1.1 Task Background
A U.S. American veterans organization regularly conducts direct marketing campaigns, asking their members for donations (called gifts in the documentation) for the treatment of veterans with spinal injuries. The goal for the organization is to maximize net profit from their campaigns.
Only a small proportion of the members donate in reply to a campaign, while each letter sent out has a unit cost of 0.68 $US. In order to maximize profit, it is therefore desirable to only mail members who are likely to donate.
The members are grouped, among other criteria, by the recency of their last gift. Of these groups, the so-called lapsed donors are of particular interest. These are members who made their last gift to the organization 13 to 24 months prior to a given campaign. This group is important for two reasons: Firstly, the probability of a member donating decreases with the time the member has not donated. Enticing these lapsed donors to give again therefore maintains an active member base. Secondly, the organization has found that there is a negative correlation between the dollar amount donated and the probability to respond to a campaign. This means it is important to include the most unlikely donors in future mailings because if they donate, the net revenue is particularly large. If these unlikely donors would be suppressed from future campaigns, the gains from additional small dollar lapsed donors would not offset the losses from the potential high dollar donors.
The data at hand was distributed for the purpose of the KDD-CUP of the year 19981. The cup was until recently held yearly under the aegis of the special interest group on Knowledge Discovery and Data Mining (SIGKDD), which itself is part of the Association for Computing Machinery2(ACM).
For an archive of past cups, see SIGKDD - KDD Cup↩