A Fast Hybrid Spatial Index with External Memory Support

Despite claiming to support external memory access, very few learned indices implement their indices on secondary storage. It is challenging to extend learned models to secondary storage while preserving high levels of index performance. In this paper, we propose a Fast Hybrid Spatial Index with External Memory Support (FHSIE). Its core idea is to learn a model which can group spatial objects in a top-down manner with few parameters. We use a height-balanced hierarchical structure, which recursively uses simple unsupervised models to group (i.e., cluster) spatial objects. We associate the learned structure with a grid for accurate query processing, and we utilize the relationship among clusters and grid cells to estimate the grid layout and optimize grid performance. Extensive experiments on real and synthetic data sets with more than 200 million points show that FHSIE is highly efficient and precise no matter where it works, i.e., in memory or on a disk.