This study proposes the spatial auto regressive scalar-on-function neural network (SARSoFNN) to simultaneously address non-linear relationships, spatial dependency, and functional covariates. SARSoFNN employs basis expansion for functional inputs and ...
This study proposes the spatial auto regressive scalar-on-function neural network (SARSoFNN) to simultaneously address non-linear relationships, spatial dependency, and functional covariates. SARSoFNN employs basis expansion for functional inputs and a decorrelation loss function via iterative optimization of the spatial autoregressive parameter ρ. Simulations verify its superior predictive performance and less sensitivity to spatial dependence compared to the traditional functional linaer model(FLM) and the functional neural network(FNN). Empirical analysis of Seoul commercial district data further validates its practical utility. This framework offers a novel deep learning approach for integrally modeling spatial and functional complexities.