A1651
Title: Flexible spatial modeling with variable selection: A higher-order nonparametric spatial autoregressive framework
Authors: Sheng-Yueh Chang - National Cheng Kung University (Taiwan) [presenting]
Kuo-Jung Lee - National Cheng-Kung University (Taiwan)
Abstract: Deep neural networks have attracted increasing attention because of their strong predictive performance and flexibility in modeling complex nonlinear relationships. A higher-order nonparametric spatial autoregressive framework with variable selection for spatial data analysis is proposed. Unlike conventional spatial autoregressive models that mainly rely on prespecified spatial weight matrices, the proposed method uses neural networks to learn multiple spatial weighting mechanisms from the data and to model nonparametric endogenous spatial effects. To perform variable selection while preserving nonlinear modeling flexibility, an L1 penalty and a hierarchical sparsity constraint are imposed on the main-effect branch. A regularization-path training procedure with a hierarchical proximal update is developed to estimate the model and identify important variables. Simulation studies show that the proposed method can recover relevant variables and achieve competitive predictive performance under nonlinear and multiple spatial dependence structures. An empirical study using the California Housing Prices data further demonstrates its ability to improve prediction and identify meaningful explanatory variables. Overall, the proposed framework integrates flexible spatial mechanism learning, nonparametric endogenous spatial effects, and interpretable variable selection in a unified deep learning model.
