revers test effici framework select classifi sampl select bia wei fan ian davidson import assumpt made classif algorithm train test set drawn distribut call stationari distribut assumpt futur past data set ident probabilist standpoint domain real world applic market solicit fraud detect drug test loan approv popul survei school enrol rare case label sampl train bias wai due varieti practic reason limit circumst tradit method evalu expect gener error classif algorithm structur risk minim ten fold cross valid leav valid return poor estim classif algorithm train bias dataset accur futur unbias dataset number compet candid estim order learn algorithm accuraci poor random guess method determin accur learner need data mine sampl select bia real world applic present approach determin learner perform unbias test set possibl bias train set fraction comput cost cross valid base approach 