A Scalable Hierarchical Algorithm for Unsupervised Clustering

Top-down hierarchical clustering can be done in a scalable way. Here we describe a scalable unsupervised clustering algorithm designed for large datasets from a variety of applications. The method constructs a tree of nested clusters top-down, where each cluster in the tree is split according to the leading principal direction. We use a fast principal direction solver to achieve a fast overall method. The algorithm can be applied to any dataset whose entries can be embedded in a high dimensional Euclidean space, and takes full advantage of any sparsity present in the data. We show the performance of the method on text document data, in terms of both scalability and quality of clusters. We demonstrate the versatility of the method in different domains by showing results from text documents, human cancer gene expression data, and astrophysical data. For that last domain, we use an out of core variant of the underlying method which is capable of efficiently clustering large datasets using only a relatively small memory partition.