Tree-Structured Classification via Generalized Discriminant Analysis

Abstract

Linear techniques are used recursively to construct classification rules which can be represented as k-nary decision trees. The method has been implemented in a computer program called FACT. It can handle ordered and unordered variables, unequal priors, variable misclassification costs, and missing observations. Besides the tree structure, it also yields an importance ranking of the variables and a cross-validation estimate of error. FACT is compared with CART (a procedure proposed recently by Breiman et al., which gives a binary tree) in a series of examples. The conclusion is that FACT and CART are usually comparable in terms of classification accuracy and interpretative capability, but FACT runs many times faster. (Author)