Prediction of Urban Population-Facilities Interactions with Graph Neural Network

Abstract

The urban population interacts with service facilities on a daily basis. The information on population-facilities interactions is considered when analyzing the current city organization and revealing gaps in infrastructure at the neighborhood level. However, often this information is limited to several observation areas. The paper presents a new graph-based deep learning approach to reconstruct population-facilities interactions. In the proposed approach, graph attention neural networks learn latent nodes’ representation and discover interpretable dependencies in a graph of interactions based on observed data of one part of the city. A novel normalization technique is used to balance doubly-constrained flows between two locations. The experiments show that the proposed approach outperforms classic models in a bipartite graph of population-facilities interactions.