A novel attributes partition method for decision tree

Abstract

In the decision tree's making phase, it is frequent to find the optimal partition of elements with different values of a category attribute at a node. This needs to search over all the partitions for the one with the minimal impurity, which is exponential in n. We present a new heuristic search algorithm, SORT_DP, to find an effective partition, which is polynomial in n. The method uses the mapping from the class probability space to the sub-spaces and the technique of dynamic programming. By comparing the performance against other methods through experiments, we demonstrated the effectiveness of the new method. © Springer-Verlag Berlin Heidelberg 2013.