A domain knowledge graph construction method and system based on a graph neural network

Abstract

The application provides a domain knowledge graph construction method and system based on a graph neural network, and relates to the field of knowledge graphs.The method comprises: obtaining basic data of a target domain, performing data analysis on the basic data, extracting knowledge nodes and node paths composed of the knowledge nodes from the basic data;dividing a plurality of target domain node groups according to the context association features of the knowledge nodes and the matching degrees with the target domain;mining potential association paths between the knowledge nodes in the target domain node groups, constructing a knowledge graph, determining core nodes in the knowledge graph, obtaining preliminary push content from a preset database according to the core nodes, determining a push priority for the preliminary push content according to the preferences of users, and obtaining priority push content.The application accurately evaluates the association degree of the knowledge graph and the target domain, accurately identifies core concepts and key relationships, and realizes intelligent and personalized knowledge push.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of knowledge graph, and in particular to a domain knowledge graph construction method and system based on graph neural network. BACKGROUND

[0002] In the information age, constructing a domain knowledge graph and implementing precise knowledge pushing is particularly crucial. This is not only an important cornerstone for promoting intelligent information processing, but also a core support for improving the efficiency of users obtaining effective information. By structuring and correlating scattered knowledge, decision support and deep insights can be provided for various industries.

[0003] However, the current knowledge graph has obvious deficiencies in practical application. The application of the knowledge graph still stays at the level of simple knowledge storage and retrieval, lacking deep mining of the correlation between knowledge and the target domain, and it is also difficult to accurately determine the importance of knowledge, resulting in that the pushed content often does not match the user's needs and is inefficient.

[0004] Therefore, how to strengthen the correlation between the knowledge domain and the target domain on the basis of constructing the knowledge graph, accurately identify core concepts and key relationships, and thereby optimize the response speed of frequently accessed content, is a key problem that the industry needs to solve. SUMMARY

[0005] Therefore, the present application proposes a domain knowledge graph construction method and system based on graph neural network, to solve the problem that the constructed knowledge graph lacks deep mining of the correlation between knowledge and the target domain, and it is also difficult to accurately determine the importance of knowledge.

[0006] The technical solution of the present application is as follows:

[0007] According to a first aspect, the present application embodiment provides a domain knowledge graph construction method based on graph neural network, which comprises:

[0008] Obtaining the basic data of the target domain, performing data analysis on the basic data, extracting knowledge nodes and node paths composed of knowledge nodes from the basic data; the knowledge nodes form node paths through logical relationships;

[0009] According to the context correlation characteristics of the knowledge nodes and the matching degree with the target domain, a plurality of target domain node groups are divided; each target domain node group includes a center node and knowledge nodes having a logical relationship with the center node;

[0010] The potential association paths between the knowledge nodes in the target field node group are mined, a knowledge graph is constructed according to the knowledge nodes, the logical relationship and the potential association paths, the core nodes in the knowledge graph are determined, the preliminary push content is obtained from the preset database according to the core nodes, the push priority of the preliminary push content is determined according to the user preference, and the priority push content is obtained.

[0011] With reference to the first aspect, in a first implementation manner of the first aspect, the target field node group is divided according to the context association features of the knowledge nodes and the matching degrees with the target field, and specifically includes:

[0012] The matching degrees of the knowledge nodes with the target field are determined, and the knowledge nodes with the matching degrees exceeding a first preset matching degree are reserved;

[0013] The context association features of the knowledge nodes are determined, the center nodes are determined according to the context association features, and the candidate node groups are divided according to the logical relationship of the center nodes;

[0014] The context association features of the knowledge nodes in each candidate node group are fused to obtain node fusion features, the matching degrees of the node fusion features with the target field are determined, and the candidate node groups with the matching degrees exceeding a second preset matching degree are taken as the target field node groups.

[0015] With reference to the first implementation manner of the first aspect, in a second implementation manner of the first aspect, the potential association paths between the knowledge nodes in the target field node group are mined, a knowledge graph is constructed according to the knowledge nodes, the logical relationship and the potential association paths, the core nodes in the knowledge graph are determined, the preliminary push content is obtained from the preset database according to the core nodes, the push priority of the preliminary push content is determined according to the user preference, and the priority push content is obtained, and specifically includes:

[0016] The semantic similarity and the context dependency matrix between the knowledge nodes in the target field node group are obtained;

[0017] The connection strength between the knowledge nodes in the target field node group is output by a graph neural network model according to the semantic similarity and the context dependency matrix and by using the graph neural network model, and the potential association paths between the knowledge nodes are mined according to the connection strength;

[0018] An initial framework of the knowledge graph is constructed according to the logical relationship and the potential association paths between the knowledge nodes in all the target field node groups, and the centrality of each knowledge node in the initial framework is determined according to the connection strength and the connection quantity;

[0019] The knowledge nodes with the centrality exceeding a preset centrality are taken as the core nodes, and the preliminary push content is obtained from the preset database according to the core nodes;

[0020] Obtaining interaction data of the user within a first preset time period, determining the interaction frequency of the user with each preliminary push content according to the interaction data, and determining portrait data representing the user's behavior pattern according to the interaction frequency. The interaction data includes historical record data and real-time behavior data of the user;

[0021] According to the portrait data, the push priority of the preliminary push content is determined, the preliminary push content is pushed to the user according to the push priority, and the interaction data of the user in a second preset time period is obtained;

[0022] By comparing the interaction data of the first time period and the second time period, the preferred content of the user is obtained, and the push priority is adjusted according to the preferred content. The priority push content is obtained according to the priority of the preliminary adjustment content.

[0023] In combination with the second embodiment of the first aspect, in the third embodiment of the first aspect, before the step of taking the knowledge node with a center degree exceeding a preset center degree as a core node and obtaining preliminary push content from a preset database, the method further comprises:

[0024] Adjusting the potential association path according to the target field; the adjustment mode includes adding the potential association path and deleting the potential association path.

[0025] In combination with the first aspect, in the fourth embodiment of the first aspect, the method further comprises:

[0026] The priority push content is stored by distributed storage, and the priority push content is allocated cache space according to the priority of the priority push content; the priority push content with higher priority is allocated more cache space.

[0027] In combination with the fourth embodiment of the first aspect, in the fifth embodiment of the first aspect, the method further comprises:

[0028] Obtaining the access request log of each priority push content, extracting the access distribution characteristics in the access request log, and adjusting the cache space allocated to the priority push content according to the access distribution characteristics.

[0029] In combination with the fifth embodiment of the first aspect, in the sixth embodiment of the first aspect, the method further comprises:

[0030] According to the adjusted storage space of the priority push content, the running data of the knowledge node corresponding to the priority push content is obtained, and the access request of the priority push content is handled according to the running data. The running data includes response time and resource utilization.

[0031] With reference to the sixth embodiment of the first aspect, in a seventh embodiment of the first aspect, the method further comprises:

[0032] According to the access request log of the target field and each priority push content, the logical relationship between the knowledge nodes is reconstructed, and the initial architecture of the knowledge graph is adjusted.

[0033] According to the second aspect, an embodiment of the present application provides a domain knowledge graph construction system based on a graph neural network, the system comprising:

[0034] A data acquisition module is configured to acquire basic data of a target field, perform data analysis on the basic data, extract knowledge nodes and node paths composed of the knowledge nodes from the basic data, and form node paths between the knowledge nodes through logical relationships;

[0035] A data matching module is configured to divide a plurality of target field node groups according to the context association features of the knowledge nodes and the matching degrees with the target field, wherein each target field node group comprises a center node and knowledge nodes having logical relationships with the center node;

[0036] A graph construction module is configured to mine potential association paths between the knowledge nodes in the target field node groups, construct a knowledge graph according to the knowledge nodes, logical relationships and potential association paths, determine core nodes in the knowledge graph, obtain preliminary push content from a preset database according to the core nodes, determine push priorities of the preliminary push content according to user preferences, and obtain priority push content.

[0037] The domain knowledge graph construction method and system based on the graph neural network have the following beneficial effects relative to the prior art:

[0038] After acquiring the knowledge nodes, a plurality of target field node groups are divided according to the context association features of the knowledge nodes and the matching degrees with the target field, so that the divided target field node groups are highly associated with the target field, ensuring that the knowledge content is highly matched with the requirements of the target field, potential association paths between the knowledge nodes in the target field node groups are mined, and a knowledge graph is constructed based on the potential association paths, core nodes in the knowledge graph are determined according to the centrality of the knowledge nodes, preliminary push content is obtained from a preset database according to the core nodes, so that core concepts and key relationships can be accurately identified, the adaptation ability of the system in a specific scenario can be enhanced, the application pertinence of the knowledge nodes can be improved, the preference content of the user can be acquired to timely capture the dynamic changes of the interest of the user, so that the push priorities can be adjusted according to the preference content to obtain priority push content, the response speed of frequently accessed content can be optimized through an efficient storage architecture, and intelligent and personalized knowledge push is realized. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of these drawings.

[0040] Figure 1 One of the flowchart of the domain knowledge graph construction method based on the graph neural network of the present application;

[0041] Figure 2 The second flowchart of the domain knowledge graph construction method based on the graph neural network of the present application;

[0042] Figure 3 The third flowchart of the domain knowledge graph construction method based on the graph neural network of the present application;

[0043] Figure 4 The fourth flowchart of the domain knowledge graph construction method based on the graph neural network of the present application;

[0044] Figure 5 The structural diagram of the domain knowledge graph construction system based on the graph neural network of the present application. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be described below in a clear and complete manner. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort fall within the scope of the present application.

[0046] In the information age, constructing a domain knowledge graph and realizing accurate knowledge pushing is particularly crucial. This is not only an important cornerstone for promoting intelligent information processing, but also a core support for improving the efficiency of users obtaining effective information. By structuring and correlating scattered knowledge, decision support and deep insight can be provided for various industries.

[0047] However, the current knowledge graph has obvious deficiencies in practical application. The application of the knowledge graph still stays at the level of simple knowledge storage and retrieval, lacks deep mining of the correlation between knowledge and target fields, and is difficult to accurately judge the importance of knowledge, resulting in that the pushed content often does not match the user's needs and is inefficient.

[0048] The main reasons for the above problems are: the assessment of the degree of association between knowledge and the target field becomes a primary problem, as there is a lack of effective measurement mechanism, the system is difficult to determine which knowledge is truly suitable for the needs of a specific scenario; in a complex knowledge system, it is difficult to identify core concepts and key relationships, and it is difficult to accurately extract the most valuable knowledge points, as important knowledge has not been effectively screened and integrated, the system often faces the dilemma of uneven resource allocation and slow response when processing frequently accessed content, which affects user experience.

[0049] In summary, how to strengthen the degree of association between the knowledge field and the target field on the basis of building a knowledge graph, accurately identify core concepts and key relationships, and optimize the response speed of frequently accessed content is a key problem that needs to be solved in the current industry.

[0050] The field knowledge graph construction method based on graph neural network provided by the present application aims to accurately assess the degree of association between the knowledge graph and the target field, accurately identify core concepts and key relationships, and optimize the response speed of frequently accessed content through an efficient storage architecture, realizing intelligent and personalized knowledge pushing and improving the pushing effect and system running efficiency.

[0051] Please refer to Figure 1 , Figure 1 The flowchart of the field knowledge graph construction method based on graph neural network of the present application, which can include the following steps:

[0052] S101, obtaining the basic data of the target field, performing data analysis on the basic data, and extracting knowledge nodes and node paths composed of knowledge nodes from the basic data.

[0053] In the present embodiment, the above-mentioned basic data is obtained through a pre-established field knowledge base, which can be supplemented by storing the basic data of the target field by continuously collecting and summarizing relevant field literature, record materials and interview materials, etc.

[0054] More specifically, step S101 includes:

[0055] S1011, obtaining the basic data of the target field, and converting the basic data into text data.

[0056] For example, the basic data includes data in the form of tables, texts, pictures, audios, etc., which is first converted into text data, which helps subsequent semantic analysis, reduces ambiguity, and improves the accuracy of data processing.

[0057] S1012, pre-processing the text data, and performing word segmentation processing on the pre-processed text data to obtain text segments.

[0058] The text data often mixes invalid characters or inconsistent formats. In this embodiment, the basic data is preliminarily preprocessed, that is, data cleaning, before data analysis, for example, using a general denoising tool to remove redundant spaces, special symbols or irrelevant labels in the text data, to ensure the basic readability of the data. The converted text data is then processed by word segmentation and splitting to obtain a plurality of text segments, that is, a text segment set.

[0059] S1013, performing semantic analysis on the text segments, determining the context correlation degree between the text segments, and dividing the text segments with a context correlation degree exceeding a preset correlation degree into the same knowledge node. The knowledge nodes are connected by logical relationships to form node paths.

[0060] By semantic analysis, semantic understanding information of all text segments can be determined, and context correlation degrees between different text segments can be obtained according to the semantic understanding information. If the context correlation degree between two text segments exceeds a preset correlation degree, the two text segments can be classified into a structured knowledge unit group. If a text segment has a context correlation degree with a text segment in the knowledge unit group exceeding a preset correlation degree, the text segment is classified into the knowledge unit group. In this way, all text segments can be divided into a plurality of knowledge unit groups, each knowledge unit group being a corresponding knowledge node. The classified knowledge unit groups provide a clear logical basis for the construction of knowledge nodes. This process effectively integrates scattered information into meaningful knowledge blocks, facilitating subsequent applications.

[0061] S1014, determining the logical relationship between different knowledge nodes, and taking the logical relationship as node connection information connecting the knowledge nodes, and establishing a node path composed of knowledge nodes according to the node connection information.

[0062] By associating and mapping the knowledge nodes, the knowledge nodes can be integrated, and the nodes with logical relationships can be connected in series to form a plurality of node paths. Different node paths can be connected by at least one knowledge node, and different node paths can also have the same knowledge node. For example, knowledge node B, knowledge node C and knowledge node D are connected in series to form node path A, knowledge node A, knowledge node B, knowledge node E and knowledge node F are connected in series to form node path B, node path A and B have the same knowledge node B, knowledge node F, knowledge node G, knowledge node H, knowledge node I, knowledge node J and knowledge node K are connected in series to form node path C, node path B and B are connected by knowledge node F, and the knowledge node F is also the same knowledge node between node path B and C.

[0063] The logical relationship between knowledge nodes is the key information for logically connecting two different knowledge nodes, for example, knowledge node B and knowledge node C are connected through node connection information "extraction mode", and these node connection information is also the prompt information / retrieval information of the node path formed by these knowledge nodes, forming a complete chain from some information to other information. This way of association mapping not only improves the traceability of knowledge, but also provides data support for subsequent intelligent question answering or decision support systems. When users are looking for related information in a target field, they can quickly find the required content through knowledge nodes, reducing search time. At the same time, structured knowledge units and node sets also provide a reliable data foundation for artificial intelligence auxiliary processing, significantly improving the intelligent level of the system.

[0064] It should be noted that the knowledge node A can be connected with multiple knowledge nodes, such as knowledge node B, knowledge node C and knowledge node D. At the same time, not all knowledge nodes have a logical relationship, for example, knowledge node D and knowledge node I can not have a logical relationship.

[0065] S102, according to the context association characteristics of the knowledge nodes and the matching degree with the target field, a plurality of target field node groups are divided, wherein each target field node group includes a center node and a knowledge node having a logical relationship with the center node.

[0066] By calculating the matching degree of the knowledge nodes and the target field, it can be ensured that the selected knowledge nodes are highly related to the core demand, so that a complete logical chain is formed from the association calculation of the knowledge nodes to the final grouping, ensuring that the knowledge content is highly consistent with the demand of the target field, and laying a solid foundation for subsequent application.

[0067] S103, mining the potential association path between each knowledge node in the target field node group, constructing a knowledge graph according to the knowledge nodes, logical relationships and potential association paths, determining the core nodes in the knowledge graph, obtaining preliminary push content from the preset database according to the core nodes, and determining the push priority of the preliminary push content according to the user's preference to obtain priority push content.

[0068] Mining the potential association path between knowledge nodes through strength helps to focus on core associations and avoid irrelevant information interference. The core node is a knowledge node with high centrality in the knowledge graph, which can quickly filter out node information that is crucial to knowledge push through centrality. Obtaining user preference content can timely capture the dynamic changes of user interest, so as to adjust the push priority according to the preference content, thereby obtaining the priority push content.

[0069] The field knowledge graph construction method based on the graph neural network of the application divides a plurality of target field node groups according to the context association features of the knowledge nodes and the matching degrees with the target field after the knowledge nodes are acquired, so that the target field node groups divided in this way are highly associated with the target field, ensuring that the knowledge content is highly matched with the requirements of the target field, the potential association paths between the knowledge nodes in the target field node groups are further mined, and a knowledge graph is constructed based on the potential association paths, the core nodes in the knowledge graph are determined according to the centrality of the knowledge nodes, and the preliminary push content is obtained from the preset database according to the core nodes, so that the core concepts and key relationships can be accurately identified, the adaptive ability of the system in a specific scenario is enhanced, the application pertinence of the knowledge nodes is improved, the preference content of the user can be obtained to timely capture the dynamic changes of the interests of the user, so that the push priority can be adjusted according to the preference content to obtain the priority push content, and the response speed of frequently accessed content is optimized through the efficient storage architecture, and intelligent and personalized knowledge push is realized.

[0070] Please refer to Figure 2 The method can further include the following steps:

[0071] S201, acquire the basic data of the target field, perform data analysis on the basic data, and extract knowledge nodes and node paths composed of the knowledge nodes from the basic data. For specific content, refer to step S101.

[0072] S2021, determine the matching degree of the knowledge nodes with the target field, and retain the knowledge nodes with a matching degree exceeding a first preset matching degree.

[0073] In this embodiment, the connection relationship (i.e., whether connected and logical relationship) between the knowledge nodes can be evaluated through a graph structure tool, the connection path and frequency with the requirements of the target field are analyzed, and then the matching degree of the knowledge nodes with the target field is obtained, the knowledge nodes with a matching degree exceeding a preset matching degree are retained, and the knowledge nodes with a matching degree lower than the preset matching degree are deleted, so that it can be ensured that the knowledge nodes selected are highly relevant to the core requirements.

[0074] S2022, determine the context association features of the knowledge nodes, determine the center nodes according to the context association features, and divide the candidate node groups according to the logical relationship of the center nodes.

[0075] In the embodiment, the context association feature refers to the peripheral connection environment of the knowledge node in the node path, that is, the knowledge nodes having a logical relationship with a certain knowledge node and the number of the knowledge nodes. Since the preliminary screening of the knowledge nodes has been performed in step S2021, the path relationship is analyzed according to the context association feature of the retained knowledge nodes in step S2022. For example, according to the fact that the knowledge node A has a shorter path with the knowledge node B, the knowledge node L and other nodes, the weight value of the context association feature of the knowledge node A is higher, the importance of the knowledge node A in the target field requirement is confirmed, and the knowledge node A is determined as the center node (the center node is also a knowledge node). Such analysis helps to identify the actual role of the node in the specific scenario.

[0076] Then, according to the determined center node, the knowledge nodes having a logical relationship with the center node and the center node are divided into a group of candidate node groups. For example, a group of knowledge nodes around the knowledge node A constitutes a candidate node group A, and a group of knowledge nodes around the knowledge node B constitutes a candidate node group B. Such grouping can focus on the core content in the field.

[0077] It should be noted that the same knowledge node can be divided into multiple candidate node groups.

[0078] S2023, the context association features of the knowledge nodes in each candidate node group are fused to obtain node fusion features, the matching degree of the node fusion features with the target field is determined, and the candidate node group whose matching degree exceeds a second preset matching degree is taken as a target field node group.

[0079] In the embodiment, the value of the second preset matching degree exceeds the value of the first preset matching degree, and further screening is performed. In the embodiment, the priority of all target field node groups can also be sorted according to the node fusion features. The target field node group with a higher matching degree of the node fusion features with the target field has a higher priority. Such sorting mechanism ensures that the knowledge content most meeting the requirements is presented preferentially, and the pertinence of knowledge management is improved.

[0080] S303, the potential association paths between the knowledge nodes in the target field node group are mined, a knowledge graph is constructed according to the knowledge nodes, the logical relationship and the potential association paths, a core node in the knowledge graph is determined, the preliminary push content is obtained from the preset database according to the core node, the push priority of the preliminary push content is determined according to the user's preference, and the priority push content is obtained. For specific content, refer to step S103.

[0081] Please refer to Figure 3 The method can further include the following steps:

[0082] S301, acquire basic data of a target field, perform data analysis on the basic data, extract knowledge nodes and node paths composed of the knowledge nodes from the basic data. For details, refer to step S101.

[0083] S302, divide a plurality of target field node groups according to the context association characteristics of the knowledge nodes and the matching degrees with the target field. For details, refer to step S102.

[0084] S3031, acquire semantic similarity and context dependency matrix between each knowledge node in the target field node group.

[0085] For example, in a knowledge management platform, a target field node group contains a plurality of knowledge nodes related to "intelligent recommendation system", such as "user preference" node, "content label" node and "recommendation algorithm" node. Through semantic analysis, it is obtained that the semantic similarity between the "user preference" node and the "content label" node is 0.85, and the semantic similarity between the "user preference" node and the "recommendation algorithm" node is only 0.45. At the same time, the context dependency data shows that the "user preference" node often appears in the context of the "content label" node, forming a strong context dependency relationship. This quantitative processing provides basic data for subsequent association mining.

[0086] S3032, according to the semantic similarity and the context dependency matrix and by using a graph neural network model, output the connection strength between each knowledge node in the target field node group from the graph neural network model, and mine the potential association path between the knowledge nodes according to the connection strength.

[0087] The semantic similarity and the context dependency matrix are input into the trained graph neural network model as input data, and the connection strength between a certain knowledge node and other knowledge nodes output by the graph neural network model can be obtained.

[0088] It should be noted that if there is a logical relationship between a certain knowledge node and another knowledge node, the connection strength reflects the association between the knowledge nodes; if there is no logical relationship between a certain knowledge node and another knowledge node, the connection strength is the possible node association between the two nodes.

[0089] Afterwards, the connection strength exceeding the preset value is taken as a potential association path, for example, the connection strength from the "user preference" node to the "content label" node is 0.9, which is higher than the preset value 0.6, so this path is determined as a potential association path; the connection strength from the "user preference" node to the "recommendation algorithm" node is only 0.5, which does not reach the preset value, and is temporarily excluded. Mining the potential association path between knowledge nodes through strength helps to focus on core associations and avoid interference from irrelevant information.

[0090] S3033, adjusting the potential association path according to the target field.

[0091] In the embodiment, the adjustment manner includes adjusting the connected knowledge nodes, i.e. adding a potential association path, deleting a potential association path, etc. According to the target field, the context logical relationship between two knowledge nodes with a potential association path can be verified, for example, the context logic of the potential association path between the "user preference" node and the "content label" node is verified, and the two nodes often jointly act on user experience improvement in the recommendation scene, so the verification is passed, while another potential association path can be excluded due to unclear context logic; at the same time, if it is found according to the target field that the node association of two knowledge nodes is very high, a potential association path between the two nodes can be established.

[0092] Through the fitting degree with the target field, the actual application value of the path is ensured, so that the comprehensiveness of the result is ensured from multiple dimensions, not only covering the core elements of the recommendation system, but also reflecting the logical dependence between nodes through path association, providing a solid foundation for subsequent knowledge application, improving the application pertinence, rationality and adaptability to actual needs of the knowledge nodes, and enhancing the adaptation ability of the system in a specific scene.

[0093] S3034, constructing an initial framework of the knowledge graph according to the logical relationship and the potential association path between each knowledge node in all target field node groups, and determining the centrality of each knowledge node in the initial framework according to the connection strength and the connection number.

[0094] After obtaining the logical relationship and the potential association path between each knowledge node in the target field node group, the initial framework of the knowledge graph can be constructed according to these information, and after the processing of step S3032, the logical relationship (also the potential association path) and the potential association path have corresponding connection strengths. By giving a certain calculation weight to the connection strength of each path and the number of other knowledge nodes connected, the centrality of each knowledge node in the initial framework can be determined.

[0095] S3035, taking the knowledge node with a centrality exceeding a preset centrality as a core node, and obtaining preliminary push content from a preset database according to the core node.

[0096] When a certain knowledge node, for example, the "machine learning" node, has a calculated centrality exceeding a preset centrality, the knowledge node is marked as a core node, so that the node information that is crucial to knowledge pushing can be quickly screened out. Next, preliminary pushing content related to the "machine learning" node is obtained, assuming that the preset database contains the latest academic paper abstracts, tutorial resources, etc., and knowledge content directly related to the "machine learning" node is extracted therefrom to form structured preliminary pushing content, such as being divided into "basic theory" and "application case" two categories, so as to facilitate more targeted distribution in subsequent distribution.

[0097] In this embodiment, all core nodes can also be prioritized according to the centrality. According to the specific numerical value of the centrality, the higher the centrality of the core node, the higher the priority, and the higher the priority of the corresponding preliminary pushing content in pushing. In this way, the preliminary pushing content not only covers the core knowledge points, but also improves the efficiency and convenience of knowledge acquisition of users through the sorting of the preliminary pushing content, while ensuring the relevance and practicality of the content.

[0098] Subsequently, the preliminary pushing content can be processed in sequence according to the priority order.

[0099] S3036, obtain the interaction data of the user in the first preset time period, determine the interaction frequency of the user with each preliminary pushing content according to the interaction data, and determine the portrait data representing the user behavior pattern according to the interaction frequency. The interaction data includes historical record data and real-time behavior data of the user.

[0100] In this embodiment, the first preset time period is a specified period of time, for example, one month. In this way, by obtaining the historical record data and real-time behavior data of the user in the preset time period, the interaction frequency of the user with each preliminary pushing content can be determined.

[0101] For example, the user frequently searches for records related to the "deep learning" topic in the past month, and recently clicks on articles related to "neural networks". By constructing the portrait data representing the user behavior pattern through the interaction frequency, the portrait data may show that the user's interest in the "deep learning" field accounts for 60%, while the interest in other branches such as "data mining" is only 20%.

[0102] S3037, determine the pushing priority of the preliminary pushing content according to the portrait data, push the preliminary pushing content to the user according to the pushing priority, and obtain the interaction data of the user in the second time period.

[0103] Since the profile data is obtained based on the user's interaction frequency, the push priority of all initial push content can be determined from this. Then, the initial push content is pushed to the user according to the push priority, and the user's interaction data of the content in the second preset time period is obtained. The interaction data of the content in the second preset time period is the user's preference feedback data in the second preset time period.

[0104] S3038. By comparing the interaction data of the first time period and the second time period, the user's preferred content is obtained, and the push priority is adjusted according to the preferred content. The priority is then adjusted based on the initial adjustment content to obtain the priority push content.

[0105] In this embodiment, the second preset time period is also a specified period of time, such as one week. The interaction data of the content in the second preset time period can be as follows: after the initial content is pushed, the user clicks on the first two "application cases" content and stays on the platform for more than 5 minutes, and also saves a related article. By comparing the historical records, it is found that the user's attention to "application cases" has increased by 30% recently compared with before, indicating that there is a change in behavior pattern, which reflects the user's preferred content.

[0106] By capturing user preferences, we can promptly identify dynamic changes in user interests. This allows us to adjust push notification priorities based on these preferences, resulting in prioritized content. Understandably, prioritized content also has a corresponding priority level; higher-priority content becomes the primary focus of push notifications. This prioritization method ensures targeted content delivery, guaranteeing that the content better matches the user's current needs, while also improving the efficiency of information retrieval through dynamic adjustments.

[0107] Please see Figure 4 The method may also include the following steps:

[0108] S401. Obtain basic data for the target domain, perform data analysis on the basic data, and extract knowledge nodes and node paths composed of knowledge nodes from the basic data. Refer to step S101 for details.

[0109] S402. Based on the contextual association features of knowledge nodes and their matching degree with the target domain, divide them into several target domain node groups. For details, refer to step S102.

[0110] S403, excavate the potential association path between each knowledge node in the target field node group, construct a knowledge graph according to the knowledge node, the logical relationship and the potential association path, determine the core node in the knowledge graph, obtain the preliminary push content from the preset database according to the core node, determine the push priority of the preliminary push content according to the user's preference, and obtain the priority push content. For specific content, refer to step S103.

[0111] S404, store the priority push content through distributed storage, and allocate cache space for the priority push content according to the priority of the priority push content. Wherein, the higher the priority of the priority push content, the more cache space is allocated.

[0112] In this embodiment, for the final priority push content, a cache-based data storage framework is provided, in which more cache space is allocated to priority push content with high priority, and the data reading path is optimized through a distributed storage mechanism.

[0113] Specifically, the cache space of the distributed storage system is divided into resources by the distributed storage tool, and according to the priority of the priority push content, if the priority of the priority push content is higher, it is allocated to the high priority area, and the adjusted allocation result is determined, so as to ensure that high-frequency content can be quickly called, and the user waiting time is reduced. According to the adjusted allocation result, the path information of data reading is obtained again, the path information is dynamically adjusted by using the path optimization tool, the path length in the path information is shortened for frequently accessed content, an optimized path scheme is obtained, assuming that the original path information needs to jump through multiple nodes and takes a long time, the path length is shortened from 5 nodes to 2 nodes by the path optimization tool. This adjustment enables users to quickly load resources when accessing high-frequency priority push content, thereby improving access efficiency. For the optimized path scheme, the storage location of the priority push content is finally adjusted by the storage management tool, combined with the characteristics of distributed storage, it is determined whether it meets the access frequency requirement, and the final storage configuration is obtained. Assuming that a relatively high-frequency priority push content is allocated to a storage node close to the user, and a relatively low-frequency priority push content is stored in a remote node, the final storage configuration is confirmed. This configuration can effectively balance the use of storage resources while ensuring fast response of high-frequency content.

[0114] Preferably, the data storage framework can also be optimized according to the distribution of user groups. Assuming that most users are concentrated in a certain area, the data storage framework will preferentially store related priority push content in nodes near the area to ensure access speed.

[0115] S405, acquire the access request log of each priority push content, extract the access distribution feature in the access request log, and adjust the cache space allocated to the priority push content according to the access distribution feature.

[0116] In this embodiment, the detailed record of the access request log of each priority push content is acquired from the operation log of the knowledge graph, then the access distribution feature in the access request log is extracted, the access distribution feature represents the actual access frequency distribution data of the priority push content allocated with priority and cache space, so that the cache space of each priority push content is dynamically adjusted according to the actual access frequency distribution data.

[0117] When dynamically adjusting the cache space, optimization can be combined with the current cache usage state, assuming that the total cache space is 100GB, and the cache space of the high-frequency access priority push content occupies 40GB, but according to the access distribution feature, the cache space of the high-frequency access priority push content should be increased to 60GB. The storage management tool will re-allocate the cache space ratio according to this demand, compress the cache space of the low-frequency access priority push content, and expand the storage area of the high-frequency access priority push content. This dynamic adjustment of cache space ensures the rational use of physical resources.

[0118] Preferably, when the dynamic adjustment of cache space, i.e. the resource allocation scheme, is implemented into the real-time strategy, the change of the access request log can be recorded by a real-time monitoring tool, and the log recording tool will continuously track this change. If it is found that the access request log suddenly decreases, it will be fed back to the system to trigger a new round of dynamic adjustment of cache space.

[0119] S406, according to the adjusted storage space of the priority push content, acquire the operation data of the knowledge node corresponding to the priority push content, and perform shunting processing on the access request of the priority push content according to the operation data. The operation data includes response time and resource utilization.

[0120] In this embodiment, the operation data of each knowledge node is monitored in real time. The operation data can reflect whether the knowledge node is a high-load node. If a knowledge node is determined to be a high-load node, a load balancing tool is used to perform shunting processing on the access request of the knowledge node. Combined with the logic of request distribution, part of the request is transferred from the high-load node to the node with lower resource utilization, avoiding response delay caused by overload of the high-load node, ensuring that the knowledge node quickly adapts to load changes, and ensuring the smoothness of system operation.

[0121] It can be understood that after the shunt processing, the running data of the knowledge node corresponding to the priority push content is continuously monitored. If the running data of the knowledge node represents that the knowledge node is in a relatively balanced state of response time and resource utilization when running, it means that the adjustment is successful, and it is not necessary to adjust again temporarily.

[0122] In the embodiment, the access request log of the user is also obtained after the shunt processing. The processing of steps S405 to S406 is circularly performed according to the access distribution characteristics in the access request log, so as to optimize the knowledge graph structure, until the priority push content dynamically adapts to the change of the user demand, and the long-term attraction of the recommended content is ensured.

[0123] S407, according to the target field and the access request log of each kind of priority push content, the logical relationship between the knowledge nodes is reconstructed, and the initial architecture of the knowledge graph is adjusted.

[0124] The mapping relationship between the knowledge nodes is redefined by using the graph construction tool. The connection strength is strengthened, and a new push content mapping relationship is generated. This adjustment can make the recommended content more focused on the user preference, and the targeting of the push is improved.

[0125] The system provided by the embodiment of the application is described below. The system described below can be correspondingly referred to the method described above.

[0126] Please refer to Figure 5 , Figure 5 The structure schematic diagram of the field knowledge graph construction system based on the graph neural network of the embodiment of the application is shown. The system can include:

[0127] The data acquisition module 10 is used for acquiring the basic data of the target field, performing data analysis on the basic data, and extracting the knowledge nodes and the node path composed of the knowledge nodes from the basic data.

[0128] In the embodiment, the above basic data is acquired through the pre-established field knowledge base. The field knowledge base can supplement the basic data of the target field stored therein by continuously collecting and summarizing literature materials, record materials and interview materials of the related field.

[0129] The data association module 20 is used for dividing a plurality of target field node groups according to the context association characteristics of the knowledge nodes and the matching degree with the target field. Each target field node group includes a center node and a knowledge node having a logical relationship with the center node.

[0130] By calculating the matching degree of the knowledge node and the target field, it can be ensured that the screened knowledge node is highly relevant to the core demand, and a complete logical chain is formed from the association calculation of the knowledge node to the final division of the group, ensuring that the knowledge content is highly consistent with the demand of the target field, and laying a solid foundation for subsequent application.

[0131] The atlas construction module 30 is used to mine the potential association path between each knowledge node in the target field node group, construct a knowledge graph according to the knowledge node, the logical relationship and the potential association path, determine the core node in the knowledge graph, obtain the preliminary push content from the preset database according to the core node, determine the push priority of the preliminary push content according to the user's preference, and obtain the priority push content.

[0132] Mining the potential association path between the knowledge nodes by strength helps to focus on the core association and avoid the interference of irrelevant information. The core node is a knowledge node with high centrality in the knowledge graph, and the centrality can be used to quickly screen out node information that is crucial to knowledge push. Obtaining the preference content of the user can timely capture the dynamic changes of the user's interest, so that the push priority can be adjusted according to the preference content, and the priority push content is obtained.

[0133] The field knowledge graph construction system based on the graph neural network of the application divides a plurality of target field node groups according to the context association characteristics of the knowledge nodes and the matching degree with the target field after obtaining the knowledge nodes, so that the divided target field node groups are highly associated with the target field, ensuring that the knowledge content is highly consistent with the demand of the target field. The potential association path between each knowledge node in the target field node group is mined and a knowledge graph is constructed therefrom. The core node in the knowledge graph is determined according to the centrality of the knowledge node. The preliminary push content is obtained from the preset database according to the core node. This can accurately identify the core concept and the key relationship, and can enhance the adaptive ability of the system in a specific scene, improve the application pertinence of the knowledge node, obtain the preference content of the user, and timely capture the dynamic changes of the user's interest. Thus, the push priority can be adjusted according to the preference content to obtain the priority push content. And the response speed of frequently accessed content is optimized through the efficient storage architecture, realizing intelligent and personalized knowledge push.

[0134] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the application.

Claims

1. A method for constructing a domain knowledge graph based on graph neural networks, characterized in that: The method includes: Acquire basic data in the target domain, perform data analysis on the basic data, and extract knowledge nodes and node paths composed of knowledge nodes from the basic data; the knowledge nodes form node paths through logical relationships. Based on the contextual association features of knowledge nodes and their matching degree with the target domain, several target domain node groups are divided; each target domain node group includes a central node and knowledge nodes that have a logical relationship with the central node; The system explores potential connections between knowledge nodes in a target domain node group, constructs a knowledge graph based on knowledge nodes, logical relationships, and potential connections, identifies core nodes in the knowledge graph, obtains initial push content from a pre-set database based on the core nodes, and determines the push priority of the initial push content based on user preferences to obtain priority push content. Specifically, the process of dividing knowledge nodes into several target domain node groups based on their contextual association features and matching degree with the target domain includes: Determine the matching degree between knowledge nodes and the target domain, and retain knowledge nodes whose matching degree exceeds the first preset matching degree; Determine the contextual association features of knowledge nodes, identify the central node based on the contextual association features, and divide the candidate node groups according to the logical relationships of the central nodes; The contextual association features of each knowledge node in each candidate node group are fused to obtain node fusion features. The matching degree between the node fusion features and the target domain is determined, and the candidate node group with the matching degree exceeding the second preset matching degree is taken as the target domain node group. Specifically, the process involves mining potential connection paths between knowledge nodes in the target domain node group, constructing a knowledge graph based on knowledge nodes, logical relationships, and potential connection paths, identifying core nodes in the knowledge graph, obtaining initial push content from a preset database based on the core nodes, and determining the push priority of the initial push content according to user preferences to obtain priority push content. This includes: Obtain the semantic similarity and context dependency matrix among the knowledge nodes in the target domain node group; Based on semantic similarity and context dependency matrix, and using graph neural network model, the graph neural network model outputs the connection strength between each knowledge node in the target domain node group, and the potential association path between knowledge nodes is mined based on the connection strength. Based on the logical relationships and potential association paths between knowledge nodes in all target domain node groups, an initial framework for the knowledge graph is constructed, and the centrality of each knowledge node in the initial framework is determined according to the connection strength and the number of connections. Knowledge nodes with centrality exceeding the preset centrality are designated as core nodes, and initial push content is obtained from the preset database based on the core nodes; The system acquires user interaction data within a first preset time period, determines the frequency of user interaction with each type of initial push content based on the interaction data, and determines profile data representing user behavior patterns based on the interaction frequency; wherein, the interaction data includes user historical data and real-time behavior data. The initial push priority of the content is determined based on the user profile data. The initial push content is pushed to the user according to the push priority, and the user's interaction data in the second time period is obtained. By comparing the interaction data of the first and second time periods, the user's preferred content is obtained, and the push priority is adjusted according to the preferred content. The priority is then determined based on the adjusted priority. The process of acquiring basic data in the target domain, performing data analysis on the basic data, and extracting knowledge nodes and node paths composed of these knowledge nodes specifically includes: Acquire basic data in the target domain and convert it into text data; The text data is preprocessed, and then the preprocessed text data is segmented to obtain text fragments; Semantic analysis is performed on text fragments to determine the contextual relevance between them, and text fragments with a relevance exceeding a preset threshold are classified as the same knowledge node; Determine the logical relationships between different knowledge nodes and use these logical relationships as node connection information to connect the knowledge nodes. Based on the node connection information, establish node paths composed of knowledge nodes.

2. The domain knowledge graph construction method based on graph neural networks as described in claim 1, characterized in that: Before the step of designating knowledge nodes with centrality exceeding a preset centrality as core nodes and obtaining initial push content from a preset database based on the core nodes, the method further includes: Adjust potential related paths according to the target domain; the adjustment methods include adding potential related paths and deleting potential related paths.

3. The domain knowledge graph construction method based on graph neural networks as described in claim 1, characterized in that: The method further includes: Priority-driven content is stored using distributed storage, and cache space is allocated to priority-driven content based on its priority; higher-priority content receives more cache space.

4. The domain knowledge graph construction method based on graph neural networks as described in claim 3, characterized in that: The method further includes: Obtain access request logs for each priority level of pushed content, extract access distribution characteristics from the access request logs, and adjust the cache space allocated to priority-pushed content based on the access distribution characteristics.

5. The domain knowledge graph construction method based on graph neural networks as described in claim 4, characterized in that: The method further includes: Based on the adjusted storage space according to the priority push content, obtain the runtime data of the knowledge nodes corresponding to the priority push content, and perform traffic distribution processing on the access requests of the priority push content according to the runtime data; wherein, the runtime data includes response time and resource utilization.

6. The domain knowledge graph construction method based on graph neural networks as described in claim 5, characterized in that: The method further includes: Based on the access request logs of the target domain and the content pushed for each priority, the logical relationships between knowledge nodes are reconstructed, and the initial architecture of the knowledge graph is adjusted.

7. A domain knowledge graph construction system based on graph neural networks, used to implement the domain knowledge graph construction method based on graph neural networks as described in any one of claims 1 to 6, characterized in that: The system includes: The data acquisition module allows users to acquire basic data in the target domain, perform data analysis on the basic data, and extract knowledge nodes and node paths composed of knowledge nodes from the basic data; the knowledge nodes are connected by logical relationships to form node paths. The data matching module is used to divide several target domain node groups according to the contextual association features of knowledge nodes and their matching degree with the target domain; each target domain node group includes a central node and knowledge nodes that have a logical relationship with the central node; The knowledge graph construction module is used to mine the potential connection paths between various knowledge nodes in the target domain node group. It constructs a knowledge graph based on knowledge nodes, logical relationships, and potential connection paths, determines the core nodes in the knowledge graph, obtains preliminary push content from a preset database based on the core nodes, and determines the push priority of the preliminary push content based on user preferences to obtain priority push content.

Images