Selecting models focussing on the modeller's purpose

Abstract

Model selection is a difficult task for which it is often profitable to take into account the modeller point of view. Hidden structure models are a good example for which this point of view can be dealt with in a simple way. In the model-based clustering context, we present model selection criteria focussing on the clustering purpose. Their rationale and theoretical features are given and their practical behavior in comparison with classical penalized likelihood criteria is discussed from numerical experiments.