Automatic Distance-Based Interpolating Unit Detection and Pruning in Self-Organizing Maps

The self-organizing map (SOM) is an unsupervised neural network that uses neuron weight vectors to model training data. Some neurons, called interpolating units, have weight vectors that do not model training data, and instead represent boundaries between emergent data clusters. Interpolating units are useful for distinguishing such clusters using SOM visualizations. However, automatic (non-visual) detection of interpolating units would be advantageous for SOM analysis. This paper proposes a novel algorithm, based on inter-neuron distances in weight vector space, for identifying and possibly pruning interpolating units. Existing methods for interpolating unit detection are surveyed, highlighting drawbacks not associated with the proposed algorithm. Focusing on classification task performance, SOMs which are pruned using the proposed algorithm are compared to unpruned SOMs. This analysis demonstrates that interpolating unit pruning does not adversely affect SOM model quality, and suggests that the proposed algorithm warrants further investigation for application in data science.