Temporal Knowledge Graph Embedding for Link Prediction

Abstract

Link prediction aims to infer the behavior of the network evolution process by predicting missed or future relationships based on currently observed connections. It has become an attractive area of research since it allows us to understand how networks will evolve. Early studies cast the link prediction task as an entity identifying problem on graphs and adopt vertex representation strategies to perform predictive analysis. Although these methods are effective to some extent, they overlook the special properties of network evolution. In this paper, we propose a new method named TKGE, short for Temporal Knowledge Graph Embedding, to learn the evolutional representations of temporal knowledge graph for link prediction task. Specifically, we employ the self-attention mechanism to incorporate the static structural information and dynamic temporal information by aggregating the context from related entities. By introducing the position embedding characterizing the dynamic information of temporal knowledge graph, TKGE can generate the evolutional embedding of entities and relations for downstream applications, such as link prediction, recommender system, and so on. We conduct experiments on several real datasets. Both quantitative results and qualitative analysis verify the effectiveness and rationality of our TKGE method.