Combining classification trees using MLE.

Abstract

We propose a probability distribution for an equivalence class of classification trees (that is, those that ignore the value of the cutpoints but retain tree structure). This distribution is parameterized by a central tree structure representing the true model, and a precision or concentration coefficient representing the variability around the central tree. We use this distribution to model an observed set of classification trees exhibiting variability in tree structure. We propose the maximum likelihood estimate of the central tree as the best tree to represent the set. This MLE retains the interpretability of a single tree model and has excellent generalizability. We implement an ascent search for the MLE tree structure using a data set of 13 classification trees that predict the presence or absence of cancer based on immune system parameters.