Investigation of bayesian networks for classification problems involving binary data

Abstract

In this dissertation, the main focus is on Bayesian Networks. In particular, we are examining the effectiveness of Na??ve Bayes, Tree Augmented Na??ve Bayes and Bayesian Networks in terms of their performance on the eventual prediction on test set data. This work looks at the various complexities of the three Networks. The criterion for the type of data used is that they should model those in the real world. Hence, two characteristics are determined to be representative of real world data a?? limited training examples and data corrupted with noise. To assist understanding of the results, binary data is used. It is found that under certain conditions, a simple algorithm like Na??ve Bayes performs better even though the underlying network structure is more complex. A real world dataset is used to verify the results. Finally, a systematic method is developed to tackle real world binary data.