Extended Bayesian Regression Models: A Symbiotic Application of Belief Networks and Multilayer Perceptrons for the Classification of Ovarian Tumors

Abstract

We describe a methodology based on a dual Belief Network-Multilayer Perceptron representation to build Bayesian classifiers. This methodology combines efficiently the prior domain knowledge and statistical data. We overview how this Bayesian methodology is able (1) to define constructively a valuable “informative” prior for black-box models, (2) to provide uncertainty information with predictions and (3) to handle missing values based on an auxiliary domain model. We assume that the prior domain model is formalized as a Belief Network (since this representation is a practical solution to acquiring prior domain knowledge) while we use black-box models (such as Multilayer Perceptrons) for learning to utilize the statistical data. In a medical task of predicting the malignancy of ovarian masses we demonstrate these two symbiotic applications of Belief Network models and summarize the practical advantages of the Bayesian approach.