Decision Trees for Prediction and Data Mining
Abstract
Three tree-structured algorithms are developed: GUIDE, CRUISE, and LOTUS. GUIDE is an algorithm for constructing least-squares, quantile, and Poisson regression trees. It can fit piecewise constant, piecewise polynomial, and piecewise multiple linear models -- with and without stepwise variable selection. CRUISE is a multiple-branching classification tree algorithm sensitive to pairwise variable interactions at each node. It can fit a single class label or fit the best two-variable linear discriminant model to each leaf node. LOTUS is a logistic regression tree algorithm. It can fit either simple linear or multiple linear logistic regression models at the leaf nodes. All three algorithms are unique in being practically free of variable selection bias at the splits. Empirical comparisons with a variety of real datasets show that the algorithms possess both high prediction accuracy and fast computational speed. Computer software for the algorithms are available for free download from the internet for the Windows, Linux, and (in the case of GUIDE) Macintosh platforms.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 10, 2005
- Accession Number
- ADA430178
Entities
People
- Yin-loh Wei
Organizations
- University of Wisconsin–Madison