modified normal WEKA training to have a fallback in case to few instances of a certain class (defect-prone, non-defect-prone) are available: it now uses ZeroR in that case, i.e., a trivial classifier that always predicts the class that appears more often.