Local fuzzy geographically weighted clustering: a new method for geodemographic segmentation

Abstract

Fuzzy geographically weighted clustering has been proposed as an approach for improving fuzzy c-means algorithm when applied to geodemographic analysis. This clustering method allows a spatial entity to belong to more than one cluster with varying degrees, namely, membership values. Although fuzzy geographically weighted clustering attempts to create geographically aware clusters, it partially fails to trace spatial dependence and heterogeneity because, as a global metric, the membership values are calculated across the entire set of spatial entities. Here we introduce the first local version of fuzzy geographically weighted clustering, ‘local fuzzy geographically weighted clustering.’ In local fuzzy geographically weighted clustering, the membership values of a spatial entity are updated only according to the membership values of the spatial entities within its neighborhood and not across the entire set of entities, as originally proposed by the global metric. Additionally, we apply particle swarm optimization meta-heuristic to overcome the random initialization problem regarding the fuzzy c-means algorithm. To evaluate our method we compare local fuzzy geographically weighted clustering to global fuzzy geographically weighted clustering using a cancer incident benchmark dataset for Manhattan, New York. The results show that local fuzzy geographically weighted clustering outperforms the global version in all experimental settings.