Edge selection by CART for bayes networks

Abstract

It is common to handle large data containing more than 100 variables in the real world like biological analysis or social statistical analysis. However, handling 100 variables at once is often impossible without the super-computer. It takes very long time to compute and build its model at once although using the super-computer. One of the good solutions is a tree-structured approach. This approach can analyze regardless of dimensionality and the sample size. Specially, the CART(classification and regression trees), statistical tree regression algorithm, gives users many options for users as well as useful results in detail. We have to pay attention to the variable importance in many cases, which the CART offers, because it reflects the contribution each variable makes in classifying or predicting the target variable. When the true model is given in the form of a Bayes network, it is most desirable that the grouping is made in such a way that every variable in a subset has at least one other variable which is connected by an arrow in the Bayes network. Additionally, a data set of 50,000 cases from a Bayes network model of 100 binary variables and found the variable importance is a useful index of grouping the random variables.