Detecting Irregularly Shaped Significant Spatial and Spatio-Temporal Clusters

Detecting significant overdensity or underdensity clusters in spatio-temporal data is critical for many real-world applications. Most existing approaches are designed to deal with regularly shaped clusters such as circular, elliptic and rectangular ones, but cannot work well on irregularly shaped clusters. In this paper, we propose GridScan, a grid-based approach for detecting irregularly shaped spatial clusters. In GridScan, a cluster is asymptotically described by a set of connected grid cells and is computed by a fast greedy regiongrowing algorithm with elaborating cluster merging in the process. The time complexity of GridScan is linear to the number of grids, making it scalable to very large datasets. A prospective spatio-temporal cluster detection approach, GridScan-Pro, is also proposed by extending GridScan. Experiments and a case study in the epidemic scenario demonstrate that our approaches greatly outperform existing ones in terms of accuracy, efficiency, and scalability.