Temporal Knowledge Graph Embedding Learning through Multi-Aspect

Temporal Knowledge Graph Completion (TKGC) aims at inferring the missing facts in Temporal Knowledge Ggraphs (TKGs), where facts are stored along with significant temporal information. However, lots of temporal knowledge graph models are extended from static knowledge graph methods, which usually consider message passing in TKGs either in predicate aspect or entity aspect. Leading to a result that entity embedding or predicate embedding obtained by these methods may lose a lot of information. To fill this gap, we propose a novel framework, MANet, to capture message passing between facts with the aspect of time, predicate and entity. We also note the irregular fact and introduce a residual component to study the irregularity of facts. Extensive experiments on five benchmarks demonstrate that our model outperforms the-state-of-art baselines, and the main indicators have improved obviously, among which MRR and H@1 have improved by four to eight percentage points on GDELT and WIKI datasets.

I. Introduction

Knowledge Graphs (KGs) are used to display the complex relations between entities. As a multi-relational and directed graph, Knowledge Graph is constructed on a specific set of entities and predicates. Entities in KG can be real-world objects or abstract concepts and predicates are the possible relations between two entities. Traditional KGs do not take into account the temporal validity of facts. However, most facts only exist at a specific timestamp or time interval in the real world. Some researchers incorporate time element into KGs and extend static KGs to TKGs. Each fact in TKGs can be represented by a quadruple, (subject, action, object, time node), i.e., (s, p, o, t). Compared with SKGs, TKGs are more informative and closer to the real world, which leads to a wider range of application scenarios.