Extending the Bnclassify R package: Bayesian network classifiers with continuous variables

Abstract

Nowadays, most of the open-source software for Bayesian networks only handles discrete variables. This limitation makes it difficult to use Bayesian networks in real life. The original bnclassify package provides the code that deals with the Bayes multinomial network (BMN). This kind of Bayesian network is not capable of handling continuous variables, if a data set contains continuous variables, it will be discretized, therefore, a loss of information will be introduced. The goal of this work is to extend the package and make it able to deal with continuous variables. To avoid the loss, this work assumes all variables follow the Gaussian distribution, therefore, new functionalities are added to the package including structural learning, parameter learning, and prediction with conditional Gaussian network (CGN). In order to determine the correctness and the efficiency of the new implementations called predictCGNs and GaussianImplement, two experiments are designed in this work. The experimental results demonstrate the usability of the program and there is no inconsistency between the new implementations and Bayesian networks.