Heterogeneous Graph Neural Networks with Post-hoc Explanations for Multi-modal and Explainable Land Use Inference

Explainability is essential for urban planning decisions: The paper demonstrates that traditional deep learning models lack the transparency needed for long-term policy decisions in city planning. By combining graph neural networks with explainable AI techniques, the framework provides both accurate predictions and interpretable reasoning—a critical requirement when decisions affect entire communities.

Spatial relationships matter for understanding cities: The research shows that existing approaches often treat urban data points independently, missing important connections between neighboring areas. By using heterogeneous graph neural networks that capture spatial correlations and different types of services, the model achieved significantly better performance, particularly for predicting office and sustenance locations.

Multi-modal data reveals human activity patterns: The study leverages multiple data sources (sensor and location technologies) to understand daily activity patterns, demonstrating that combining different types of information provides richer insights into how cities actually function—information that can improve evidence-based urban planning.

Feature attribution explanations validate real-world logic: The analysis of which factors the model uses to make predictions revealed that "residence" and "work" categories showed symmetrical patterns aligned with actual commuter behavior, suggesting the model learned meaningful urban dynamics rather than spurious correlations—an important validation for trustworthy AI in policy contexts.