The Use of Graph Neural Networks in Rail Transport Planning

This work explores how graph theory and graph neural networks can support the strategic planning of rail network expansions using only publicly available city data, applied to the São Paulo Metropolitan Region. The methodology consolidates information from multiple public sources, develops a catchment-area formula to estimate potential passenger demand, applies Random Forest to identify the most relevant demographic features, and implements a GraphSAGE model that derives predictive capability from network topology together with socioeconomic features and origin–destination trips. The demand approximation was checked against observed station boardings, with predicted and observed rankings in agreement. The GraphSAGE model achieved an R2 of 0.874 ± 0.042 when predicting the proxy demand indicator, with minimal overfitting, outperforming the Random Forest baseline and achieving accuracy comparable to an XGBoost baseline while overfitting substantially less; this performance remained stable under spatial cross-validation. The model is computationally efficient and requires no rail-system-specific information beyond topology, making it suitable for the fast, low-cost comparison of expansion proposals rather than as a replacement for detailed transport demand models. It was used to evaluate eleven real projects and proposals for the São Paulo Metropolitan Region. Employment, residences, and destinations where people go to eat together represent about 65% of the model’s predictive capacity.