Visualisation of multi-class ROC surfaces

The Receiver Operating Characteristic (ROC) has become a standard tool for the analysis and comparison of binary classifiers when the costs of misclassification are unknown. Although there has been relatively little work in examining ROC for more than two classes – there has been growing interest in the area, and in recent studies we have formulated it in terms of misclassification rates. Although techniques exist for the numerical comparison of the fronts generated by these new methods, the useful visualisation of these fronts to aid the selection of a final operating point are still very much in their infancy. Methods exist for the visualisation of similar surfaces, Pareto fronts, which we discuss, however the particular properties of the ROC front that the practitioner is interested in may also direct us to new and more suitable visualisation methods. This paper briefly outlines what is currently in use, and what avenues may be of interest to examine in the future.