Type-Enhanced Temporal Knowledge Graph Representation Learning Model

Knowledge graph representation learning, which aims to represent elements (entities and relations) of knowledge graphs in low-dimensional continuous vector spaces, is able to effectively perform knowledge graph completion and improve computational efficiency. And it has been a significant issue throughout the whole process from knowledge graph construction to application. However, most existing approaches of knowledge graph representation learning have been developed only on the basis of static structured triples and ignor the time attribute of knowledge and entity type features, which limits their performance on knowledge graph completion and semantic computation. To conquer this problem, based on a famous tensor factorization technique, we propose a type-enhanced temporal knowledge graph representation learning model called T-Temp. The T-Temp model explicitly incorporates different forms of time information into the process of knowledge graph representation learning, and takes advantage of type compatibility between entities and their associated relations to fully explore the implicit type features of entities, so as to further improve the performance of knowledge graph representation learning. In addition, we prove that T-Temp is fully expressive and compared with competitive models, our model has lower time and space complexity. Extensive experiments on several real-world temporal knowledge graphs demonstrate the merits and advancement of our proposed model.