A New Measure of Feature Selection Algorithms' Stability

Stability or robustness of feature selection methods is a topic of recent interest. A new stability measure based on the Shannon entropy is proposed in this paper to evaluate the overall occurrence of individual features in selected subsets of possibly varying cardinality. We compare the new measure to stability measures proposed recently by Somol et al. The new measure is computationally very efficient and adds another type of insight into the stability problem. All considered measures have been used to compare the stability of several feature selection methods (individually best ranking, sequential forward selection, sequential forward floating selection and dynamic oscillating search) on a set of examples.