Measuring unit relevance and stability in hierarchical spatio-temporal clustering

Abstract

Understanding the significance of individual data points within clustering structures is critical to effective data analysis. Traditional stability methods, while valuable, often overlook the nuanced impact of individual units, particularly in spatial contexts. In this paper, we explore the concept of unit relevance in clustering analysis, emphasizing its importance in capturing the spatio-temporal nature of the clustering problem. We propose a simple measure of unit relevance, the Unit Relevance Index (URI), and define an overall measure of clustering stability based on the aggregation of computed URIs. Considering two experiments on real datasets with geo-referenced time series, we find that the use of spatial constraints in the clustering task yields more stable results. Therefore, the inclusion of the spatial dimension can be seen as a way to stabilize the clustering.