A maximum margin approach for semi-supervised ordinal regression clusteri