Identifying non-hierarchical spatial clusters

Investigating relationships in information is an important component of statistical and spatial analysis. Clustering techniques have proven to be effective in helping to find patterns or relationships which otherwise may not have been detected. One of the most popular partitioning based clustering approaches is k-means. In this paper we explore the application of the k-means approach to spatial information in order to assess its characteristic features. One aspect of concern is implementation in commercial statistical software packages. In particular, the invariable sub-optimality (or poor solution quality) of identified clusters will be investigated. Another aspect of evaluation is a direct comparison to an alternative, spatially-based clustering approach. Although numerous authors over the past 30 or so years have discussed shortcomings associated with the use of the k-means approach, empirically based work which illustrates biasing effects is lacking. This paper contributes to the further understanding of clustering techniques.