Geographically Neural Network Weighted Regression

• GNNWR is a hybrid spatial regression model that replaces fixed kernels with a neural network-driven weighting mechanism for localized parameter estimation.
• It enhances flexibility by learning variable-specific spatial weights that capture non-monotonic and multi-scale spatial influences, outperforming classical GWR.
• Empirical studies, such as the Shenzhen housing valuation, demonstrate GNNWR’s superior performance with reduced RMSE and improved interpretability of spatial heterogeneity.

Searching arXiv for the cited GNNWR-related papers to ground the article in the current literature. Geographically Neural Network Weighted Regression (GNNWR), also rendered in the cited literature as Geographical Neural Network Weighted Regression, is a hybrid spatial regression framework that replaces the fixed spatial kernel of Geographically Weighted Regression (GWR) with a learned weighting mechanism implemented by a neural network, while preserving a locally weighted regression interpretation. In the Shenzhen housing study, GNNWR was introduced as a response to spatial heterogeneity in real-estate markets and as a way to avoid the rigidity of Gaussian or bi-square kernels and the difficulty of bandwidth selection in classical GWR (Wang et al., 2022). Subsequent work situated GNNWR within a broader family of geographically weighted neural models, including geographically and temporally weighted neural networks and geographically weighted layers embedded in fully nonlinear architectures (Li et al., 2018, Cao et al., 1 Apr 2025).

1. Conceptual position within spatial regression

GNNWR is best understood against the contrast between two established modeling traditions. Classical GWR augments ordinary least squares with location-specific weighting, thereby allowing coefficients to vary over space, but it remains linear in the predictors and depends on a pre-specified distance-decay kernel with a scalar bandwidth. Standard artificial neural networks can approximate nonlinear predictor-response relationships, yet they ordinarily treat observations as exchangeable and do not encode location-driven parameter variation. The GNNWR formulation described in the literature combines these traditions by learning geography-dependent weights from data rather than fixing them a priori (Wang et al., 2022, Cao et al., 1 Apr 2025).

The original Shenzhen application emphasizes several motivations for this substitution. GWR’s kernel function is described as non-linear, elusive, and complex to opt bandwidth, and its predictive power is presented as improvable. GNNWR therefore uses neural networks to learn a more flexible spatial weighting function, with the stated goal of improving real-estate appraisal accuracy and materializing weight distributions of different variables that classical GWR is difficult to recover. The same study explicitly frames the method as extending application from natural to socioeconomic geospatial data (Wang et al., 2022).

Related spatiotemporal work prefigures and complements this development. Geographically and temporally weighted neural networks (GTWNNs) were developed for satellite-based PM$_{2.5}$ estimation by combining nonlinear neural prediction with a continuous Gaussian weight in both space and time. Later work on crime prediction and PM$_{2.5}$ estimation further generalized the geographically weighted neural idea through weighted losses, history-dependent modules, and specialized geographically weighted layers (Li et al., 2018, Fisk et al., 27 Mar 2025, Cao et al., 1 Apr 2025).

2. Core mathematical formulation

In the Shenzhen formulation, let $n$ denote the number of observations, $p$ the number of explanatory variables, $X \in \mathbb{R}^{n \times (p+1)}$ the design matrix with leading column of ones, and $y \in \mathbb{R}^n$ the response. For location $i$ with coordinates $(u_i,v_i)$, define the distance vector

$$d_i = [d_{i1}, d_{i2}, \ldots, d_{in}]^\top \in \mathbb{R}^n,$$

where $d_{ij}$ is the Euclidean distance between observations $_{2.5}$0 and $_{2.5}$1. Classical GWR forms

$_{2.5}$2

and computes

$_{2.5}$3

GNNWR replaces the fixed kernel $_{2.5}$4 with a neural network $_{2.5}$5 that takes $_{2.5}$6 as input and outputs $_{2.5}$7 nonnegative local weights,

$_{2.5}$8

These are assembled into

$_{2.5}$9

yielding

$n$0

Equivalently, the model may be viewed as weighted least squares in which the weight assigned to variable $n$1 at location $n$2 is $n$3 (Wang et al., 2022).

This parameterization is notable because the learned weights are variable-specific and location-dependent. The Shenzhen paper argues that this allows GNNWR to capture non-monotonic or multi-scale spatial influences, whereas GWR’s kernel is monotonic in distance and rigidly controlled by a scalar bandwidth. The reported theoretical advantage is therefore not merely kernel replacement, but replacement of a single prescribed decay law by a learned mapping $n$4 that adapts automatically to the data (Wang et al., 2022).

Later literature broadens the term “GNNWR” beyond this specific algebraic form. One strand describes it as a hybrid OLS $n$5 neural network structure that mainly learns nonlinear spatial weights while retaining a linear output layer; another describes a weighted neural network trained by geographically or spatiotemporally weighted losses; still another embeds geographical weighting directly inside hidden layers. This suggests that, in subsequent usage, GNNWR denotes a family of geographically weighted neural formulations rather than a single immutable estimator (Cao et al., 1 Apr 2025, Fisk et al., 27 Mar 2025).

3. Network architecture, loss, and estimation procedure

The Shenzhen implementation uses a compact fully connected network to generate spatial weights. Its input layer receives the $n$6-dimensional Euclidean-distance vector $n$7. The network contains four fully connected hidden layers of sizes $n$8, with PReLU and Batch-Normalization in each hidden layer. Dropout with $n$9 is applied after each hidden layer to improve generalization, the output layer contains $p$0 neurons with a non-negative activation such as softplus or ReLU, and weights are initialized by He’s method (Wang et al., 2022).

Once $p$1 has been produced, prediction follows

$p$2

The trainable parameters $p$3 are estimated by minimizing the root-mean-square error over a training batch $p$4: $p$5 No additional regularization term was used beyond dropout. Optimization was performed with Adam, with learning rate approximately $p$6, $p$7, $p$8, batch size $p$9, early stopping after $X \in \mathbb{R}^{n \times (p+1)}$0 iterations without validation-loss improvement, and a maximum of $X \in \mathbb{R}^{n \times (p+1)}$1 epochs (Wang et al., 2022).

The training protocol is also tightly specified. Coordinates were transformed to WGS1984/50N projection. Independent variables were normalized to zero mean and unit variance, and the response variable, reference price in RMB/m$X \in \mathbb{R}^{n \times (p+1)}$2, was also normalized for training. The dataset of $X \in \mathbb{R}^{n \times (p+1)}$3 samples was split by randomly holding out $X \in \mathbb{R}^{n \times (p+1)}$4 observations ($X \in \mathbb{R}^{n \times (p+1)}$5) as a final test set, while the remaining $X \in \mathbb{R}^{n \times (p+1)}$6 were used for 10-fold cross-validation. Each fold used approximately $X \in \mathbb{R}^{n \times (p+1)}$7 points, with 9 folds for training the spatially weighted neural network and 1 fold for validation; losses over the validation folds were summed for hyperparameter selection, including network size, dropout, and learning rate. Comparator GWR models used bi-square and Gaussian kernels, with adaptive bandwidth selected by AICc through golden-section search over neighbor counts $X \in \mathbb{R}^{n \times (p+1)}$8 (Wang et al., 2022).

4. Shenzhen housing valuation and spatial heterogeneity

The Shenzhen case study is the canonical socioeconomic application of GNNWR in the cited literature. The dataset consists of $X \in \mathbb{R}^{n \times (p+1)}$9 residential complexes across nine districts. The normalized endogenous variables are Age of building (AB), Number of parking spots (NPS), Management fee (MF), Number of buildings in complex (NB), Green ratio (GR), and Plot ratio (PR). The normalized environmental variables are Distance to nearest coast (SD), Quality of public school (QAPS), Number of subway stations within $y \in \mathbb{R}^n$0 km (NSS), and Distance to nearest subway station (DSS). OLS and GWR were implemented in ArcGIS Pro, and GNNWR in TensorFlow 1.15 on Python 3.6 (Wang et al., 2022).

On merged validation results from 10-fold cross-validation, OLS achieved $y \in \mathbb{R}^n$1, RMSE $y \in \mathbb{R}^n$2, MAE $y \in \mathbb{R}^n$3, MAPE $y \in \mathbb{R}^n$4, and Pearson $y \in \mathbb{R}^n$5. GWR improved these figures to $y \in \mathbb{R}^n$6, RMSE $y \in \mathbb{R}^n$7, MAE $y \in \mathbb{R}^n$8, MAPE $y \in \mathbb{R}^n$9, and Pearson $i$0. GNNWR further improved them to $i$1, RMSE $i$2, MAE $i$3, MAPE $i$4, and Pearson $i$5. On the held-out test set, GWR obtained $i$6, RMSE $i$7, MAE $i$8, MAPE $i$9, and Pearson $(u_i,v_i)$0, whereas GNNWR obtained $(u_i,v_i)$1, RMSE $(u_i,v_i)$2, MAE $(u_i,v_i)$3, MAPE $(u_i,v_i)$4, and Pearson $(u_i,v_i)$5. The study reports that GNNWR reduces RMSE by approximately $(u_i,v_i)$6 relative to GWR and approximately $(u_i,v_i)$7 relative to OLS (Wang et al., 2022).

Model	10-fold CV	Held-out test
GWR	$(u_i,v_i)$8, RMSE $(u_i,v_i)$9, MAE $$d_i = [d_{i1}, d_{i2}, \ldots, d_{in}]^\top \in \mathbb{R}^n,$$0, MAPE $$d_i = [d_{i1}, d_{i2}, \ldots, d_{in}]^\top \in \mathbb{R}^n,$$1, $$d_i = [d_{i1}, d_{i2}, \ldots, d_{in}]^\top \in \mathbb{R}^n,$$2	$$d_i = [d_{i1}, d_{i2}, \ldots, d_{in}]^\top \in \mathbb{R}^n,$$3, RMSE $$d_i = [d_{i1}, d_{i2}, \ldots, d_{in}]^\top \in \mathbb{R}^n,$$4, MAE $$d_i = [d_{i1}, d_{i2}, \ldots, d_{in}]^\top \in \mathbb{R}^n,$$5, MAPE $$d_i = [d_{i1}, d_{i2}, \ldots, d_{in}]^\top \in \mathbb{R}^n,$$6, $$d_i = [d_{i1}, d_{i2}, \ldots, d_{in}]^\top \in \mathbb{R}^n,$$7
GNNWR	$$d_i = [d_{i1}, d_{i2}, \ldots, d_{in}]^\top \in \mathbb{R}^n,$$8, RMSE $$d_i = [d_{i1}, d_{i2}, \ldots, d_{in}]^\top \in \mathbb{R}^n,$$9, MAE $d_{ij}$0, MAPE $d_{ij}$1, $d_{ij}$2	$d_{ij}$3, RMSE $d_{ij}$4, MAE $d_{ij}$5, MAPE $d_{ij}$6, $d_{ij}$7

The same study uses GNNWR to diagnose spatial non-stationarity. A global $d_{ij}$8 test rejects spatial smoothness at $d_{ij}$9, and variable-wise $_{2.5}$00 tests show that all nine coefficients vary significantly in space at $_{2.5}$01. Among average normalized local weights, the largest absolute mean weights are reported for distance to coast, management fee, age, number of parking spots, and distance to nearest subway station. The greatest spatial variation, measured by the standard deviation of local weights, occurs for DSS, MF, SD, and AB (Wang et al., 2022).

The mapped local effects further illustrate how GNNWR is used as an interpretive device. Intercept premium peaks in Nanshan’s Houhai area and Futian’s Xiangmi Lake shore. Building age has its strongest negative effect in new-supply corridors of Nanshan and Futian, but flips mildly positive around the oldest urban cores. Management fee is strongest in southwestern Nanshan and coastal Bao’an and negative near border districts. Sea-distance has its largest negative effect in central districts and a smaller effect in suburbs. Distance to nearest subway station is generally negative, but in employment hubs such as Huaqiang North and Shekou, more distant buildings carry a premium attributed in the study to fewer nuisances. The weight pattern of NSS is almost the mirror of DSS, and school quality has a strong effect only in the old-city pocket of western Luohu, with modest effects elsewhere (Wang et al., 2022).

5. Spatiotemporal variants and later architectural generalizations

A closely related development is the geographically and temporally weighted neural network, introduced for satellite-based PM$_{2.5}$02 mapping. In that formulation, prediction at grid cell $_{2.5}$03 on day $_{2.5}$04 is a locally adaptive function

$_{2.5}$05

implemented by a generalized regression neural network whose training samples are weighted by spatiotemporal proximity. The weight is

$_{2.5}$06

and prediction multiplies this weight by the GRNN radial-basis activation. The reported final hyperparameters were $_{2.5}$07, adaptive $_{2.5}$08 corresponding to average $_{2.5}$09 km·day, and GRNN spread $_{2.5}$10. Sample-based and site-based 10-fold cross-validation produced $_{2.5}$11 and $_{2.5}$12, with RMSE $_{2.5}$13 and $_{2.5}$14 $_{2.5}$15g/m$_{2.5}$16, outperforming GWGRNN, original GRNN, GTWR, and daily GWR (Li et al., 2018).

A further line of work extends the GTWNN framework for crime prediction. In this family of models, the prediction is expressed through a coefficient-network $_{2.5}$17, a feature-network $_{2.5}$18, their element-wise combination, and a final output function. Training minimizes a weighted MSE

$_{2.5}$19

with Gaussian spatial and temporal kernels. Three mathematical extensions are described: a spatially expanded output with spatially weighted loss (“GTWNN$_{2.5}$20”), a spatiotemporal output and spatiotemporal loss (“GTWNN$_{2.5}$21”), and a history-dependent module (“HDGTWNN”) that allows coefficients at time $_{2.5}$22 to depend explicitly on coefficients and factors at $_{2.5}$23. On monthly London crime data, GTWNN reduced MSE to approximately $_{2.5}$24 with $_{2.5}$25 up to $_{2.5}$26, while HDGTWNN achieved $_{2.5}$27 up to $_{2.5}$28 with MSE approximately $_{2.5}$29. On daily Detroit crime data, GTWNN$_{2.5}$30 achieved the best MSE at approximately $_{2.5}$31, whereas the full spatiotemporal expansion was reported as largely redundant when the history module was included. The paper’s summary states that the history-dependent module generally surpassed one of the other extensions and that model suitability depended strongly on geographic and temporal context (Fisk et al., 27 Mar 2025).

By 2025, GNNWR was also recast inside a broader deep-learning framework termed the Artificial Geographically Weighted Neural Network (AGWNN). AGWNN is described as generalizing both GWR and ANN by retaining a fully nonlinear deep network backbone, inserting a Geographically Weighted Layer (GWL), and introducing trainable Geographical Weight Adjusters (GWA). For geographical neuron $_{2.5}$32 and sample $_{2.5}$33, the GWL output is

$_{2.5}$34

with Gaussian kernel

$_{2.5}$35

and adjusted weight

$_{2.5}$36

where $_{2.5}$37 is trainable. In synthetic experiments, training RMSE and $_{2.5}$38 were reported as approximately $_{2.5}$39 and $_{2.5}$40 for GNNWR versus approximately $_{2.5}$41 and $_{2.5}$42 for AGWNN; on $_{2.5}$43 new datasets, mean RMSE and $_{2.5}$44 were approximately $_{2.5}$45 and $_{2.5}$46 for GNNWR versus approximately $_{2.5}$47 and $_{2.5}$48 for AGWNN. In a real-world PM$_{2.5}$49 case over the contiguous United States, AGWNN also improved seasonal RMSE and adjusted $_{2.5}$50 relative to GNNWR and comparator methods (Cao et al., 1 Apr 2025).

6. Inductive bias, interpretation, and open directions

Later work explicitly frames GNNWR as a problem of inductive bias. The critique is that existing implementations, although more flexible than classical GWR, often remain restricted by fixed distance-based schemes and limited architectural priors. In response, a hybrid model was proposed that augments the learned weighting function with local receptive fields from convolutional neural networks, sequential spatial context from bidirectional recurrent networks, and global self-attention from Transformer-style blocks. In that design, neighbor features are embedded, processed through CNN, RNN, and attention branches, fused by an MLP, and normalized by a softmax to produce spatial weights (Chen, 14 Jul 2025).

The reported synthetic benchmark isolates heterogeneity, noise, and sample size. Average test RMSE under low, medium, and high heterogeneity was $_{2.5}$51, $_{2.5}$52, and $_{2.5}$53 for GWR; $_{2.5}$54, $_{2.5}$55, and $_{2.5}$56 for GNNWR with an MLP weighting module; and $_{2.5}$57, $_{2.5}$58, and $_{2.5}$59 for the hybrid CNN+RNN+attention model. Across settings, the hybrid model reportedly reduced RMSE by $_{2.5}$60–$_{2.5}$61 relative to GNNWR and $_{2.5}$62–$_{2.5}$63 relative to GWR. The paper further states that local models excel in highly heterogeneous or small-sample scenarios, whereas global models perform better with larger, more homogeneous data (Chen, 14 Jul 2025).

These later analyses sharpen two recurring interpretive points. First, GNNWR is not merely “an ANN with coordinates”; its distinctive feature is a geography-aware weighting mechanism, whether implemented as learned local coefficients, spatiotemporally weighted sample contributions, geographically weighted losses, or dedicated geographically weighted layers. Second, flexibility alone does not resolve the modeling problem: performance depends on how inductive bias matches the data regime. The cited future directions therefore emphasize learnable spatial weighting functions beyond Euclidean distance, hybrid neural architectures, dynamic gating across inductive biases, and improved interpretability through visualization of learned filters and attention maps (Chen, 14 Jul 2025).

Taken together, the literature presents GNNWR as a bridge between local spatial regression and neural function approximation. In its original Shenzhen form it preserves a local-linear structure while learning variable-specific spatial weights; in its spatiotemporal and later deep variants it becomes a broader design pattern for encoding nonstationarity inside neural predictors. The consistent theme is the replacement of rigid spatial weighting assumptions with learned, geography-aware weighting operators, while retaining the analytical objective of mapping how predictor effects vary across space or across space-time (Wang et al., 2022, Li et al., 2018, Cao et al., 1 Apr 2025).