Title: Stability Assessment for Trees and other Supervised Statistical Learning
Results
Authors: Michel Philipp, Carolin Strobl, Thomas Rusch, Kurt Hornik, Achim Zeileis
Affiliation: UZH Zuerich
Abstract:
Classification trees are an example of a statistical learning algorithm that is
known to be "unstable", because small changes in the learning data can lead to
substantially different trees. Ensemble methods, like random forests, are more
stable but lack the interpretability of a single tree. Therefore, from a user's
perspective, the question is: When is it OK to interpret a single tree and when
should it be considered with caution? In a first attempt to address this
question, in this talk we illustrate a toolbox of summary statistics and plots
for assessing stability, that will be available in the R package stablelearner.
Furthermore, we will outline how the ideas can be generalized to a framework for
measuring the stability of supervised statistical learning results in general.