The Research of Link Prediction in Knowledge Graph based on Distance Constraint

Abstract

Large-scale knowledge graphs have a lot of hidden knowledge which has not been discovered, so the link prediction of the knowledge graph is an important topic. Translation models represented by TransE are the well-researched algorithms of link prediction. They project the entities and the relations in the knowledge graphs into some continuous vector spaces, and adjust the vector representations of the relations and the entities according to each piece of knowledge. However, in the case of a non-1-to-1 relationship, multiple entity vectors will compete for the same coordinate position in the space. Aiming at this problem, this paper proposes an improved method. By imposing a distance constraint on the competitive entities of a non-1-to-1 relationship, we can narrow the differences between them. Each entity will consider the other competitive entities while adapting itself to fit a triplet, so as to reach the status that each competitive entity is close to the coordinate point of the competition as a whole. Distance constraint can be applied to the existing translation models as a means of optimization. Experiments are conducted on the datasets: FB15K and WN18, and the experimental results show that the method we proposed is effective.