Nonlinear river distance effects on housing prices: Integrating XGBoost and GeoShapley for spatially explicit valuation

Abstract

Urban river landscapes not only provide recreational and aesthetic benefits but also generate tangible market value. Quantifying the externalities of such amenities on housing prices is critical for evidence-based policy making and spatial planning. Focusing on Tianjin's six central districts, this study integrates conventional spatial models with interpretable machine learning, to identify the nonlinear influence of river water systems on housing prices in Tianjin and the threshold effect. Using 2020 data from the early pandemic period captures pre-regulation market conditions while avoiding structural distortions from post-pandemic policy interventions during 2021–2022. We have compared OLS, GWR and XGBoost; the latter increased R 2 by 10.4 % (CV-R 2 = 0.712). The winning XGBoost model is subsequently combined with GeoShapley to visualize local marginal effects. Results reveal a three-regime, non-linear distance–price gradient for proximity to the Haihe River and its tributaries, emergent transition zones approximately within 1.0–1.5 km, with suggestive inflection points around 1.0 km and 1.3 km: within 1 km, a 1 % reduction in distance raises prices by 0.155 %; with a 0.074 % increase per 1 % additional distance in the 1–1.3 km sub-zone. These findings provide quantitative benchmarks for tiered river buffer management in Tianjin's Territorial Spatial Master Plan (2021–2035): intensive amenity investment within 1 km, standard regulations for transitional zones (1–1.5 km), and accessibility-focused planning beyond 1.5 km. Spatial heterogeneity across river segments further indicates that uniform zoning is suboptimal, with eastern and northern corridors warranting targeted investment and western and southern segments requiring functional repositioning. Future work will incorporate resident surveys and temporal variation through XGBoost-STSHAP to enable dynamic, space-time analysis.