GeoAggregator: Spatial Data Aggregation

• GeoAggregator is a framework that aggregates and models spatial data using formal GIS models and efficient OLAP integration.
• It employs precomputed overlays, grid-based caching, and trie-like data structures to accelerate complex polygonal queries and reduce computation time.
• Transformer-based models and optimized pipelines enhance spatial regression accuracy and interpretability while managing spatial heterogeneity and autocorrelation.

GeoAggregator (GA) refers to several technical frameworks and models developed for the efficient aggregation and modeling of spatial data. It encompasses formal data models and query frameworks for spatial aggregation in GIS, deep transformer-based models for geospatial tabular data regression, and pre-aggregating data structures for scalable polygonal queries. GA frameworks enable rapid, precise, and explainable spatial data analysis across both relational and machine learning paradigms.

1. Formal Models for Spatial Aggregation

GeoAggregator’s foundational approach to spatial aggregation is grounded in a formal, multi-layered data model integrating GIS and OLAP systems (0707.4304). It assumes a finite set of thematic layers $L_1, ..., L_k$, each comprising:

• Algebraic part: Point-level, $\operatorname{dom}(\mathrm{Point}) = \mathbb{R}^2$, described by linear constraints.
• Geometric part: Relations of identifiers for $$G \in \{\mathrm{Point}, \mathrm{Line}, \mathrm{Polyline}, \mathrm{Polygon}, ...\}$$.
• OLAP part: Hierarchical dimension levels (e.g., state → region → all).

The model supports roll-up relations $$r_L^{G \rightarrow G'} \subset \operatorname{dom}(G) \times \operatorname{dom}(G')$$ and mappings $\alpha_{L,D}^{A \rightarrow G}$ to associate OLAP dimension levels to geometric identifiers. Fact tables exist at the point level ($ft(x, y, L, G)$) as well as higher geometry levels ($ft(g_{\mathrm{id}}, L)$) for measures like density or elevation.

A geometric aggregation query has the form:

$Q = \iint_{(x, y) \in C} h(x, y) \, dx\, dy$

where $C \subseteq \mathbb{R}^2$ is a semi-algebraic region defined by first-order logical predicates, and $h(x, y)$ is built from constants and fact-table functions.

2. Efficient Summable Aggregation and Overlay Precomputation

A subset of queries, the summable or summation-reducible queries, are efficiently evaluated by precomputing geometric overlays (0707.4304). These are characterized by:

• $\operatorname{dom}(\mathrm{Point}) = \mathbb{R}^2$0 for finite $\operatorname{dom}(\mathrm{Point}) = \mathbb{R}^2$1, and
• Existence of an $\operatorname{dom}(\mathrm{Point}) = \mathbb{R}^2$2 such that $\operatorname{dom}(\mathrm{Point}) = \mathbb{R}^2$3 with $\operatorname{dom}(\mathrm{Point}) = \mathbb{R}^2$4.

Overlay precomputation generates sub-polygonizations, storing for each sub-cell the mapping to contributing geometries. This reduces per-query complexity from $\operatorname{dom}(\mathrm{Point}) = \mathbb{R}^2$5 spatial joins to a scan over the overlay table ($\operatorname{dom}(\mathrm{Point}) = \mathbb{R}^2$6), accelerating both geometric and aggregate queries. Benchmarks demonstrate 4–8× speedups over R-tree-based systems for typical aggregate queries, with greater efficiency for overlay-aligned operations (0707.4304).

3. Transformer-based GeoAggregator Models for Tabular Data

The term GeoAggregator has also been adopted for a family of neural architectures designed for regression over geospatial tabular data (Deng et al., 20 Feb 2025, Deng et al., 23 Jul 2025). These models address spatial autocorrelation and heterogeneity directly in the model architecture:

Model architecture highlights:

• Encoder-Processor-Decoder framework with explicit spatial neighborhood selection for each query point.
• Gaussian-biased attention: Softmax attention scores are biased by learnable spatial kernel terms of the form $\operatorname{dom}(\mathrm{Point}) = \mathbb{R}^2$7, where $\operatorname{dom}(\mathrm{Point}) = \mathbb{R}^2$8 is spatial distance. This enables handling both spatial autocorrelation and nonstationarity.
• 2D Rotary positional encodings: Encodes absolute and relative 2D positions, capturing global spatial heterogeneity within attention layers.
• Multi-head Cartesian Product Attention (MCPA): Increases effective expressiveness via virtual head combinations, maintaining low parameter counts.
• Inducing point compression: Reduces quadratic attention costs to linear in neighborhood size.

Empirical evaluations on synthetic (spatial lag, GWR, Durbin) and real datasets (PM25 pollution, housing, and poverty data) demonstrate that GeoAggregator variants achieve best or second-best MAE and $\operatorname{dom}(\mathrm{Point}) = \mathbb{R}^2$9 on nearly all benchmarks, with parameter counts (4k–8k) and FLOPs per inference (1–6 million) much lower than GCNN or SRGCNN models (Deng et al., 20 Feb 2025).

4. Pipeline Optimization, Ensembling, and Explainability

Advanced iterations of the GeoAggregator pipeline address computational efficiency and model interpretability (Deng et al., 23 Jul 2025):

• Optimized data pipeline: Precomputes and caches $$G \in \{\mathrm{Point}, \mathrm{Line}, \mathrm{Polyline}, \mathrm{Polygon}, ...\}$$0-nearest neighbor indices for all queries, reducing neighbor-lookup cost during training and inference from $$G \in \{\mathrm{Point}, \mathrm{Line}, \mathrm{Polyline}, \mathrm{Polygon}, ...\}$$1 to $$G \in \{\mathrm{Point}, \mathrm{Line}, \mathrm{Polyline}, \mathrm{Polygon}, ...\}$$2.
• Stochastic ensembling: At inference, stochastic neighbor subsampling across $$G \in \{\mathrm{Point}, \mathrm{Line}, \mathrm{Polyline}, \mathrm{Polygon}, ...\}$$3 ensemble members yields modest but consistent gains in $$G \in \{\mathrm{Point}, \mathrm{Line}, \mathrm{Polyline}, \mathrm{Polygon}, ...\}$$4 (e.g., $$G \in \{\mathrm{Point}, \mathrm{Line}, \mathrm{Polyline}, \mathrm{Polygon}, ...\}$$5 increases $$G \in \{\mathrm{Point}, \mathrm{Line}, \mathrm{Polyline}, \mathrm{Polygon}, ...\}$$6 by $$G \in \{\mathrm{Point}, \mathrm{Line}, \mathrm{Polyline}, \mathrm{Polygon}, ...\}$$7 vs. $$G \in \{\mathrm{Point}, \mathrm{Line}, \mathrm{Polyline}, \mathrm{Polygon}, ...\}$$8) and supports epistemic uncertainty quantification.
• GeoShapley explanation: Post-hoc model interpretation decomposes predictions into spatial and feature components, following Kernel SHAP weighting. Components include $$G \in \{\mathrm{Point}, \mathrm{Line}, \mathrm{Polyline}, \mathrm{Polygon}, ...\}$$9 (global mean), $$r_L^{G \rightarrow G'} \subset \operatorname{dom}(G) \times \operatorname{dom}(G')$$0 (spatial effect), $$r_L^{G \rightarrow G'} \subset \operatorname{dom}(G) \times \operatorname{dom}(G')$$1 (feature effect), and interaction terms. Caching renders explanations feasible on deep models, although computation remains nontrivial without GPU acceleration.

Experiments confirm that the optimized pipeline accelerates inference by 36%, and GeoAggregator explanations recover smoother, more accurate spatial effect estimates than tree-based baselines (XGBoost) on controlled synthetic data (Deng et al., 23 Jul 2025).

5. Query Languages and System Implementations

GeoAggregator frameworks introduce custom query languages and system architectures to bridge GIS and OLAP ecosystems (0707.4304):

• GISOLAP-QL: Composed queries pair a GIS selection: $\alpha_{L,D}^{A \rightarrow G}$1 with an optional OLAP part (MDX query). Spatial predicates include intersection and containment evaluated at various spatial granularities.
• Integrated architecture: Key system components include frontends (Piet-JUMP), overlay precomputation plugins, OLAP/GIS association, and backend storage on PostgreSQL + PostGIS. JTS and extended spatial libraries deliver robust geometric computations.
• Benchmarks: Sub-polygonization and overlay precomputation (20×50 grid, 4 layers) complete in 8 h 4 m. Storage overheads are moderate relative to modern server capacities (raw: 166 MB; overlay: 621 MB).

6. Data Structures for Fast Interactive Polygonal Aggregation

GeoAggregator also refers to a query-accelerated data structure approach based on fine-grained grid-cell pre-aggregation and trie-like caching (Winter et al., 2019):

• Grid covering and error bounds: Arbitrary polygons $$r_L^{G \rightarrow G'} \subset \operatorname{dom}(G) \times \operatorname{dom}(G')$$2 are approximated by the union of quadtree cells at level $$r_L^{G \rightarrow G'} \subset \operatorname{dom}(G) \times \operatorname{dom}(G')$$3 ($$r_L^{G \rightarrow G'} \subset \operatorname{dom}(G) \times \operatorname{dom}(G')$$4), with spatial error $$r_L^{G \rightarrow G'} \subset \operatorname{dom}(G) \times \operatorname{dom}(G')$$5. This enables explicit granularity-accuracy trade-off: halving cell size doubles spatial precision but quadruples computation.
• AggregateTrie cache: A trie encodes frequently encountered grid cell aggregates, accelerating repeated or skewed queries. Cache is periodically rebuilt based on observed query frequencies, with insertion based on cell- and ancestor-hit counts.
• Complexity: Pure on-the-fly SELECT queries run in $$r_L^{G \rightarrow G'} \subset \operatorname{dom}(G) \times \operatorname{dom}(G')$$6; cache-accelerated queries achieve $$r_L^{G \rightarrow G'} \subset \operatorname{dom}(G) \times \operatorname{dom}(G')$$7, where $$r_L^{G \rightarrow G'} \subset \operatorname{dom}(G) \times \operatorname{dom}(G')$$8 is cache hit rate and $$r_L^{G \rightarrow G'} \subset \operatorname{dom}(G) \times \operatorname{dom}(G')$$9 trie depth.
• Empirical results: On NYC taxi data, GeoBlocks achieve 100–1000× speedup over naive scans and remain sub-second at moderate query selectivities. Memory overhead is about 1.5× raw data, and error bounds are guaranteed by construction.

7. Topological Aggregation and Advanced Geometric Analytics

GA frameworks support topological queries (adjacency, connectivity) invariant under isotopy (0707.4304). Full common sub-polygonizations enable extraction of vertex, edge, and face relations, region-cell mappings, boundary traversals, and cyclic orders for efficient, coordinate-invariant topological analyses. These representations occupy $\alpha_{L,D}^{A \rightarrow G}$0 space and enable rapid evaluation of "touches" and higher-order topological joins.

---

References:

• "Spatial Aggregation: Data Model and Implementation" (0707.4304)
• "GeoAggregator: An Efficient Transformer Model for Geo-Spatial Tabular Data" (Deng et al., 20 Feb 2025)
• "Improving the Computational Efficiency and Explainability of GeoAggregator" (Deng et al., 23 Jul 2025)
• "GeoBlocks: A Query-Cache Accelerated Data Structure for Spatial Aggregation over Polygons" (Winter et al., 2019)