mlr: A new Package to conduct Machine Learning Experiments in R

The mlr package [1] provides a generic, object-oriented interface to many machine learning methods in R for classification and regression and can easily be extended with further ones. It enables the researcher to rapidly conduct complex experiments or implement his own meta-methods using the package’s building blocks. Resampling like cross-validation, bootstrapping and subsampling are used to assess the generalization performance. Hyperparameters of learners can be tuned by grid search or more sophisticated deterministic or stochastic search methods like e.g. Nelder-Mead or CMA-ES [3]. The same holds true for variable selection. Here, mainly the wrapper approach [4] is currently implemented, but the package will be extended with filters and a combination of the two. Benchmark experiments with two levels of resampling, e.g. nested crossvalidation, can be specified with few lines of code to compare different classes of learning algorithms. An interface to the benchmark package by Eugster and Leisch [2] is provided, which enables exploratory and inferential analysis of the results. Parallel high-performance computing is supported through the snowfall, nws andmulticore packages [5]. Experiments can be converted to parallelized versions with a simple configuration command, without touching any further code. The job granularity of scheduled tasks can be changed, so jobs don’t complete too early, providing a better scale-up for problems of different sizes. The talk will include short use cases to explain the package and its optimization algorithms, some general remarks about its programming methodology and a live R demo.