Sampling methods in decision trees

CART is widely used by researchers in data mining applications. However, for a very large data set building CART is nearly an impossible task. We argue that a tree built entirely from a large simple random sample from the data set is reasonable. In addition, we propose two new algorithms; the parabola method that relies on the property that the Gini-index is a smooth function in any continuous attribute thus allowing accurate approximation of the minimum Gini-index and the double sampling method useful when a data set is very large. Experimental results show that these two method perform extremely well.