Pazzani, M.,   Murphy, P.,   Ali,  K. and  Schulenburg, D.
 (1994). Trading off coverage for accuracy in forecasts:
 Applications to clinical data analysis. AAAI Symposium on AI
 in Medicine  (pp 106-110). Stanford, CA.
 Machine learning algorithms that perform classification tasks
 create concept descriptions (e.g., rules, decision trees, or
 weights on a neural net)  guided by a set of classified training
 examples.  The accuracy of the resulting concept description
 can be evaluated empirically by asking a classifier  to  use
 the concept description to classify a set of test examples.  One
 goal of this evaluation is to determine the accuracy of the
 learning algorithm if it were applied in a "real" application.
  In such an application, it is assumed that a classifier is
  produced on a set of training cases and a decision  is made
 automatically on  each new case based upon a forecast of the
 classification of the case.  However, in some applications, it
 is advantageous not to produce a classification on every
 example.  In particular,  when a machine learning program is
 being used to assist a person to perform some task, it might be
 desirable to have the machine automatically make some
 decisions while allowing others to be made by the person.   The
 general idea is that in some cases, a user of a machine
 learning system might be willing to allow the learning  system
 to produce no classification on some examples.  This may be
 acceptable if it means that when the system makes a
 prediction, the prediction is more accurate.