Research And Application Of Link Prediction Algorithm Based On N-ary Relations

In today’s era of big data,the information and knowledge that people acquire is increasing at an exponential rate,and the knowledge graph is proposed as a carrier for storing and using this knowledge.However,only binary relations cannot represent all the facts in the real world,so knowledge hypergraphs consisting of n-ary relations representing associations between any number of entities is required to organize knowledge.However,due to the limitations of the knowledge base construction process,there are a lot of knowledge missing,so it is of great significance to use the link prediction algorithm to complete the knowledge hypergraph.N-ary relational link prediction is based on known links in the knowledge hypergraph to predict potential facts,and promotes downstream tasks based on the knowledge hypergraph.Although the existing link prediction techniques have achieved good results,there are still many problems.For example,when modeling structural information,the neighborhood information of entities and the information interaction between fact tuples are not fully considered.In view of the above problems,the main work completed in the thesis is as follows:1.When considering neighborhood structure information,graph attention network is used to gather entity neighborhood structure information,the tuple-entity pair is used instead of the neighbor entity pair,so that the model can gather complete neighborhood information,and extract corresponding information according to the location of the entity,so that there are differences in the information that can be obtained by different entities in the tuple.At the same time,combined with the modeling of the directional constraints of the knowledge hypergraph,the directional information is also learned into the entity embedding,so that the relations can better perceive the compatibility with the entity.Finally,compared with other baseline models,this model has certain improvements in multiple evaluation indicators on two public datasets.2.When considering the global information interaction among n-ary relational facts,an edge-level self-supervised auxiliary task is integrated into the training process of the model in this paper.The original knowledge hypergraph is converted into the hyperedge relationship graph,and the global interaction between tuple nodes is learned in the transformation graph,so as to enhance the modeling of knowledge hypergraph.By maximizing the interaction information between the tuple representations learned by the two networks,the rich features between the tuple nodes are extracted,the interaction information between the tuples is fully learned,and the embedded representation is enhanced.Finally,through a series of experimental results on two public datasets,the improvement effect of the self-supervised learning framework on the mainline model is verified.3.Combining the above models,thesis designs and implements a knowledge discovery system based on knowledge hypergraph.Realize discovering the knowledge associated with the input information in the knowledge base,and be able to query the specified knowledge.The usability of the model is verified while satisfying the relevant system requirements.