A new relational reflection graph convolutional network for the knowledge representation

The goal of the knowledge representation is to embed entities and relationships in the facts into consecutive low-dimensional dense vectors. Although shallow embedding methods can directly map entities or relations into vectors, they lose information about the structure of the knowledge graph network during the learning process. Alternatively, deep embedding methods can be used to learn rich structural information. As a practical matter, existing deep embedding methods rely too heavily on simple logical operations, such as subtraction and multiplication, between entities and relations. This paper proposes a method for deep embedding, RRGCN. Unlike the traditional method of knowledge graph convolution, this method does not rely on logical transformations to determine aggregation information. In RRGCN, aggregation information is defined as a mapping projection of neighboring features on a unique hyperplane that corresponds to the relation. Furthermore, RRGCN constructs a residual neural network between two graph convolution layers in order to reduce the amount of information loss as a consequence of superimposing graph convolution layers on top of each other. Results from experiments show that RRGCN is capable of performing well on the publicly available benchmark datasets FB15k-237 and WN18RR in the knowledge graph link prediction task, which outerforms the state-of-the-art relevent models. © 2023, The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature.