The Visual SuperTree: similarity-based multi-scale visualization

Abstract

Similarity-based exploration of multi-dimensional data sets is a difficult task, in which most techniques do not perform well with large data sets, particularly in handling clutter that invariably happens as data sets grow larger. In this paper, we introduce the Visual SuperTree (VST), a method to build a multi-scale similarity tree that can deal with large data sets at interactive rates, maintaining most of the accuracy and the data organization capabilities of other available methods. The VST is built on top of a clustered multi-level configuration of the data that allows the user to quickly explore data sets by similarity. The method is shown to be useful for both unlabeled and labeled data, and it is capable of revealing external and internal cluster structures. We demonstrate its application on artificial and real data sets, showing additional advantages of the approach when exploring data that can be summarized meaningfully.