Two models for single cluster clustering

Boris Mirkin

(Birbeck College, London)

Single cluster clustering is a discipline devoted to the problem of separating a homogeneous cluster (a pattern) from a data set. This type of clustering can both complement and supplement more traditional clustering problems such as partitioning or building a hierarchy.

Two models for single cluster clustering will be presented; one based on the traditional data format, entity-to-feature table, and the other on a more flexible format of image and signal data, the entity-to-subset linkage function.

The first model leads to a clustering method, which is akin to k-means clustering and gives the user more advice with regard to data pre-processing, initial setting, and interpretation.

The second model leads to a method for structuring data in a nested manner by separating the dense core, its less dense shell, then more shells, each sparser than the previous one. Mathematically, the problem emerging is that of maximization of a specific set function, which admits a polynomial-time greedy like solution.
Thursday 14th March 2002, 15:00
Robert Recorde Room
Department of Computer Science