A Study of Classification by using Bayesian Networks with Various Background Distributions

Abstract

Model-based decision support systems are preferred due to consistency in decision-making and due to time-efficiency in model evaluation and modification. Constructing a model may take time if a number of random variables are involved in the model and the model structure is not simple. Classification is a form of decision making under a certain loss structure. We consider a decision support system based on a Bayesian network(BN) where all the variables involved are binary, each taking on 0 or 1 and have various bakcround distributions. We explore classification agreement between Bayesian network models by applying the concept of model similarity. We use a beta distribution and its variation at each variable for its conditional probability given its parent variables in a given Bayesian network model. A main result of this work is that we may use a Bayesian network model to make a robust classification when the true Bayesian network model satisfies the positive association condition among the variables involved in the model.