Heuristics for Multiple Class Classification Problems via ROC Hypersurface

Receiver operating characteristics (ROC) graphs are useful for organizing binary classifiers and visualizing their performance. But it doesn´t work on the corresponding multi-class classifiers problem. We analyse a series of multi-class classifiers and present a new algorithm based on its ROC parameters in which the goal is to minimise the cost of Q(Q - 1) misclassification. Empirical results suggest that our algorithm is more stable than the several existing popular methods.