Research On Link Prediction Algorithm Of Knowledge Graph Based On Few-shot Relational Learning

Knowledge Graph(KG)is an important method to describe the knowledge and model the relation between the entities with a graph model.Due to the incompleteness of the data sources for constructing KG,KG is often incomplete.Link prediction is used to predict the unknown triples by the existing triples for KG completion.Traditional method rely on relations that correspond to a large number of triples.However,in the real world knowledge graph,relations conform to long-tail distribution,that is,a large number of relations correspond to few triple instances.Recently,the research on Few-Shot Knowledge Graph Link Prediction(FSKGLP)faces great challenges.and proposes the knowledge representation method of dense neighbor coding,which effectively improves entity representation ability.The existing research on FSKGLP is not mature,and only simply introduces the few-shot learning algorithm into the knowledge graph link prediction task,whose knowledge representation ability and measurement ability need to be improved.This thesis proposes two improved Few-Shot Relational Learning(FSRL)algorithm to improve knowledge representation capability and metric capability respectively for the research on FSKGLP algorithm based on large-scale general KG and Chinese medicine industry KG.(1)Few-Shot Relational Learning considering Neighbor’s Degree Distribution(FSRLNDD)algorithm is proposed.Local neighbor encoding utilizes the information of neighbors by random sampling,which doesn’t consider the influence of sparsity of neighbor nodes on entities.Aiming at the problem of poor entity representation ability,this thesis analyzes the sparsity of the entity based on the degree centrality in the complex network,and use the knowledge representation method of dense neighbor encoding to effectively improves entity’s representation ability.(2)Few-Shot Relational Learning based on Deep Metric Network(FSRLDMN)algorithm is proposed.Cosine similarity,a fixed metric function,cannot effectively measure the semantic similarity of triples in KG.Aiming at the problem of poor metric ability,Deep Metric Network(DMN)is proposed to measure the similarity of entity pairs.It uses the automatically learned nonlinear metric function to effectively improve the metric ability between entity pairs.Evaluation in link prediction on public KG dataset NELL and private industry KG dataset for Traditional Chinese Medicine shows the above two proposed algorithms achieves new state-of-the-art results.