Understanding and improving ontology reasoning efficiency through learning and ranking

Abstract

Ontologies are the fundamental building blocks of the Semantic Web and Linked Data. Reasoning is critical to ensure the logical consistency of ontologies, and to compute inferred knowledge from an ontology. It has been shown both theoretically and empirically that, despite decades of intensive work on optimising ontology reasoning algorithms, performing core reasoning tasks on large and expressive ontologies is time-consuming and resource-intensive. In this paper, we present the meta-reasoning framework R2O2* to tackle the important problems of understanding the source of TBox reasoning hardness and predicting and optimising TBox reasoning efficiency by exploiting machine learning techniques. R2O2* combines state-of-the-art OWL 2 DL reasoners as well as an efficient OWL 2 EL reasoner as components, and predicts the most efficient one by using an ensemble of robust learning algorithms including XGBoost and Random Forests. A comprehensive evaluation on a large and carefully curated ontology corpus shows that R2O2* outperforms all six component reasoners as well as AutoFolio, a robust and strong algorithm selection system.