Uncertainty Quantification of Knowledge Graph Embedding with Statistical Guarantees

Yuqicheng Zhu, Nico Potyka, Daniel Hernández, Jiarong Pan, Bo Xiong, Yunjie He, Yuan He, Zifeng Ding, Evgeny Kharlamov, Steffen Staab

Keywords: Knowledge Graphs, Uncertainty Quantification

Abstract: Knowledge Graph Embedding (KGE) methods have been widely used in downstream tasks such as link prediction and question answering. However, a critical limitation remains: their predictions lack reliable uncertainty estimates, which poses significant risks in high-stakes applications such as medical decision support. In this extended abstract, we summarize our recent work accepted at NAACL and ACL, which introduces principled methods for uncertainty quantification in KGE based on conformal prediction. Our approach provides statistical coverage guarantees by constructing answer sets that include the true answer with a user-specified confidence level. Intuitively, we quantify uncertainty by determining how many candidate answers are needed to ensure the true answer is covered with desired confidence level.

Track: Knowledge Graphs, Ontologies and Neurosymbolic AI

Paper Type: Extended Abstract

Resubmission: No

Submission Number: 5