Algorithmic speedups in growing classification trees by using an additive split criterion

Abstract

We propose a new split criterion to be used in building classification trees. This criterion called weighted accuracy or wacc has the advantage that it allows the use of divide-and-conquer algorithms when minimizing the split criterion. This is useful when more complex split families, such as intervals corners and rectangles, are considered. The split criterion is derived to imitate the Gini function as closely as possible by comparing preference regions for the two functions. The wacc function is evaluated in a large empirical comparison and is found to be competitive with the traditionally used functions.