Study on Geographically Attribute Neural Network Weighted Regression

Geographically weighted regression (GWR) uses the Euclidean distance between observation points to characterize the proximity relationship. Its spatial weighting kernels are simple, making it insufficient to calculate complicated geographical interactions. Faced with the limitations of GWR, a geographically attribute neural network weighted regression (GANNWR) was proposed. Specifically, utilizing the attention mechanism, we designed an attribute weighted neural network (AWNN) to compute the spatial weight matrix. Fusing the spatial features and attribute features by attention mechanism, AWNN could characterize the attribute spatial proximity relationship. To examine the effectiveness of GANNWR, we investigated the correlations between socioeconomic conditions and violent crime in the city of Chicago. Compared with ordinary linear regression (OLR), GWR, Multiscale Geographically Weighted Regression (MGWR), and geographically neural network weighted regression (GNNWR), the results show that GANNWR could capture interpretable attribute and spatial features and have the best predictive ability.