Implementation and Performance Optimization of Dynamic Random Forest

Abstract

Bernard combines the weight updating of the boosting algorithm with the Random Forest (RF), and proposes a new RF induction algorithm called Dynamic Random Forest (DRF). The idea with DRF is to grow only trees that would fit the sub-forest already built, use the existing forest to update the weight of each randomly selected training instance, force the next tree to pay attention to those samples that can not be classified well by the current forest, thus improving the RF accuracy. However, this weight updating method is still flawed, which does not make a good distinction between the samples classified correctly and the samples classified wrongly by the current forest. In this paper, we implement the DRF algorithm, and propose a new weight update method, that is, giving higher weight to the samples classified wrongly by the current forest, giving lower weight to the samples classified correctly by the current forest, so that the next tree will be more concerned with those misclassified samples. Experimental results show that our method is better than DRF algorithm and traditional RF algorithm.