GeoAI for dynamic population distribution forecasting with graph convolutional networks using mobile phone data

Keywords: GeoAI, human mobility, GCN, deep learning
Abstract Type: Paper Abstract

Abstract

Fine-grained dynamic population distribution forecasting benefits smart transportation operations and management, such as public transport dispatch, traffic demand prediction, and transport emergency response. Considering the co-evolutionary patterns of crowd distribution, the spatial interactions among places are essential for modelling crowd distribution variations. However, two issues remain. First, the lack of sampling design in passive big data acquisition makes the spatial interaction characterizations of less crowded places insufficient. Second, the multi-order spatial interactions among places can help forecasting crowd distribution but are rarely considered in the existing literature. To address these issues, a novel GeoAI approach using graph convolutional networks for crowd distribution forecasting method with multi-order spatial interactions is proposed. In particular, a weighted random walk algorithm is applied to generate simulated trajectories for improving the interaction characterizations derived from sparse mobile phone data. The multi-order spatial interactions among contextual non-adjacent places are modelled with an embedding learning technique. The future crowd distribution is forecasted via a graph-based deep neural network. The proposed method is verified using a real-world country-scale mobile phone dataset, and the results show that both the multi-order spatial interactions and the trajectory data enhancement algorithm can help improve the crowd distribution forecasting performance. The proposed method can be utilized for capturing fine-grained dynamic population distribution, which supports various applications such as intelligent transportation management and public health decision making.