Clustering high-dimensional data using growing SOM

The self-organizing map (SOM) is a very popular unsupervised neural-network model for analyzing of high-dimensional input data as in scientific data mining applications. However, to use the SOM, the network structure must be predetermined, this often leads constrains on potential applications. When the network is unfit to the data model, the resulting map will be of poor quality. In this paper, an intuitive and effective SOM is proposed for mapping high-dimensional data onto the two-dimensional SOM structure with a growing self-organizing map. In the training phase, an improved growing node structure is used. In the procedure of adaptive growing, the probability distribution of sample data is also a criterion to distinguish where the new nodes should to be added or deleted besides the maximal quantization error (mqe) of a unit. The improved method is demonstrated on a data set with promising results and a significantly reduced network size.