LorenTzE: Temporal Knowledge Graph Embedding Based on Lorentz Transformation
Lists
Only you see this
Abstract
Inferring missing facts in knowledge graphs(KGs) has been proved to be a significant task due to the incompleteness problem. Temporal knowledge graphs(TKGs) in which each fact is incorporated with a timestamp to model the dynamic properties of facts, suffer from the incompleteness either. Thus temporal knowledge graph completion(TKGC) is introduced to infer missing dynamic facts. However, most of the existing TKGC models are proposed based on existing static KG embedding models and then try to learn embeddings for timestamps and incorporate them into the score functions, which, in several cases, impairs models’ expressivity and interpretabilities. In this paper, we provide a novel physical view for TKGC and propose LorenTzE, an embedding based model inspired by Lorentz transformation [1] in which each embedding is endowed with physical meanings. We also prove that LorenTzE is capable of modeling various relation patterns under temporal settings. Experimental results on three main TKGC benchmarks demonstrate the effectiveness and superiorities to several static and temporal models of our proposed method.
BibTeX
@inproceedings{li2023lorentze,
author = {Ningyuan Li and Haihong E and Li Shi and Xueyuan Lin and Meina Song and Yuhan Li},
title = {{LorenTzE: Temporal Knowledge Graph Embedding Based on Lorentz Transformation}},
booktitle = {ICANN},
year = {2023},
doi = {10.1007/978-3-031-44223-0\_38},
}Loading related content...