Latent-class statistical relational learning with uncertain formal ontologies

Ontologies are a well researched formalism in computer science to represent knowledge in a machine processable manner. Recently, with the growing semantic web and social networks, ontologies with large amounts of data are becoming available. While such knowledge representations are intended for data integration and deduction of implicit knowledge, they are in general not designed to induce uncertain knowledge. This thesis focuses on the statistical analysis of known as well as deducible facts in ontologies and the induction of uncertain knowledge from both. Hereby, the uncertainty of induced knowledge originates not only from a lack of information but also from potentially untrustworthy sources providing the facts. We outline common ontology languages, their (formal) semantics and expressivity from a perspective of existing real world data sources. Then, existing learning methods are discussed that could be used for automated statistical analysis of facts given in such ontologies and the induction of facts that are not explicitly given in the ontology. Hereby, two fundamental approaches to artificial intelligence need to be combined: The deductive reasoning by logical consequence and the inductive learning by statistical generalization. The main contribution of this work is a machine learning approach that enforces hard constraints during the learning phase. In short, this is achieved by checking description logic satisfiability of statements inductively inferred by an infinite latent-class multi-relational Bayesian learning method based on Dirichlet process mixture models. This guarantees the compliance of induced facts with deductive arguments and can lead to an improved cluster analysis and predictive performance. To demonstrate the effectiveness of our approach we provide experiments using real world social network data in form of an OWL DL ontology. In addition, the learning from untrustworthy facts in formal knowledge representations is discussed. We show how to model and learn context-sensitive trust using our learning method. The efficiency and performance of our approach is evaluated empirically e. g. , with user-ratings gathered from online auctions. In summary, this thesis contributes to the combination of inductive and deductive reasoning approaches and more specifically to latent-class statistical learning with description logic ontologies containing information from potentially untrustworthy sources.