Ontology Modeling for Probabilistic Knowledge Graphs

Abstract

Reasoning about data given uncertain future conditions is a crucial component of modern decision support systems. While ontologies allow domain-specific data to be represented in knowledge graphs, they generally represent data in an absolute way and do not support reasoning under uncertain conditions. In this paper, we present a novel technique for using a probabilistic ontology to extend knowledge graphs as Bayesian networks, which is a step towards enabling probabilistic reasoning under uncertainty within a knowledge graph. Although prior work has focused on using the semantic relationships in an ontology to automatically infer the structure of a Bayesian network, we additionally emphasize the ontological modeling of the Bayesian network as a layer in the knowledge graph itself. We define a custom probabilistic ontology that describes the requisite probabilistic elements, including Random Variables, the conditional dependencies between them, and their distributions. It also includes graph structures for representing decision optimization under uncertainty. Our technique is generalized to work regardless of the domain of the data. In the evaluation, we demonstrate that our approach can scale over large data volumes, and that integrating the probabilistic elements with the domain ontology allows for flexible queries that can answer competency questions more efficiently than queries over the domain ontology alone.