Exploratory Analysis of Point Proximity in Subspaces

We consider clustering as computation of a structure of proximity relationships within a data set in a feature space or its subspaces. We propose a data structure to represent such relationships, and show that, despite unavoidable arbitrariness in the clustering algorithms, constructive uses of their results can be made by studying correlations between multiple proximity structures computed from the same data. We describe a software tool that facilitates such explorations and example applications.