OWL-Miner: Concept Induction in OWL Knowledge Bases

Abstract

The Resource Description Framework (RDF) and Web Ontology Language (OWL) have been widely used in recent years, and automated methods for the analysis of data and knowledge directly within these formalisms are of current interest. Concept induction is a technique for discovering descriptions of data, such as inducing OWL class expressions to describe RDF data. These class expressions capture patterns in the data which can be used to characterise interesting clusters or to act as classifica- tion rules over unseen data. The semantics of OWL is underpinned by Description Logics (DLs), a family of expressive and decidable fragments of first-order logic. Recently, methods of concept induction which are well studied in the field of Inductive Logic Programming have been applied to the related formalism of DLs. These methods have been developed for a number of purposes including unsuper- vised clustering and supervised classification. Refinement-based search is a concept induction technique which structures the search space of DL concept/OWL class expressions and progressively generalises or specialises candidate concepts to cover example data as guided by quality criteria such as accuracy. However, the current state-of-the-art in this area is limited in that such methods: were not primarily de- signed to scale over large RDF/OWL knowledge bases; do not support class lan- guages as expressive as OWL2-DL; or, are limited to one purpose, such as learning OWL classes for integration into ontologies. Our work addresses these limitations by increasing the efficiency of these learning methods whilst permitting a concept language up to the expressivity of OWL2-DL classes. We describe methods which support both classification (predictive induction) and subgroup discovery (descrip- tive induction), which, in this context, are fundamentally related. We have implemented our methods as the system called OWL-Miner and show by evaluation that our methods outperform state-of-the-art systems for DL learning in both the quality of solutions found and the speed in which they are computed. Furthermore, we achieve the best ever ten-fold cross validation accuracy results on the long-standing benchmark problem of carcinogenesis. Finally, we present a case study on ongoing work in the application of OWL-Miner to a real-world problem directed at improving the efficiency of biological macromolecular crystallisation.