Model selection for dynamic processes

In machine learning, ROC (Receiver Operating Characteristic) analysis is widely used in model selection when we consider both class distribution and misclassification costs that must be given at test time. In this paper we consider the case of a dynamic process, such that the class distributions are different in different time periods or states. The main problem is then to decide when to change models according to the different states of the generating process. In this paper we use a control chart to choose models for the process when misclassification costs are considered. Four strategies are considered and model selection approaches are discussed.