Joint ranking and clustering based on Markov Chain transition probabilities learned from data

Abstract

The focus of this thesis is to develop a Markov Chain based framework for joint ranking and clustering of a dataset without the need for critical user-defined hyper-parameters. Joint ranking and clustering may be useful in several respects, and may give additional insight for the data analyst, as opposed to the traditional separate ranking and clustering procedures. By coupling Markov chain theory with recent advances in kernel methods using the so-called probabilistic cluster kernel, we are able to learn the transition probabilities from the inherent structures in the data in a near parameter-free approach. The theory developed in this thesis is applied to several real world datasets of different types with promising results.