Optimizing Sequential Testing Performance Using Classification Trees

• Activity Number: 561
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Christine Schubert Kabban*+ and Brandon Greenwell
• Companies: Air Force Institute of Technology and AFIT
• Keywords: classification tree ; sequential testing ; ROC curve ; Youden's Index ; impurity ; optimal point

Abstract

We consider common sequential testing scenarios such as believe the positive and believe the negative. These testing scenarios classify subjects without necessitating that each subject be classified by each test. Therefore, these scenarios may be less costly yet perform as accurately as when every test is combined for classification. ROC curves and optimal point criteria may be used to sequence the tests and determine the associated thresholds that assure maximum accuracy (correct classification). This process may be inhibited by the length of the sequence and the assumptions for the underlying class distributions. Thus, we examine an equivalence relation between the structure of the sequential tests and classification trees. By structuring each sequence as a tree, we may efficiently utilize algorithms to find the optimal tree, and thereby the optimal thresholds and sequence of tests. We demonstrate the equivalence between ROC optimal point criteria, such as Youden's Index, and measures of the classification tree performance, such as impurity for a fixed length test sequence. Finally, this equivalence is further demonstrated using application data.