Quadratic-assignment-like problem formulation in feature selection

Quadratic-assignment-like problem formulation in feature selection is proposed. We select a group of m most similar features out of n features by solving quadratic-assignment-like problem using an effective genetic algorithm. As feature similary metrics the correlation coefficient of individual features is used. The proposed method is validated on NIPS2003 FS benchmark data-sets. The method finds stable feature subsets in which the identities of the discovered features do not vary much. Our proceduce can easily be applied in multi-class problems, including other measures of feature similarityKauno technologijos universiteta