Knowledge-based Bayesian networks for discriminant analysis

This thesis examines the utility of an expert's prior knowledge in developing a Bayesian network (BN) to perform discriminant analysis. The expert's prior knowledge will take the form of a graphical description of the relationships between variables in the BN. Expert knowledge will also be used to discretize variables in ways that are meaningful to a decision maker. Lastly, expert knowledge will provide the information necessary to gain a thorough understanding of the domain. We attempt to demonstrate the benefits of building a BN to gain important insights, using probabilistic inference, which are not available using other techniques. We will do this with three different data sets and will discuss circumstances where the application of BNs will be beneficial.; There are a number of discriminant analysis techniques that exist today. Expert knowledge is not typically emphasized in the implementation of existing discriminant analysis methods or software packages. An expert's prior knowledge is also not emphasized in the analysis of data sets.; A graphical representation of the domain will lead to an intuitive understanding of the data set and the insights gleaned from the data set. This thesis will contribute to the body of knowledge in discriminant analysis by investigating the utility of using expert knowledge to develop a BN. We will build the BN and then compare its performance with other methods.