A1406
Title: Geographically weighted quantile XGBoost: Nonlinear quantile learning for spatial heterogeneity
Authors: Vivian Yi-Ju Chen - National Chengchi University (Taiwan) [presenting]
Yi-Ling Dai - National Chengchi University (Taiwan)
Abstract: Spatial heterogeneity and distributional heterogeneity are two characteristics frequently considered in spatial data analysis. Geographically weighted quantile regression (GWQR) has been developed to jointly address these issues by allowing regression relationships to vary across locations while estimating conditional quantiles of the response variable. However, GWQR relies on linear modeling assumptions, which may limit its flexibility in capturing complex nonlinear relationships and high-order interactions in spatial data. Motivated by this limitation, geographically weighted quantile XGBoost (GWQXGB) is proposed as a novel spatial machine learning approach to extend the capabilities of GWQR. The proposed approach integrates spatial kernel weighting into the quantile boosting optimization process, enabling localized nonlinear quantile learning for spatially heterogeneous data. To enhance interpretability, SHapley Additive exPlanations (SHAP) values are incorporated to quantify localized variable contributions across different quantile levels. The performance of GWQXGB is evaluated through simulation experiments and compared with existing geographically weighted and tree-based quantile regression methods. Finally, the practical utility of GWQXGB is demonstrated using empirical spatial datasets.
