Alina Tkachenko

Topic

Tracking Uncertainty in Knowledge Graphs Using Kalman Filtering

Department of Computer Science

Date & location

• Thursday, December 18, 2025

• 1:30 P.M.

• Virtual Defence

Reviewers

Supervisory Committee

• Dr. Alex Thomo, Department of Computer Science, University of Victoria (Supervisor)

• Dr. Kevin Stanley, Department of Computer Science, UVic (Member) 

External Examiner

• Dr. Xiaodai Dong, Department of Electrical and Computer Engineering, University of Victoria 

Chair of Oral Examination

• Prof. Freya Kodar, Faculty of Law, UVic 

Abstract

Knowledge graphs (KGs) represent structured knowledge as networks of entities and relations, forming a foundation for reasoning in artificial intelligence. To make these symbolic structures usable by machine learning systems, knowledge graph em beddings (KGEs) map entities and relations into continuous vector spaces. However, traditional KGE models are typically static and deterministic. They treat all facts as equally certain and require full retraining when new data arrive, making them unsuitable for evolving, uncertain knowledge.

This thesis introduces a new framework that reframes knowledge graph embedding as an online state estimation problem. By integrating the Kalman filter, a recursive algorithm that updates beliefs under uncertainty, into KGE training, the proposed approach enables continuous and uncertainty aware learning of entity and relation representations. The framework treats each embedding as a probabilistic latent state, updated incrementally as new triples arrive, blending prior knowledge with new, possibly noisy, observations.

Two models instantiate this framework. KalmanKG2E extends the probabilistic Gaussian embedding model KG2E with Kalman-based online updates of means and covariances. KalmanComplEx adapts the non-probabilistic, complex-valued ComplEx model to a dynamic, uncertainty-tracking setting. Together, these demonstrate the frameworks generality across fundamentally different embedding architectures.

Extensive experiments on six benchmark datasets show consistent improvements over static baselines. The Kalman-based models converge faster, achieve higher predictive accuracy, and exhibit greater robustness in sparse, evolving graphs. These results validate Kalman filtering as a principled and efficient mechanism for online knowledge graph learning. Overall, this work bridges classical state estimation and modern representation learning, advancing knowledge graph embeddings from static snapshots to dynamic, continuously adaptive systems that better reflect the evolving nature of real-world knowledge.