Joint framework for tensor decomposition-based temporal knowledge graph completion

Abstract

Knowledge graphs (KGs) usually contain a lot of missing information, static knowledge graph completion (KGC) is widely used to solve their incompleteness. In recently, since the incompleteness of temporal KGs which typically contain lots of temporal facts, like (Barack Obama, Is-president-of, USA, 2008), temporal knowledge graph completion (TKGC) is proposed to predict the missing part of (?, Is-president-of, USA, 2008) or (Barack Obama, Is-president-of, ?, 2008). In particular, tensor decomposition method has shown excellent performance in KGC and therefore many existing methods extend it to TKGC. One of the key challenges is how to make full use of static information and effectively fuse static (non-temporal) and temporal information for improving the performance of TKGC. The existing models usually concatenate or sum the static and temporal embeddings directly. Moreover, we observe that neighborhood information of temporal facts may be very useful for inferring the missing parts of temporal facts while has not been fully explored and utilized.To address these challenges, we propose a joint framework composed of temporal and static modules for tensor decomposition-based TKGC. In temporal module, we propose a new neighborhood time sharing technology for aggregating richer temporal semantic information. The static module is proposed as an auxiliary model to further learn embedding representation of static information based on our proposed two training strategies (named typed entity strategy and adjacent active entity strategy). Finally, we propose a novel way of fusing static and temporal information through joint learning of two modules based on our proposed entity sharing technology. A series of comparison and ablation experiments show that our model can make better and full use of static information while capturing richer temporal semantic information, coupled with an effective fusion method, thus achieving state-of-the-art (SOTA) performance compared to existing tensor decomposition-based TKGC methods on four benchmark datasets.