On stability of ensemble gene selection

When the feature selection process aims at discovering useful knowledge from data, not just producing an accurate classifier, the degree of stability of selected features is a very crucial issue. In the last years, the ensemble paradigm has been proposed as a primary avenue for enhancing the stability of feature selection, especially in high-dimensional/small sample size domains, such as biomedicine. However, the potential and the implications of the ensemble approach have been investigated only partially, and the indications provided by recent literature are not exhaustive yet. To give a contribution in this direction, we present an empirical analysis that evaluates the effects of an ensemble strategy in the context of gene selection from high-dimensional micro-array data. Our results show that the ensemble paradigm is not always and necessarily beneficial in itself, while it can be very useful when using selection algorithms that are intrinsically less stable.