Research On Representation Model Of Knowledge Graph Completion Based On Tensor Decomposition

In the information age,the explosive growth of data directly leads to information overload.How to extract accurate information from complex information has become a key issue.In this context,knowledge graphs are proposed to optimize existing search engines.A knowledge graph is a large-scale semantic network designed to describe objective world facts.But knowledge graphs are usually incomplete,which makes users unable to obtain information accurately and completely.Thus,knowledge graph completion has become a research hotspot in the field of artificial intelligence.Generally,knowledge graph completion includes static knowledge graph completion and temporal knowledge graph completion.Static KGs is a semantic network composed of entities and relations.Due to the increasingly serious problem of incomplete triples in static KGs,manual completion will not be adequate.With the close integration of mathematical theory and computer,tensor decomposition has been well applied in KGs.Thesis proposes Block Decomposition Based On Relational Interaction For Knowledge Graph Completion(BDRI).BDRI is a special variant of Block Term Decomposition that more accurately decomposes the formed factors.Aiming at the problem that existing models only learn inverse relations independently,BDRI achieves sufficient binding of the forward and inverse relations in an enhanced way.Thesis alleviates the overfitting problem by improving the nuclear norm regularization,and further proves that the BDRI is fully expressive.Experiment on public datasets with good results.Temporal KGs is a semantic network composed of timestamps,entities and relations,which is more in line with the development laws of the objective world.Due to the non-stationarity and heterogeneity of data,the task of completing the temporal knowledge graph is extremely complicated.Aiming at the problem of inaccurate fact descriptions caused by the lack of temporal information,most of the existing models are only built around static knowledge graphs and ignore the lack of temporal information.Thesis proposes Block Decomposition Based On Relational Interaction For Temporal Knowledge Graph Completion(TBDRI)Tucker Decomposition For Temporal Knowledge Graph Completion(TTucker).Aiming at the insufficient fusion of time information in the existing models,TBDRI studies time information more fully by embedding time information into core tensors on the basis of BDRI.Based on Tucker decomposition,TTucker is based on Tucker decomposition,and realizes the binding of time and relation by considering any inter-entity relation as a relation at a specific time.Thesis proves that both models are fully expressive.Validation on multiple different public time series datasets demonstrates the effectiveness of both models.