Issues related to data mining with self organizing maps

This thesis demonstrates that the clustering by Kohonen's Self-Organizing Map algorithm (KSOM) can be significantly improved by magnification control. As a second contribution, the thesis proposes a fully automated technique for clustering the prototypes of the data (SOM weights). The motivation for this work comes from a serious need for effective, precise, and detailed knowledge discovery (clustering) for complex, high-dimensional data that are encountered in a variety of important applications such as remote sensing, medical imaging, etc. While many conventional clustering methods may fail to handle such data, modifications of KSOM show promise. We analyze the performance of one such advanced modification, the magnification control algorithm by Bauer, Der and Herrmann [13] to determine its scope. By magnification control, different clustering criteria can be optimized. We also demonstrate that negative magnification improves the detectability of rare classes in the data potentially leading to discovery of new classes