Clustering Methods in Geometric Modeling and Scientific Visualization

Date: Wednesday, May 6, 2009
Time: 4:00pm - 5:00pm
Location: LBNL Bldg. 50F, Room 1647

Speaker:
Professor Hans Hagen
Department of Computer Science
University of Kaiserslautern, Germany

Abstract:

Clustering and classification of datasets is an important problem in
many application fields, most often to be solved in order to
determine specific regions of interest. In Geometric Modeling and
Scientific Visualization, clustering methods based on generalized
Voronoi Diagrams developed over the last couple of years into
standard techniques. Commonly, the topology of the dataset has to be
preserved, implying that unconnected parts of a cluster (so-called
holes) have to be avoided. We develop a new Voronoi clustering
technique based on the k-means scheme. For this purpose, we
construct generalized distance functions that guarantee topological
correctness. The distance functions are given by a custom definition
of the unit neighborhood of seed points. Using a region-growing
algorithm that is independent of the used distance function and
specific weighting, we arrive at a flexible and general
topology-preserving clustering method. We apply our scheme to
several both multi-dimensional and scattered data examples from
simulation based modeling and in urban planning. However, the domain
independent algorithm can easily be used in many other clustering
and classification tasks.

Host of Seminar:

    Gunther Weber