An Interpretable Spatio-Temporal Graph Neural Framework Integrating Urban Economic Theory for Data-Driven Investment Prioritization in Regeneration

Abstract

This study designs and implements a novel data-driven framework for dynamically mapping investment alignment in urban regeneration, bridging the gap between urban economic theory and computational planning practice. The framework integrates parcel-level microdata on ownership, land use, building characteristics, and structural condition with contextual economic, social, and morphological data, modeling the city as a dynamic, multi-scale system. Methodologically, this research advances the state-of-the-art by combining Spatio-Temporal Graph Neural Networks with Explainable AI to simultaneously capture spatial dependencies, temporal dynamics, and provide transparent decision pathways. The proposed approach reveals four distinct investment archetypes—Heritage-Led Development Cores, Strategic Redevelopment Corridors, Transition Zones, and Stable Residential Areas—through unsupervised learning applied to Spatio-Temporal Graph Neural Networks embeddings. A key innovation lies in operationalizing theoretical concepts from Rent Gap theory and Institutional Economics into quantifiable machine learning features, enabling causal-like explanations via SHAP analysis. The framework outputs dynamic priority maps and an interactive geospatial dashboard that enables scenario testing and evidence-based resource allocation. Empirical validation demonstrates that model-generated investment archetypes showed 85% spatial concordance with areas independently identified as high-priority by a panel of senior urban planners. Additionally, the model achieved a Silhouette Score of 0.62 in cluster validation and R² of 0.92 in surrogate model performance, indicating strong analytical robustness. This research contributes to both urban science and planning practice by providing a scalable, interpretable approach for investment prioritization in complex urban environments.