Learning Bayesian network classifiers from data with missing values

Learning accurate Bayesian network (BN) classifiers from complete databases is a very active research topic in data mining and machine learning. However, in practice, databases are rarely complete. This affects their real world data mining applications. This paper investigates the methods for learning four types well-known Bayesian network classifiers from incomplete databases. These four types BN classifiers are: Naive-Bayes, tree augmented Naive-Bayes, BN augmented Naive-Bayes, and general BN, where the latter two are learned using dependency analysis based algorithms that work only on the database completeness assumption. In order to enable this kind of algorithms to handle with missing data, this paper introduces a novel deterministic method to estimate the (conditional) mutual information from incomplete databases, which can be used to do CI tests, a fundamental step in the dependency analysis based algorithms. The experimental results show that our algorithm is efficient and reliable.