EC-MMR: Revised exemplar-based clustering with automatic parameter estimation techniques

In this study, we discuss recent advances in the theory and practice of exemplar-based clustering. In the context of clustering, exemplars are those representative objects in the data sets. A recently proposed approach called convex clustering with exemplar-based models, referred as (CCE), adopts a convex objective function with a global solution. Although the existing frame work of CCE is attractive, the parameter sensitivity problem may make the original CCE infeasible to be used for some real applications. In this paper, we propose an improved version called exemplar-based clustering with minimal marginal redundancy (EC-MMR). In EC-MMR, the shape parameter is estimated automatically based on the data. Further more, the finally exemplars are selected in an improved way in which both the representativeness of each individual object and the whole exemplar set are considered. Our experiment results show that with these procedures incorporated, the new approach improves the CCE approach greatly with respect to producing higher quality of clusters in a fully automatical manner.