Learning the Number of Clusters in Self Organizing Map
Guénaël Cabanès, Younès Bennani · InTech eBooks · 2010
We proposed here a density-based simultaneous two-level clustering method. It uses SOM as dimensionality reduction technique and achieves an improved final clustering in the second level, using both distance and density information. The proposed algorithm DS2LSOM locates regions of high density that are separated from one another by regions of low density. The performance of DS2L-SOM have been evaluated on a set of critical clustering problems, and compared to other two-level clustering algorithms. The experimental results demonstrate that the proposed clustering method achieves a better clustering quality than classical approaches. The results also demonstrate that DS2L-SOM is able to discover irregular and intertwined clusters, while conventional partitional clustering algorithms can deal with convex clusters only. Finally, the number of clusters in our approach is determined automatically during the learning process, i.e., no a priori hypothesis for the number of clusters is required. In the future we plan to incorporate in the DS2L-SOM algorithm some plasticity property, to evaluate its impact on the cluster quality and stability.