On Averaging ROC Curves
Abstract: Receiver operating characteristic (ROC) curves are a popular method of summarising the performance of classifiers. The ROC curve describes the separability of the distributions of predictions from a two-class classifier. There are a variety of situations in which an analyst seeks to aggregate multiple ROC curves into a single representative example. A number of methods of doing so are available; however, there is a degree of subtlety that is often overlooked when selecting the appropriate one. An important component of this relates to the interpretation of the decision process for which the classifier will be used. This paper summarises a number of methods of aggregation and carefully delineates the interpretations of each in order to inform their correct usage. A toy example is provided that highlights how an injudicious choice of aggregation method can lead to erroneous conclusions.
Certifications: Survey Certification
License: Creative Commons Attribution 4.0 International (CC BY 4.0)
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Hsuan-Tien_Lin1
Submission Number: 811