STKE: Temporal Knowledge Graph Embedding in the Spherical Coordinate System
Source of Publication
Communications in Computer and Information Science
Knowledge graph embedding (KGE) aims to learn the representation of entities and predicates in low-dimensional vector spaces which can complete the missing parts of the Knowledge Graphs (KGs). Nevertheless, temporal knowledge graphs (TKGs) that include time information are more consistent with real-world application scenarios. Meanwhile, the facts with time constraints make the results of reasoning over time more accurate. Because of these, we propose a novel temporal knowledge graph embedding (TKGE) model, namely Spherical Temporal Knowledge Graph Embedding (STKE), which embeds facts into a spherical coordinate system. We treat each fact as a rotation from the subject to the object. The entities and predicates in STKE are divided into three parts--the radial part, the azimuth part, and the polar part. The radial part aims to resize the modulus between two entities. The azimuth part is mainly used to distinguish entities with the same module length and the polar part aims to represent the transformation of the time embedding with the change of polar angle. We evaluate the proposed model via the link prediction task on four typical temporal datasets. Experiments demonstrate that STKE achieves a significant surpass compared with the state-of-the-art static knowledge graph embedding (SKGE) model and TKGE model. In addition, we analyze the representation ability of different facts in the spherical coordinate system and confirm that our model can better represent time-constrained facts.
Springer International Publishing
Temporal knowledge graph embedding, The Azimuth part, The polar part, The radial part, The spherical coordinate system
Wang, Shibo; Liu, Ruinan; Shen, Linshan; and Khattak, Asad Masood, "STKE: Temporal Knowledge Graph Embedding in the Spherical Coordinate System" (2022). All Works. 5200.
Indexed in Scopus