T2V-TransD: A Method of Temporal Knowledge Representation Learning Based on Mapping Matrix

The purpose of knowledge representation learning(KRL) is to represent the symbolic knowledge graph with a numerical vector. The traditional knowledge representation learning method learn the representation of triples without temporal information. While on temporal knowledge graphs with temporal information, traditional methods cannot effectively utilize temporal information. To address this problem, researchers have proposed temporal knowledge representation learning(TKRL) methods, but most of the existing methods represent entities and relations in the same vector space, which limits the effectiveness of the models. In addition, some methods model for point time information, which cannot effectively handle temporal information of time intervals. Therefore, this paper proposes a mapping matrix based learning method for temporal knowledge graph representation. In this method, temporal information is represented by vector, which is used to construct mapping matrix with relation vector, and entities are mapped into the relation space through the mapping matrices. The experimental results show that on the link prediction task of the temporal knowledge graph, the method in this paper can effectively utilize temporal information. For example, on the dataset Wikidata12k with time interval information, compared with the model TransE, Hit@10 increased by 10.8 percentage points, Hit@1 increased by 4.4 percentage points.