TKGFrame: A Two-Phase Framework for Temporal-Aware Knowledge Graph Completion

Jiasheng Zhang, Yongpan Sheng, Zheng Wang, Jie Shao

Abstract: In this paper, we focus on temporal-aware knowledge graph (TKG) completion, which aims to automatically predict missing links in a TKG by making inferences from the existing temporal facts and the temporal information among the facts. Existing methods conducted on this task mainly focus on modeling temporal ordering of relations contained in the temporal facts to learn the low-dimensional vector space of TKG. However, these models either ignore the evolving strength of temporal ordering relations in the structure of relational chain, or discard more consideration to the revision of candidate prediction results produced by the TKG embeddings. To address these two limitations, we propose a novel two-phase framework called TKGFrame to boost the final performance of the task. Specifically, TKGFrame employs two major models. The first one is a relation evolving enhanced model to enhance evolving strength representations of pairwise relations pertaining to the same relational chain, resulting in more accurate TKG embeddings. The second one is a refinement model to revise the candidate predictions from the embeddings and further improve the performance of predicting missing temporal facts via solving a constrained optimization problem. Experiments conducted on three popular datasets for entity prediction and relation prediction demonstrate that TKGFrame achieves more accurate prediction results as compared to several state-of-the-art baselines.