Description

In this project, we study the case of Bayesian Network statistical classifiers, which utilize the properties of traditional Bayesian Networks aiming to classify random variable observations to a specific class, by implementing algorithms of structure learning and parameter estimation. More specifically, we implement two well known Bayes based classifiers, the Naive Bayes and the Tree Augmented Naive Bayes classifier. Both classifiers make some independence assumptions between the attributes, which in both cases are treated as random variables.

The algorithms above where tested in the networks that are shown below. The first on is the widely used alarm network and the second one a medical network that can be used to classify patients. We firstly created two artificial datasets using prior sampling on both the networks and then - based on these data- we learnt the parameters and the structure of two new networks, whose performance was then examined.