[Rd] [SoC09-Idea] Party On!
Torsten.Hothorn at stat.uni-muenchen.de
Fri Feb 27 11:45:05 CET 2009
find our SoC proposal below.
Torsten & Achim
Party On! New Recursive Partytioning Tools.
Mentor: Torsten Hothorn & Achim Zeileis
The aim of the project is the implementation of recursive partitioning
methods ("trees") which aren't available in R at the moment. The student
can choose a method to begin with from a larger set of interesting algorithms.
Recursive partitioning methods, or simply "trees", are simple yet powerful
methods for capturing regression relationships. Since the publication of the
automated interaction detection (AID) algorithm in 1964, many extensions,
modifications, and new approaches have been suggested in both the statistics
and machine learning communities. Most of the standard algorithms are
available to the R user, e.g., through packages rpart, party, mvpart, and
However, no common infrastructure is available for representing trees
fitted by different packages. Consequently, the capabilities for extraction
of information - such as predictions, printed summaries, or
visualizations - vary between packages and come with somewhat different user
interfaces. Furthermore, extensions or modifications often require considerable
programming effort, e.g., if the median instead of the mean of a numerical
response should be predicted in each leaf of an rpart tree.
Similarly, implementations of new tree algorithms might also require new
infrastructure if they have features not available in the above-mentioned
packages, e.g., multi-way splits or more complex models in the leafs.
To overcome these difficulties, the partykit package has been started on
R-Forge. It is still being developed but already contains a stable class
"party" for representing trees. It is a very flexible class with unified
predict(), print(), and plot() methods, and can, in principle, capture
all trees mentioned. But going beyond that, it can also accommodate
multi-way or functional splits, as well as complex models in (leaf) nodes.
We aim at making more recursive partitioning methods available
to the R community. A first step in this direction is the CHAID package
(also hosted on R-Forge). Much more prominent procedures come to mind,
for example exhaustive CHAID, C4.5, GUIDE, CRUISE, LOTUS, and many others.
Students can choose among these and other recursive partitioning methods
they want to implement based on the partykit infrastructure.
Good R programming skills, depending on the complexity of the chosen
algorithm C programming might be required as well. A basic understanding
of statistics and machine learning would be helpful.
Consider the "GlaucomaM" dataset from package ipred. Write a small R function
that searches for the best binary split in variable "vari" when "Class" is
the response variable. Implement any method you like but without using any
More information about the R-devel