Spatial Queries Based on Learned Index

Abstract

With the popularity of location-based services, the scale of spatial data is increasing. Spatial indexes play an important role in spatial databases, and their performance determines the efficiency of data access and query processing. Most of the traditional spatial indexes divide data space or data objects without considering the distribution characteristics of data. In this paper, we design a spatial index structure, named learned Hilbert Model (HM) index. We combine the Hilbert space-filling curve and the two-stage model to build the spatial index. We propose algorithms for point query and range query according to data distribution rules. Experimental results show that the learned HM index can reduce the storage cost by 99% compared with R-tree and Grid Index. Point query efficiency is 40% higher than R-tree and 51% higher than Grid Index. The efficiency of range query is up to 50% higher than R-tree and 57% higher than Grid Index.