A Review of Parameter Learning Methods in Bayesian Network

Abstract

Bayesian network (BN) is one of the most classical probabilistic graphical models. It has been widely used in many areas, such as artificial intelligence, pattern recognition, and image processing. Parameter learning in Bayesian network is a very important topic. In this study, six typical parameter learning algorithms were investigated. For the completeness of dataset, there are mainly two categories of methods for parameter estimation in BN: one is suitable to deal with the complete data, and another is for incomplete data. We mainly focused on two algorithms in the first category: maximum likelihood estimate, and Bayesian method; Expectation - Maximization algorithm, Robust Bayesian estimate, Monte - Carlo method, and Gaussian approximation method were discussed for the second category. In the experiment, all these algorithms were applied on a classic example to implement the inference of parameters. The simulating results reveal the inherent differences of these six methods and the effects of the inferred parameters of network on further probability calculation. This study provides insight into the parameter inference strategies of Bayesian network and their applications in different kinds of situations.