An investigation of using SHAP to interpret machine learning models of spatial data

Keywords: machine learning, spatial statistics, GeoAI, XAI
Session Type: Virtual Paper Abstract
Day: Saturday
Session Start / End Time: 2/26/2022 03:40 PM (Eastern Time (US & Canada)) - 2/26/2022 05:00 PM (Eastern Time (US & Canada))
Room: Virtual 2

Abstract

Machine learning (ML) often has superior predictive power than traditional statistical approaches but is criticized as being an inexplicable black box. In recent years, new methods (e.g., LIME, SHAP) have been proposed to interpret ML model parameters and provide opportunities to derive statistical associations and inferences. Even though these approaches are gaining popularity in the ML community, there is limited understanding of using them for spatial data. This research, therefore, aims to provide examples and guidelines on how to leverage interpretable ML methods to explain models of spatial data. Specifically, the use of SHAP to explain xgboost model for regression tasks will be examined. Both simulation and empirical evidence are presented with comparisons to spatial statistical models. The potential and limitations of SHAP are also discussed.