A visual multi-scale spatial clustering method based on graph-partition

Spatial data mining is the process of discovering interesting and previously unknown, but potentially useful patterns from large spatial databases. Extracting interesting and useful patterns from spatial datasets is more difficult than extracting patterns from traditional numeric and categorical data due to the complexity of spatial data types, spatial relationships, and spatial autocorrelation. In this paper, we provide a new visual hierarchical clustering based on graph-partitioning algorithm called VSG-CLUST, which groups and visualizes cluster hierarchies consisting of both non-spatial and spatial attributes. Our method is fundamentally different from conventional clustering algorithms, that usually do not take into account the spatial structure, which refers to the distance between patterns, topology, density, and other spatial distribution characteristics, and lack efficient level-of-detail strategy for visualization. In contrast, VSG-CLUST is able to recognize spatial patterns that involve neighbors. With the help of tree graph our method converts a multidimensional spatial clustering problem to a graph partitioning (tree partitioning) problem. We provide a theoretical basis for the approach and demonstrate the capability of the graph for maintaining the spatial structure. VSG-CLUST is implemented in a fully open and interactive manner, and it supports various visualization techniques including data mining algorithm visualization. A Web-based working demo with Fujian province environmental monitoring data is presented to illustrate the usability and effectiveness of VSG-CLUST and the proposed scheme.