An evaluation of automated structure learning with Bayesian networks

Abstract

We develop a Bayesian network (BN) model that describes estuarine chlorophyll dynamics in the upper section of the Neuse River Estuary in North Carolina, using automated constraint based structure learning algorithms. We examine the functionality and usefulness of the structure learning algorithms in building model topology with real-time data under different scenarios. Generated BN models are evaluated and a final model is selected. Model results indicate that although the effect of water temperature and river flow on chlorophyll dynamics has remained unchanged following the implementation of the nitrogen Total Maximum Daily Load (TMDL) program; the response of chlorophyll levels to nutrient concentrations has been altered. The results stress the importance of incorporating expert defined constraints and links in conjunction with the automated structure learning algorithms to generate more plausible structures and minimize the sensitivity of the learning algorithms. This hybrid approach towards structure learning allows for the incorporation of existing knowledge while limiting the scope of the learning algorithms to defining the links between environmental variables for which the expert has little or no information.