Title: Stability Assessment of Tree Ensembles and Psychotrees

Authors: Lennart Schneider, Achim Zeileis, Carolin Strobl

Affiliation: Ludwig Maximilian University of Munich

Abstract:
While individual trees (classification and regression trees as well as
model-based trees) are easy to interpret, their results can be instable in the
sense that small changes in the data can lead to very different looking trees.
The R package "stablelearner" allows for assessing the stability of individual
trees by providing a toolbox of plots and descriptive statistics. In this
talk, we demonstrate how "stablelearner" can now also be used to visually
assess the stability of tree ensembles (bagging and random forests). These
tree ensembles can either be fit in "stablelearner" directly relying on base
learners of the "partykit" package or externally relying on the R packages
"party", "partykit" or "randomForest". Finally, we also demonstrate how to
assess the stability of model-based trees of item response models, e.g.,
raschtrees fitted via the "psychotree" package, which can be added to
"stablelearner"'s LearnerList.