Clustering, Noise Reduction and Visualization Using Features Extracted from the Self-Organizing Map

摘要

This paper presents an analysis of a feature space generated by extracting properties related to pattern density and Euclidean distances between neurons from the self-organizing map network. Hence, along with the weight vector, each neuron has a 2-D feature vector associated with it, whose components are extracted from the U-matrix and a hit matrix, where latter is based on hyperspheres centered on each neuron. This collection of feature vectors, that represents the neurons of the network, is partitioned into different groups, and their labels are carried back to the data space as well as the neuron grid, in order to perform the tasks of clustering, noise reduction and visualization. Experiments were carried out using synthetic and real world data sets.