Estimation of Area under the Multi-Class ROC for Non-Normal Data

In modelling ROC curves, there are several bi-distributional ROC models available in the literature. These are developed in the context of normal and non-normal data patterns and in the framework of binary classification. However, in most of the practical data at hand may exhibit multi-model patterns or it may be of multi-class, then the existing bi-distributional ROC forms are not viable to apply and fit the curve. So, in this paper, we made an attempt to address the above mentioned situations using finite mixtures. We proposed a mixture Exponential ROC model and its measures like AUC, FPR, TPR and optimal cut-offs are derived. The methodology is supported with simulated and real data sets.