This paper introduces a comprehensive methodological framework for analyzing non-linear unbalanced network panels, addressing challenges in complex non-linear spatio-temporal dependencies, unbalanced panel structures, and network endogeneity and heterogeneity. The Adaptive Spatio-Temporal Graph Network (ASTGN) architecture integrates graph neural networks with temporal modeling to capture dynamic network effects, including mechanisms for handling missing data, non-linear dynamics, and cross-sectional heterogeneity.
Key findings
ASTGN architecture captures dynamic network effects through unified spatio-temporal aggregation.
Novel learned imputation mechanism addresses unbalanced panels by leveraging temporal continuity and network structure.
Theoretical foundations provide convergence guarantees and consistency results under realistic assumptions.
Rigorous experimental validation framework evaluates performance across varying degrees of non-linearity and missing data rates.
Limitations & open questions
Further research needed to extend the framework to other types of network data and structures.