Interpretable Graph Deep Learning Reveals Ecological Risks and Attribution Patterns of Polycyclic Aromatic Hydrocarbons in Urban Greenspace Soils across China

Abstract

Characterizing the complex environmental and socioeconomic associations underlying risks related to polycyclic aromatic hydrocarbons (PAHs) in urban greenspace soils is important for ecosystem health, yet conventional tabular models do not explicitly represent these cross-domain dependencies. In this study, we present an interpretable heterogeneous graph attention network (Hetero-GAT) framework for PAH-related risk assessment in urban greenspace soils across China. By integrating SHapley additive explanations (SHAP) with attention-derived graph salience, our dual-perspective framework reveals clear heterogeneity in model-derived PAH risk attribution patterns. Geographical and meteorological variables form recurrent background attribution signals across the five risk end points, whereas energy, infrastructure, and traffic/logistics variables contribute more selectively to end point-specific patterns. Under grouped validation and unified benchmarking, Hetero-GAT achieved the highest mean overall predictive performance, although its numerical advantage over GCN was modest, indicating that its principal value lies in structured representation and attribution rather than predictive superiority alone. Furthermore, a city-level typology analysis, based on the aggregation of sample-level observations before clustering, identifies three urban structural profiles with distinct risk stratifications and attribution patterns. Mixed-profile and logistics/service-dominated cities exhibit higher risk distributions than infrastructure-provision-dominated cities, with the logistics/service-dominated group showing the highest median values across all five end points. These findings indicate that interpretable graph modeling can complement concentration-based monitoring with risk-oriented contextual information and support differentiated risk screening across urban structural profiles.