Adaptive Multilevel Cluster Analysis by Self-Organizing Box Maps
Tobias Galliat · Universitätsbibliothek der FU Berlin Hochschulschriftenstelle u. Dokumentenserver · 2002
The aim of this thesis is a fruitful combination of Perron Cluster analysis and self-organized neural networks within an adaptive multilevel clustering approach that allows a fast and robust identification and an efficient description of clusters in high-dimensional data. In a general variant that needs a correct number of clusters k as an input, this new approach is relevant for a great number of cluster problems since it uses a cluster model that covers geometrically, but also dynamically based clusters. Its essential part is a method called representative clustering that guarantees the applicability to large cluster problems: Based on an adaptive decomposition of the object space via self-organized neural networks, the original problem is reduced to a smaller cluster problem. The general clustering approach can be extended by Perron Cluster analysis so that it can be used for large reversible dynamic cluster problems, even if a correct number of clusters k is unknown a priori. The basic application of the extended clustering approach is the conformational analysis of biomolecules, with great impact in the field of Drug Design. Here, for the first time the analysis of practically relevant and large molecules like an HIV protease inhibitor becomes possible.