Research And Application Of Knowledge Graph Embedding Technique For Link Prediction

Knowledge graphs organize and represent data in a structured form,which utilizes triplets to describe the attributes of transactions and the semantic links between transactions.They have been widely used in the field of search engines,question answering systems,recommendation systems and etc.However,the existing knowledge graphs are incomplete,in which several links between entities are missing,and thus the link prediction task is proposed.The knowledge graph embedding technique is one of the mainstream approaches to solve the link prediction task by learning the representations of entities and relations in a low-dimensional continuous vector space to predict the potential semantic links between entities and relations.In order to handle the link prediction task in the knowledge graphs,this thesis researches and utilizes the knowledge graph embedding technique to complete the missing links in the knowledge graphs,and the main research work and contributions are summarized as follows:1.Only one type of convolution filters cannot model the feature interactions of entities and relations in the same and different dimensions simultaneously.To address this problem,this thesis proposes a convolution-based knowledge graph embedding model Simul E,which uses two paths to extract the interaction information of entities and relations in the same and different dimensions respectively,and generates more expressive feature embeddings by combining one-dimensional convolution and three-dimensional convolution.Experimental results of link prediction show that Simul E improves the MRR metric by 2.9%,2.9%,9.8% and 2.8% for the WN18 RR,FB15k-237,YAGO3-10 and DB100K datasets,respectively,compared with the baseline model Conv E,which proves the effectiveness and robustness of the model.2.To address the problem that the existing models based on convolutional neural network do not make full use of the path information,this thesis proposes a pathbased knowledge graph embedding model Simul E-Path.The model first samples relation paths,then encodes and aggregates sequential and semantic features on relation paths,and finally employs convolutional neural network to capture feature interactions of entity,relation,and path embeddings simultaneously.Experimental results of link prediction show that after incorporating the path information,Simul E-Path improves the Hits@10 metric by 1.3% and 2.8% for the FB15k-237 and YAGO3-10 datasets,respectively,further enhancing the performance of the model.3.To address the data sparsity problem faced by recommendation system,this thesis proposes Simul E-PRS,a recommendation system model based on knowledge graph embedding.The model applies a multi-task training framework to alternately train the knowledge graph embedding module and the recommendation system module,and uses the semantic information embedded in the knowledge graph embedding to enrich the feature embeddings of items and users in the recommendation system.Experimental results of click prediction show that the Simul E-PRS improves the ACC metric by 0.7%,1.9% and 1.8% for the Movie Lens-1M,Book-Crossing and Last.FM datasets,respectively,compared with the baseline model MKR,which enhances the performance of the recommendation model.