We stress that the field of machine learning would benefit significantly if more work were focused on explaining learning behavior by attending to the connection between learning strategies and the different types of example distributions. The goal of this paper is to exemplify the benefits of such kind of studies through a simple decomposition of classes into clusters. We show how the decomposition process is instrumental in explaining the behavior of Naive Bayes. We then exploit the same decomposition process to propose a new approach to learning that adapts to the characteristics of the dataset under analysis. Specifically, the idea is to fit models of different complexity over separate regions of the input space.