[Paper Review] DeepSPACE: Approximate Geospatial Query Processing with Deep Learning

Dimitri Vorona, Andreas Kipf+2 more|arXiv (Cornell University)|Jun 14, 2019

Data Management and Algorithms18 references4 citations

TL;DR

DeepSPACE is a deep learning-based approximate query processing system for geospatial data that uses neural distribution estimation and space-filling curves to model spatial data distributions with minimal memory (a few hundred KiBs). It enables fast, interactive exploration of large geospatial datasets on lightweight devices by accurately estimating aggregation queries—outperforming sampling-based methods in precision, especially under high selectivity.

ABSTRACT

The amount of the available geospatial data grows at an ever faster pace. This leads to the constantly increasing demand for processing power and storage in order to provide data analysis in a timely manner. At the same time, a lot of geospatial processing is visual and exploratory in nature, thus having bounded precision requirements. We present DeepSPACE, a deep learning-based approximate geospatial query processing engine which combines modest hardware requirements with the ability to answer flexible aggregation queries while keeping the required state to a few hundred KiBs.

Motivation & Objective

• Address the growing challenge of processing massive geospatial datasets with low-latency, low-resource requirements for interactive data exploration.
• Overcome limitations of traditional indexing and sampling-based approximate query processing, especially poor performance on highly selective queries.
• Enable efficient, accurate estimation of complex geospatial aggregation queries (e.g., COUNT, SUM, MEAN) on resource-constrained devices like smartphones.
• Develop a modular, extensible architecture that supports heterogeneous data types and user-defined distribution families for improved modeling fidelity.
• Demonstrate that a compact learned model can maintain high query accuracy across varying dataset sizes and selectivities, unlike sampling-based approaches.

Proposed method

• Employ neural autoregressive modeling via normalizing flows to learn the joint distribution of geospatial data points, enabling efficient conditional probability estimation for arbitrary queries.
• Use space-filling curves (specifically, a quad-tree-based discretization) to map 2D geographic coordinates into a 1D sequence, preserving spatial locality and enabling effective distribution learning.
• Train the model in an unsupervised manner on raw geospatial data enriched with attribute metadata, without requiring prior query logs or labeled examples.
• Support flexible query interfaces by computing conditional distributions over attributes given spatial and attribute-based predicates, enabling estimation of aggregates like COUNT, SUM, and MEAN.
• Allow integration of domain-specific distribution families (e.g., Gaussian Mixture Models for multimodal continuous attributes) to improve modeling accuracy for specific data types.
• Maintain a compact model state (a few hundred KiBs) independent of dataset size, enabling deployment on edge devices.

Experimental results

Research questions

• RQ1 Can a deep learning model approximate complex geospatial aggregation queries with high precision while using minimal memory and no access to historical query logs?
• RQ2 How does the performance of DeepSPACE compare to sampling-based approximate query processing under varying selectivity and dataset sizes?
• RQ3 To what extent can a single compact learned model generalize across different data distributions and query types in geospatial workloads?
• RQ4 How does the model maintain accuracy when the underlying dataset size increases, especially for highly selective queries?
• RQ5 Can the integration of user-defined distribution families improve estimation accuracy for specific attributes, such as multimodal or skewed distributions?

Key findings

• DeepSPACE achieves a median q-error of 1.25 on COUNT(*) queries, outperforming 10% sampling (median q-error 1.28) and significantly reducing error on highly selective queries.
• For continuous aggregates (SUM and MEAN), DeepSPACE achieves a median sMAPE of 0.04 for queries with over 1000 qualifying tuples, outperforming 10% sampling (median sMAPE 0.05) and showing superior robustness under high selectivity.
• The model maintains near-constant performance across dataset sizes—when data doubled, DeepSPACE’s median q-error remained stable at 1.11 (vs. 445 for 0.05% sample), confirming its scalability and state-size invariance.
• On queries with fewer than 100 qualifying tuples, DeepSPACE achieved a median q-error of 1.11, significantly lower than 0.5% sampling (median q-error 332), demonstrating its advantage in low-selectivity regimes.
• The model’s performance is robust to data distribution shifts: when trained on January and February NYC taxi data, it generalized well to the same workload, maintaining high accuracy without retraining.
• DeepSPACE’s sMAPE for MEAN and SUM aggregates remained below 0.11 for queries with >1000 results, while sampling at 10% only achieved a median sMAPE of 0.06 for MEAN and 0.05 for SUM—indicating superior precision in high-accuracy regimes.