Benchmark for filter methods for feature selection in high-dimensional classification data

Feature selection is one of the most fundamental problems in machine learning and has drawn increasing attention due to high-dimensional data sets emerging from different fields like bioinformatics. For feature selection, filter methods play an important role, since they can be combined with any machine learning model and can heavily reduce run time of machine learning algorithms. The aim of the analyses is to review how different filter methods work, to compare their performance with respect to both run time and predictive accuracy, and to provide guidance for applications. Based on 16 high-dimensional classification data sets, 22 filter methods are analyzed with respect to run time and accuracy when combined with a classification method. It is concluded that there is no group of filter methods that always outperforms all other methods, but recommendations on filter methods that perform well on many of the data sets are made. Also, groups of filters that are similar with respect to the order in which they rank the features are found. For the analyses, the R machine learning package mlr is used. It provides a uniform programming API and therefore is a convenient tool to conduct feature selection using filter methods.