The Research And Implementation Of Knowledge Graph Completion Based On Graph Neural Network

With the arrival of Big Data era,in order to better manage and analyze data,knowledge graphs(KGs)came into being.KGs provide the capability to organize,manage and analyze massive amounts of information,and are the basis for many downstream applications,such as,search referral and user portraits.However,most of the existing knowledge graphs are incomplete,which limits the development of knowledge graphs in downstream applications.Thus,it is urgent to improve the completeness of knowledge graphs,i.e.,to solve knowledge graph completion(KGC)problem.Knowledge graph completion aims to supplement the missing facts in knowledge graphs based on existing facts.According to whether existing facts are time-sensitive,this thesis divides knowledge graphs into static knowledge graphs and dynamic knowledge graphs.Thus,this thesis will focus on addressing the static knowledge graph completion problem and the dynamic knowledge graph completion problem and details research contents as follows.For the static knowledge graph completion problem,most of the existing methods lack a comprehensive consideration of the information of neighborhood nodes and directions of relations,which leads to losses of some important information in representations of target nodes.In order to solve the above problem,this thesis proposes a static knowledge graph completion method based on graph attention network,named,a Multirelation Graph Attention Network(MGAT).Firstly,this thesis designs graph attention-based sub-modules to aggregate the information of relations in different directions and entities to learn the embeddings of target entities.Then,this thesis designs a loss function based on validity and consistency of facts to guide the training process of the model.Experiments on two benchmark datasets show that MGAT improves the accuracy of static knowledge graph completion.For the dynamic knowledge graph completion problem,most of the existing methods lack a comprehensive consideration of the direction of relations in knowledge graphs and the timeliness of facts,which leads to declines of model capabilities for representing the timeliness in knowledge graphs.In order to solve this problem,this thesis proposes a dynamic knowledge graph completion method based on time graph network,called,Multi-relation Temporal Graph Attention Network(MTGAT).Firstly,this thesis models the dynamic knowledge graph as a collection of temporal graphs.Then,this thesis proposes a graph information aggregation method based on time decay,which achieves the aggregation of graph structure features at each moment and the aggregation of temporal features,and models the timeliness of facts through the time decay.This thesis conducts experiments on two benchmark datasets.Experiments on two benchmark datasets show that MTGAT improves the accuracy of dynamic knowledge graph completion.