Research On Link Prediction Method Based On Entity Description Hyperplane

Link prediction is an important task of knowledge graph completion,which aims to model the possibility of links between nodes and complete the missing knowledge.And knowledge representation learning is a typical method to achieve knowledge graph link pre-diction at present,including triple-based link prediction models that learn structural infor-mation,and information fusion-based link prediction models that fuse structural information with other information.However,the application of latter is limitted by several factors,such as the high cost of external information,the lack of internal features and inefficient integra-tion process,etc.Therefore,it is necessary to consider introducing general information and efficiently fusing information to improve the accuracy of link prediction through information fusion.To address the above limitation,this thesis analyzes the information fusion-based link prediction models,and further propose a link prediction model based on entity descrip-tion hyperplane.Firstly,entity neighbor are adopted as the entity description information,which can be easily accessed within the knowledge graph.The local frequency and global frequency in the entity neighbor statistical information with automatic keyword extraction technology are combined to select the more important key neighbors.By introducing entity neighbor as descriptive information,reliance on external information is avoided and making the general link prediction model based on information fusion possible.Secondly,to com-plete the link prediction,this thesis uses hyperplane projection technology to represent the entity description as hyperplane and introduces the entity description information into the triple-based link prediction model.In addition,based on entity neighbor as the entity de-scription,this thesis proposes a link prediction model based on entity joint representation so as to improve the utilization of parameters and exploit the potential of entity neighbors.In this model,the entity structure representation corresponding to the entity neighbor is taken as the introduced description representation.Then,the model uses linear transformation to fuse information from different sources,and establishes a shortcut for the transmission of structural information to maintain its integrity.By introducing entity neighbor as entity description information,this thesis proposes two general link prediction models based on information fusion.We also conduct experi-ment invoiving several standard datasets to evaluate the performance of the proposed model in knowledge graph link prediction task.The results show that the proposed model can com-bine descriptive information and structural information effectively,and can be generalized to other knowledge graphs.