På svenska

Presentation Information     2008-04-21 (15:15)   •  The seminar room at Vi2

Speaker Patrik Malm
Title Multisectional and optimized hierarchical clustering
Abstract Hierarchical clustering is a widely employed iterative method for the identification of subcategories in large and multivariate datasets and the resulting hierarchical structure is often visualized in the form of dendrogram-type representations. Traditional hierarchical clustering is, however, not designed to optimize categorization using a flexible performance criterion.

A framework for multisectional and optimized hierarchical clustering (MOHIC) is outlined. At each partitioning level (of a cluster into subclusters), MOHIC aims to optimize the subcluster configuration as regards both number of subclusters, as well as their composition. Initiation is performed with k-means clustering, whereas optimization procedure relies on silhouette analysis. Outputs are presented in the form of dendrograms with several features to support interpretation of the hierarchical structure. Results from three separate areas in biomedicine are presented to demonstrate applicability of the novel algorithm.