Density based learned spatial index for clustered data

Abstract

Retrieving spatial points, such as GPS records or Point of Interests, that satisfy specific location-based query criteria is a core operation in location-based services. Recent studies have shown that learned indexes can outperform traditional indexing methods in both query performance and space efficiency by leveraging data distribution to construct compact predictive models. On the other hand, traditional indexes typically make minimal assumptions about the underlying data distribution. In real-world spatial databases, data is often non-uniformly distributed and tends to cluster in specific regions or along road networks. Adaptivity to such data patterns may bring performance benefits.

In this paper, we explore the construction of efficient learned indexes that exploit the clustering characteristics of spatial datasets. Specifically, we propose a Density-based Grid Learning Spatial Index (DGLSI), which partitions the spatial domain based on point density and utilizes learned models, including multiple recursive model indexes to predict the grid cell IDs of query points. We evaluate DGLSI’s performance on real-world GPS datasets and demonstrate that the proposed methods outperform analogous grid-based indexes across various query workloads, including nearest point queries and range queries while maintaining high space efficiency.