Construction method of knowledge graph of autism multi-source heterogeneous big data

Abstract

The present application relates to the technical field of knowledge graph construction, especially to a construction method of knowledge graph of autism multi-source heterogeneous big data, comprising the following steps: step 1: acquiring original data from electronic medical record systems, gene databases, research literature databases and social media multi-source heterogeneous data sources in real time or timing mode through a pluggable data collector; step 2: cleaning, denoising, missing value completion and de-identification of the original data, and standardizing mapping of terms based on the unified medical language system, the fifth edition of the diagnostic and statistical manual of mental disorders and the human phenotype medical ontology; step 3: expanding a subset of autism-related concepts and performing version control based on the above medical ontology to form a special autism ontology, in the present application, high-precision fusion of multi-source entities is realized, and technical bottlenecks such as difficulty in standardization mapping, low extraction accuracy and difficulty in fusion alignment are fundamentally solved.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph construction technology, specifically a method for constructing a knowledge graph of multi-source heterogeneous big data on autism. Background Technology

[0002] The data sources in the field of autism are abundant, including a variety of heterogeneous big data such as electronic medical records, genomic databases, literature, and social media.

[0003] Traditional methods for constructing knowledge graphs often rely on rule-driven or statistical learning, which often require manual design of mapping rules or reliance on large-scale labeled corpora. Some studies also use deep models to extract entities and relationships, but most of them are limited to a single data source and lack the ability to integrate multiple heterogeneous sources, making it difficult to support the comprehensive analysis of autism across disciplines and modalities.

[0004] In existing autism knowledge graph construction: First, due to inconsistent terminology naming conventions across data sources and low coverage of general ontology, the mapping rate of synonymous and ambiguous terms is insufficient. Second, traditional statistical or shallow learning models perform poorly in long-tail relationships and small-sample scenarios, making it difficult to mine deep semantic and complex structural information. Third, the multi-source entity alignment process is prone to errors and omissions, and lacks a quantitative evaluation and optimization mechanism for conflicting triples. In addition, traditional batch construction methods cannot respond promptly to updates of massive heterogeneous data, and the online incremental update mechanism is still imperfect, limiting the system's scalability and timeliness. Finally, clinical and social media data often contain sensitive information, and existing solutions struggle to balance de-identification and full-process compliance auditing, posing significant challenges to security and compliance. Therefore, this paper proposes a method for constructing a knowledge graph of multi-source heterogeneous big data related to autism to address these issues. Summary of the Invention

[0005] The purpose of this invention is to provide a method for constructing a knowledge graph from multi-source heterogeneous big data on autism, addressing the following issues in existing autism knowledge graph construction: First, due to inconsistent terminology naming conventions across data sources and low coverage of general ontology, the mapping rate of synonymous and ambiguous terms is insufficient; second, traditional statistical or shallow learning models perform poorly in long-tail relationships and small-sample scenarios, making it difficult to mine deep semantic and complex structural information; third, the multi-source entity alignment process is prone to errors and omissions, and lacks a quantitative evaluation and optimization mechanism for conflicting triples; furthermore, traditional batch construction methods cannot respond promptly to updates of massive heterogeneous data, the online incremental update mechanism is still imperfect, and the system's scalability and timeliness are limited; finally, clinical and social media data often contain sensitive information, and existing solutions struggle to strike a balance between de-identification and full-process compliance auditing, posing significant challenges to security and compliance.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The method for constructing a knowledge graph of multi-source heterogeneous big data on autism includes the following steps:

[0008] Step 1: Using a pluggable data acquisition device, raw data is acquired from multiple heterogeneous data sources, including electronic medical record systems, gene databases, research literature repositories, and social media, in real-time or at scheduled times.

[0009] Step 2: Clean, denoise, fill in missing values ​​and de-identify the raw data, and standardize the terminology based on the Unified Medical Language System, Diagnostic and Statistical Manual of Mental Disorders, 5th Edition and Human Phenotypic Ontology.

[0010] Step 3: Based on the above medical ontology, expand the subset of autism-related concepts and implement version control to form a dedicated autism ontology, and provide query services to the outside world through the presentation layer state transition interface;

[0011] Step 4: Based on the fusion of rules and deep learning models, entity recognition is performed using a fine-tuned biomedical bidirectional encoder representation model, and relation extraction is performed by combining dependency syntax rules and scientific text bidirectional encoder representation models. The confidence of the extraction results is evaluated, and results below the preset threshold are marked for expert review.

[0012] Step 5: For the same entity extracted from multiple sources, calculate the attribute cosine similarity and the context embedding similarity obtained from the graph neighborhood convergence model, merge the same entity using a weighted discrimination strategy, and optimize the storage of conflicting triples based on the trust level of the data source.

[0013] Step 6: Store the merged entities and relations in the form of triples into a graph database that supports both ontology query language and graph query language interfaces, and build an index based on Neo4j or JanusGraph to accelerate the query.

[0014] Step 7: Employ a cross-relation translation model to embed cross-type relations, and use a graph neighborhood convergence model to learn the representation of node context, and use a link prediction algorithm to automatically complete potential relations.

[0015] As a further optimization of the present invention, it also includes the following steps:

[0016] Step 8: Monitor the change logs of each data source and trigger the incremental update process of data preprocessing, knowledge extraction and fusion; periodically perform graph consistency and drift detection. In addition to the regular connectivity and pattern drift detection, the entity conflict rate is also calculated and used as the basis for data quality early warning.

[0017] Step 9: Based on the Unified Authorization Protocol 2 and the role-based access control mechanism, manage access and operation permissions for different roles, and record all query and update logs to meet GDPR and HIPAA compliance.

[0018] Step 10: Provide an interactive knowledge graph visualization interface based on D3.js, as well as a RESTful / GraphQL query and analysis service interface for external systems.

[0019] As a further optimization of the present invention, in step 2, the data preprocessing specifically includes:

[0020] For text data, first apply regular expression rules to filter out noisy characters;

[0021] Stop words are removed using a stop word list and a multi-pattern matching algorithm based on a prefix tree;

[0022] Fine-grained word segmentation is performed using Jieba word segmentation combined with a sub-word segmentation algorithm;

[0023] For missing values, the k-nearest neighbor interpolation algorithm is used to complete the missing values ​​based on patient group characteristics;

[0024] The candidate set of synonyms for terms is calculated based on edit distance, and the ontology term mapping is completed by combining it with the cosine similarity threshold.

[0025] As a further optimization of the present invention, in step 3, the ontology management specifically includes:

[0026] The DSM-5 and UMLS source ontology is parsed through the OWL application programming interface, and autism-related concepts are automatically extracted using a description logic-based inference engine.

[0027] A concept discovery algorithm based on pattern matching is used to identify potential new concepts from clinical texts and generate ontology expansion candidates.

[0028] The extended ontology is managed with version numbers according to a semantic version control strategy, and incremental differences are automatically recorded in the code repository.

[0029] As a further optimization of the present invention, in step 4, the knowledge extraction specifically includes:

[0030] Entity recognition: The model is represented by a fine-tuned biomedical bidirectional encoder. The output feature sequence is followed by a conditional random field layer and trained using the cross-entropy loss function.

[0031] Relation extraction: Tree-LSTM path vectors are constructed based on syntactic dependency trees, and sentence-level features are extracted using a scientific text bidirectional encoder representation model. After concatenation, the relation type is predicted by a fully connected classifier.

[0032] All extraction results are scored using a normalized exponential function confidence score, and results below 0.6 are automatically marked for manual review.

[0033] As a further optimization of the present invention, in step 5, the entity alignment and fusion specifically includes:

[0034] Attribute similarity calculation: After vectorizing the entity attributes, calculate the cosine similarity and set the threshold to 0.8 for initial screening;

[0035] Contextual embedding similarity: The entity neighbor subgraph is encoded using a graph neighborhood convergence model and the cosine similarity of the embedding vectors is calculated, with a threshold of 0.75.

[0036] Based on a bottom-up hierarchical agglomerative clustering algorithm, entity clusters with both double similarity exceeding the threshold are merged into like-point nodes.

[0037] Among triples with multiple source conflicts, the triple with the highest score is selected and added to the database after weighting by source trust level.

[0038] As a further optimization of the present invention, in step 6, the graph storage and retrieval specifically includes: constructing a composite B+ tree index in the Neo4j database for entity unique identifiers, entity type labels and timestamp attributes;

[0039] For high-frequency query patterns, pre-compute path summaries and use an improved FloydWarshall algorithm to generate a shortest path cache;

[0040] It provides parameterized query templates based on Cypher and ontology query interfaces based on SPARQL, and shares a unified permission management layer.

[0041] As a further optimization of this invention, the knowledge embedding and completion in step 7 specifically includes: The cross-relational translation model is trained with L as the objective function: ; In the formula, h represents the head entity in the triple, t represents the tail entity in the triple, r represents the relation type described by the triple, and S represents the set of positive samples, i.e., the set of real triples that already exist in the graph. This represents the set of negative samples, a set of incorrect triples constructed by replacing the head or tail entity. This is a marginal hyperparameter used to control the distance difference between positive and negative samples; it is set to 1.0. The calculation is the distance difference between the true triplet and the negative triplet, used to minimize the true triplet distance and maximize the negative triplet distance. The Adam optimizer is used, the learning rate is set to 0.001, the positive to negative sample ratio is 1:5, and the training iteration is 100 rounds. The graph neighborhood aggregation model uses an average aggregater with a neighbor sampling size of 25. After aggregation, the vector is concatenated with its own features and then processed by a multilayer perceptron to generate the final node representation. Based on the trained entity and relation embeddings, link prediction is performed, and candidate edges with a confidence of 0.7 are automatically completed into the graph and their sources are labeled.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] 1. In this invention, a unified medical ontology mapping is achieved by combining a pluggable data collector with a multi-mode preprocessing algorithm; an entity recognition and relation extraction technology based on rule-based and deep model fusion is adopted, and confidence assessment and manual review are introduced; a dual alignment strategy of attribute cosine similarity and graph neighborhood convergence model, along with a confidence weighting mechanism, is used to achieve high-precision fusion of multi-source entities, fundamentally solving the technical bottlenecks such as difficulty in standardized mapping, low extraction accuracy, and difficulty in fusion alignment;

[0048] 2. In this invention, a change log monitoring and incremental update process is designed, which combines consistency and drift detection to realize automated online updates of the knowledge graph; pre-computed path summaries and composite indexing technology greatly improve query performance, enabling the system to stably support daily updates of massive triples, meeting the timeliness and scalability requirements of autism research and clinical applications;

[0049] 3. In this invention, by using refined permission management based on the unified authorization protocol 2 and role-based access control, combined with de-identification and log auditing strategies, this method provides a D3.js visual interface and RESTful / GraphQL service interface while ensuring data privacy and compliance requirements, greatly improving the platform's security, trustworthiness, and user-friendliness. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the method for constructing a knowledge graph of multi-source heterogeneous big data on autism according to the present invention.

[0051] Figure 2 This is the overall framework diagram for constructing a knowledge graph of multi-source heterogeneous big data on autism in this invention. Detailed Implementation

[0052] Please see Figures 1-2 The present invention provides a technical solution:

[0053] The method for constructing a knowledge graph of multi-source heterogeneous big data on autism includes the following steps:

[0054] Step 1: Using a pluggable data acquisition device, raw data is acquired from multiple heterogeneous data sources, including electronic medical record systems, gene databases, research literature repositories, and social media, in real-time or at scheduled times.

[0055] Step 2: Clean, denoise, fill in missing values ​​and de-identify the raw data, and standardize the terminology based on the Unified Medical Language System, Diagnostic and Statistical Manual of Mental Disorders, 5th Edition and Human Phenotypic Ontology.

[0056] Step 3: Based on the above medical ontology, expand the subset of autism-related concepts and implement version control to form a dedicated autism ontology, and provide query services to the outside world through the presentation layer state transition interface;

[0057] Step 4: Based on the fusion of rules and deep learning models, entity recognition is performed using a fine-tuned biomedical bidirectional encoder representation model, and relation extraction is performed by combining dependency syntax rules and scientific text bidirectional encoder representation models. The confidence of the extraction results is evaluated, and results below the preset threshold are marked for expert review.

[0058] Step 5: For the same entity extracted from multiple sources, calculate the attribute cosine similarity and the context embedding similarity obtained from the graph neighborhood convergence model, merge the same entity using a weighted discrimination strategy, and optimize the storage of conflicting triples based on the trust level of the data source.

[0059] Step 6: Store the merged entities and relations in the form of triples into a graph database that supports both ontology query language and graph query language interfaces, and build an index based on Neo4j or JanusGraph to accelerate the query.

[0060] Step 7: Employ a cross-relation translation model to embed cross-type relationships, and use a graph neighborhood convergence model to learn the representation of node context. Use a link prediction algorithm to automatically complete potential relationships. Through an end-to-end, modular process design, systematically solve the complete link problems of multi-source heterogeneity, standardization, extraction, fusion, storage and completion of data, and build a high-quality and scalable knowledge graph platform for the autism field.

[0061] As a further technical solution for implementing this solution, the following steps are also included: Step 8: Monitor the change logs of each data source and trigger the incremental update process of data preprocessing, knowledge extraction and fusion; periodically perform graph consistency and drift detection. In addition to the regular connectivity and pattern drift detection, the entity conflict rate is also calculated and used as the basis for data quality early warning. Step 9: Based on the Unified Authorization Protocol 2 and the role-based access control mechanism, manage access and operation permissions for different roles, and record all query and update logs to meet GDPR and HIPAA compliance. Step 10: Provide an interactive knowledge graph visualization interface based on D3.js, as well as RESTful / GraphQL query and analysis service interfaces for external systems. Ensure the timeliness and consistency of the graph through incremental updates and continuous monitoring; meet regulatory requirements by adopting granular permission and audit control; and improve user experience and system availability through visualization and open interfaces. As a further implementation of this solution, in step 2, the data preprocessing specifically includes: For text data, first apply regular expression rules to filter out noisy characters; Stop words are removed using a stop word list and a multi-pattern matching algorithm based on a prefix tree; Fine-grained word segmentation is performed using Jieba word segmentation combined with a sub-word segmentation algorithm; For missing values, the k-nearest neighbor interpolation algorithm is used to complete the missing values ​​based on patient group characteristics; The candidate set of synonyms for terms is calculated based on edit distance, and the ontology term mapping is completed by combining cosine similarity threshold. This step significantly improves the accuracy of text cleaning and word segmentation through the collaboration of multiple algorithms. The interpolation strategy based on patient group features and high threshold mapping ensure data integrity and mapping accuracy. As a further implementation of this solution, in step 3, the ontology management specifically includes: The DSM-5 and UMLS source ontology is parsed through the OWL application programming interface, and autism-related concepts are automatically extracted using a description logic-based inference engine. A concept discovery algorithm based on pattern matching is used to identify potential new concepts from clinical texts and generate ontology expansion candidates. The extended ontology is managed with version numbers according to a semantic version control strategy, and incremental differences are automatically recorded in the code repository. This enables automated extension and fine-grained management of the domain ontology, ensuring that the ontology keeps pace with the times and that every change can be traced to improve maintenance efficiency. As a further implementation of this solution, in step 4, the knowledge extraction specifically includes: Entity recognition: The model is represented by a fine-tuned biomedical bidirectional encoder. The output feature sequence is followed by a conditional random field layer and trained using the cross-entropy loss function. Relation extraction: Tree-LSTM path vectors are constructed based on syntactic dependency trees, and sentence-level features are extracted using a scientific text bidirectional encoder representation model. After concatenation, the relation type is predicted by a fully connected classifier. All extraction results are scored using a normalized exponential function confidence score. Results below 0.6 are automatically marked for manual review. The fusion of rules and deep learning models improves extraction performance. The application of Conditional Random Field (CRF) and Tree-LSTM can better capture sequence and structural information. The manual review mechanism balances accuracy and efficiency. As a further implementation of this solution, in step 5, the entity alignment and fusion specifically include: Attribute similarity calculation: After vectorizing the entity attributes, calculate the cosine similarity and set the threshold to 0.8 for initial screening; Contextual embedding similarity: The entity neighbor subgraph is encoded using a graph neighborhood convergence model and the cosine similarity of the embedding vectors is calculated, with a threshold of 0.75. Based on a bottom-up hierarchical agglomerative clustering algorithm, entity clusters with both double similarity exceeding the threshold are merged into like-point nodes. In triples with multiple source conflicts, the triple with the highest score is selected and entered into the database after being weighted by source trust level. The combination of dual similarity and hierarchical clustering effectively reduces alignment error; the trust level weighting mechanism ensures the reliability of the fused data. As a further implementation of this solution, in step 6, the graph storage and retrieval specifically includes: constructing a composite B+ tree index in the Neo4j database for entity unique identifiers, entity type labels and timestamp attributes; For high-frequency query patterns, pre-compute path summaries and use an improved FloydWarshall algorithm to generate a shortest path cache; It provides parameterized query templates based on Cypher and ontology query interfaces based on SPARQL, and shares a unified permission management layer. Composite indexes and path caching significantly improve query performance; the dual query interfaces meet diverse retrieval needs and can be seamlessly integrated with the security layer. As a further implementation of this solution, the knowledge embedding and completion described in step 7 specifically includes: The cross-relational translation model is trained with L as the objective function: ; In the formula, h represents the head entity in the triple, t represents the tail entity in the triple, r represents the relation type described by the triple, and S represents the set of positive samples, i.e., the set of real triples that already exist in the graph. This represents the set of negative samples, a set of incorrect triples constructed by replacing the head or tail entity. This is a marginal hyperparameter used to control the distance difference between positive and negative samples; it is set to 1.0. The calculation is the distance difference between the true triplet and the negative triplet, used to minimize the true triplet distance and maximize the negative triplet distance. The Adam optimizer is used, the learning rate is set to 0.001, the positive to negative sample ratio is 1:5, and the training iteration is 100 rounds. The graph neighborhood aggregation model uses an average convergent vector with a neighbor sampling size of 25. After aggregation, the vector is concatenated with its own features and then processed by a multilayer perceptron to generate the final node representation. Based on the trained entity and relation embeddings, link prediction is performed, and candidate edges with a confidence of 0.7 are automatically completed into the graph and their sources are labeled. The TransR cross-relation translation model is used to refine the projection and marginal loss constraints, so that real relations and negative samples have significant distinguishability in the embedding space. The GraphSAGE graph neighborhood sampling aggregation model enhances the semantic representation of nodes, thereby improving the overall accuracy and coverage of graph completion.

[0062] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only for the purpose of helping to understand the method and core ideas of the present invention. The above descriptions are only preferred embodiments of the present invention. It should be noted that due to the limitations of textual expression, while there are objectively infinite specific structures, those skilled in the art can make several improvements, modifications, or changes without departing from the principles of the present invention, and can also combine the above technical features in an appropriate manner. These improvements, modifications, changes, or combinations, or the direct application of the inventive concept and technical solution to other situations without modification, should all be considered within the scope of protection of the present invention.

Claims

1. A method for constructing a knowledge graph of multi-source heterogeneous big data on autism, characterized in that, Includes the following steps: Step 1: Using a pluggable data acquisition device, raw data is acquired from multiple heterogeneous data sources, including electronic medical record systems, gene databases, research literature repositories, and social media, in real-time or at scheduled times. Step 2: Clean, denoise, fill in missing values ​​and de-identify the raw data, and standardize the terminology based on the Unified Medical Language System, Diagnostic and Statistical Manual of Mental Disorders, 5th Edition and Human Phenotypic Ontology. Step 3: Based on the above medical ontology, expand the subset of autism-related concepts and implement version control to form a dedicated autism ontology, and provide query services to the outside world through the presentation layer state transition interface; Step 4: Based on the fusion of rules and deep learning models, entity recognition is performed using a fine-tuned biomedical bidirectional encoder representation model, and relation extraction is performed by combining dependency syntax rules and scientific text bidirectional encoder representation models. The confidence of the extraction results is evaluated, and results below the preset threshold are marked for expert review. Step 5: For the same entity extracted from multiple sources, calculate the attribute cosine similarity and the context embedding similarity obtained from the graph neighborhood convergence model, merge the same entity using a weighted discrimination strategy, and optimize the storage of conflicting triples based on the trust level of the data source. Step 6: Store the merged entities and relations in the form of triples into a graph database that supports both ontology query language and graph query language interfaces, and build an index based on Neo4j or JanusGraph to accelerate the query. Step 7: Employ a cross-relation translation model to embed cross-type relations, and use a graph neighborhood convergence model to learn the representation of node context, and use a link prediction algorithm to automatically complete potential relations.

2. The method for constructing a knowledge graph of multi-source heterogeneous big data on autism according to claim 1, characterized in that: It also includes the following steps: Step 8: Monitor the change logs of each data source and trigger the incremental update process of data preprocessing, knowledge extraction and fusion; periodically perform graph consistency and drift detection. In addition to the regular connectivity and pattern drift detection, the entity conflict rate is also calculated and used as the basis for data quality early warning. Step 9: Based on the Unified Authorization Protocol 2 and the role-based access control mechanism, manage access and operation permissions for different roles, and record all query and update logs to meet GDPR and HIPAA compliance. Step 10: Provide an interactive knowledge graph visualization interface based on D3.js, as well as a RESTful / GraphQL query and analysis service interface for external systems.

3. The method for constructing a knowledge graph of multi-source heterogeneous big data on autism according to claim 1, characterized in that: In step 2, the data preprocessing specifically includes: For text data, first apply regular expression rules to filter out noisy characters; Stop words are removed using a stop word list and a multi-pattern matching algorithm based on a prefix tree; Fine-grained word segmentation is performed using Jieba word segmentation combined with a sub-word segmentation algorithm; For missing values, the k-nearest neighbor interpolation algorithm is used to complete the missing values ​​based on patient group characteristics; The candidate set of synonyms for terms is calculated based on edit distance, and the ontology term mapping is completed by combining it with the cosine similarity threshold.

4. The method for constructing a knowledge graph of multi-source heterogeneous big data on autism according to claim 1, characterized in that: In step 3, the ontology management specifically includes: The DSM-5 and UMLS source ontology is parsed through the OWL application programming interface, and autism-related concepts are automatically extracted using a description logic-based inference engine. A concept discovery algorithm based on pattern matching is used to identify potential new concepts from clinical texts and generate ontology expansion candidates. The extended ontology is managed with version numbers according to a semantic version control strategy, and incremental differences are automatically recorded in the code repository.

5. The method for constructing a knowledge graph of multi-source heterogeneous big data on autism according to claim 1, characterized in that: In step 4, the knowledge extraction specifically includes: Entity recognition: The model is represented by a fine-tuned biomedical bidirectional encoder. The output feature sequence is followed by a conditional random field layer and trained using the cross-entropy loss function. Relation extraction: Tree-LSTM path vectors are constructed based on syntactic dependency trees, and sentence-level features are extracted using a scientific text bidirectional encoder representation model. After concatenation, the relation type is predicted by a fully connected classifier. All extraction results are scored using a normalized exponential function confidence score, and results below 0.6 are automatically marked for manual review.

6. The method for constructing a knowledge graph of multi-source heterogeneous big data on autism according to claim 1, characterized in that: In step 5, the entity alignment and fusion specifically includes: Attribute similarity calculation: After vectorizing the entity attributes, calculate the cosine similarity and set the threshold to 0.8 for initial screening; Contextual embedding similarity: The entity neighbor subgraph is encoded using a graph neighborhood convergence model and the cosine similarity of the embedding vectors is calculated, with a threshold of 0.

75. Based on a bottom-up hierarchical agglomerative clustering algorithm, entity clusters with both double similarity exceeding the threshold are merged into like-point nodes. Among triples with multiple source conflicts, the triple with the highest score is selected and added to the database after weighting by source trust level.

7. The method for constructing a knowledge graph of multi-source heterogeneous big data on autism according to claim 1, characterized in that: In step 6, the graph storage and retrieval specifically includes: constructing a composite B+ tree index in the Neo4j database for entity unique identifiers, entity type labels, and timestamp attributes; For high-frequency query patterns, pre-compute path summaries and use an improved FloydWarshall algorithm to generate a shortest path cache; It provides parameterized query templates based on Cypher and ontology query interfaces based on SPARQL, and shares a unified permission management layer.

8. The method for constructing a knowledge graph of multi-source heterogeneous big data on autism according to claim 1, characterized in that: The knowledge embedding and completion described in step 7 specifically include: The cross-relational translation model is trained with L as the objective function: ； In the formula, h represents the head entity in the triple, t represents the tail entity in the triple, r represents the relation type described by the triple, and S represents the set of positive samples, i.e., the set of real triples that already exist in the graph. This represents the set of negative samples, a set of incorrect triples constructed by replacing the head or tail entity. This is a marginal hyperparameter used to control the distance difference between positive and negative samples; it is set to 1.

0. The calculation is the distance difference between the true triplet and the negative triplet, used to minimize the true triplet distance and maximize the negative triplet distance. The Adam optimizer is used, the learning rate is set to 0.001, the positive to negative sample ratio is 1:5, and the training iteration is 100 rounds. The graph neighborhood aggregation model uses an average aggregater with a neighbor sampling size of 25. After aggregation, the vector is concatenated with its own features and then processed by a multilayer perceptron to generate the final node representation. Based on the trained entity and relation embeddings, link prediction is performed, and candidate edges with a confidence of 0.7 are automatically completed into the graph and their sources are labeled.

Images