A scholar relationship network storage construction method and system based on an artificial intelligence graph database

Abstract

The application relates to the technical field of data storage, and particularly discloses a scholar relationship network storage construction method and system based on an artificial intelligence graph database, which comprises the following steps: constructing a heterogeneous scholar relationship knowledge graph of each heterogeneous academic data source; extracting a joint feature vector of each scholar node in the heterogeneous scholar relationship knowledge graph; performing coarse screening on two scholar nodes in the heterogeneous scholar relationship knowledge graph of different heterogeneous academic data sources based on the self attribute vector of each scholar node, so as to obtain a plurality of to-be-aligned scholar node combinations; calculating the similarity of the joint feature vectors of all the scholar nodes in each to-be-aligned scholar node combination; fusing all the heterogeneous scholar relationship knowledge graphs based on the similarity scores of all the to-be-aligned scholar node combinations, so as to obtain a unified scholar relationship network and store the unified scholar relationship network into a distributed native graph database; and the unified scholar relationship network can meet the efficient access demand of scholar relationship network data.

Description

Technical Field

[0001] This invention relates to the field of data storage technology, and in particular to a method and system for constructing a scholar relationship network storage based on an artificial intelligence graph database. Background Technology

[0002] In the field of academic research, in-depth analysis of scholar relationship networks helps reveal important information such as academic collaboration patterns, disciplinary development trends, and the dissemination paths of academic influence. With the accelerated digitization of academic resources, a large amount of academic data is scattered across multiple heterogeneous academic data sources. Effectively integrating this data and constructing accurate and complete scholar relationship networks has become a key challenge. In this context, a method for storing and constructing scholar relationship networks based on an artificial intelligence graph database is particularly important. With the continuous growth of academic data and the increasing demands for depth and breadth in academic research, this type of scholar relationship network construction and storage method based on advanced technology has broad application prospects and is expected to promote academic research towards a more efficient and in-depth direction.

[0003] In constructing scholar relationship networks, existing technologies face numerous challenges. On one hand, the fusion of multi-source heterogeneous academic data presents difficulties. Inconsistent entity definitions and relationship descriptions across different data sources, coupled with a lack of unified logic for constructing heterogeneous knowledge graphs, lead to fragmented scholar relationship networks, hindering effective integration of multi-source data for cross-data source scholar association analysis. On the other hand, scholar node feature extraction is overly simplistic, relying solely on single features for entity alignment while neglecting semantic and network structure features. The failure to integrate multi-dimensional features to form a unified representation results in low entity alignment accuracy, impacting network integrity. Furthermore, heterogeneous knowledge graph fusion lacks a precise screening mechanism, blindly comparing nodes without effective alignment target selection, similarity quantification, or precise matching logic. This results in massive computational demands, low fusion efficiency, and a susceptibility to misfusion and redundancy. Moreover, poor storage architecture adaptability is a significant issue. Traditional relational databases or ordinary graph databases cannot adapt to the "high node density, multiple relationship types" characteristics of the fused large-scale unified network. The lack of distributed architecture support leads to high storage redundancy and inefficient multi-hop queries and relationship traversal operations, failing to meet the demands for efficient access and analysis.

[0004] Therefore, this invention proposes a method and system for constructing a scholar relationship network storage based on an artificial intelligence graph database. Summary of the Invention

[0005] This invention provides a method and system for constructing and storing scholar relationship networks based on artificial intelligence graph databases. By leveraging the advantages of artificial intelligence technology and graph databases, this method can deeply mine and integrate multi-source heterogeneous academic data, accurately construct scholar relationship networks, and achieve efficient storage. This not only helps researchers comprehensively understand the collaborative networks and scholar influence within the academic field, but also provides strong data support for talent evaluation and discipline planning in academic institutions.

[0006] This invention provides a method for constructing a scholar relationship network storage based on an artificial intelligence graph database, comprising: Based on all entities and relationships between all entities in the scholar relationship data from multiple heterogeneous academic data sources, a heterogeneous scholar relationship knowledge graph is constructed for each heterogeneous academic data source. The joint feature vector of each scholar node in the heterogeneous scholar relationship knowledge graph is extracted based on the graph neural network model. The joint feature vector is formed by fusing the semantic features of the corresponding scholar node and the multi-hop structure features in the corresponding heterogeneous scholar relationship knowledge graph. Based on the self-attribute vectors of each scholar node, a coarse screening of alignment targets is performed on pairs of scholar nodes in the heterogeneous scholar relationship knowledge graph from different heterogeneous academic data sources to obtain multiple combinations of scholar nodes to be aligned. Calculate the similarity of the joint feature vectors of all scholar nodes in each scholar node combination to be aligned, and use it as the similarity score of the corresponding scholar node combination to be aligned; Based on the similarity scores of all scholar node combinations to be aligned, all heterogeneous scholar relationship knowledge graphs are fused to obtain a unified scholar relationship network, which is then stored in a distributed native graph database.

[0007] Preferably, based on all entities and all relationships between entities in scholar relationship data from multiple heterogeneous academic data sources, a heterogeneous scholar relationship knowledge graph is constructed for each heterogeneous academic data source, including: Acquire scholar relationship data from multiple heterogeneous academic data sources; The scholars, paper titles, academic institutions and research fields in the scholar relationship data of each heterogeneous academic data source are processed into entities to obtain all entities of each heterogeneous academic data source. Based on all entities in each heterogeneous academic data source and the co-authorship, affiliation, publication, and belonging relationships among all entities, a heterogeneous scholar relationship knowledge graph is constructed for each heterogeneous academic data source.

[0008] Preferably, the joint feature vector of each scholar node in the heterogeneous scholar relationship knowledge graph is extracted based on the graph neural network model, including: Based on the pre-trained language model, the attribute information of the associated text of each scholar node in the heterogeneous scholar relationship knowledge graph is extracted to obtain the semantic features of the corresponding scholar node. Based on the graph neural network model, the connection patterns and neighbor attribute information of each scholar node in the heterogeneous scholar relationship knowledge graph are extracted and fused within the multi-order neighbor range, so as to obtain the multi-hop structure features of the corresponding scholar node in the corresponding heterogeneous scholar relationship knowledge graph. The semantic features of each scholar node are fused with the multi-hop structural features in the corresponding heterogeneous scholar relationship knowledge graph to obtain the joint feature vector of the corresponding scholar node.

[0009] Preferably, the process of obtaining the self-attribute vector of each scholar node includes: Obtain the relationship formation time, associated paper DOI identifier, and relationship type values ​​for each scholar node, and summarize them to generate the basic attribute vector for each scholar node; Analyze all collaboration relationships of each scholar node, and calculate the subject matching degree of all collaboration relationships of each scholar node based on the similarity of keywords of all papers of the collaborating parties in each collaboration relationship; Obtain all cited texts of all papers signed by each scholar node, and analyze the sentiment values ​​of all cited texts of all papers signed by each scholar. The semantic attribute vector of each scholar node is generated by summarizing the subject matching degree of all collaborations of each scholar node and the sentiment value of all cited texts of all papers under the signature. Generate the intensity attribute vector for each scholar node; The self-attribute vector includes the basic attribute vector, the semantic attribute vector, and the intensity attribute vector.

[0010] Preferably, the strength attribute vector for each scholar node is generated, including: The collaboration intensity of each scholar node is calculated based on the number of collaborative papers and the journal impact factor weight of the corresponding papers. The citation intensity of each scholar node is calculated based on the citation frequency of all papers signed by each scholar node and the citation weight of all cited papers. The citation intensity of each scholar node is calculated based on the citation frequency of all papers signed by each scholar node. The citation balance factor of each scholar node is calculated based on the citation strength and the cited strength of each scholar node. Based on the cooperation strength, citation strength, cited strength, and citation relationship balance factor of each scholar node, a strength attribute vector for each scholar node is generated.

[0011] Preferably, based on the self-attribute vectors of each scholar node, a coarse screening of alignment targets is performed on pairwise scholar nodes in the heterogeneous scholar relationship knowledge graph from different heterogeneous academic data sources, resulting in multiple combinations of scholar nodes to be aligned, including: The first similarity between basic attribute vectors, the second similarity between semantic attribute vectors, and the third similarity between intensity attribute vectors are calculated for each pair of scholar nodes in the heterogeneous scholar relationship knowledge graph from different heterogeneous academic data sources. In heterogeneous scholar relationship knowledge graphs from different heterogeneous academic data sources, any two scholar nodes whose first similarity is not less than the first similarity threshold, whose second similarity is not less than the second similarity threshold, and whose third similarity is not less than the third similarity threshold are combined to obtain multiple combinations of scholar nodes to be aligned.

[0012] Preferably, based on the similarity scores of all scholar node combinations to be aligned, all heterogeneous scholar relationship knowledge graphs are fused to obtain a unified scholar relationship network, including: Based on the similarity scores of all scholar node combinations to be aligned and the position of each scholar node in the heterogeneous scholar relationship knowledge graph, the three-dimensional projection coordinates of each scholar node are constructed. Identify scholar node combinations that belong to different heterogeneous scholar relationship knowledge graphs and have the highest similarity scores with each other, and use them as matching scholar node combinations. Mark the three-dimensional projection coordinates of all scholar nodes in each heterogeneous scholar relationship knowledge graph in a preset three-dimensional coordinate system, and combine all mutually matched scholar node combinations to determine at least one maximum hexahedron such that the scholar node matching density of the maximum hexahedron is not less than a preset density threshold, where the scholar node matching density of each maximum hexahedron = the volume of the maximum hexahedron ÷ the maximum value among the total number of mutually matched scholar node combinations of the two heterogeneous scholar relationship knowledge graphs it encloses; Based on all the largest hexahedrons, all heterogeneous scholar relationship knowledge graphs are merged to obtain a unified scholar relationship network.

[0013] Preferably, based on the similarity scores of all scholar node combinations to be aligned and the position of each scholar node in its respective heterogeneous scholar relationship knowledge graph, the three-dimensional projected coordinates of each scholar node are constructed, including: Based on the similarity scores of all scholar node combinations to be aligned, the maximum similarity score of each scholar node in each heterogeneous scholar relationship knowledge graph is determined, and all scholar nodes with a maximum similarity score greater than a preset score threshold are selected from each heterogeneous scholar relationship knowledge graph as all high-precision scholar nodes. Based on the distribution of all high-precision scholar nodes in each heterogeneous scholar relationship knowledge graph, a central scholar node is identified in each heterogeneous scholar relationship knowledge graph. Based on the central scholar node, multiple concentric structures that can cover all scholar nodes in the corresponding heterogeneous scholar relationship knowledge graph and the ordinal number of each concentric structure are identified, wherein each concentric structure covers at least one high-precision scholar node. The high-precision scholar node with the highest similarity score in each layer structure is taken as the starting scholar node of each layer structure. Based on the preset statistical direction, the number of node intervals between each scholar node in the heterogeneous scholar relationship knowledge graph and the starting scholar node of the corresponding layer structure is determined. Based on the ordinal number of the layer structure to which each scholar node belongs in the heterogeneous scholar relationship knowledge graph, the number of node intervals between the scholar node and the starting scholar node of the layer structure, and the maximum similarity score, the three-dimensional projection coordinates of the corresponding scholar node are constructed.

[0014] Preferably, based on all maximal hexahedrons, all heterogeneous scholar relationship knowledge graphs are fused to obtain a unified scholar relationship network, including: The two heterogeneous scholar relationship knowledge graphs corresponding to the maximum value among the total number of mutually matching scholar node combinations of the pairwise heterogeneous scholar relationship knowledge graphs enclosed by each maximal hexahedron are used as the coarse matching knowledge graph combination corresponding to each maximal hexahedron. In each coarse matching knowledge graph combination, the partial graph structure enclosed by all scholar nodes whose corresponding three-dimensional projection coordinates are covered by the corresponding maximum hexahedron in the two heterogeneous scholar relationship knowledge graphs are summarized and regarded as the graph structure combination to be aligned. The spectral structure combination to be aligned is corrected and aligned to obtain the spectral correction and alignment result; The graph correction and alignment results are updated to the corresponding heterogeneous scholar relationship knowledge graph to obtain a new heterogeneous scholar relationship knowledge graph. Based on all the new heterogeneous scholar relationship knowledge graphs, new graph structure combinations to be aligned are determined. The current heterogeneous scholar relationship knowledge graph is updated based on the new graph structure combinations to be aligned until there are no new graph structure combinations to be aligned. Then, all the currently obtained heterogeneous scholar relationship knowledge graphs are aligned and fused to obtain a unified scholar relationship network.

[0015] This invention provides a scholar relationship network storage and construction system based on an artificial intelligence graph database, comprising: The graph construction module is used to construct a heterogeneous scholar relationship knowledge graph for each heterogeneous academic data source based on all entities and all relationships between entities in the scholar relationship data from multiple heterogeneous academic data sources. The feature fusion module is used to extract the joint feature vector of each scholar node in the heterogeneous scholar relationship knowledge graph based on the graph neural network model. The joint feature vector is formed by fusing the semantic features of the corresponding scholar node and the multi-hop structure features in the corresponding heterogeneous scholar relationship knowledge graph. The target coarse screening module is used to perform alignment target coarse screening on pairs of scholar nodes in heterogeneous scholar relationship knowledge graphs from different heterogeneous academic data sources based on the self-attribute vectors of each scholar node, and obtain multiple combinations of scholar nodes to be aligned. The similarity calculation module is used to calculate the similarity of the joint feature vector of all scholar nodes in each scholar node combination to be aligned, which is used as the similarity score of the corresponding scholar node combination to be aligned. The heterogeneous fusion and storage module is used to fuse all heterogeneous scholar relationship knowledge graphs based on the similarity scores of all scholar node combinations to be aligned, obtain a unified scholar relationship network, and store the unified scholar relationship network in a distributed native graph database.

[0016] The beneficial effects of this invention compared to existing technologies are as follows: By constructing a knowledge graph of heterogeneous scholar relationships, it clarifies entity types and core relationships, standardizes and integrates multi-source heterogeneous data, breaks down information silos in the network, lays the foundation for cross-data source fusion, significantly improves the integration efficiency of multi-source academic data, and constructs a network covering more comprehensive scholar association information. It utilizes a graph neural network model to extract joint feature vectors of semantic features and multi-hop structural features, and combines this with coarse screening using self-attribute vectors. Entity alignment is performed from multiple dimensions, including text semantics, network structure, and attribute features, significantly improving alignment accuracy, effectively distinguishing between cases of identical names and different names belonging to the same person, enhancing network integrity, and providing reliable data support for subsequent academic analysis. Furthermore, it employs a two-stage fusion logic of coarse screening and similarity score quantification. Coarse screening using self-attribute vectors filters out a large number of low-similarity node combinations, reducing computational load and improving fusion efficiency; precise matching based on similarity scores using joint feature vectors avoids redundancy and misassociations, improving the relationship accuracy of the unified network. Finally, the unified scholar relationship network is stored in a distributed native graph database to adapt to the characteristics of large-scale networks with high number of nodes and multiple relationships, reduce storage redundancy, support dynamic expansion, optimize multi-hop queries and relationship traversal, control query latency, and meet the requirements for efficient access.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the overall process of constructing a scholar relationship network storage based on an artificial intelligence graph database, as described in this embodiment of the invention. Figure 2 This is a flowchart illustrating the joint feature vector extraction and entity alignment process in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] like Figure 1 and Figure 2 As shown, this invention provides an embodiment of a method for constructing a scholar relationship network storage based on an artificial intelligence graph database, comprising: Based on all entities and relationships between all entities in the scholar relationship data from multiple heterogeneous academic data sources, a heterogeneous scholar relationship knowledge graph is constructed for each heterogeneous academic data source. The joint feature vector of each scholar node in the heterogeneous scholar relationship knowledge graph is extracted based on the graph neural network model. The joint feature vector is formed by fusing the semantic features of the corresponding scholar node and the multi-hop structure features in the corresponding heterogeneous scholar relationship knowledge graph. Based on the self-attribute vectors of each scholar node, a coarse screening of alignment targets is performed on pairs of scholar nodes in the heterogeneous scholar relationship knowledge graph from different heterogeneous academic data sources to obtain multiple combinations of scholar nodes to be aligned. Calculate the similarity of the joint feature vectors of all scholar nodes in each scholar node combination to be aligned, and use it as the similarity score of the corresponding scholar node combination to be aligned; Based on the similarity scores of all scholar node combinations to be aligned, all heterogeneous scholar relationship knowledge graphs are fused to obtain a unified scholar relationship network, which is then stored in a distributed native graph database.

[0022] In this embodiment, heterogeneous academic data sources refer to multiple providers of academic data of different types, formats, or sources. These data sources may differ in terms of data structure, storage method, and data specifications. For example, some data sources may come from well-known academic journal databases, while others may be local databases compiled by an academic institution itself, such as DBLP, Google Scholar, and AMiner.

[0023] In this embodiment, the scholar relationship data is obtained from heterogeneous academic data sources and is information related to scholars and reflects the relationships between scholars and between scholars and other academic elements. This includes scholars' personal information, their published papers, their affiliated academic institutions, their research fields, and collaborative relationships among scholars.

[0024] In this embodiment, an entity is an abstract representation of a specific thing in the scholar relationship data. In this scenario, the scholars, paper titles, academic institutions, and research fields are represented as entities. For example, each scholar is considered an entity, and each paper title, each academic institution, and each research field are also considered as entities. These are the basic elements for constructing a knowledge graph.

[0025] In this embodiment, the heterogeneous scholar relationship knowledge graph is a set of knowledge representing scholar relationships from various heterogeneous academic data sources using a graphical structure. Nodes in the graph represent entities (such as scholars, papers, academic institutions, etc.), and edges represent relationships between entities (such as co-authorship, affiliation, publication, and belonging relationships). Due to the heterogeneity of data sources, the constructed knowledge graph may differ in structure, node attributes, and relationship types.

[0026] In this embodiment, the graph neural network model is a neural network specifically designed for processing graph-structured data. It can learn from the nodes and edges in the graph to extract features from the graph data. In this embodiment, it is used to extract the multi-hop structural features of each scholar node in the heterogeneous scholar relationship knowledge graph. The construction process of the graph neural network model includes: Acquire a large amount of sample data containing a heterogeneous scholar relationship knowledge graph with references (used as model input during training) and manually labeled connection patterns and neighbor attribute information within the multi-order neighbor range of each scholar node in the graph (used as model output during training). These sample data are divided into training set, validation set and test set.

[0027] Based on task requirements and data characteristics, a suitable model architecture is selected from common architectures such as graph convolutional networks and graph attention networks. After determining the architecture, the corresponding model parameters are set, such as the kernel size, number of layers, and output dimension of each layer in graph convolutional networks, and the attention mechanism parameters in graph attention networks.

[0028] The model is then trained using a training set. During training, the model learns the features of nodes and edges in the training set samples based on the selected architecture and parameters, continuously adjusting its parameters to minimize the loss function value. The loss function measures the difference between the model's predictions and the actual situation; for example, when predicting the features of a scholar node, it calculates the difference between the predicted feature vector and the true feature vector. Simultaneously, a validation set is used to monitor model performance and prevent overfitting or underfitting. Training stops when the model's performance on the validation set no longer improves. Finally, the model's performance on unseen data is evaluated using a test set to ensure good generalization ability. Through these steps, a graph neural network model is finally obtained for extracting connection patterns and neighbor attribute information within the multi-order neighborhood range of scholar nodes.

[0029] In this embodiment, scholar nodes are a type of node in a heterogeneous scholar relationship knowledge graph, representing specific individual scholars. Each scholar node contains attribute information related to that scholar and is connected to other nodes (such as paper nodes, academic institution nodes, etc.) through edges to reflect the relationship between scholars and other academic elements.

[0030] In this embodiment, the semantic features of scholar nodes are obtained by extracting attribute information from the associated text of scholar nodes through a pre-trained language model. These associated texts may include abstracts of papers published by the scholar, personal profiles, etc. Semantic features reflect the characteristics of the scholar in terms of academic content and help to understand the scholar's research direction, academic contributions, and other information from a semantic level.

[0031] In this embodiment, the multi-hop structure feature in the corresponding heterogeneous scholar relationship knowledge graph refers to the feature extracted by the graph neural network model, which is the fusion of the connection pattern and neighbor attribute information within the multi-order neighbor range of a scholar node in its heterogeneous scholar relationship knowledge graph. For example, the multi-hop structure feature of a scholar node may include information about its direct collaborating scholars (first-order neighbors), as well as the connection methods and attribute information of multi-order neighbors such as the collaborating scholars of these collaborating scholars (second-order neighbors).

[0032] In this embodiment, the scholar node combinations to be aligned are multiple pairs or sets of scholar nodes obtained after coarsely screening each pair of scholar nodes in a heterogeneous scholar relationship knowledge graph from different heterogeneous academic data sources based on their self-attribute vectors. The scholar nodes in these combinations are initially considered to possibly represent the same scholar, and further verification is needed by calculating the similarity of joint feature vectors to determine whether they are truly matched and aligned.

[0033] In this embodiment, the distributed native graph database is a distributed database specifically designed for storing graph structure data (such as Neo4j Cluster), possessing distributed storage and processing capabilities. It can store and manage a unified scholar relationship network using a native graph data model, supporting efficient graph query and analysis operations. Its distributed nature enables it to process large-scale graph data, meeting the storage requirements for scholar relationship networks formed by the fusion of numerous heterogeneous academic data sources.

[0034] To acquire scholar relationship data from heterogeneous academic data sources, entityize relevant elements, and construct heterogeneous scholar relationship knowledge graphs based on the relationships between entities, thus laying the foundation for subsequent scholar relationship network construction, this paper proposes to construct heterogeneous scholar relationship knowledge graphs for each heterogeneous academic data source based on all entities and all relationships between entities in the scholar relationship data from multiple heterogeneous academic data sources. This includes: Acquire scholar relationship data from multiple heterogeneous academic data sources; The scholars, paper titles, academic institutions and research fields in the scholar relationship data of each heterogeneous academic data source are processed into entities to obtain all entities of each heterogeneous academic data source. Based on all entities in each heterogeneous academic data source and the co-authorship, affiliation, publication, and belonging relationships among all entities, a heterogeneous scholar relationship knowledge graph is constructed for each heterogeneous academic data source.

[0035] In this embodiment, the scholars, paper titles, academic institutions, and research fields in the scholar relationship data of various heterogeneous academic data sources are entityized to obtain all entities of each heterogeneous academic data source. This means transforming these different types of information into identifiable and operable basic units in a knowledge graph. A scholar, as an entity, contains their personal information; a paper title represents a specific paper entity and may be associated with various attributes of the paper; an academic institution entity contains the relevant characteristics of the institution; and a research field entity reflects a specific academic direction. For example, the scholar "Zhang San," the paper title "On the Development of Artificial Intelligence," the academic institution "School of Computer Science, XX University," and the research field "Artificial Intelligence" are respectively transformed into entities that can be used for relationship construction and analysis in the knowledge graph, facilitating subsequent sorting and modeling of academic relationships.

[0036] In this embodiment, co-authorship, affiliation, publication, and belonging relationships between entities are ways to describe the connections between different entities. Co-authorship refers to multiple scholars jointly writing the same paper, reflecting a collaborative relationship between scholars. For example, if scholar A and scholar B jointly publish a paper, they have a co-authorship relationship. Affiliation indicates the association between a scholar and their affiliated academic institution. For example, if scholar C belongs to "XX University," this indicates that scholar C has an affiliation with that university. Publication relationship refers to the correspondence between a scholar and their published papers; that is, a scholar has published a particular paper. Belonging relationship is used to indicate the affiliation between a research field and a scholar, paper, or academic institution. For example, the research field of "artificial intelligence" is a key research direction of a certain academic institution, or a paper belongs to the research field of "artificial intelligence."

[0037] In this embodiment, a heterogeneous scholar relationship knowledge graph is constructed based on all entities in each heterogeneous academic data source and the co-authorship, affiliation, publication, and belonging relationships between all entities. This involves drawing a graph structure that displays academic relationships, with entities as nodes and these relationships as edges. For example, a scholar entity is centered, connected to other scholars through co-authorship, to paper entities through publication, to academic institutions through affiliation, and to research fields through belonging. In this way, each heterogeneous academic data source can construct a knowledge graph reflecting the scholar relationships within its own data. Due to the heterogeneity of the data sources, different knowledge graphs may differ in node details, the number and type of relationships, but they can all present the academic relationship network within a specific data source.

[0038] like Figure 2 As shown, to extract semantic features and multi-hop structural features of scholar nodes using pre-trained language models and graph neural network models respectively, and to fuse the two to obtain a joint feature vector for each scholar node, providing an effective feature basis for entity alignment, a method based on graph neural network models is proposed to extract the joint feature vector of each scholar node in heterogeneous scholar relationship knowledge graphs, including: Based on the pre-trained language model, the attribute information of the associated text of each scholar node in the heterogeneous scholar relationship knowledge graph is extracted to obtain the semantic features of the corresponding scholar node. Based on the graph neural network model, the connection patterns and neighbor attribute information of each scholar node in the heterogeneous scholar relationship knowledge graph are extracted and fused within the multi-order neighbor range, so as to obtain the multi-hop structure features of the corresponding scholar node in the corresponding heterogeneous scholar relationship knowledge graph. The semantic features of each scholar node are fused with the multi-hop structural features in the corresponding heterogeneous scholar relationship knowledge graph to obtain the joint feature vector of the corresponding scholar node.

[0039] In this embodiment, the pre-trained language model is a model that is pre-trained on large-scale text data. It is constructed by learning from a large amount of sample data containing a knowledge graph of heterogeneous scholar relationships (used as model input during training) and attribute information of the associated text of each scholar node (used as model output during training). For example, common models such as BERT and GPT can effectively capture semantic information in text after being trained on large-scale text.

[0040] In this embodiment, the attribute information of the associated text of a scholar node refers to the information contained in the text associated with the scholar node that reflects the scholar's characteristics. This associated text may include the titles, abstracts, and personal profiles of the scholar's published papers. The attribute information may involve aspects such as research topics, research methods, and academic contributions. For example, the research methods and innovative points mentioned in a paper's abstract constitute the attribute information of the text associated with that scholar node.

[0041] In this embodiment, the attribute information of the associated text of each scholar node in the heterogeneous scholar relationship knowledge graph is extracted based on the pre-trained language model to obtain the semantic features of the corresponding scholar node. The specific process is as follows: the text related to each scholar node is input into the pre-trained language model, the model analyzes and processes these texts, extracts the key attribute information from them, and converts them into semantic features represented in vector form.

[0042] In this embodiment, the connection patterns and neighbor attribute information within the multi-level neighbor range of a scholar node describe the connection methods of multi-level neighbor nodes surrounding a particular scholar node in a heterogeneous scholar relationship knowledge graph, as well as the attribute information of these neighbor nodes. First-level neighbors are nodes directly connected to the scholar node, second-level neighbors are nodes connected to first-level neighbors, and so on. Connection patterns reflect the association structure between neighbor nodes, such as chain connections or mesh connections; neighbor attribute information refers to the characteristics contained within the neighbor nodes themselves, such as the scholar neighbor's research field, number of published papers, and the publication year and citation count of paper neighbors.

[0043] In this embodiment, a graph neural network model is used to extract and fuse the connection patterns and neighbor attribute information within the multi-order neighbor range of each scholar node in the heterogeneous scholar relationship knowledge graph. This yields the multi-hop structure features of the corresponding scholar node in its respective heterogeneous scholar relationship knowledge graph, meaning that a graph neural network model is used to analyze each scholar node and its multi-order neighbors. The model learns the connection patterns between neighbor nodes and their attribute information, and then fuses this information together to form a comprehensive feature representation, namely, the multi-hop structure feature.

[0044] In this embodiment, the semantic features of each scholar node and the multi-hop structural features in the corresponding heterogeneous scholar relationship knowledge graph are fused to obtain the joint feature vector of the corresponding scholar node. This involves combining the semantic features obtained through the pre-trained language model with the multi-hop structural features extracted by the graph neural network model using mathematical methods such as vector addition and concatenation to form a new joint feature vector. This vector more comprehensively and accurately represents the characteristics of the scholar in the academic relationship network, providing richer and more effective information for subsequent tasks such as scholar node alignment and knowledge graph fusion.

[0045] To obtain various attribute information of scholar nodes, including relationship formation time and paper DOI identifiers, and to calculate subject matching degree and sentiment tendency value, thereby generating a self-attribute vector containing basic, semantic, and intensity attribute vectors, and providing multi-dimensional attribute basis for coarse screening, this paper proposes a process for obtaining the self-attribute vector of each scholar node, including: Obtain the relationship formation time, associated paper DOI identifier, and relationship type values ​​for each scholar node, and summarize them to generate the basic attribute vector for each scholar node; Analyze all collaboration relationships of each scholar node, and calculate the subject matching degree of all collaboration relationships of each scholar node based on the similarity of keywords of all papers of the collaborating parties in each collaboration relationship; Obtain all cited texts of all papers signed by each scholar node, and analyze the sentiment values ​​of all cited texts of all papers signed by each scholar. The semantic attribute vector of each scholar node is generated by summarizing the subject matching degree of all collaborations of each scholar node and the sentiment value of all cited texts of all papers under the signature. Generate the intensity attribute vector for each scholar node; The self-attribute vector includes the basic attribute vector, the semantic attribute vector, and the intensity attribute vector.

[0046] In this embodiment, the formation time of all relationships, DOI identifiers of associated papers, and relationship type values ​​of each scholar node are obtained and summarized to generate a basic attribute vector for each scholar node. Specifically, starting from each scholar node, the time when it establishes relationships with other nodes is collected, which reflects the timeline of the scholar's academic activities. The DOI identifier (i.e., unique identifier of digital objects) of associated papers can uniquely identify related papers, facilitating the location and query of specific research content. Relationship types (such as co-authorship, affiliation, etc.) are assigned values, and this information is integrated to form a vector. This vector describes some key attributes of the scholar node at a basic level, providing basic information for subsequent analysis. For example, scholar A and scholar B have a co-authorship relationship in 2020 (assigned a value of 1), and the DOI of the associated paper is "12345". This information is arranged in a specific order to form a vector [2020,12345,1]. Multiple similar pieces of information are summarized into a basic attribute vector.

[0047] In this embodiment, the similarity of keywords in all papers of both parties = the number of identical keyword groups in all papers of both parties ÷ the total number of the largest number of keywords in the papers of both parties.

[0048] In this embodiment, all cooperative relationships of a scholar node refer to other scholar nodes that have cooperative relationships with the scholar node in the heterogeneous scholar relationship knowledge graph, as well as the specific information of such cooperative relationships.

[0049] In this embodiment, the subject matching degree of all cooperative relationships of each scholar node is calculated based on the similarity of the keywords of all papers of both parties in each cooperative relationship. In other words, the average of the similarity of the keywords of all papers of both parties in each cooperative relationship of each scholar node is taken as the subject matching degree of all cooperative relationships of the corresponding scholar node.

[0050] In this embodiment, all cited text in a paper refers to the specific textual content of other literature cited within the paper. These cited texts contain relevant information about the cited literature, as well as the citer's understanding and application of the cited content, and can reflect, to some extent, the research background, theoretical basis, and relevance to other research within the paper. For example, if a paper cites a viewpoint from another paper and explains in the cited text how that viewpoint relates to this research, this part of the cited text constitutes the paper's cited text.

[0051] In this embodiment, the sentiment analysis of all cited texts from all papers authored by each scholar is performed by inputting the cited texts from each scholar's papers into a pre-trained sentiment analysis model. This model analyzes the texts, determining whether the sentiment expressed is positive, negative, or neutral, and assigns a numerical sentiment value between -1 and 1. For example, 0.6 indicates a positive sentiment, -0.4 indicates a negative sentiment, and 0 indicates a neutral sentiment. This allows understanding of the scholar's emotional attitude when citing other literature. This sentiment analysis model is obtained through deep learning on a large sample containing paper texts (used as model input during deep learning) and manually labeled corresponding sentiment values ​​(used as model output during deep learning).

[0052] In this embodiment, the strength attribute vector of a scholar node is a vector that comprehensively reflects the scholar's influence and activity in academic collaboration, citation, and other aspects. For example, scholar I has a collaboration strength of 0.6, a citation strength of 0.7, a cited strength of 0.8, and a citation balance factor of 0.75. These values ​​are arranged in a specific order to form a vector [0.6, 0.7, 0.8, 0.75], which is the strength attribute vector of scholar I, used to measure the strength level of scholar I in the academic relationship network.

[0053] To generate the strength attribute vector for each scholar node by calculating cooperation strength, citation strength, cited strength, and citation balance factor, and to improve the composition of the self-attribute vector for a more comprehensive evaluation of scholar node characteristics, this paper proposes a method for generating the strength attribute vector for each scholar node, including: The collaboration intensity of each scholar node is calculated based on the number of collaborative papers and the journal impact factor weight of the corresponding papers. The citation intensity of each scholar node is calculated based on the citation frequency of all papers signed by each scholar node and the citation weight of all cited papers. The citation intensity of each scholar node is calculated based on the citation frequency of all papers signed by each scholar node. The citation balance factor of each scholar node is calculated based on the citation strength and the cited strength of each scholar node. Based on the cooperation strength, citation strength, cited strength, and citation relationship balance factor of each scholar node, a strength attribute vector for each scholar node is generated.

[0054] In this embodiment, the number of collaborative papers for each scholar node refers to the number of papers co-authored with that scholar, reflecting the scholar's activity level in academic collaboration. The journal impact factor weight for each paper is a weighted value assigned to each collaborative paper based on the journal's influence in the academic community. Generally, the higher the journal's influence, the higher its impact factor, and the greater the weight of the corresponding paper. For example, if a scholar has participated in the collaboration of 5 papers, the weight of the paper published in a high-impact journal might be set to 0.8, and the weight of the paper published in a journal with moderate impact might be set to 0.5.

[0055] In this embodiment, the collaboration strength of each scholar node is calculated based on the number of collaborative papers and the journal impact factor weight of the corresponding papers. It can be the ratio of the sum of the journal impact factor weights of all collaborative papers to the number of collaborative papers.

[0056] In this embodiment, the citation frequency of a paper refers to the number of times a paper is cited by other documents.

[0057] In this embodiment, a cited paper refers to a paper that cites other literature in academic research. When one paper cites another, the former is the cited paper, which conducts further research or elaboration based on the viewpoints, methods, etc., of the cited paper. For example, if paper A cites the research results of paper B in its research content, then paper A is the cited paper.

[0058] In this embodiment, the citation count weight is a weight value assigned based on the number of times the paper itself has been cited, which is generally the ratio of the number of citations to the maximum number of citations.

[0059] In this embodiment, the citation strength of each scholar node is calculated based on the citation frequency of all papers signed by each scholar node and the citation weight of all cited papers. This is achieved by taking the ratio of the sum of the citation weights of all cited papers signed by each scholar node to the citation frequency of all papers as the citation strength of each scholar node.

[0060] In this embodiment, the citation frequency of a paper refers to the number of times a paper is cited by other papers.

[0061] In this embodiment, the citation intensity of each scholar node is calculated based on the citation frequency of all papers signed by each scholar node, and the ratio of the average citation frequency of all papers signed by each scholar node to the maximum citation frequency is taken as the citation intensity of each scholar node.

[0062] In this embodiment, the citation relationship balance factor of each scholar node is calculated based on the citation strength and the cited strength of each scholar node. That is, the ratio of the citation strength to the cited strength of each scholar node is used as the citation relationship balance factor of each scholar node.

[0063] In this embodiment, the strength attribute vector of each scholar node is generated based on the cooperation strength, citation strength, cited strength, and citation relationship balance factor of each scholar node. This integrates these indicators that reflect different aspects of the scholar's academic strength into one vector.

[0064] To calculate the similarity between the self-attribute vectors of pairwise scholar nodes from different heterogeneous academic data sources, and to group scholar nodes that meet the similarity threshold into pairs of scholar nodes to be aligned, thus initially screening potentially matching scholar nodes and reducing subsequent computational cost, this paper proposes a method based on the self-attribute vectors of each scholar node to perform a coarse screening of alignment targets for pairwise scholar nodes in heterogeneous scholar relationship knowledge graphs from different academic data sources, obtaining multiple pairs of scholar nodes to be aligned, including: The first similarity between basic attribute vectors, the second similarity between semantic attribute vectors, and the third similarity between intensity attribute vectors are calculated for each pair of scholar nodes in the heterogeneous scholar relationship knowledge graph from different heterogeneous academic data sources. In heterogeneous scholar relationship knowledge graphs from different heterogeneous academic data sources, any two scholar nodes whose first similarity is not less than the first similarity threshold, whose second similarity is not less than the second similarity threshold, and whose third similarity is not less than the third similarity threshold are combined to obtain multiple combinations of scholar nodes to be aligned.

[0065] In this embodiment, the first similarity between the basic attribute vectors, the second similarity between the semantic attribute vectors, and the third similarity between the intensity attribute vectors of each pair of scholar nodes in the heterogeneous scholar relationship knowledge graph of different heterogeneous academic data sources are calculated. The similarity is calculated by the cosine similarity between vectors.

[0066] In this embodiment, the first similarity threshold is a pre-set numerical standard used to determine whether the similarity between the basic attribute vectors of two scholar nodes in a heterogeneous scholar relationship knowledge graph from different heterogeneous academic data sources meets the requirements for further processing. For example, if the first similarity threshold is set to 0.6, and the first similarity between the basic attribute vectors of two scholar nodes is 0.7, then the further screening condition is met; if it is 0.5, then it is not met.

[0067] In this embodiment, the second similarity threshold is also a pre-set standard, used to judge the similarity between semantic attribute vectors. It is used to measure whether the similarity between two scholar nodes at the academic semantic level meets the requirements. For example, if the second similarity threshold is set to 0.55, and the second similarity between the semantic attribute vectors of two scholar nodes is 0.6, it can be further examined; if it is 0.5, it does not meet the requirements.

[0068] In this embodiment, the third similarity threshold is also a pre-set value used to evaluate the similarity between intensity attribute vectors. It measures the similarity between two scholar nodes in terms of academic influence and activity. For example, if the third similarity threshold is set to 0.65, a third similarity of 0.7 between the intensity attribute vectors of two scholar nodes satisfies the condition; if it is 0.6, it does not. By using these three similarity thresholds, scholar nodes are filtered from different dimensions, improving the accuracy of subsequently identifying the same scholar.

[0069] To construct 3D projected coordinates based on the similarity scores of scholar node combinations to be aligned and the positions of scholar nodes, determine mutually matching scholar node combinations, and thus determine the maximum hexahedron, providing a spatial structural foundation for the fusion of heterogeneous knowledge graphs, this paper proposes to fuse all heterogeneous scholar relationship knowledge graphs based on the similarity scores of all scholar node combinations to be aligned, obtaining a unified scholar relationship network, including: Based on the similarity scores of all scholar node combinations to be aligned and the position of each scholar node in the heterogeneous scholar relationship knowledge graph, the three-dimensional projection coordinates of each scholar node are constructed. Identify scholar node combinations that belong to different heterogeneous scholar relationship knowledge graphs and have the highest similarity scores with each other, and use them as matching scholar node combinations. The three-dimensional projection coordinates of all scholar nodes in each heterogeneous scholar relationship knowledge graph are marked in a preset three-dimensional coordinate system. Combined with all mutually matched scholar node combinations, at least one maximal hexahedron is determined such that the scholar node matching density of the maximal hexahedron is not less than a preset density threshold. The scholar node matching density of each maximal hexahedron is the ratio of the volume of the maximal hexahedron to the maximum value among the total number of mutually matched scholar node combinations in the two surrounding heterogeneous scholar relationship knowledge graphs. Based on all the largest hexahedrons, all heterogeneous scholar relationship knowledge graphs are merged to obtain a unified scholar relationship network.

[0070] In this embodiment, scholar node pairs belonging to different heterogeneous scholar relationship knowledge graphs and having the highest similarity scores are identified as mutually matched scholar node pairs. This means that in multiple heterogeneous scholar relationship knowledge graphs, all scholar node pairs between different knowledge graphs are traversed, and the similarity score of each pair of scholar nodes is calculated. For each scholar node, the scholar node with the highest similarity score in other knowledge graphs is found. When the similarity scores of these two scholar nodes are the highest among all node pairs in their respective knowledge graphs, these two scholar nodes constitute a mutually matched scholar node pair. For example, in knowledge graphs A and B, scholar A calculates the similarity between all scholar nodes in knowledge graph A and knowledge graph B, and finds that the similarity score with scholar B is the highest; at the same time, scholar B calculates the similarity between all scholar nodes in knowledge graph B and knowledge graph A, and also finds that the similarity score with scholar A is the highest. Therefore, scholar A and scholar B form a mutually matched scholar node pair.

[0071] In this embodiment, highly similar scholar node pairs from different heterogeneous knowledge graphs have similar coordinates in three-dimensional space. Among them, scholar nodes belonging to the same heterogeneous scholar relationship knowledge graph have a uniform distribution. Only scholar nodes from different heterogeneous scholar relationship knowledge graphs that cannot be highly similar will deviate from the group in three-dimensional space. If the three-dimensional projection coordinates of these deviated scholar nodes are also included in the hexahedron, the matching density of scholar nodes in the hexahedron will decrease. Therefore, by limiting the volume of the hexahedron and its scholar node matching density, the spatial range of the three-dimensional projection coordinates of highly similar scholar node pairs in three-dimensional space can be defined.

[0072] In this embodiment, the preset density threshold is a pre-defined value used to evaluate the tightness of scholar node matching within the max-hexahedron. This threshold serves as a standard for determining whether the max-hexahedron meets specific requirements. If the scholar node matching density within the max-hexahedron reaches or exceeds the preset density threshold, it indicates that the polyhedron contains numerous matching relationships between scholar nodes from different knowledge graphs, thus possessing high fusion value. Conversely, if it falls below the preset density threshold, adjustments to the polyhedron's construction or fusion strategy may be necessary. The preset density threshold can be set to 0.8 pairs per unit space volume.

[0073] In this embodiment, the scholar node matching density of each maxima hexahedron is the ratio of the volume of the maxima hexahedron to the maximum value among the total number of mutually matching scholar node combinations in the pairwise heterogeneous scholar relationship knowledge graphs it encloses. A quantitative method is used to measure the tightness of scholar node matching within the maxima hexahedron. The volume of the maxima hexahedron can be calculated based on its geometric shape in three-dimensional space and node coordinates, and the total number of mutually matching scholar node combinations in the pairwise heterogeneous scholar relationship knowledge graphs enclosed by the maxima hexahedron can also be statistically determined. When calculating the matching density, the total number of mutually matching scholar node combinations in each pair of heterogeneous scholar relationship knowledge graphs covered by the maxima hexahedron is first counted, and then the maximum value is selected from the total number for calculation. For example, assuming the volume of a maxima hexahedron is V, and the total number of mutually matching scholar node combinations in the pairwise heterogeneous scholar relationship knowledge graphs it encloses are N1, N2, N3, ..., the maximum value Nmax is taken, then the scholar node matching density of the maxima hexahedron = V / Nmax.

[0074] To determine high-precision scholar nodes based on the similarity scores of scholar node combinations to be aligned, the layer structure and starting scholar node are calibrated. The 3D projected coordinates of each scholar node are constructed based on the node interval number and related attributes. A precise 3D coordinate system for graph fusion and localization is then built. This paper proposes constructing the 3D projected coordinates of each scholar node based on the similarity scores of all scholar node combinations to be aligned and the position of each scholar node in its respective heterogeneous scholar relationship knowledge graph. The coordinates include: Based on the similarity scores of all scholar node combinations to be aligned, the maximum similarity score of each scholar node in each heterogeneous scholar relationship knowledge graph is determined, and all scholar nodes with a maximum similarity score greater than a preset score threshold are selected from each heterogeneous scholar relationship knowledge graph as all high-precision scholar nodes. Based on the distribution of all high-precision scholar nodes in each heterogeneous scholar relationship knowledge graph, a central scholar node is identified in each heterogeneous scholar relationship knowledge graph. Based on the central scholar node, multiple concentric structures that can cover all scholar nodes in the corresponding heterogeneous scholar relationship knowledge graph and the ordinal number of each concentric structure are identified, wherein each concentric structure covers at least one high-precision scholar node. The high-precision scholar node with the highest similarity score in each layer structure is taken as the starting scholar node of each layer structure. Based on the preset statistical direction, the number of node intervals between each scholar node in the heterogeneous scholar relationship knowledge graph and the starting scholar node of the corresponding layer structure is determined. Based on the ordinal number of the layer structure to which each scholar node belongs in the heterogeneous scholar relationship knowledge graph, the number of node intervals between the scholar node and the starting scholar node of the layer structure, and the maximum similarity score, the three-dimensional projection coordinates of the corresponding scholar node are constructed.

[0075] In this embodiment, the preset score threshold is a pre-set numerical standard used to filter out scholar nodes with high similarity scores. For example, if the preset score threshold is set to 0.7, and the similarity score of a scholar node with other nodes forming a group of scholar nodes to be aligned reaches 0.75, then the scholar node meets the further filtering conditions; otherwise, it will be excluded from certain subsequent operations.

[0076] In this embodiment, based on the similarity scores of all scholar node combinations to be aligned, the maximum similarity score of each scholar node in each heterogeneous scholar relationship knowledge graph is determined. This means that for each scholar node in each heterogeneous scholar relationship knowledge graph, the maximum similarity score is found among all scholar node combinations to be aligned formed by it and scholar nodes in other knowledge graphs. This maximum value represents the highest similarity of that scholar node among all possible matching relationships.

[0077] In this embodiment, based on the distribution position of all high-precision scholar nodes in each heterogeneous scholar relationship knowledge graph, the central scholar node is identified in each heterogeneous scholar relationship knowledge graph. That is, the scholar node with a relatively central position and a high maximum similarity score in each heterogeneous scholar relationship knowledge graph is regarded as the central scholar node.

[0078] In this embodiment, multiple concentric structures covering all scholar nodes in the corresponding heterogeneous scholar relationship knowledge graph are identified based on the central scholar node, along with the ordinal number of each concentric structure. Using the determined central scholar node as the core, other scholar nodes in the graph are divided into multiple concentric structures according to rules such as distance or connection to the central scholar node. Scholar nodes closer to the central scholar node form the inner concentric layer, while those farther away form the outer concentric layer. Each concentric structure is assigned an ordinal number, increasing sequentially from the inside out, such as the first layer, the second layer, etc. For example, scholar nodes directly connected to the central scholar node form the first layer, scholar nodes indirectly connected through one layer form the second layer, and so on, facilitating hierarchical analysis of the graph.

[0079] In this embodiment, the high-precision scholar node with the highest similarity score in each concentric structure is taken as the starting scholar node of each concentric structure. Within each concentric structure, the node with the highest similarity score is found from among many high-precision scholar nodes and defined as the starting scholar node of that concentric structure.

[0080] In this embodiment, based on a preset statistical direction, the number of node intervals between each scholar node in the heterogeneous scholar relationship knowledge graph and the starting scholar node of its respective layer structure is determined. This preset statistical direction may be clockwise or counterclockwise. Following this direction, within each layer structure, the number of node intervals between each scholar node and the starting scholar node of that layer is calculated. For example, starting from the starting scholar node and moving clockwise along the layer structure, the number of nodes traversed to reach another scholar node A is counted; this number is the node interval count.

[0081] In this embodiment, the three-dimensional projected coordinates of each scholar node are constructed based on the ordinal number of the layer structure to which each scholar node belongs in the heterogeneous scholar relationship knowledge graph, the number of node intervals between the scholar node and the starting scholar node in the same layer structure, and the maximum similarity score. These three pieces of information are used as three dimensions to determine the position of the scholar node in three-dimensional space. The ordinal number of the layer structure represents the level of the scholar node, reflecting its macroscopic position in the overall graph structure; the number of node intervals reflects the relative position of the scholar node within the layer relative to the starting scholar node; and the maximum similarity score provides a quantitative indicator for determining the position of the scholar node from a similarity perspective. For example, if a scholar node is located in the third layer, has a node interval of 5 with the starting scholar node, and a maximum similarity score of 0.8, then its three-dimensional projected coordinates can be represented as (3, 5, 0.8).

[0082] To determine the coarse-matching knowledge graph combination corresponding to each maximal hexahedron, summarize and correct the alignment of the graph structures to be aligned, continuously update the heterogeneous scholar relationship knowledge graph until no new graph structures to be aligned are found, and finally align and fuse to obtain a unified scholar relationship network, this paper proposes to fuse all heterogeneous scholar relationship knowledge graphs based on all maximal hexahedrons to obtain a unified scholar relationship network, including: The two heterogeneous scholar relationship knowledge graphs corresponding to the maximum value among the total number of mutually matching scholar node combinations of the pairwise heterogeneous scholar relationship knowledge graphs enclosed by each maximal hexahedron are used as the coarse matching knowledge graph combination corresponding to each maximal hexahedron. In each coarse matching knowledge graph combination, the partial graph structure enclosed by all scholar nodes whose corresponding three-dimensional projection coordinates are covered by the corresponding maximum hexahedron in the two heterogeneous scholar relationship knowledge graphs are summarized and regarded as the graph structure combination to be aligned. The spectral structure combination to be aligned is corrected and aligned to obtain the spectral correction and alignment result; The graph correction and alignment results are updated to the corresponding heterogeneous scholar relationship knowledge graph to obtain a new heterogeneous scholar relationship knowledge graph. Based on all the new heterogeneous scholar relationship knowledge graphs, new graph structure combinations to be aligned are determined. The current heterogeneous scholar relationship knowledge graph is updated based on the new graph structure combinations to be aligned until there are no new graph structure combinations to be aligned. Then, all the currently obtained heterogeneous scholar relationship knowledge graphs are aligned and fused to obtain a unified scholar relationship network.

[0083] In this embodiment, the two heterogeneous scholar relationship knowledge graphs corresponding to the maximum value among the total number of mutually matching scholar node combinations among all pairs of heterogeneous scholar relationship knowledge graphs enclosed by each maximal hexahedron are taken as the coarse matching knowledge graph combination corresponding to each maximal hexahedron. This step is to find knowledge graph pairs with potentially high matching degrees among multiple heterogeneous scholar relationship knowledge graphs based on the mutually matching scholar node combinations enclosed by the maximal hexahedron. For example, for a certain maximal hexahedron, the number of mutually matching scholar node combinations between different pairs of heterogeneous knowledge graphs it encloses is counted. Suppose there are knowledge graphs A, B, C, etc., and there are 10 mutually matching scholar node combinations between A and B, 8 between A and C, and 6 between B and C. 10 is the maximum value among these combinations. Then, knowledge graphs A and B constitute the coarse matching knowledge graph combination corresponding to the maximal hexahedron.

[0084] In this embodiment, the partial graph structures formed by all scholar nodes whose corresponding 3D projection coordinates are covered by the corresponding maximum hexahedron in the two heterogeneous scholar relationship knowledge graphs of each coarse-matching knowledge graph combination are summarized and treated as the graph structure combination to be aligned. For the previously determined coarse-matching knowledge graph combination, in these two knowledge graphs, the scholar nodes whose 3D projection coordinates fall within the range of the maximum hexahedron are found. The partial graph structures formed by these nodes and their connections are the objects that need further processing. These partial graph structures are summarized to form the graph structure combination to be aligned. For example, in the coarse-matching knowledge graphs X and Y, the maximum hexahedron covers some scholar nodes a, b, c in X and their connections, and scholar nodes d, e, f in Y and their connections. Summarizing these partial graphs containing nodes and connections yields the graph structure combination to be aligned.

[0085] In this embodiment, the graph structure combination to be aligned is corrected and aligned to obtain the graph correction and alignment result. For the graph structure combination to be aligned, the graph structures from different knowledge graphs are adjusted and matched to make them as consistent as possible in terms of nodes and relationships, eliminating differences caused by heterogeneous data sources. For example, the attributes of nodes may be standardized and the representation of relationships may be adjusted. Through these operations, the corrected and aligned result is obtained, which allows corresponding parts of different knowledge graphs to be better integrated. For example, nodes representing the same scholar in different knowledge graphs are merged, and the attribute information of nodes is unified, making the relationships between nodes clearer and more accurate. This is the process and result of graph correction and alignment.

[0086] When correcting and aligning assemblies of aligned spectrogram structures, the following common algorithms or rules may be used: Algorithms based on attribute matching: Attribute similarity calculation: This involves calculating the similarity of attribute values ​​between nodes in different knowledge graphs. For example, string similarity algorithms (such as edit distance algorithms) can be used to compare the name attribute of scholar nodes. If the edit distance between two names is within a certain threshold, they are considered to likely represent the same entity. For instance, the similarity between "Zhang Jianguo" and "Zhang Jianguo" is calculated using the edit distance algorithm. If the similarity is higher than the set threshold, other attributes are further examined.

[0087] Attribute type matching: Ensure that attributes of the same type have consistent meanings and representations across different knowledge graphs. For example, for the "publication year of the paper" attribute in all knowledge graphs, unify its data format to four digits.

[0088] Algorithms based on graph structures: Subgraph isomorphism algorithms identify isomorphic relationships between different subgraphs in a graph structure combination to be aligned; that is, they find subgraph parts with similar structures and consistent node-edge correspondences. By identifying these isomorphic subgraphs, it is determined which nodes and edges can be aligned. For example, if two knowledge graphs both contain a triangular subgraph structure formed by three scholar nodes connected by co-authorship edges, then the nodes and edges of these two subgraphs can be aligned accordingly.

[0089] Shortest path algorithms: Shortest path algorithms (such as Dijkstra's algorithm) are used to analyze the connection paths between nodes. If the shortest path lengths and structures from two nodes to other key nodes are similar in different knowledge graphs, then these two nodes may represent the same entity. For example, in knowledge graph A, the shortest path length and structure from scholar X to "domain authority scholar" Y and "renowned academic institution" Z are similar to the shortest path from scholar M to the same "domain authority scholar" Y and "renowned academic institution" Z in knowledge graph B. Therefore, aligning scholars X and M can be considered.

[0090] Machine learning-based methods: Classification Model: Train a classification model that takes a node's attributes and its structural features in the graph as input and predicts whether the node matches a node in another knowledge graph. For example, using a Support Vector Machine (SVM) model, the model is trained to determine whether two nodes should be merged, taking the node's attribute vectors (including self-attributes, semantic attributes, strength attributes, etc.) and its connection features with neighboring nodes as input.

[0091] Clustering algorithms: Cluster all nodes to be aligned, grouping similar nodes into one class. Nodes in each class may be the objects to be aligned. For example, the K-means clustering algorithm calculates the distance between nodes based on their combined characteristics (attributes and structural features), grouping nodes that are close in distance into the same class, thereby determining possible combinations of aligned nodes.

[0092] If there are conflicts in the information between different heterogeneous scholar relationship knowledge graphs, the attribute information of the scholar node in all heterogeneous scholar relationship knowledge graphs can be used for generalization, and the generalized information can be used as the information represented in the fused graph.

[0093] In this embodiment, new graph structure combinations to be aligned are determined based on all new heterogeneous scholar relationship knowledge graphs. The current heterogeneous scholar relationship knowledge graph is then updated based on these new combinations until no new combinations exist. At this point, all the obtained heterogeneous scholar relationship knowledge graphs are aligned and fused to obtain a unified scholar relationship network. After one correction and alignment, an updated heterogeneous scholar relationship knowledge graph is obtained. Then, following the previous method, new graph structure combinations to be aligned are determined again from these new knowledge graphs, and they are corrected and aligned, repeating this process continuously. Each update gradually reduces the differences between knowledge graphs and improves the matching degree. When no new graph structure combinations to be aligned can be found after multiple iterations, it indicates that the differences between knowledge graphs have been largely eliminated. At this point, all heterogeneous scholar relationship knowledge graphs are finally aligned and fused to form a unified scholar relationship network. This unified network integrates scholar relationship information from multiple heterogeneous data sources, eliminating inconsistencies caused by heterogeneity and providing a complete and consistent relationship graph for subsequent academic research and analysis.

[0094] This invention provides an embodiment of a scholar relationship network storage and construction system based on an artificial intelligence graph database, comprising: The graph construction module is used to construct a heterogeneous scholar relationship knowledge graph for each heterogeneous academic data source based on all entities and all relationships between entities in the scholar relationship data from multiple heterogeneous academic data sources. The feature fusion module is used to extract the joint feature vector of each scholar node in the heterogeneous scholar relationship knowledge graph based on the graph neural network model. The joint feature vector is formed by fusing the semantic features of the corresponding scholar node and the multi-hop structure features in the corresponding heterogeneous scholar relationship knowledge graph. The target coarse screening module is used to perform alignment target coarse screening on pairs of scholar nodes in heterogeneous scholar relationship knowledge graphs from different heterogeneous academic data sources based on the self-attribute vectors of each scholar node, and obtain multiple combinations of scholar nodes to be aligned. The similarity calculation module is used to calculate the similarity of the joint feature vector of all scholar nodes in each scholar node combination to be aligned, which is used as the similarity score of the corresponding scholar node combination to be aligned. The heterogeneous fusion and storage module is used to fuse all heterogeneous scholar relationship knowledge graphs based on the similarity scores of all scholar node combinations to be aligned, obtain a unified scholar relationship network, and store the unified scholar relationship network in a distributed native graph database.

[0095] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for constructing a scholar relationship network storage based on an artificial intelligence graph database, characterized in that, include: Based on all entities and relationships between all entities in the scholar relationship data from multiple heterogeneous academic data sources, a heterogeneous scholar relationship knowledge graph is constructed for each heterogeneous academic data source. The joint feature vector of each scholar node in the heterogeneous scholar relationship knowledge graph is extracted based on the graph neural network model. The joint feature vector is formed by fusing the semantic features of the corresponding scholar node and the multi-hop structure features in the corresponding heterogeneous scholar relationship knowledge graph. Based on the self-attribute vectors of each scholar node, a coarse screening of alignment targets is performed on pairs of scholar nodes in the heterogeneous scholar relationship knowledge graph from different heterogeneous academic data sources to obtain multiple combinations of scholar nodes to be aligned. Calculate the similarity of the joint feature vectors of all scholar nodes in each scholar node combination to be aligned, and use it as the similarity score of the corresponding scholar node combination to be aligned; Based on the similarity scores of all scholar node combinations to be aligned, all heterogeneous scholar relationship knowledge graphs are fused to obtain a unified scholar relationship network, which is then stored in a distributed native graph database. Specifically, the joint feature vector of each scholar node in the heterogeneous scholar relationship knowledge graph is extracted based on the graph neural network model, including: Based on the pre-trained language model, the attribute information of the associated text of each scholar node in the heterogeneous scholar relationship knowledge graph is extracted to obtain the semantic features of the corresponding scholar node. Based on the graph neural network model, the connection patterns and neighbor attribute information of each scholar node in the heterogeneous scholar relationship knowledge graph are extracted and fused within the multi-order neighbor range, so as to obtain the multi-hop structure features of the corresponding scholar node in the corresponding heterogeneous scholar relationship knowledge graph. The semantic features of each scholar node are fused with the multi-hop structural features in the corresponding heterogeneous scholar relationship knowledge graph to obtain the joint feature vector of the corresponding scholar node. The process of obtaining the intrinsic attribute vector of each scholar node includes: Obtain the relationship formation time, associated paper DOI identifier, and relationship type values ​​for each scholar node, and summarize them to generate the basic attribute vector for each scholar node; Analyze all collaboration relationships of each scholar node, and calculate the subject matching degree of all collaboration relationships of each scholar node based on the similarity of keywords of all papers of the collaborating parties in each collaboration relationship; Obtain all cited texts of all papers signed by each scholar node, and analyze the sentiment values ​​of all cited texts of all papers signed by each scholar. The semantic attribute vector of each scholar node is generated by summarizing the subject matching degree of all collaborations of each scholar node and the sentiment value of all cited texts of all papers under the signature. Generate the intensity attribute vector for each scholar node; The self-attribute vector includes the basic attribute vector, the semantic attribute vector, and the intensity attribute vector.

2. The method for constructing a scholar relationship network based on an artificial intelligence graph database according to claim 1, characterized in that, Based on all entities and relationships between entities in scholar relationship data from multiple heterogeneous academic data sources, a heterogeneous scholar relationship knowledge graph is constructed for each heterogeneous academic data source, including: Acquire scholar relationship data from multiple heterogeneous academic data sources; The scholars, paper titles, academic institutions and research fields in the scholar relationship data of each heterogeneous academic data source are processed into entities to obtain all entities of each heterogeneous academic data source. Based on all entities in each heterogeneous academic data source and the co-authorship, affiliation, publication, and belonging relationships among all entities, a heterogeneous scholar relationship knowledge graph is constructed for each heterogeneous academic data source.

3. The method for constructing a scholar relationship network based on an artificial intelligence graph database according to claim 1, characterized in that, Generate the intensity attribute vector for each scholar node, including: The collaboration intensity of each scholar node is calculated based on the number of collaborative papers and the journal impact factor weight of the corresponding papers. The citation intensity of each scholar node is calculated based on the citation frequency of all papers signed by each scholar node and the citation weight of all cited papers. The citation intensity of each scholar node is calculated based on the citation frequency of all papers signed by each scholar node. The citation balance factor of each scholar node is calculated based on the citation strength and the cited strength of each scholar node. Based on the cooperation strength, citation strength, cited strength, and citation relationship balance factor of each scholar node, a strength attribute vector for each scholar node is generated.

4. The method for constructing a scholar relationship network based on an artificial intelligence graph database according to claim 1, characterized in that, Based on the self-attribute vectors of each scholar node, a coarse screening of alignment targets is performed on pairwise scholar nodes in heterogeneous scholar relationship knowledge graphs from different heterogeneous academic data sources, resulting in multiple combinations of scholar nodes to be aligned, including: The first similarity between basic attribute vectors, the second similarity between semantic attribute vectors, and the third similarity between intensity attribute vectors are calculated for each pair of scholar nodes in the heterogeneous scholar relationship knowledge graph from different heterogeneous academic data sources. In heterogeneous scholar relationship knowledge graphs from different heterogeneous academic data sources, any two scholar nodes whose first similarity is not less than the first similarity threshold, whose second similarity is not less than the second similarity threshold, and whose third similarity is not less than the third similarity threshold are combined to obtain multiple combinations of scholar nodes to be aligned.

5. The method for constructing a scholar relationship network based on an artificial intelligence graph database according to claim 1, characterized in that, Based on the similarity scores of all scholar node combinations to be aligned, all heterogeneous scholar relationship knowledge graphs are fused to obtain a unified scholar relationship network, including: Based on the similarity scores of all scholar node combinations to be aligned and the position of each scholar node in the heterogeneous scholar relationship knowledge graph, the three-dimensional projection coordinates of each scholar node are constructed. Identify scholar node combinations that belong to different heterogeneous scholar relationship knowledge graphs and have the highest similarity scores with each other, and use them as matching scholar node combinations. The three-dimensional projection coordinates of all scholar nodes in each heterogeneous scholar relationship knowledge graph are marked in a preset three-dimensional coordinate system. Combined with all mutually matched scholar node combinations, at least one maximum hexahedron is determined such that the scholar node matching density of the maximum hexahedron is not less than a preset density threshold. The scholar node matching density of each maximum hexahedron is the ratio of the maximum value among the total number of mutually matched scholar node combinations in the pairwise heterogeneous scholar relationship knowledge graphs enclosed by the volume of the maximum hexahedron to the volume of the corresponding maximum hexahedron. Based on all the largest hexahedrons, all heterogeneous scholar relationship knowledge graphs are merged to obtain a unified scholar relationship network.

6. The method for constructing a scholar relationship network storage based on an artificial intelligence graph database according to claim 5, characterized in that, Based on the similarity scores of all scholar node combinations to be aligned and the position of each scholar node in its respective heterogeneous scholar relationship knowledge graph, the three-dimensional projected coordinates of each scholar node are constructed, including: Based on the similarity scores of all scholar node combinations to be aligned, the maximum similarity score of each scholar node in each heterogeneous scholar relationship knowledge graph is determined, and all scholar nodes with a maximum similarity score greater than a preset score threshold are selected from each heterogeneous scholar relationship knowledge graph as all high-precision scholar nodes. Based on the distribution of all high-precision scholar nodes in each heterogeneous scholar relationship knowledge graph, a central scholar node is identified in each heterogeneous scholar relationship knowledge graph. Based on the central scholar node, multiple concentric structures that can cover all scholar nodes in the corresponding heterogeneous scholar relationship knowledge graph and the ordinal number of each concentric structure are identified, wherein each concentric structure covers at least one high-precision scholar node. The high-precision scholar node with the highest similarity score in each layer structure is taken as the starting scholar node of each layer structure. Based on the preset statistical direction, the number of node intervals between each scholar node in the heterogeneous scholar relationship knowledge graph and the starting scholar node of the corresponding layer structure is determined. Based on the ordinal number of the layer structure to which each scholar node belongs in the heterogeneous scholar relationship knowledge graph, the number of node intervals between the scholar node and the starting scholar node of the layer structure, and the maximum similarity score, the three-dimensional projection coordinates of the corresponding scholar node are constructed.

7. The method for constructing a scholar relationship network storage based on an artificial intelligence graph database according to claim 5, characterized in that, Based on all maximal hexahedrons, all heterogeneous scholar relationship knowledge graphs are fused to obtain a unified scholar relationship network, including: The two heterogeneous scholar relationship knowledge graphs corresponding to the maximum value among the total number of mutually matching scholar node combinations of the pairwise heterogeneous scholar relationship knowledge graphs enclosed by each maximal hexahedron are used as the coarse matching knowledge graph combination corresponding to each maximal hexahedron. In each coarse matching knowledge graph combination, the partial graph structure enclosed by all scholar nodes whose corresponding three-dimensional projection coordinates are covered by the corresponding maximum hexahedron in the two heterogeneous scholar relationship knowledge graphs are summarized and regarded as the graph structure combination to be aligned. The spectral structure combination to be aligned is corrected and aligned to obtain the spectral correction and alignment result; The graph correction and alignment results are updated to the corresponding heterogeneous scholar relationship knowledge graph to obtain a new heterogeneous scholar relationship knowledge graph. Based on all the new heterogeneous scholar relationship knowledge graphs, new graph structure combinations to be aligned are determined. The current heterogeneous scholar relationship knowledge graph is updated based on the new graph structure combinations to be aligned until there are no new graph structure combinations to be aligned. Then, all the currently obtained heterogeneous scholar relationship knowledge graphs are aligned and fused to obtain a unified scholar relationship network.

8. A scholar relationship network storage and construction system based on an artificial intelligence graph database, characterized in that, The method for constructing a scholar relationship network storage based on an artificial intelligence graph database according to any one of claims 1 to 7 includes: The graph construction module is used to construct a heterogeneous scholar relationship knowledge graph for each heterogeneous academic data source based on all entities and all relationships between entities in the scholar relationship data from multiple heterogeneous academic data sources. The feature fusion module is used to extract the joint feature vector of each scholar node in the heterogeneous scholar relationship knowledge graph based on the graph neural network model. The joint feature vector is formed by fusing the semantic features of the corresponding scholar node and the multi-hop structure features in the corresponding heterogeneous scholar relationship knowledge graph. The target coarse screening module is used to perform alignment target coarse screening on pairs of scholar nodes in heterogeneous scholar relationship knowledge graphs from different heterogeneous academic data sources based on the self-attribute vectors of each scholar node, and obtain multiple combinations of scholar nodes to be aligned. The similarity calculation module is used to calculate the similarity of the joint feature vector of all scholar nodes in each scholar node combination to be aligned, which is used as the similarity score of the corresponding scholar node combination to be aligned. The heterogeneous fusion and storage module is used to fuse all heterogeneous scholar relationship knowledge graphs based on the similarity scores of all scholar node combinations to be aligned, obtain a unified scholar relationship network, and store the unified scholar relationship network in a distributed native graph database.

Images