Each Snapshot to Each Space: Space Adaptation for Temporal Knowledge Graph Completion

Abstract

Temporal knowledge graphs (TKGs) organize and manage the dynamic relations between entities over time. Inferring missing knowledge in TKGs, known as temporal knowledge graph completion (TKGC), has become an important research topic. Previous models handle all facts with different timestamps in an identical latent space, even though the semantic space of the TKG changes over time. Therefore, they are not effective to reflect the temporality of knowledge. To effectively learn the time-aware information of TKGs, different latent spaces are adapted for temporal snapshots at different timestamps, which yields a novel model, i.e., Space Adaptation Network (SANe). Specifically, we extend convolutional neural networks (CNN) to map the facts with different timestamps into different latent spaces, which can effectively reflect the dynamic variation of knowledge. Meanwhile, a time-aware parameter generator is designed to explore the overlap of latent spaces, which endows CNN with specific parameters in term of the context of timestamps. Therefore, knowledge in adjacent time intervals is efficiently shared to boost the performance of TKGC, which can learn the validity of knowledge over a period of time. Extensive experiments demonstrate that SANe achieves state-of-the-art performance on four well-established benchmark datasets for temporal knowledge graph completion.