Structure-augmented knowledge graph embedding for sparse data with rule learning

Abstract In recent years, knowledge graphs have received considerable attention because of their ability to express rich information and their potential for use in knowledge-based reasoning. For example, they can assist in-depth knowledge discovery related to user associations, switching policies, and traffic content in mobile services. Knowledge graph embeddings project the entities and relations of knowledge graphs into vectors that are dense and low dimensional, thus allowing the complex semantic information and relations between these entities to be measured efficiently. However, traditional knowledge graph embedding methods consider only direct facts, making it difficult to achieve reasonable embedding learning of entities and relations when faced with sparse data. To settle this issue, this paper proposes a novel knowledge graph embedding method based on tensor decomposition combined with rule learning. First, rules are inferred and scored based on the initial embeddings of entities and relations. Then, new triples for sparse entities are inferred from rules with high scores. Finally, these new triples are iteratively embedded into the model. Experimental results obtained on the WN18 and FB15k datasets indicate that the proposed model achieves significantly better performance than other state-of-the-art knowledge graph embedding models when faced with sparse data.