Compact Ensemble Trees for Imbalanced Data

This paper introduces a novel splitting criterion parametrized by a scalar ‘α’ to build a class-imbalance resistant ensemble of decision trees. The proposed splitting criterion generalizes information gain in C4.5, and its extended form encompasses Gini(CART) and DKM splitting criteria as well. Each decision tree in the ensemble is based on a different splitting criterion enforced by a distinct α. The resultant ensemble, when compared with other ensemble methods, exhibits improved performance over a variety of imbalanced datasets even with small numbers of trees.