Geographical Graph Attention Networks: Spatial Deep Learning Models for Spatial Prediction and Exploratory Spatial Data Analysis

Some recent geospatial artificial intelligence (GeoAI) models have contributed to bridging the gap between artificial intelligence (AI) and spatial analysis. However, existing models struggle with handling small sample sizes for spatial prediction tasks across large areas. For exploratory spatial data analysis (ESDA), they are susceptible to distortion from local outliers and lack reliable interpretability methods that consider causal relationships. This study proposes Geographical Graph Attention Networks (GeoGATs), which are spatial deep learning models based on the principle of spatial (geographic) similarity. Two variants of the model are designed, namely GeoGAT-P for spatial prediction and GeoGAT-E for ESDA. Case studies using U.S. election data and homicide data demonstrate that GeoGAT-P can achieve more accurate predictions over a large spatial extent with a small sample size than existing models. GeoGAT-E can achieve decent performance in comparison with existing models and understand complex spatial relationships. Our study demonstrates how spatial similarity can be integrated with the latest deep learning models, offering valuable insights for the future direction of GeoAI research.