Ontology Design for Scientific Theories That Make Probabilistic Predictions
by David Poole, Clinton Smyth, and Rita Sharma
Scientific theories that make predictions about observable quantities can be evaluated by their fit to existing data and can be used for predictions on new cases. The authors' goal is to publish such theories along with observational data and the ontologies needed to enable the interoperation of the theories and the data. This article is about designing ontologies that take into account the defining properties of classes. The authors show how a multidimensional design paradigm based on Aristotelian definitions is natural, can easily be represented in OWL, and can provide random variables that provide structure for theories that make probabilistic predictions. They also show how such ontologies can be the basis for representing observational data and probabilistic theories in their primary application domain of geology.
Imagine having a number of expert systems that provide predictions—for example, diagnoses of what is wrong with patients based on their symptoms, or predictions of whether there will be a landslide at some particular location. Which of these predictions should we believe most? Apparently, many of Google’s queries are people typing in symptoms and wanting diagnoses. Google’s ranking system, based on page rank, essentially measures popularity. Other recommender systems base their predictions explicitly on some measure of how authoritative sources are. Scientists (and the rest of us) should be suspicious of both answers. We would prefer the prediction that best fits the available evidence. To this end, semantic science can provide a way to have explicit theories that make predictions together with the data upon which to test the predictions.
Read more ...
Share this article
