Generating population migration flow data from inter-regional relations using graph convolutional network

Spatial and socioeconomic structures of geographical units produce various inter-regional relations, which impose a direct effect on origin–destination flows. Currently, most flow prediction models only lay emphasis on regional attributes ignoring the inter-regional relations, which limits their abilities to estimate spatial flows more accurately. In this research, we apply the graph convolutional network (GCN) architecture to generate flow data based on inter-regional relations, providing a promising perspective for spatial flow modeling. We develop a relation-to-flow graph convolutional network (R2F-GCN) model to learn the latent representations of regions in an inter-regional relation graph for flow intensity estimation. The relational graph is constructed using the k-nearest neighbor method. We validate the feasibility and effectiveness of our model with experiments based on a mobility dataset of 281 Chinese cities and inter-city relations regarding spatial proximity and transport connectivity. We also discuss the impacts of hyperparameters on the model's performance.