An investigation into the effect of unlabeled neurons on Self-Organizing Maps

Self-Organizing Maps (SOMs) are unsupervised neural networks that build data models. Neuron labeling attaches descriptive textual labels to the neurons making up a SOM, and is an important component of SOM-based exploratory data analysis (EDA) and data mining (DM). Several neuron labeling approaches tend to leave some neurons unlabeled. The interaction between unlabeled neurons and SOM model accuracy affect the choice of labeling algorithm for SOM-based EDA and DM, but has not been previously investigated. This paper applies the widely used example-centric neuron labeling algorithm to several classification problems, and empirically investigates the relationship between the percentage of neurons left unlabeled and classification accuracy. Practical recommendations are also presented, which address the treatment of unlabeled neurons and the selection of an appropriate neuron labeling algorithm.