GeoAggregator: Geospatial Data Aggregation

• GeoAggregator is a framework that integrates transformer-based deep learning with spatial bias techniques for efficient and interpretable geospatial regression.
• It employs an optimized data-loading pipeline and fused Gaussian bias computations, achieving approximately 36% faster inference on large-scale datasets.
• Its inherent ensembling with GeoShapley explainability provides principled uncertainty quantification and robust spatial attribution for enhanced model trustworthiness.

GeoAggregator refers to a suite of methodologies and specialized systems for efficient, expressive, and explainable aggregation of geospatial data. The term encompasses transformer-based deep learning architectures for geospatial tabular data regression, optimized pipelines for spatial data loading and model inference, formal aggregation strategies for spatial statistics, and ensembling and explainability frameworks tightly integrated with geospatial inductive biases. The most mature instantiation, as captured in the GeoAggregator system and its subsequent computational and explainability enhancements, demonstrates state-of-the-art predictive accuracy and computational efficiency alongside advanced spatial model interpretability (Deng et al., 20 Feb 2025, Deng et al., 23 Jul 2025).

1. Core Architecture and Geospatial Inductive Biases

GeoAggregator is an attention-based architecture explicitly designed for supervised regression on geospatial tabular data (GTD). Each observation (row) is treated as a token, with the model directly attending to the K spatially nearest neighbors—eschewing proxy grids, explicit graphs, or heavy preprocessing. The vanilla attention kernel is extended with a Gaussian spatial bias to encode spatial autocorrelation and heterogeneity:

$$\alpha_{ij} = \mathrm{softmax} \, ( (W^Q e_i)^\top (W^K e_j) - \lambda d_{ij}^2 )$$

$e_i' = \sum_j \alpha_{ij} W^V e_j$

$\tilde{e}_i = W^O e_i'$

where $W^{Q,K,V,O}$ are learnable projections, $d_{ij}$ is the Euclidean distance between points $p_i$ and $p_j$, and $\lambda$ is a learnable attention bias factor. Notably, the most recent version supports $\lambda^{(h)}$ per attention head for increased expressivity:

$$\alpha_{ij}^{(h)} = \mathrm{softmax} \, ( (W_{(h)}^Q e_i)^\top W_{(h)}^K e_j - \lambda^{(h)} d_{ij}^2 )$$

Rotary positional embeddings are incorporated into queries and keys, encoding continuous spatial coordinates without artificial discretization (Deng et al., 20 Feb 2025, Deng et al., 23 Jul 2025).

2. Computational Optimization and Scalability

Scalability for large GTD is achieved through two primary engineering improvements:

• Optimized Data-Loading: Datasets are partitioned into a context pool and a query pool. Instead of performing a $k$-d-tree search on each forward pass (which incurs $O(N\,m\,\log N)$ I/O complexity), nearest neighbor sets are precomputed and cached for each query point. Neighbor lookup is thus reduced to a constant-time ($O(1)$) table lookup; overall I/O complexity becomes $O(N\,m)$.
• Streamlined Forward Pass: All per-head Gaussian bias computations are fused into a batched matrix-multiplication, eliminating Python loops and optimizing multi-head attention. Furthermore, the use of induced (“global”) tokens reduces the effective attention complexity from quadratic ($O(m^2 d)$) to near-linear ($O(m\,r\,d)$), where $r \ll m$ is the number of global tokens. This ensures end-to-end inference scales nearly linearly in sequence length $m$ (Deng et al., 23 Jul 2025).

Compared to naïve implementations, these optimizations resulted in ∼36% faster inference and superior scaling properties for synthetic spatial regression benchmarks.

3. Model Ensembling and Uncertainty Quantification

GeoAggregator introduces an intrinsic model ensembling mechanism at inference time:

• For each prediction, $M$ context sets are drawn by slightly perturbing the search radius and subsampling to $K$ neighbors.
• Each context set results in a model evaluation $f_k(x)$, with the ensemble prediction given by

$$\hat{y}_{\mathrm{ens}}(x) = \frac{1}{M} \sum_{k=1}^M f_k(x)$$

Variance of the ensemble estimator reduces by a factor of $1/M$ (i.e., $$\mathrm{Var}[\hat{y}_{\mathrm{ens}}] = \mathrm{Var}[f]/M$$), while bias remains basically unchanged. Empirically, this reduced mean absolute error (MAE) and slightly improved $R^2$ across multiple spatial synthetic datasets (MAE from 1.149 → 1.135, $R^2$ from 0.841 → 0.844 for $M=1$ → 8) (Deng et al., 23 Jul 2025). This ensembling thus provides principled uncertainty quantification.

4. Explainability via GeoShapley and Post-Hoc Decomposition

GeoAggregator incorporates GeoShapley, a novel adaptation of the Shapley-value framework for spatial models. Model predictions are post-hoc decomposed into:

$$\hat{y} = \phi_0 + \phi_{\mathrm{GEO}} + \sum_{j=1}^p \phi_j + \sum_{j=1}^p \phi_{(\mathrm{GEO},j)}$$

where $\phi_0$ is a baseline, $\phi_{\mathrm{GEO}}$ captures the pure spatial effect, $\phi_j$ captures marginal non-spatial effects, and $\phi_{(\mathrm{GEO},j)}$ encodes spatially-varying interactions. Each term is computed via kernel SHAP weighting to ensure Shapley consistency.

A practical predict-wrapper ensures that, even when spatial features are masked, the neighborhood structure is preserved, making the GeoShapley decomposition actionable in large-scale GTD settings. Experiments demonstrated that GeoAggregator’s GeoShapley explanations smoothly recovered spatial coefficient surfaces, whereas alternative approaches (e.g., XGBoost’s SHAP) resulted in noisy or discontinuous attribution (Deng et al., 23 Jul 2025).

5. Empirical Performance and Comparative Evaluation

GeoAggregator achieves or matches state-of-the-art results on diverse spatial regression challenges:

• On synthetic datasets representing complex spatial processes (e.g., spatial lag, geographically weighted regression), the optimized GeoAggregator outperforms or ties established baselines (XGBoost, spatial GCNs, GWR) in both MAE and $R^2$ metrics (Deng et al., 20 Feb 2025, Deng et al., 23 Jul 2025).
• On real datasets—PM$_{2.5}$ (China), US county poverty, and King County housing—GeoAggregator exhibits competitive or best MAE and $R^2$ versus deep learning and statistical models, with substantially lower parameter counts and FLOPs.
• Efficiency benchmarks indicate the parameter count (e.g. 4.3K–6.3K parameters) and inference cost are one to two orders of magnitude lower than graph-CNN/spatial CNN approaches.

Ablation analyses reveal that optimal performance is achieved for intermediate values of spatial bias ($\lambda \approx 0.5$–5), and that performance is robust to increasing receptive field size ($\ell_{\max}$), up to the limit of marginal return (Deng et al., 20 Feb 2025).

6. Practical Relevance and Software Ecosystem

The GeoAggregator framework is distributed via the open-source GA-sklearn package, which integrates:

• Optimized data-loading and forward pass for fast batching.
• Drop-in scikit-learn compatibility.
• Built-in model ensembling capability.
• Turnkey GeoShapley explainability functions.

This architecture makes transformer-grade geospatial regression feasible on commodity hardware, supports robust spatial prediction under computational constraints, and provides interpretable and uncertainty-aware outputs critical for environmental and social-science applications (Deng et al., 23 Jul 2025).

7. Significance and Outlook

GeoAggregator uniquely bridges spatial statistics and modern attention-based deep learning. By encoding spatial autocorrelation and heterogeneity directly in the attention kernel, supporting efficient computation and principled explanation, GeoAggregator establishes a practical, theoretically grounded tool for next-generation geospatial science.

Empirical evidence supports claims that GeoAggregator not only offers the highest or comparable predictive accuracy to both statistical and machine learning competitors but does so with superior computational efficiency, model compactness, and spatial explainability. A plausible implication is the standardization of transformer-based geospatial pipelines for applied spatial analysis, especially in policy-relevant contexts where robust uncertainty quantification and interpretability are as important as point accuracy (Deng et al., 20 Feb 2025, Deng et al., 23 Jul 2025).