Modeling Relational Data with Graph Convolutional Networks

This paper introduces Relational Graph Convolutional Networks (R-GCNs), a neural network model developed specifically to handle the highly multi-relational data characteristic of realistic knowledge bases. Motivated by the fact that even the largest knowledge graphs, such as Yago, DBpedia and Wikidata, remain incomplete, the authors apply R-GCNs to two standard knowledge base completion tasks: link prediction, the recovery of missing subject-predicate-object triples, and entity classification, the recovery of missing entity attributes. R-GCNs are related to a recent class of neural networks that operate directly on graphs.

The authors demonstrate that R-GCNs are effective as a stand-alone model for entity classification. For link prediction, they show that factorization models such as DistMult can be significantly improved by using an R-GCN encoder to accumulate evidence over multiple inference steps in the graph. This combination yields a large improvement of 29.8% on the FB15k-237 dataset over a decoder-only baseline, establishing R-GCNs as a strong approach for encoding relational structure in knowledge base completion.