Reproduce “Explainable AI in Spatial Analysis” on OpenResearch

About this paper

Ziqi Li

This paper investigates integrating Explainable AI techniques with spatial analysis to enhance the interpretability of machine learning models in geographic contexts. It introduces GeoShapley, an approach extending Shapley values to decompose feature contributions into non-geographical and geographical components. A case study on 2020 US Presidential election voting patterns demonstrates that GeoShapley identifies key sociodemographic predictors and reveals spatially varying relationships consistent with traditional spatial statistical methods.

Key insights

• Explainable AI (XAI) methods, particularly model-agnostic approaches like GeoShapley, can bridge the gap between high-performance machine learning models and the need for interpretability in spatial analysis.
• It is possible to decompose feature contributions in spatial machine learning models into non-geographical (intrinsic variable) and geographical (location-specific) components using extended Shapley values.
• Interpretable machine learning models can capture spatially varying relationships and intrinsic contextual effects that are consistent with findings from traditional spatial statistical methods.

Summary by alphaXiv.