CP Tensor Factorization for Knowledge Graph Completion

The problem of incomplete knowledge caused by the lack of relations in large-scale knowledge graphs increases the difficulty of downstream application tasks. Predicting the missing relations between entities according to the existing facts is the main means of knowledge graph completion. The triple of knowledge graph can be seen as a third-order binary tensor element that linearly transforms entities and relations into low-dimensional vectors through tensor decomposition to determine the probability that the triple of missing relations is true. However, the non-deterministic polynomiality in determining the tensor rank can lead to overfitting and unfavorable to the generation of low-rank models. Aiming at this problem, we propose to use CP decomposition to decompose the third-order tensor into the sum of multiple rank-one tensors, which is the sum of the outer products of the head entity embedding, relation embedding, and tail entity embedding for each triple, and convert it into a super-diagonal tensor product the factor matrix of each mode, and use scoring function calculate the probability that the triple of missing relation is true. Link prediction experimental results from four different domains of benchmarks knowledge graph datasets show that the proposed methods are better than other comparison methods, it also can express the complex relations of knowledge graph, and the decomposition has uniqueness, reduces the total amount of calculations and parameters, avoids overfitting.