GeoShapley: A Game Theory Approach to Measuring Spatial Effects in Machine Learning Models
Presenter: Dr. Ziqi Li (Florida State University)
Abstract: GeoShapley is a game theory approach to measuring spatial effects in machine learning models. The approach extends the Nobel Prize–winning Shapley value framework by conceptualizing location as a player in a model prediction game, which enables the quantification of the importance of location and the synergies between location and other features in a model. GeoShapley is model-agnostic and can be applied to models of various structures. The webinar will discuss the principles of GeoShapley, its effectiveness with both synthetic and empirical data, and a GeoShapley Python package. Lastly, the webinar will discuss some best practices and caveats when applying GeoShapley to real-world data.