Proactive data mining using decision trees

Most of the existing data mining algorithms are `passive'. That is, they produce models which can describe patterns, but leave the decision on how to react to these patterns in the hands of the user. In contrast, in this work we describe a proactive approach to data mining, and describe an implementation of that approach, using decision trees. We show that the proactive role requires the algorithms to consider additional domain knowledge, which is exogenous to the training set. We also suggest a novel splitting criterion, termed maximal-utility, which is driven by the proactive agenda.