Greedy Optimization for Contiguity-Constrained Hierarchical Clustering

Abstract

The discovery and construction of inherent regions in large spatial datasets is an important task for many research domains such as climate zoning, eco-region analysis, public health mapping, and political redistricting. From the perspective of cluster analysis, it requires that each cluster is geographically contiguous. This paper presents a contiguity constrained hierarchical clustering and optimization method that can partition a set of spatial objects into a hierarchy of contiguous regions while optimizing an objective function. The method consists of two steps: contiguity constrained hierarchical clustering and two-way fine-tuning. The above two steps are repeated to create a hierarchy of regions. Evaluations and comparison show that the proposed method consistently and significantly outperforms existing methods by a large margin in terms of optimizing the objective function. Moreover, the method is flexible to accommodate different objective functions and additional constraints (such as the minimum size of each region), which are useful to for various application domains.