Goal-Oriented Clustering

We introduce goal-oriented clustering, a process that clusters items with the explicit knowledge that the ultimate use of the clusters is prediction. In this approach, we use data on a set of target variables (those we want to predict) and a set of input variables (those we do not want to predict) to learn a graphical (generative) model with a single hidden layer of discrete variables H. The states of H correspond to clusters. We describe a generalized EM algorithm for learning the parameters of this class of models and provide a convergence guarantee. We compare our goal-oriented approach to a standard clustering approach on the task of targeted advertising on a web site.