Cluster Connections: A visualization technique to reveal cluster boundaries in self-organizing maps

Abstract

The self-organizing map is one of the most prominent unsupervised learning architectures used to visualize the similarities of high-dimensional input structures. What remains by no means straight-forward, is an explicit representation of cluster boundaries in the final two-dimensional map display. The detection of these boundaries rather requires some amount of insight into the inherent structure of the input data which may not be expected in real-world application scenarios. In this paper we address this deficiency by suggesting an extension to the standard map representation that leads to an easy recognition of cluster boundaries. The general idea is the visualization of clusters within the input data items by connecting units representing similar data items while disconnecting units representing dissimilar data items. As a result we get a grid of connected nodes where the intensity of the connection mirrors the similarity of the underlying data items. Such a representation allows intuitive analysis of the similarities inherent in the input data without the necessity of substantial prior knowledge, and an intuitive recognition of cluster boundaries.