Induction of decision rules that fulfil user-specified performance requirements

Existing techniques that automatically derive a classifier from a set of labelled objects try to minimize the error rate. Diagnostic classifiers used in medicine, however, are often intended to be optimal for other performance measures, such as sensitivity and specificity. We developed an algorithm that induces decision rules which fulfil user-specified performance requirements. Basically, the algorithm generates all possible rules and selects the one which is best. The algorithm was tested on various commonly used data sets, showing that its performance is at least as good as that of other induction methods. Further, we established the effect of different heuristics to speed up the induction process. Finally, we demonstrated the algorithm's capability to generate rules that satisfy different user needs. (C) 1997 Elsevier Science B.V.