Self Organized Swarms for cluster preserving Projections of high-dimensional Data

Abstract

A new approach for topographic mapping, called Swarm-Organized Projection (SOP) is presented. SOP has been inspired by swarm intelligence methods for clustering and is similar to Curvilinear Component Analysis (CCA) and SOM. In contrast to the latter the choice of critical parameters is substituted by self-organization. On several crucial benchmark data sets it is demonstrated that SOP outperforms many other projection methods. SOP produces coherent clusters even for complex entangled high dimensional cluster structures. For a nontrivial dataset on protein DNA sequence Multi Dimensional Scaling (MDS) and CCA fail to represent the clusters in the data, although the clusters are clearly defined. With SOP the correct clusters in the data could be easily detected.