A method for avoiding bias from feature selection with application to naive Bayes classification models