Estimating intercity human mobility flow from city attributes and intercity relations in physical space and cyberspace via graph attention network

Accurate intercity human mobility estimation is essential for urban socioeconomic development. Studies have confirmed that city attributes and intercity relations are vital for estimating intercity mobility flow. Moreover, latest graph neural network (GNN) models show great potential for improving flow estimation accuracy. However, most GNN models rely more on city attributes but ignore intercity relations, especially in cyberspace, and they cannot explain or quantify how city attributes and intercity relations influence GNN outputs. To solve these problems, this study proposes a novel intercity flow estimation model that combines city attributes and intercity relations in physical space and cyberspace under graph attention network architecture (ARPC2F-GAT). Then, a gradient-based saliency map and ablation experiments are used to quantify importance and contribution of attributes and relations to flow estimation. Results from an empirical study of estimating intercity mobility flow in Guangdong province prove that the proposed ARPC2F-GAT outperforms all baseline models. Among city attributes, average wage and resident population are the two most important features. Among intercity relations, relations in cyberspace are more prominent than those in physical space and can bring significant performance gains. This study improves intercity flow estimation by integrating city attributes and intercity relations and enhances model explainability.