45 ROC Trajectory Measures for Classifier Accuracy and Robustness

The Receiver Operating Characteristic (ROC) curve can be thought of as a projection of a 3 dimensional trajectory through some True positive, False positive, and decision threshold space onto the True positive, False positive plane. Using this frame of reference, Alsing, et a1. (2002) developed a metric for the comparison of classifiers. It was shown that this metric appears to capture information concerning the robustness of a classifier. In this paper we investigate further the notions proposed in Alsing et al. relative to the robustness of total classification accuracy at a point of interest. Investigations are conducted with the application of several common classification methods on binary classification problems using breast cancer data from the University of Wisconsin.