Improving self organizing maps method for data clustering and classification

The unsupervised learning of Self Organizing Map (SOM) is an effective computational tool in data mining exploration processes. It provides topology preserved data mapping from high-dimensional input space into low-dimensional representation such as two-dimensional map. The visualization and classification of clustered data even with good topological preservation between input and output spaces however are not always easy to be interpreted especially when the data are unknown. The boundaries between the clusters and sub-clusters mapped on SOM map are occasionally not clear. In this paper, we develop an improved SOM (iSOM) method to produce an alternative SOM clustering, classification and visualization. The proposed method firstly groups data into their own classes and then arrange them on the third axis according to their computational distances to winning neurons. The method is demonstrated by computing iSOM clustering and classification on Iris Flowers dataset. The computational results have shown that iSOM was able to provide additional inherent information compared to SOM method. The separation of classes, the positions of a data with respect to other data, the existence of sub-clusters have been clearly presented while classification accuracy has increased.