Tensor Decompositions For Temporal Knowledge Graph Completion with Time Perspective

Jinfa Yang, Xianghua Ying, Yongjie Shi, Ruibin Wang, Taiyan Chen, Bowei Xing

Keywords: Knowledge Graph, Temporal Knowledge Graph Completion

Abstract: Facts in the real world are often tied to time, such as the spread of diseases, and the state of military affairs. Therefore, knowledge graphs combined with temporal factors have gained growing attention. In the temporal knowledge graph, most researchers focus on the original facts and pay attention to their changes over time. The temporal factors are only used as auxiliary information for representation learning. In this paper, we try to observe from the perspective of time and find some interesting properties of temporal knowledge graph: (1) Simultaneousness. Various facts occur at the same time; (2) Aggregation. The facts may aggregately occur for a certain individual, organization, or location; (3) Associativity. Some specific relations tend to occur at specific times, such as celebrations at festivals. Based on the above three properties, we add a simple time-aware module to the existing tensor decomposition-based temporal knowledge graph model TComplEx Lacroix et al. (2020), which obtains impressive improvements and achieves state-of-the-art results on four standard temporal knowledge graph completion benchmarks. Specifically, in terms of mean reciprocal rank (MRR), we advance the state-of-the-art by +21.8% on ICEWS14, +16.9% on ICEWS05-15, +20.7% on YAGO15k, and 13.1% on GDELT.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning

TL;DR: Instead of focusing on facts and their evolution, we observe temporal knowledge graphs through time perspective, and improve the current tensor decomposition model based on the observed properties.

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/tensor-decompositions-for-temporal-knowledge/code)