The mlr Package: Machine Learning in R

Written by: Bernd Bischl, Max Wornowizki, Katharina Borg

Very much work in progess, planning to publish a first version soon.

Motivation: There is no unifying mechanism for the many classification algorithms in R. Non-trivial experiments force you to write lengthy, tedious and error-prone code. Although there are methods / packages which try to improve upon this, they often lack in other areas like extensibility.

Features:

• Clear S4 interface to R classification methods
• Easy extension mechanism through S4 inheritance for new methods
• Resampling strategies like boostrapping, cross-validation and subsampling ("paired" - same splits for different algorithms - as the experiment setup is S4 object itself and can be reused)
• Tuning of hyperparamters
• Benchmarking of different algorithms (e.g. by double cross-validation with tuning on inner cross-validation)
• Use benchmark package for analysis and visualization
• Supports different parallelization frameworks in R (like snowfall)

• Have a look at other visualization packages (e.g. rgobi, ROCR, classifly) and include useful plots
• Investigate and implement better tuning techniques than grid search
• GUI

install.packages("mlr",repos="http://R-Forge.R-project.org")

• Project page at R-Forge (bugtracker, mailing list, svn)
• Documentation