A two-stage SOM for noise reduction in SOM map

A Self-organizing map (SOM) is a neural network model for clustering high-dimensional data into a low-dimensional space that could be used for applications of data clustering and visualization. One of the major problems of the SOM algorithm is the difficulty for non-expert users to interpret information from a variety of clusters in a map of SOM especially, when the data set consists of noise. Noise is a sample data which has a value different from the other members and distributed in the general cluster of SOM map. In this paper, we propose two stages SOM for investigating into this problem by removing the abnormal data - noise - from each cluster. The two stages are : to calculate the cluster and the members of cluster by using the training step of standard SOM; and to calculate the variance distance of the members in each cluster and find the noise of data by retaining stage of SOM without updating the weights but to comparing the data by using variance. The experiment shows the proposed method that can detect the noise and remove the abnormal data from each cluster. Furthermore, it reveals that, with a new chunk of the data set, the results from this experiment came out in good visualization. The SOM map also became more smooth than the former. The results of the proposed methods are compared and discussed on five medical data sets.