Beyond Geographic Contiguity: A Hybrid Spatiotemporal Deep Learning–Spatial Econometric Framework for Detecting Functional Economic Networks in Algeria

Introduction: Official Algerian GDP statistics are released with delays and limited subnational detail. Monthly Nighttime Light (NTL) intensity therefore provides a validated, high-resolution proxy for regional economic activity. Conventional spatial models, however, rely on fixed geographic weight matrices (e.g.,contiguity or distance) that may not capture latent functional ties driven by production systems, infrastructure corridors, and interregional market linkages.
Methods: We propose a hybrid framework that links spatiotemporal deep learning with structural spatial econometrics. Using monthly log-transformed NTL data (2012–2024), we estimate a customized Graph WaveNet on the non-hydrocarbon subsystem (40 wilayas) and extract a learned interaction matrix with constrained self-loops (diag(W)=0), quantile sparsification, and row standardization. We then compare the learned matrix with geographic baselines using dynamic Moran’s I and integrate it into SAR specifications, including a richer-control robustness model.
Results: Geographic baselines show weak-to-negative global autocorrelation (mean Moran’s I: KNN =-0.1071, Queen =-0.0256, distance-decay =-0.0256), while the learned functional matrix yields a strong positive structure (mean Moran’s I = 0.3619) and significance across all evaluated time points. The learned network is moderately sparse (density = 0.400), which is consistent with concentrated but non-local spillover channels. In the baseline SAR, the spatial lag parameter is large and significant (rho = 0.6026, p < 0.001; pseudo-R2 = 0.2793; AIC = 71.985). Under the richer-control specification, model fit increases markedly (pseudo-R2 = 0.9991; AIC =-189.002), with a smaller but still positive residual spatial dependence (rho = 0.0366).
Conclusion: Algerian regional dynamics are not fully explained by administrative proximity alone; functional economic connectivity provides substantial additional signals. The proposed workflow offers a reproducible bridge between data-driven network discovery and interpretable spatial econometrics, supporting policy that targets strategic corridors, spillover hubs, and regionally coordinated investment priorities.