VTG: Learning historical information with variable time granularity
Download as PDF
DOI: 10.23977/ICAMCS2023.005
Corresponding Author
Hao Xiong
ABSTRACT
Temporal knowledge graph inference is pivotal for under- standing knowledge graph evolution. However, current approaches often underutilize temporal information. Moreover, there exists room for improvement in the critical task of link prediction within knowledge graph inference. This study introduces an innovative inference architecture that employs multi-domain attention and dynamic time strategies, enabling adaptive selection of historical data and temporal spans to enhance the timeliness and accuracy of temporal knowledge graph inference. Empiri- cal results demonstrate substantial performance enhancements in knowl- edge graph state prediction. Furthermore, the model excels in the task of link prediction across multiple temporal knowledge graph datasets. Our approach showcases substantial potential across diverse domains, including social networks, scientific literature, and intelligent recommendation systems, contributing to the advancement of knowledge graph research.
KEYWORDS
Temporal knowledge graphs, Variable time granularity, Hierarchical weighted aggregation, Link prediction, Embedding model