Explaining Intra-Urban Spatial Interaction with Theory-Informed Interpretable Machine Learning: Nonlinear Contributions of Complementarity, Intervening Opportunities, and Transferability

Abstract

Understanding intra-urban spatial interaction is essential for context-sensitive urban planning. While classical spatial interaction theories provide strong conceptual foundations, systematically translating these theoretical concepts into quantifiable indicators remains a significant methodological challenge. Furthermore, capturing the complex, non-linear dynamics driving urban mobility requires analytical approaches that balance predictive power with interpretability. To address this gap, this study develops a feasible, theory-informed analytical framework that bridges classical spatial interaction theory with interpretable machine learning to quantify the predictive patterns underlying intra-urban mobility. In a case study of Changchun, China, Ullman’s three core concepts, together with fundamental measures of urban scale, were systematically operationalized as a set of quantitative proxy variables based on multi-source geospatial big data. XGBoost was then used to model grid-level origin-destination flows at multiple spatial resolutions, and the SHapley Additive exPlanations (SHAP) was used to assess the contributions and dependence patterns of the theory-driven indicators. The results demonstrate the framework’s predictive robustness, with the XGBoost model consistently outperforms the traditional parametric benchmark across all evaluated spatial resolutions. The 1000 m resolution provided the best balance between predictive performance and spatial detail, yielding an R2 of 0.696, compared with 0.611 for the benchmark. The explanatory analysis indicates that functional complementarity is the most critical predictive dimension overall. It also identifies distinct nonlinear patterns, including negative associations between transfer impedance and predicted mobility flows beyond critical thresholds, positive associations between built-environment scale indicators and predicted flows only above minimum intensity thresholds, and diminishing marginal associations between intervening opportunities and predicted flows. This study provides a scalable and transferable approach for diagnosing spatial interactions in data-rich urban contexts, providing an empirical basis for calibrating future micro-level urban simulations.