Multi Branch Decision Tree: A New Splitting Criterion

In this paper, a new splitting criterion to build a decision tree is proposed. Splitting criterion specifies the best splitting variable and its threshold for further splitting in a tree. Giving the idea from classical Forward Selection method and its enhanced versions, the variable having the largest absolute correlation with the target value is chosen as the best splitting variable in each node. Then, the idea of maximizing the margin between classes in SVM is used to find the best threshold on the selected variable to classify the data. This procedure will execute recursively in each node, until reaching the leaf nodes. The final decision tree has a comparable shorter height than the previous methods, which effectively reduces more useless variables and the time of classification for future data. Unclassified regions are also generated, which can be interpreted as an advantage or disadvantage for the proposed method. Simulation results demonstrate this improvement in the proposed decision tree. Keyword: Decision Tree, Splitting Criterion, Support Vector Machine, Correlation, Unclassified Region