We study an extension of active learning in which the learning algorithm may
ask the annotator to compare the distances of two examples from the boundary of
their label-class. For example, in a recommendation system application (say for
restaurants), the annotator may be asked whether