Dynamic dual quaternion knowledge graph embedding

Knowledge graph aims to describe the concepts, the entities and the complex relations of them in the real world. Recently, a series of quaternion rotation models that usually considering the relation vector as the rotation between head and tail entities, have been extensively studied. The results showed these models had the advantages of simplicity and efficiency. However, they are quite weak in capturing the representation and the feature interaction between entities and relations, resulting in insufficient expressiveness of the underlying models, because these models only focus on the linear relations between entities and relations. In order to solve this problem, this paper proposes a novel knowledge graph embedding model called DualDE, which dynamically maps the dual quaternions to the knowledge graph. Specifically, DualDE uses a dynamic mapping mechanism to construct the entity transition vector and the relation transition vector, and continuously adjusts the embedding position of the entity vector in the dual quaternion space according to the dual quaternion multiplication rules. In addition, DualDE can dynamically construct a variety of complex relation types, such as one-to-many, many-to-one and many-to-many. The experimental results based on three standard data sets show that the DualDE model is superior to the existing knowledge graph embedding models on many metrics.