"Pairwise Data Clustering and Multidimensional Scaling"

Abstract:

THOMAS HOFMANN and JOACHIM BUHMANN, University of Bonn

Euclidian embedding and partitioning a data set which is characterized
by pairwise dissimilarities of the data is a difficult combinatorial
optimization problem. Algorithms for embedding such a data set in a
Euclidian space, for clustering these data and for actively selecting
data items to support the clustering process are discussed in the
maximum entropy framework. The algorithms implement a new strategy for
nonlinear dimension reduction and visualization.  To yield a
clustering solution of predefined quality, active data selection
reduces the number of required data considerably.