Binary Trees for Dependence Structure

Abstract

In a data set with many categorical variables and several continuous valuables, the relationship between continuous random variables may differ from category to category for a given categorical variable. To study how categorical variables may affect the dependent structure of continuous variables, we proposed two splitting criteria constructed based on copula entropy to build decision trees serving for different purposes. One type of tree can be used to identify the attributes or combinations of them under which the continuous variables have a strong relationship. The other type of tree is used to classify regions with different strength of relationship. Applying these methods to the survey data on the status of poor families of Sichuan province, it is found that the method successfully evaluated the effectiveness of the poverty alleviation policies.