Neural network scholar representation method and system based on heterogeneous graph
By applying adaptive contrast learning and overlapping cluster deep clustering methods in academic paper networks, the label dependence and community boundary blurring problems in scholar representation learning and community detection are solved, and more accurate and robust scholar representation and Community detection effect.
Patent Information
- Application Number
- CN202411403937.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-09
- Publication Date
- 2025-02-14
AI Technical Summary
The prior art has problems with strong label dependence and poor robustness in the characterization learning of academic paper networks. At the same time, community overlap and clustering feature information are ignored in scholar community detection, resulting in fuzzy community boundaries and low accuracy.
A scholar representation method based on heterogeneous graph neural network is proposed. Heterogeneous graph representation learning model is constructed through adaptive contrast learning, and the perturbation view and contrast mechanism are improved. A scholar community detection algorithm based on overlapping clusters and deep clustering is used. Partitioning method and clustering information serve as a fusion of high confidence soft tags to supervise the fusion of attributes and structural information.
Achieve more accurate and robust academic representation learning on academic heterogeneous diagrams, clarify the clustering boundaries of the scholar community, improve the accuracy of community detection, and visually demonstrate the field's core scholar recommendations and scholar-related multiple communities. Detection function.
Smart Images
Figure CN119443149A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of graph neural network technology, and more specifically, to a scholar representation method and system based on heterogeneous graph neural network. Background Art
[0002] Academic papers contain huge knowledge value, but academic paper networks often contain multiple heterogeneous entities and complex relationships such as papers, scholars, institutions, citations, and fields, which brings new challenges and difficulties to knowledge mining. As an emerging technology for graph representation learning, heterogeneous graph neural networks mainly use semi-supervised methods to achieve downstream tasks such as representation learning and classification of heterogeneous graphs, and have achieved good results on heterogeneous graphs such as social networks and biomedical networks. As a typical heterogeneous graph network, the academic paper network has problems such as label dependence and poor robustness when learning the representation of the network; at the same time, the existing methods mostly ignore community overlap and clustering feature information in the community detection of heterogeneous scholar networks, and have problems such as fuzzy community boundaries and low accuracy. In order to solve the above problems, a technical solution is now provided. Summary of the invention
[0003] In order to overcome the above-mentioned defects of the prior art, the present invention provides a scholar representation method and system based on heterogeneous graph neural network, which performs node embedding by using the proposed academic heterogeneous graph neural network model of adaptive contrastive learning, and performs scholar community detection based on overlapping clusters and deep clustering. On the basis of these two, a scholar field community detection system is developed, which finally realizes the functions of recommending core scholars in the field and detecting multiple communities associated with scholars, and performs a visual display to solve the problems raised in the above-mentioned background technology.
[0004] To achieve the above object, the present invention provides the following technical solutions:
[0005] A scholar representation method based on heterogeneous graph neural network includes the following steps:
[0006] Step 1: Build a heterogeneous graph representation learning model based on contrastive learning. Use contrastive learning to build a self-supervised heterogeneous graph neural network model and improve the perturbation view and contrast mechanism.
[0007] Step 2: Based on the scholar community detection algorithm of overlapping clusters and deep clustering, the overlapping network structure of scholars with cross-fields is decomposed into a scholar role graph with multiple node embeddings using the local partitioning method, so that the community has clearer clustering boundaries;
[0008] Step 3: Use clustering information as a high-confidence soft label to supervise the fusion of attribute and structure information, so that the generated scholar feature vector has a more cohesive cluster distribution and accuracy;
[0009] Step 4: Build a heterogeneous academic network representation and detection system based on open academic data sets. By integrating the above model algorithms, the core scholar detection and scholar overlapping community detection functions can be realized and visualized.
[0010] As a further solution of the present invention, in step 2, a heterogeneous graph representation learning model is constructed based on contrastive learning, and a self-supervised heterogeneous graph neural network model is constructed by using contrastive learning and improving the perturbation view and contrast mechanism, specifically:
[0011] Step 11, constructing a perturbation view through an adaptive data enhancement method on a heterogeneous graph, and randomly deleting edges according to their heterogeneity and importance, so that the perturbation view retains the main data structure;
[0012] Step 12, convert the original image into a feature matrix through a graph neural network, and convert the perturbed view into a perturbed feature matrix;
[0013] Step 13: Based on the generated perturbation view, a comparison mechanism based on adaptive relationship weights is proposed to enhance the comparison performance of the model by using structural information to constrain the correlation between positive and negative samples.
[0014] Step 14, introduce a contrast mechanism based on autoencoding, use the encoder to compress and extract important information in the perturbed view, and then restore it according to the original view features to make the features of different views consistent, and then combine the two contrast loss training models to learn attribute and structural information.
[0015] As a further solution of the present invention, in step 2, the steps of constructing the scholar community detection algorithm based on overlapping clusters and deep clustering are:
[0016] Step 21, local ego network partitioning: decouple the overlapping structure of the real scholar community through the partitioning algorithm, partition the meta-path graph of the academic heterogeneous network into multiple local ego networks, and simplify the overlapping clustering problem into a non-overlapping problem;
[0017] Step 22, global feature generation: Generate global overlapping features using a contrastive self-supervision method, so that multiple role nodes of overlapping scholars have different semantic features;
[0018] Step 23, community deep clustering: while fusing attribute and structural information, use clustering high-confidence soft labels for supervision, so that the generated scholar feature vector has a more cohesive cluster distribution, and then combine the graph neural network to learn the clustering embedding features of scholars.
[0019] A scholar representation system based on heterogeneous graph neural network, used to implement the above-mentioned scholar representation method based on heterogeneous graph neural network, including a data processing layer, a deep learning layer, a data storage layer, an intelligent computing layer and an application service layer;
[0020] The data used in the data processing layer are extracted from open academic data sets. The original data are screened and filtered, and the papers on computer science in the past five years are retained. Four secondary fields are selected as research fields, and scholars with high-impact papers and a large number of published papers are selected for community detection;
[0021] The deep learning layer uses a heterogeneous graph representation learning model based on contrastive learning and a scholar community detection algorithm based on overlapping clusters and deep clustering to generate cluster embedding representations of scholars: the contrastive self-supervised representation learning model is used to obtain the feature vectors of scholars containing heterogeneous attributes and structures, and then the overlapping community construction module is used to partition and disassemble the scholar community. Based on the community deep clustering module, some scholar field labels are used for constraints to obtain multiple vector representations of overlapping scholars;
[0022] The data storage layer uses the MySQL academic knowledge base to store the original information of the data set, including papers, scholars, fields, institutions and journals. It saves the scholar feature vectors generated by the deep learning layer based on files, and uses the DGL framework to build and store academic graph data and statistically analyze the connection relationship of the graph structure.
[0023] The intelligent computing layer uses scholar feature vectors to perform cluster analysis and business logic implementation: the core scholar ranking function in a field first uses cluster center calculation to locate the center vector of the community, and then uses cluster distance to calculate the most relevant scholars in each field community. The scholar-related community detection function first finds multiple cluster vectors of scholars, and then uses intra-class similarity calculations to obtain multi-dimensional related communities of overlapping scholars.
[0024] The application service layer provides functional modules to users and uses the front-end interface to interactively transmit results; the functional modules are cluster visualization, ranking of core scholars in the field, detection of scholar-related communities, and display of scholars and paper details.
[0025] The present invention discloses a scholar representation method and system based on heterogeneous graph neural network, and has the following technical effects and advantages: the present invention proposes a heterogeneous graph representation learning model based on contrastive learning, and solves the problems of strong label dependence and complex structure of heterogeneous academic networks by using contrastive learning to construct a self-supervised heterogeneous graph neural network model and improve the perturbation view and contrast mechanism; proposes a scholar community detection algorithm based on overlapping clusters and deep clustering, and uses a local partitioning method to decompose the overlapping network structure of scholars with cross-fields into a scholar role graph with multiple nodes embedded, so that the community has clearer cluster boundaries; at the same time, in view of the problem that the graph embedding process lacks clustering guidance, it is proposed to use clustering information as a fusion of high-confidence soft label supervision attributes and structural information, so that the generated scholar feature vector has a more condensed cluster distribution and accuracy; a heterogeneous academic network representation and detection system based on open academic data sets is developed, and by integrating the above model algorithm, the core scholar detection and scholar overlapping community detection functions are realized and visualized. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 A schematic diagram of a flow chart of a scholar representation method based on a heterogeneous graph neural network provided in Example 1 of the present invention;
[0027] Figure 2 A flowchart for building a heterogeneous graph representation learning model based on contrastive learning;
[0028] Figure 3 A schematic diagram of the structure of a heterogeneous graph representation learning model based on contrastive learning;
[0029] Figure 4 A flowchart for constructing a scholar community detection algorithm based on overlapping clusters and deep clustering;
[0030] Figure 5 It is a schematic diagram of the structure of the scholar community detection algorithm based on overlapping clusters and deep clustering;
[0031] Figure 6 A structural schematic diagram of a scholar representation system based on a heterogeneous graph neural network provided by the present invention. DETAILED DESCRIPTION
[0032] 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.
[0033] Example 1
[0034] Figure 1FIG. 1 shows a flow chart of a method for representing scholars based on a heterogeneous graph neural network provided by the first embodiment of the present invention. Figure 1 As shown, a scholar representation method based on a heterogeneous graph neural network in this embodiment includes the following steps:
[0035] Step 1: Build a heterogeneous graph representation learning model based on contrastive learning. Use contrastive learning to build a self-supervised heterogeneous graph neural network model and improve the perturbation view and contrast mechanism.
[0036] Step 2: Based on the scholar community detection algorithm of overlapping clusters and deep clustering, the overlapping network structure of scholars with cross-fields is decomposed into a scholar role graph with multiple node embeddings using the local partitioning method, so that the community has clearer clustering boundaries;
[0037] Step 3, using clustering information as a high-confidence soft label to supervise the fusion of attribute and structure information, so that the generated scholar feature vector has a more cohesive cluster distribution and accuracy;
[0038] Step 4: Build a heterogeneous academic network representation and detection system based on open academic data sets. By integrating the above model algorithms, the core scholar detection and scholar overlapping community detection functions can be realized and visualized.
[0039] Specifically, in step 2, a heterogeneous graph representation learning model is constructed based on contrastive learning. By using contrastive learning to construct a self-supervised heterogeneous graph neural network model and improve the perturbation view and contrast mechanism, specifically:
[0040] Step 11, constructing a perturbation view through an adaptive data enhancement method on a heterogeneous graph, and randomly deleting edges according to their heterogeneity and importance, so that the perturbation view retains the main data structure;
[0041] Step 12, convert the original image into a feature matrix through a graph neural network, and convert the perturbed view into a perturbed feature matrix;
[0042] Step 13: Based on the generated perturbation view, a comparison mechanism based on adaptive relationship weights is proposed to enhance the comparison performance of the model by using structural information to constrain the correlation between positive and negative samples.
[0043] Step 14, introduce a contrast mechanism based on autoencoding, use the encoder to compress and extract important information in the perturbed view, and then restore it according to the original view features to make the features of different views consistent, and then combine the two contrast loss training models to learn attribute and structural information.
[0044] like Figure 3The structural diagram of the heterogeneous graph representation learning model based on contrastive learning is shown in the following figure. The input is a heterogeneous graph (upper left corner), which is composed of multiple types of nodes and edges. The perturbation view (lower left corner) is constructed by an adaptive data enhancement method. The enhancement method randomly deletes edges based on the heterogeneity and importance of the edges. Specifically, the probability of deleting edges depends on their importance in the graph, so that the main structural information of the graph can be retained as much as possible. The original graph is converted into a feature matrix through a graph neural network (HGNN), and the perturbation view is converted into a perturbation feature matrix to represent the nodes and their structural information. By calculating the original graph feature matrix and the perturbation feature matrix, a contrast mechanism is constructed using adaptive relationship weights, which involves calculating the difference in graph structure, that is, using an encoder to extract structural information and construct positive and negative sample pairs. In the comparison process, the model compares positive sample pairs (from the same view) and negative sample pairs (from different views). By adjusting the relationship weights, the model learns how to effectively distinguish samples and maximize the similarity of positive sample pairs while minimizing the similarity of negative sample pairs.
[0045] A contrast mechanism based on autoencoders is introduced to further enhance the contrast performance of the model. The encoder compresses the perturbed feature matrix into a low-dimensional representation, and then reconstructs these low-dimensional representations into the original features through the decoder. The goal of this process is to make the perturbed view features consistent with the original graph features. The contrast loss of the autoencoder calculates the distance between the compressed representation of the perturbed view and the original representation. By minimizing this distance, the model can maintain consistent feature representations between different views; the final model combines the contrast loss based on edge relations and the contrast loss based on autoencoders, and optimizes the model through joint training. The two contrast loss functions learn the structural information and feature information of the graph respectively. During the training process, the model continuously adjusts the weights and parameters to minimize the sum of these two losses, and finally achieves the best performance on academic heterogeneous graphs.
[0046] Specifically, in step 2, the construction steps of the scholar community detection algorithm based on overlapping clusters and deep clustering are:
[0047] Step 21, local ego network partitioning: decouple the overlapping structure of the real scholar community through the partitioning algorithm, partition the meta-path graph of the academic heterogeneous network into multiple local ego networks, and simplify the overlapping clustering problem into a non-overlapping problem;
[0048] Step 22, global feature generation: Generate global overlapping features using a contrastive self-supervision method, so that multiple role nodes of overlapping scholars have different semantic features;
[0049] Step 23, community deep clustering: while fusing attribute and structural information, use clustering high-confidence soft labels for supervision, so that the generated scholar feature vector has a more cohesive cluster distribution, and then combine the graph neural network to learn the clustering embedding features of scholars.
[0050] like Figure 5 As shown in the structural diagram of the scholar community detection algorithm based on overlapping clusters and deep clustering, the input of the local self-network partition is an academic heterogeneous network graph, in which nodes represent scholars and edges represent the relationship between scholars (such as cooperation, citation relationship). The academic heterogeneous graph neural network (HGAC) based on adaptive contrastive learning is used to partition the academic heterogeneous network graph into local self-networks. The result of the partition is to generate multiple local subgraphs (or local self-networks), which represent the structure of scholars in different local contexts. The output of the local self-network partition is a neighbor graph and a role graph, which respectively represent the neighbor relationship (connections in the same field or community) and role relationship (different roles of the same scholar in different communities). Finally, the heterogeneous graph neural network encodes the local network features into a feature matrix;
[0051] For each local ego network, the global features of the scholar are further generated through the Graph Attention Network (GAT), using two types of subgraphs: the neighbor graph (neighborhood relationship graph) and the role graph (role relationship graph). In the figure, the neighbor graph and the role graph are processed by GAT respectively to generate the original view feature representation and perturbation view feature representation Z. These feature matrices represent the global embedding features of scholars under neighbor relationships and role relationships, respectively. GAT uses the attention mechanism to perform weighted aggregation of node features of the neighbor graph and role graph, and the calculated global features It reflects the semantic information of scholars under different relationships. Finally, these features are used to generate positive and negative sample pairs. Through contrastive learning mechanism, the similarity of positive sample pairs and negative sample pairs is calculated to improve the distinguishing ability of feature representation;
[0052] The K-means clustering algorithm is used to cluster scholars based on global features. The generated clustering result P represents the probability distribution of each scholar belonging to each potential community. The KL divergence (Kullback-Leibler Divergence, KL) is used to calculate the difference between the community distribution P and the original distribution Q. By minimizing this difference L clu =KL(P||Q),L clu That is, the minimum difference, which enables the model to learn more accurate community allocation results. The final output clustering results reflect the community structure information of scholars.
[0053] A scholar representation system based on heterogeneous graph neural network, used for the above-mentioned scholar representation method based on heterogeneous graph neural network, including a data processing layer, a deep learning layer, a data storage layer, an intelligent computing layer and an application service layer;
[0054] The data used by the data processing layer is extracted from the Open Academic Dataset (OAG). The original data is screened and filtered, and the papers in the computer science in the past five years are retained and four secondary fields are selected as research fields. The papers with high influence and scholars with a large number of published papers are selected for community detection. Among them, the Open Academic Dataset (OAG) is a compilation of the Tsinghua Academic Graph Dataset (AMiner) and the Microsoft Academic Graph Dataset (MAG). In order to more intuitively display the relationship between scholars, this embodiment selects papers with high influence and scholars with a large number of published papers for community detection, and finally extracts papers with more than 15 citations and 3,000 core scholars with the largest number of papers for experiments.
[0055] The deep learning layer uses a heterogeneous graph representation learning model based on contrastive learning and a scholar community detection algorithm based on overlapping clusters and deep clustering to generate cluster embedding representations of scholars: the contrastive self-supervised representation learning model is used to obtain the feature vectors of scholars containing heterogeneous attributes and structures, and then the overlapping community construction module is used to partition and disassemble the scholar community. Based on the community deep clustering module, some scholar field labels are used for constraints to obtain multiple vector representations of overlapping scholars;
[0056] The data storage layer uses MySQL academic knowledge base to store the original information of the data set, including papers, scholars, fields, institutions and journals. It saves the scholar feature vectors generated by the deep learning layer based on files, and uses the DGL framework to build and store academic graph data and statistically analyze the connection relationship of the graph structure. In order to import graph structure data into MySQL academic knowledge base, data tables corresponding to five nodes, namely papers, scholars, fields, institutions and journals, are first constructed. Then, foreign key fields are used to represent the one-to-many relationship between papers and journals and scholars and institutions. Finally, many-to-many foreign keys are used to represent the many-to-many relationships between papers and fields, papers and papers, and scholars and papers, and connection data tables are generated respectively.
[0057] The intelligent computing layer uses scholar feature vectors to perform cluster analysis and business logic implementation: the core scholar ranking function in a field first uses cluster center calculation to locate the center vector of the community, and then uses cluster distance to calculate the most relevant scholars in each field community. The scholar-related community detection function first finds multiple cluster vectors of scholars, and then uses intra-class similarity calculations to obtain multi-dimensional related communities of overlapping scholars.
[0058] The application service layer provides functional modules to users and uses the front-end interface to interactively transmit results; the functional modules are cluster visualization, ranking of core scholars in the field, detection of scholar-related communities, and display of scholars and paper details.
[0059] Specifically, in terms of technology selection, the system is a web service based on the browser / server model. The front-end part uses HTML, CSS and JavaScript-based Bootstrap4.3, the web framework uses Django3.0, the development language uses Python3.7, the development platform is Pycharm, the database uses MySQL8.0, the client uses Chrome, and the deep learning framework uses Pytorch1.8, as shown in the following table:
[0060] Development Language Python Development Platform Pycharm Web front end HTML, CSS, JavaScript, Bootstrap Web Frameworks Django database MySQL Client Chrome Deep Learning Frameworks Pytorch
[0061] The embodiment of the present invention proposes a heterogeneous graph representation learning model based on contrastive learning. By using contrastive learning to construct a self-supervised heterogeneous graph neural network model and improving the perturbation view and contrast mechanism, the problems of strong label dependence and complex structure of heterogeneous academic networks are solved; a scholar community detection algorithm based on overlapping clusters and deep clustering is proposed, and the overlapping network structure of scholars with cross-fields is decomposed into a scholar role graph with multiple nodes embedded by a local partitioning method, so that the community has clearer clustering boundaries; at the same time, in view of the problem that the graph embedding process lacks clustering guidance, it is proposed to use clustering information as a fusion of high-confidence soft label supervision attributes and structural information, so that the generated scholar feature vector has a more cohesive cluster distribution and accuracy; a heterogeneous academic network representation and detection system based on open academic data sets is developed, and by integrating the above model algorithms, the core scholar detection and scholar overlapping community detection functions are realized and visualized.
[0062] 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.
[0063] 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 scholar representation method based on heterogeneous graph neural network, characterized in that: The steps include: Step 1: Build a heterogeneous graph representation learning model based on contrastive learning. Use contrastive learning to build a self-supervised heterogeneous graph neural network model and improve the perturbation view and contrast mechanism. Step 2: Based on the scholar community detection algorithm of overlapping clusters and deep clustering, the overlapping network structure of scholars with cross-fields is decomposed into a scholar role graph with multiple node embeddings using the local partitioning method, so that the community has clearer clustering boundaries; Step 3: Use clustering information as a high-confidence soft label to supervise the fusion of attribute and structure information, so that the generated scholar feature vector has a more cohesive cluster distribution and accuracy; Step 4: Build a heterogeneous academic network representation and detection system based on open academic data sets. By integrating the above model algorithms, the core scholar detection and scholar overlapping community detection functions can be realized and visualized.
2. The method for representing scholars based on heterogeneous graph neural network according to claim 1 is characterized in that: In step 2, a heterogeneous graph representation learning model is constructed based on contrastive learning. By using contrastive learning to build a self-supervised heterogeneous graph neural network model and improve the perturbation view and contrast mechanism, specifically: Step 11, constructing a perturbation view through an adaptive data enhancement method on a heterogeneous graph, and randomly deleting edges according to their heterogeneity and importance, so that the perturbation view retains the main data structure; Step 12, convert the original image into a feature matrix through a graph neural network, and convert the perturbed view into a perturbed feature matrix; Step 13: Based on the generated perturbation view, a comparison mechanism based on adaptive relationship weights is proposed to enhance the comparison performance of the model by using structural information to constrain the correlation between positive and negative samples. Step 14, introduce a contrast mechanism based on autoencoding, use the encoder to compress and extract important information in the perturbed view, and then restore it according to the original view features to make the features of different views consistent, and then combine the two contrast loss training models to learn attribute and structural information.
3. The method for representing scholars based on heterogeneous graph neural network according to claim 1 is characterized in that: In step 2, the construction steps of the scholar community detection algorithm based on overlapping clusters and deep clustering are: Step 21, local ego network partitioning: decouple the overlapping structure of the real scholar community through the partitioning algorithm, partition the meta-path graph of the academic heterogeneous network into multiple local ego networks, and simplify the overlapping clustering problem into a non-overlapping problem; Step 22, global feature generation: Generate global overlapping features using a contrastive self-supervision method, so that multiple role nodes of overlapping scholars have different semantic features; Step 23, community deep clustering: while fusing attribute and structural information, use clustering high-confidence soft labels for supervision, so that the generated scholar feature vector has a more cohesive cluster distribution, and then combine the graph neural network to learn the clustering embedding features of scholars.
4. A scholar representation system based on heterogeneous graph neural network, used to implement a scholar representation method based on heterogeneous graph neural network as described in any one of claims 1 to 3, characterized in that: Includes data processing layer, deep learning layer, data storage layer, intelligent computing layer and application service layer; The data used in the data processing layer are extracted from open academic data sets. The original data are screened and filtered, and the papers on computer science in the past five years are retained. Four secondary fields are selected as research fields, and scholars with high-impact papers and a large number of published papers are selected for community detection; The deep learning layer uses a heterogeneous graph representation learning model based on contrastive learning and a scholar community detection algorithm based on overlapping clusters and deep clustering to generate cluster embedding representations of scholars: the contrastive self-supervised representation learning model is used to obtain the feature vectors of scholars containing heterogeneous attributes and structures, and then the overlapping community construction module is used to partition and disassemble the scholar community. Based on the community deep clustering module, some scholar field labels are used for constraints to obtain multiple vector representations of overlapping scholars; The data storage layer uses the MySQL academic knowledge base to store the original information of the data set, including papers, scholars, fields, institutions and journals. It saves the scholar feature vectors generated by the deep learning layer based on files, and uses the DGL framework to build and store academic graph data and statistically analyze the connection relationship of the graph structure. The intelligent computing layer uses scholar feature vectors to perform cluster analysis and business logic implementation: the core scholar ranking function in a field first uses cluster center calculation to locate the center vector of the community, and then uses cluster distance to calculate the most relevant scholars in each field community. The scholar-related community detection function first finds multiple cluster vectors of scholars, and then uses intra-class similarity calculations to obtain multi-dimensional related communities of overlapping scholars. The application service layer provides functional modules to users and uses the front-end interface to interact and deliver results; The functional modules include cluster visualization, ranking of core scholars in the field, detection of scholar-related communities, and display of scholar and paper details.