Selection-fusion approach for classification of datasets with missing values