Computer-supported decision making with object dependent costs for misclassifications

Abstract

It is described, how object dependent costs can be used in learning decision trees for cost optimal instead of error minimal class decisions. This is demonstrated using decision theory and the algorithm CAL5, which automatically converts real-valued attributes into discrete-valued ones by constructing intervals. Then, a cost dependent information measure is defined for selection of the (lo-cally) next maximally discriminating attribute in building the tree. Experiments with two artificial data sets and one application example show the feasibility of this approach and that it is more adequate than a method using cost matrices if cost dependent training objects are available.