Many fraud analysis systemshave at their heart a rule-based enginefor generatingalertsaboutsuspiciousbehaviors.The rules in the systemareusually basedon expert knowledge. Automatic rule discovery aims at using past examples of fraudulent and legitimateusageto find new patternsandrules to help distinguish betweenthe two. Someaspectsof the problem of finding rules suitable for fraud analysis make this problem unique. Among themarethe following: the needto find rules combining both the propertiesof the customer(e.g., credit rating) and propertiesof the specific “behavior” which indicates fraud (e.g., number of international calls in one day); and the needfor a new definition of accuracy: We need to find rules which do not necessarily classify correctly each individual “usage sample” as either fraudulentor not, but ensuretheidentification, with aminimum of wasted cost and effort, of most of the fraud “cases” (i.e., defraudedcustomers). Theseaspectsrequireaspecial-purposerule d...