Presenter Profile

TALK TITLE
Variable selection and prediction for geographically weighted regression

KEYWORDS
geographically weighted regression, sparse group Lasso, variable selection

ABSTRACT
Geographically weighted regression (GWR) is a popular method for regression of spatial data. By estimating varying coefficients for each location, GWR can describe geographical effects flexibly. This presentation deals with a variable selection for GWR via geographically weighted sparse group Lasso (GWSGL). GWR has two types of variable selection: local and global variable selections, and GWSGL can conduct the two variable selections simultaneously by sparse group Lasso. In a regression problem, a prediction accuracy of a model is one of important points. Then, I consider whether the variable selection by GWSGL can improve the prediction accuracy of the GWR model. However, GWSGL cannot provide predicted values for future observations in closed form. Since the GWSGL estimates are obtained for each observed point like GWR estimates, we need to interpolate the estimates of unobserved points for prediction. Although the exact interpolation for GWR is possible, it is not for GWSGL. Hence, I approximately interpolate GWSGL estimates for unobserved points and investigate the prediction accuracy of GWSGL.