Data enrichment of spatial databases using ontologies and Bayesian networks

Abstract

In the map generalization process the supplied data often lacks the explicit information for a proper automated approach. This is a problem that is very apparent when we want to generalize to a social construct, a description of something made by a society, both for crisp entities like "a detached house" and vernacular geographies such as "a suburban area" or "the high street". In this thesis we will explore a new way of data enrichment in spatial datasets for the use of such generalization. We will model the entity that we want to enrich the data with as an ontology using OWL, try to exploit the hierarchical nature of these entities for modeling and find these entities with the use of Bayesian networks that are generated from the ontology. We have created a Protege plug-in called BNGen as tool to convert ontologies to Bayesian networks and a code blueprint for the enrichment process framework. We will show that this approach works with an illustrative use case where we will enrich a dataset with the leafy residential area concept. While the use case is successful in showing that our approach works, it will also be shown that OWL is not good at modelling vague relationships where a relation might hold or only partially holds. To counter some of these problems, and the fact that Bayesian networks are not dynamic in structure, we introduce summary nodes in the Bayesian network and staged classification. It will also be shown that we can exploit the enriched data to deal with the vagueness of spatial concepts as social constructs in our approach.