MG-ASTN: Multi-Graph Framework with Attentive Spatial-Temporal Networks for Crowd Mobility Prediction

Predicting urban crowd patterns/flows is a challenging task due to complex spatial-temporal (ST) dependencies. In this paper, we aim to examine and report the capability and effectiveness of the current most widely used Graph Convolutional Networks (GCNs) on mobile data analysis in ST networks for crowd flow forecasting. Specifically, we propose a novel dual-stream framework leveraging Multi-Graph with Attentive Spatial-Temporal Networks (MG-ASTN) to simultaneously predict crowd in-out flow and OD flow based on the trajectory data collected by on-board devices (e.g., GPS). MG-ASTN utilizes Multi-Graph Convolutional Networks encoding non-Euclidean correlations to explore pair-wise relationships among regions. In addition, we further apply a cross-channel attention mechanism with 3D Temporal Convolutional Network to address the heterogeneity of ST features and capture more meaningful data representations for multi-task learning. In the evaluation, we conduct experiments based on 2 real-world datasets and verify most well-known state-of-the-art methods for crowd flow prediction. The results demonstrate that MG-ASTN could outperform other solutions -in-out flow prediction with lowest RMSE and MAE, and OD flow prediction beyond others in most cases, thus it has great potential in modeling the complex correlations among regions in ST networks and enabling accurate prediction in urban computing.