One of the common problems with clustering is that the generated clusters often do not match user expectations. This paper proposes a novel probabilistic framework that exploits supervised information in a discriminative and transferable manner to generate better clustering of unlabeled data. The supervision is provided by revealing the cluster assignments for some subset of the ground truth clusters and is used to learn a transformation of the data such that labeled instances form well-separated clusters with respect to the given clustering objective. This estimated transformation function enables us to fold the remaining unlabeled data into a space where new clusters hopefully match user expectations. While our framework is general, in this paper, we focus on its application to Gaussian and von Mises-Fisher mixture models. Extensive testing on 23 data sets across several application domains revealed substantial improvement in performance over competing methods.
