ABSTRACT
This chapter introduces a powerful, yet seemingly lesser known, decision tree algorithm for generalized, unbiased, interaction detection, and estimation (GUIDE), which was introduced in Loh. Unfortunately, GUIDE seems much less adopted among practitioners and researchers when compared to other algorithms with easy-to-use open source implementations, like CART. Like CTree, GUIDE is based on statistical tests of hypothesis striving to achieve unbiased split variable selection. Exhaustive search procedures, like CART, can be insensitive to local interactions; according to Loh, splits that are sensitive to two-way interaction effects can produce shorter trees. Similar to GUIDE for regression, the people can increase flexibility by fitting non-constant models in the nodes of the classification tree. It introduces the GUIDE algorithm for building classification and regression trees. In selecting the splitting variable, GUIDE also looks at two-way interaction effects that can potentially mask the importance of a split when only main effects are considered.
