As a solution to the latter, i.e. too few cases for learning a stable and accurate scoring model, we propose the semi-supervised modeling approach when selecting learning cases: an initial classifier is learned on the ideal participants from a previous campaign.