k-DT: A Multi-Tree Learning Method

랜덤 포레스트의 시초.

Abstract

This paper introduces a technique for using the randomized nature of some learning algorithms to increase their accuracy. Our method is to generate multiple classifiers and combine them with a majority voting scheme. The purpose of this technique is to overcome small errors that appear in individual classifiers. We have tested our idea on a type of randomized decision tree with real data, and found that it consistently improves the accuracy over that of average trees. We have also shown that this technique outperforms some other methods that attempt to improve accuracy by using randomization in a different way.