Combining expert knowledge with data based on belief function theory: an application in waste water treatment

This paper presents a methodology for combining expert knowledge with information from statistical data, in classification and prediction problems. The method is based on (1) a case-based approach allowing to predict a quantity of interest from past cases in the form of a belief function, (2) Bayesian networks for modelling expert knowledge and (3) a tuning mechanism allowing to optimally discount information sources by optimizing a performance criterion. This methodology is applied to the prediction of chemical oxygen demand solubility in wastewater. The approach is expected to be useful in situations where both small databases and partial expert knowledge are available.