A Class of Parametric Tree-Based Clustering Methods

Abstract

We introduce a class of tree-based clustering methods based on a single parameter W and show how to generate the full collection of cluster sets C(W), without duplication, by varying W according to conditions identified during the algorithm’s execution. The number of clusters within C(W) for a given W is determined automatically, using a graph representation in which cluster elements are represented by nodes and their pairwise con- nections are represented by edges. We identify features of the clusters produced which lead to special procedures to accelerate the computation. Finally, we introduce a related node-based variant of the algorithm based on a parameter Y which can be used to generate clusters with complementary features, and a method that combines both variants based on a parameter Z and a weight that determines the contribution of each variant.