Entity prediction is the task of predicting a missing entity that has a specific relation-ship with another given entity. Researchers usually use knowledge graphs embedding(KGE) methods to embed triples into continuous vectors for computation and perform the tasks of entity prediction. However, KGE models tend to use simple operations to refactor entities and relationships, resulting in insufficient interaction of components of knowledge graphs (KGs), thus limiting the performance of the entity prediction model. In this paper, we propose a new entity prediction model called FRS(Feature Refactoring Scoring) to alleviate the problem of insufficient interaction and solve information incom-pleteness problems in the KGs. Different from the traditional KGE methods of directly using simple operations, the FRS model innovatively provides the procedure of feature processing in the entity prediction tasks, realizing the alignment of entities and relationships in the same feature space and improving the performance of entity prediction model. Although FRS is a simple three-layer network, we find that our own model outperforms state-of-the-art KGC methods in FB15K and WN18. Through extensive experiments on FRS, we discover several insights. For example, the effect of embedding size and negative candidate sampling probability on experimental results is in reverse.
Keywords: FRS; entity prediction; knowledge graphs; knowledge graphs embedding.