On the scalability of feature selection methods on high-dimensional data

Lately, derived from the explosion of high dimensionality, researchers in machine learning became interested not only in accuracy, but also in scalability. Although scalability of learning methods is a trending issue, scalability of feature selection methods has not received the same amount of attention. This research analyzes the scalability of state-of-the-art feature selection methods, belonging to filter, embedded and wrapper approaches. For this purpose, several new measures are presented, based not only on accuracy but also on execution time and stability. The results on seven classical artificial datasets are presented and discussed, as well as two cases study analyzing the particularities of microarray data and the effect of redundancy. Trying to check whether the results can be generalized, we included some experiments with two real datasets. As expected, filters are the most scalable feature selection approach, being INTERACT, ReliefF and mRMR the most accurate methods.