An Empirical Study of Stability of Feature Selection Algorithms

摘要: Stability is an important yet under-addressed issue for feature selection from high-dimensional data. In this paper, we investigate two causes of instability of feature selection: small sample size and elimination of redundant features. We propose a general stability measure which takes into account feature correlation when assessing the similarity of two feature subsets or two sets of feature groups. We empirically evaluate the stability of several representative feature selection and feature groups selection algorithms, and discuss the merits of the proposed measure and the impact of feature redundancy on stability based on stability profiles of these algorithms on microarray data sets.