Construction and applications in data mining of Bayesian networks

Bayesian network approaches have become an important research direction in data mining. This paper discusses the structure and the construction of Bayesian networks, emphasizing the basic methods for learning the structure and probabilities of Bayesian networks from prior knowledge and sample data. Compared with other approaches used for data mining, Bayesian networks can combine prior knowledge with observed data, which is very important when data is scarce or very expensive. Moreover, Bayesian networks can discover causal relationships among data and handle incomplete data sets, which other methods can not do. The disadvantages of Bayesian networks are the high computational cost, the difficulties in determining appropriate parameters and structures, and the lack of principles to justify if the hypotheses required by the Bayesian network are actually satisfied by the problems.