attribute features of the two views respectively. The graph convolutional network (GCN) is used to perform graph convolution operations on the two views respectively, and extract the graph structure embedding and attribute feature embedding under the two views from the perturbed adjacency matrix and node features, that is, to obtain the structural view A k Graph structure embedding Embedding with attribute features Feature View A p Graph structure embedding Embedding with attribute features Combined into a unified graph structure embedding Z GAE and attribute feature embedding Z AE ; The autoencoder (AE) is used to reduce the dimension and reconstruct the perturbed data, learn the low-dimensional representation of the two views, and obtain the corresponding embedded structure view and an embedded feature view Z I , and then integrate the high-level structural information of the graph Get the final node representation Z G ; Use cross-view comparative learning to guide the learning process of node embedding, so that the similarity between nodes with the same pseudo-label is high and the similarity between nodes with different pseudo-labels is low, while ensuring that the representation of nodes in different views is consistent. After the pseudo-label-guided comparative learning community discovery model completes the learning of node embedding, the KMeans algorithm is used to cluster the nodes to obtain the community division results.
[0037] Among them, the pseudo code of the training process of the contrastive learning community discovery model based on pseudo-label guidance is:
[0038] Input: adjacency matrix A, node input feature X, number of communities C;
[0039] Output: Community division results {V1, V2, ..., V C}
[0040] Initialize node representation Z and generate contrast view G k ={X k ,A k}, G p ={X p ,A p}
[0041] for epoch=1to n
[0042] Use autoencoders to learn the attribute features of views,
[0043] Use graph neural networks to learn the structural features of views,
[0044] Fusion of embedding representations of two views
[0045] Form the initial node representation Z I =αZ AE +(1-α)Z GAE
[0046] Learning graph structure information on low-order neighbor graphs and high-order neighbor graphs
[0047] Fusion graph structure information, get the final node representation Z = βZ G +Z L
[0048] Calculate loss function, back propagate, and update model parameters
[0049] Use KMeans to cluster the node representation Z to obtain the community division result {V1, V2, …, V C}
[0050] return{V1,V2,…,V C}.
[0051] Among them, α is the control Z AE and Z GAE The weight ratio in the fusion process, β is the adjustment coefficient.
[0052] A community discovery system based on heterogeneous academic networks is used to implement the above-mentioned community discovery method based on heterogeneous academic networks, including a scholar community visualization module, a scholar information display module, a paper information display module and an important scholar discovery module in the field, such as Figure 4 The functional structure diagram of a community discovery system based on heterogeneous academic networks is shown;
[0053] The scholar community visualization module is used to reduce the dimension of the scholar embedding representation learned by the heterogeneous graph representation learning model and draw a scatter plot to display it. According to the structure of the community discovery model, each point in the scatter plot is marked with different colors to show the community to which the scholar belongs; the scholar community visualization module is the core module of the system, which facilitates users to have a clear and intuitive understanding of the scholar community structure of the academic network.
[0054] The scholar information display module includes scholar search function and scholar basic information display function. The scholar search function performs fuzzy query on scholar names based on keywords and displays the name, affiliated institution and total number of citations of each scholar in the search results in the form of a list; the scholar basic information display function mainly displays the scholar's name, affiliated institution, number of published papers, total number of citations and other inherent attributes, while displaying the main research fields calculated by the algorithm, and displaying the papers published by the scholar in the form of a list; the scholar information display also displays the papers published by the scholar in the form of a list, which is convenient for users to quickly understand all the papers in which the scholar participated.
[0055] The paper information display module includes paper search function and paper basic information display function. The paper search function performs fuzzy query on paper titles based on keywords and displays the title, number of citations, journal or conference, and abstract information of each paper in the result in list form; the paper basic information display function displays the basic attributes of the paper in detail, including the title of the paper, the journal or conference, the author, the publication time, the number of cited papers, the number of citations, the abstract, and the research field, so that users can quickly understand the basic information of a paper.
[0056] The module for discovering important scholars in a field uses the embedded representation of scholars in the academic network calculated by intelligence and the embedded representation of the community center to calculate the distance from the scholar to the community center, and classifies the scholars according to their importance according to the pre-set distance range, into three categories: important, general and weak. The display order of scholars in the field is displayed in reverse order according to the importance of the scholars.
[0057] like Figure 5 The overall architecture diagram of a community discovery system based on heterogeneous academic networks is shown in the figure. The overall architecture of the system is divided into four layers: data processing layer, data storage layer, intelligent computing layer and application service layer. The data processing layer is responsible for collecting and processing academic network data to form a structured academic network; the data storage layer is responsible for storing structured data and using distributed and other technologies to improve the availability and fault tolerance of the system; the intelligent computing layer uses heterogeneous graph representation learning models and community discovery models to learn scholar representation and divide the communities to which scholars belong, and also conducts statistical analysis related to scholars; the application service layer uses the calculation and analysis results of the intelligent computing layer to provide users with various system functions.
[0058] The embodiment of the present invention uses a heterogeneous graph representation learning algorithm based on meta-path embedding and label propagation, uses a multi-layer graph neural network to learn all meta-path embeddings related to nodes, and then fuses the meta-path embeddings to obtain node representations with sufficient semantic information, thereby solving the problem of insufficient semantic information of node representation. A pseudo-label-guided comparative learning community discovery algorithm is proposed, which solves the problem of insufficient node representation discrimination and the problem of insufficient node representation structure information of mainstream models; a scholar community discovery system in an academic network is designed and implemented, using the heterogeneous graph representation learning model and community discovery model proposed in this article to learn the representation of scholar nodes in an academic network composed of open source academic data, and to discover scholar communities, and finally the scholar community is displayed through visualization technology.
[0059] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
[0060] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A community discovery method based on heterogeneous academic networks, characterized in that: The steps include: Step 1: By constructing a heterogeneous graph representation learning algorithm based on deep learning, the semantic information of the nodes is supplemented by learning meta-path embedding using a multi-layer graph convolutional network; Step 2: Based on the pseudo-label-guided contrastive learning community discovery model, the learning process of node representation is guided by contrastive learning, so that the node representations meet the requirements that the node representations of the same category are close and the node representations of different categories are far away, thereby improving the distinguishability of node representations; the low-order and high-order structural information of the graph is learned by message passing on the low-order neighbor graph and the high-order neighbor graph, thus supplementing the structural features of the node representation; Step 3: Calculate scholar representations and discover scholar communities through heterogeneous graph representation learning algorithms and community discovery algorithms, and show users the community structure of the academic network and the importance of scholars in the community.
2. The method for representing scholars based on heterogeneous graph neural network according to claim 1 is characterized in that: In step 1, the heterogeneous graph representation learning algorithm based on deep learning specifically includes data preprocessing and model training: for a given heterogeneous academic network, the label information of the training set nodes is counted, a label set is constructed, label nodes are added to the heterogeneous academic network, and edges are added between the training set nodes and their corresponding label nodes; After the preprocessing is completed, the transformed heterogeneous academic network is used as the input of the heterogeneous graph representation learning model for model training.
3. The method for representing scholars based on heterogeneous graph neural network according to claim 2 is characterized in that: The heterogeneous graph representation learning model includes an input feature conversion module, a meta-path embedding module, a meta-path information exchange module and a meta-path fusion module; the input feature conversion module is used to map the input features of different types of nodes to the same implicit space; the meta-path embedding module learns the meta-path embedding associated with the target node through a multi-layer graph neural network; the meta-path information exchange module uses a language model to complete the information exchange between meta-path embeddings and supplement the information required for meta-path embeddings; the meta-path fusion module uses a neural network to fuse all meta-path embeddings to form an embedded representation of the target node.
4. The method for representing scholars based on heterogeneous graph neural network according to claim 1, characterized in that: The overall structure of the pseudo-label-guided contrastive learning community discovery model, including contrastive view generation, view feature extraction, and cross-view contrastive learning; For the input adjacency matrix A and node feature matrix X, adjacency matrix perturbation and feature perturbation are performed to obtain two views. The same graph convolutional network and autoencoder are used to learn the structural features and attribute features of the two views respectively, and the graph structure embedding and attribute feature embedding under the two views are obtained, which are combined into a unified graph structure embedding and attribute feature embedding; the graph structure embedding and attribute feature embedding are fused to obtain the initial node representation, and then the high-order structural information of the graph is integrated to obtain the final node representation; cross-view comparative learning is used to guide the learning process of node embedding, so that the similarity between nodes with the same pseudo-label is high and the similarity between nodes with different pseudo-labels is low, while ensuring that the representation of nodes in different views is consistent. After the pseudo-label-guided comparative learning community discovery model completes the learning of node embedding, the KMeans algorithm is used to cluster the nodes to obtain the community division result.
5. A community discovery system based on heterogeneous academic networks, used to implement a community discovery method based on heterogeneous academic networks as described in any one of claims 1 to 4, characterized in that: It includes the scholar community visualization module, scholar information display module, paper information display module and important scholar discovery module in the field; The scholar community visualization module is used to reduce the dimension of the scholar embedding representation learned by the heterogeneous graph representation learning model and draw a scatter plot to display it. According to the structure of the community discovery model, each point in the scatter plot is marked with different colors to show the community to which the scholar belongs; The scholar information display module includes the scholar search function and the scholar basic information display function. The scholar search function performs fuzzy query on the scholar's name based on keywords and displays the name, affiliated institution and total number of citations of each scholar in the search results in the form of a list; the scholar basic information display function mainly displays the scholar's name, affiliated institution, number of published papers, total number of citations and other inherent attributes, and also displays the main research fields calculated by the algorithm, and displays the papers published by the scholar in the form of a list; The paper information display module includes the paper search function and the paper basic information display function. The paper search function performs fuzzy query on the paper title based on keywords and displays the title, number of citations, journal or conference, and abstract information of each paper in the result in the form of a list; the paper basic information display function displays the basic attributes of the paper in detail, including the title of the paper, the journal or conference, the author, the publication time, the number of cited papers, the number of citations, the abstract, and the research field; The module for discovering important scholars in a field uses the embedded representation of scholars in the academic network calculated by intelligence and the embedded representation of the community center to calculate the distance from the scholar to the community center, and classifies the scholars according to their importance according to the pre-set distance range, into three categories: important, general and weak. The display order of scholars in the field is displayed in reverse order according to the importance of the scholars.