Dynamic Recursive Model Class Selection for Classifier Construction

Before applying an automated model selection procedure, one must first choose the class (or family) of models from which the model will be selected. If there is no prior knowledge about the data that indicates the best class of models, then the choice is difficult at best. In this chapter, we present an approach to automating this step in classifier construction. In addition to searching for the best model, our approach searches for the best model class using a heuristic search strategy that finds the best model class for each recursive call of a divide-and-conquer tree induction algorithm. The end result is a hybrid tree-structured classifier, which allows different subspaces of a data set to be fit by models from different model classes. During search for the best model, the method considers whether and why a model class is a poor choice, and selects a better class on that basis. We describe an implementation of the approach, the MCS system, and present experimental results illustrating the system’s ability to identify the best model (and model class) efficiently.