Evaluating feature selection robustness on high-dimensional data

With the explosive growth of high-dimensional data, feature selection has become a crucial step of machine learning tasks. Though most of the available works focus on devising selection strategies that are effective in identifying small subsets of predictive features, recent research has also highlighted the importance of investigating the robustness of the selection process with respect to sample variation. In presence of a high number of features, indeed, the selection outcome can be very sensitive to any perturbations in the set of training records, which limits the interpretability of the results and their subsequent exploitation in real-world applications. This study aims to provide more insight about this critical issue by analysing the robustness of some state-of-the-art selection methods, for different levels of data perturbation and different cardinalities of the selected feature subsets. Furthermore, we explore the extent to which the adoption of an ensemble selection strategy can make these algorithms more robust, without compromising their predictive performance. The results on five high-dimensional datasets, which are representatives of different domains, are presented and discussed.