Knowledge Graph Completion Based On Network Role

As an important part of artificial intelligence technology,Knowledge Graphs(KGs)have been widely used in intelligent question-answering,intelligent search,personalized recommendation and other fields.However,due to the dynamic growth of knowledge graph,the existence of data sparsity affects the accuracy of artificial intelligence model.The problem of incompleteness of knowledge graph needs to be solved urgently.The concept of knowledge graph completion has been widely concerned in academia and industry.Most of the existing knowledge graph completion methods rely on the relations between entities in the knowledge graph.However,some entities in the real knowledge graph are rare and have fewer relations,which leads to low quality of entity representation learned,thus affecting the accuracy of the model.In addition,due to the challenge of processing non-discrete attribute values,most of the existing advanced relational learning models ignore these attribute information,and there are few researches on attribute prediction.Moreover,existing knowledge completion models are based on the triples of the knowledge graph,which requires users who have the background of the knowledge graph to quickly understand the prediction results,ignoring the importance of comprehensibility.Based on the existing researches,a novel knowledge graph completion method based on network role is proposed to solve the above problems.The main content is as follows:(1)A knowledge graph entity embedding method based on network role is proposed.According to the different similarities between entities,the entity role is discovered from three aspects of homogeneity,attributive similarity and structural similarity.Then different entity paths are generated to obtain representations of entities which have rich semantic information,solving the problem that the representation learning is negatively affected by the sparse relation in the knowledge graph.(2)A novel knowledge graph completion method based on entity label is proposed.With entity labels as the input and output of the model,users can quickly understand the prediction results through a small number of entity labels,which enhances the comprehensibility of the results.At the same time,the non-discrete attribute value is discretized,and the fuzzy prediction is used to replace the exact prediction of the attribute value,so the result of attribute prediction is changed from the exact value to the range.(3)Design and implement a knowledge graph completion system based on network role.The system can search the target entity and display the completion result of entity profiles in the knowledge graph dynamically.At the same time,the degree of fitting of the corresponding label prediction results are shown.In summary,a novel knowledge graph completion method based on network role is proposed: Firstly,the entity representations which are generated based on the entity role similarity,are used to generate the entity profiles.Then,with the entity labels as the input and output of knowledge graph completion model,new entity labels are predicted.Experimental results on real knowledge graph datasets show that the prediction performance of the proposed method is better than most existing models.Finally knowledge graph completion system is designed and implemented.