Comparison of Classification Methods

Abstract

Classification methods are useful tools for data mining. Results of some comparisons among them are given in this talk. First, split selection methods for classification trees are compared. It is demonstrated that the usual exhaustive search method has selection bias toward variable with more split points or missing values. Statistical methods are given to correct the bias. Second, some results from an extensive study on classification methods which cover from classical and modern statistical methods to classification trees and neural networks are presented.