Lifted Query Answering in Gaussian Bayesian Networks
Proceedings of the 10th International Conference on Probabilistic Graphical Models, PMLR 138:233-244, 2020.
Gaussian Bayesian networks are widely used for modeling behaviors of continuous random variables. Lifting exploits symmetries when dealing with large numbers of isomorphic random variables to support more compact representations and more efficient query answering. This paper presents a lifted construction and representation of a joint distribution derived from a Gaussian Bayesian network and a lifted query answering algorithm on the lifted joint distribution. To lift the query answering, needed algebraic operations that work fully in the lifted space are developed. A theoretical complexity analysis and experimental results show that both the lifted joint construction and the lifted query answering significantly outperform their grounded counterparts.