A1386
Title: Large-scale comparison and clustering of spatial point patterns using the two-dimensional Kolmogorov-Smirnov statistic
Authors: Takaaki Ohnishi - Rikkyo University (Japan) [presenting]
Abstract: Quantifying spatial collocation patterns from large-scale data poses statistical and computational challenges. Cluster-based methods suffer from the modifiable areal unit problem, while distance-based methods require kernel functions, bandwidth parameters, or a reference distribution against which patterns are tested. A nonparametric, parameter-free framework for comparing spatial point patterns using the two-dimensional Kolmogorov-Smirnov statistic D is proposed. As a two-sample statistic, D requires no predefined spatial units and directly compares two empirical distributions without specifying a reference. Unlike the p-value, D directly quantifies distributional difference rather than statistical significance, making it suitable as a pairwise feature for large-scale comparison. Applying this framework to an establishment database of approximately 1.2 million records in the Greater Tokyo Area (Tokyo, Saitama, Chiba, and Kanagawa) across 579 categories, all 167,331 pairwise D values are computed, yielding a 579x579 spatial distance matrix. Ward's hierarchical clustering reveals groupings that differ substantially from conventional industry classifications, uncovering latent functional spatial structures, including residential-service and coastal-natural clusters, not captured by standard classification schemes.
