A Hybrid Visualization-Induced Self-Organizing Map for Multi Dimensional Reduction and Data Visualization

Abstract

Self-Organizing Map (SOM), being a prominent unsupervised learning algorithm, is often used for multivariate data visualization. However, SOM only preserves inter-neurons distances in the input space and not in the output space due to the rigid grid used in SOM. Visualization-induced Self-Organizing Map (ViSOM) has been proposed as a visualization-wise improved variation of the popular unsupervised SOM. However ViSOM suffers from dead neuron problem as a huge number of neurons fall outside of the data region due to the regularization effect, even when the regularization control parameter is properly chosen. In this paper, a hybrid ViSOM that employs a modified Adaptive Coordinates (AC) technique is proposed for data visualization. Empirical studies of the hybrid technique yield promising topology preserved visualizations and data structure exploration for synthetic as well as benchmarking datasets.