Some issues in building and assessing classification rules with extremely skewed datasets