A feature selection method for selecting a predetermined number of features, the method may comprise for each category, ranking features according to feature ranking for two-category problems; and while the predetermined number of features has not yet been selected: picking a category c; selecting a feature based on the ranking for category c; and adding the selecting feature to an output list if not already present in the output list and removing the selected feature from further consideration for the category c. An apparatus and computer program code are also provided.