Title: Automatic detection of clusters by lifting
Authors: Nebahat Bozkus - University of Leeds (United Kingdom) [presenting]
Stuart Barber - University of Leeds (United Kingdom)
Abstract: In clustering, the aim is to group related objects together and an important question is how to detect the number of groups in a data set. Many methods have been proposed which capture the number of groups quite well if the groups are well separated and regularly shaped. However, if groups overlap or have unusual shapes, the performance of these methods deteriorates. We propose a new method based on a multiscale technique called lifting which has recently been developed to extend the `denoising' abilities of wavelets to data on irregular structures. The method seeks for the best representation of the clustering pattern by checking all possible clustering schemes in a tree. After denoising the tree, if the leaves under a node are all close enough to their centroid for the deviations to be explained as `noise', we label those leaves as forming a cluster. The proposed method automatically decides how much departure can be allowed from the centroid of each cluster. The behaviour of the method will be illustrated using some phylogenetic data sets.