Informatik Ensemble and Constrained Clustering with Applications

The main focus of this thesis concerns the further developments in the areas of ensemble and constrained clustering. The goal of the proposed methods is to address clustering problems, in which the optimal clustering method is unknown. Additionally, by means of pairwise linkage constraints, it is possible to aggregate extra information to the clustering framework. Part I investigates the concept of ensemble clustering. It presents a comprehensive review of the state of the art in ensemble clustering. It follows by discussing the impact of the ensemble variability in the final consensual result. Visualization of ensemble variability based on multidimensional scaling is also a topic addressed in this part. A software which is able to perform ensemble clustering using various existing consensus functions is also introduced. A consensus function based on random walker originally developed for image segmentation combination is adapted to the ensemble clustering problem. A lower bound is proposed to explore how well cluster ensemble methods perform in an absolute sense, without the usage of ground-truth. Finally, a study evaluating how well the general ensemble clustering techniques perform in the context of image segmentation combination closes this part. Part II introduces an ensemble clustering method based on a new formulation for the median partition problem. The performance of this method is assessed in relation to other well known ensemble clustering methods. Part III addresses the potential of ensemble techniques in the framework of constrained clustering. It presents a comprehensive review of the state of the art in constrained clustering and discusses the impact of considering constraints locally or globally. An experiment is presented comparing both approaches. A new clustering method is introduced combining both ensemble and constrained clustering. Constraints are introduced into three consensus functions. This part closes with an experimental evaluation, in which constraints are considered in different steps of the clustering ensemble framework.