The ROC skeleton for multiclass ROC estimation

Multiclass operating characteristics are a generalisation of the two-class receiver operator characteristic. A limitation regarding this generalisation is the computational complexity with increasing numbers of classes. In this paper, the ROC skeleton approach is proposed for efficiently estimating the operating characteristic. New operating points are computed from actual training samples, versus an alternative approach involving grid generation, that is prone to redundant calculations, and poor adaptation to certain classifier architectures. An extensive experimentation with a number of datasets and classifiers as a function of the number of calculations reveals the efficiency of this approach. Also notable is how in many cases good performance can be achieved with surprisingly few calculations, but the converse may also apply.