A Mean Deviation Based Splitting Criterion for Classification Tree

For the learning of Classification Tree, many researchers have used different splitting criteria, in which most commonly impurity-based criteria are: Gini index, Entropy function and Exponent-based index. By comparing Misclassification rates, none of the splitting criterion can be declared as providing best results in every situation. In this study, a new Mean Deviation based index has been proposed and a simulation study is designed, to explore which measure gives best result in what type of situation? From simulation study, it is concluded that generally new proposed M.D-based index and Exponent-based index give excellent results in case of imbalanced data. While, in case of balanced data, Gini index and Entropy function have less misclassification rates. Keywords— "Classification Tree", "Misclassification rate", "Impurity-based criteria", "Simulation", "Imbalanced data"