Unsupervised Weight-Based Cluster Labeling for Self-Organizing Maps

Abstract

Self-organizing maps (SOMs) have been applied for practical data analysis, in the contexts of exploratory data analysis (EDA) and data mining (DM). Many SOM-based EDA and DM techniques require that descriptive labels be applied to a SOM’s neurons. Several techniques exist for labeling SOM neurons in a supervised fashion, using classification information associated with a set of labeling data examples. However, classification information is often unavailable, necessitating the use of unsupervised labeling approaches that do not require pre-classified labeling data. This paper surveys existing unsupervised neuron labeling techniques. A novel unsupervised labeling algorithm, namely unsupervised weight-based cluster labeling, is described and critically discussed. The proposed method labels emergent neuron clusters using sub-labels built from statistically significant weights. Visualizations of the labelings produced by a prototype of the proposed approach are presented.