GLIN: A Lightweight Learned Indexing Mechanism for Complex Geometries

Although spatial index structures shorten the query response time, they rely on complex tree structures to narrow down the search space. Such structures in turn yield additional storage overhead and take a toll on index maintenance. Recently, there has been a flurry on works attempting to leverage machine-Learning(ML) models to simplify the index structures. Some follow-up works extend the idea to support geospatial point data. These approaches partition the multidimensional space to cells and assign IDs to these cells using space-filling curve(e.g., Z-order curve) or mathematical equations. These approaches work well for geospatial points but are not able to handle complex geometries such as polygons and trajectories which are widely available in geospatial data. This paper introduces GLIN, a lightweight learned index for spatial range queries on complex geometries. To achieve that, GLIN transforms geometries to Z-address intervals, and builds a hierarchical model to learn the cumulative distribution function between these intervals and the record positions. The lightweight hierarchical model greatly shortens the index probing time. Furthermore, GLIN augments spatial query windows using an add-on function to guarantee the query accuracy for both Contains and Intersects spatial relationships. Our experiments on real-world and synthetic datasets show that GLIN occupies 40-70 times less storage overhead than popular spatial indexes such as Quad-Tree while still showing similar query response time in medium selectivity queries. Moreover, GLIN's maintenance speed is around 1.5 times higher on insertion and 3-5 times higher on deletion