Urban ride-hailing demand prediction with multiple spatio-temporal information fusion network

Abstract Urban ride-hailing demand prediction is a long-term but challenging task for online car-hailing system decision, taxi scheduling and intelligent transportation construction. Accurate urban ride-hailing demand prediction can improve vehicle utilization and scheduling, reduce waiting time and traffic congestion. Existing traffic flow prediction approaches mainly utilize region-based situation awareness image or station-based graph representation to capture traffic spatial dynamic while we observe that combination of situation awareness image and graph representation are also critical for accurate forecasting. In this paper, we propose the Multiple Spatio-Temporal Information Fusion Networks (MSTIF-Net), a novel deep learning approach to better fuse multiple situation awareness information and graphs representation. MSTIF-Net model integrates structures of Graph Convolutional Neural Networks (GCN), Variational Auto-Encoders (VAE) and Sequence to Sequence Learning (Seq2seq) model to obtain the joint latent representation of urban ride-hailing situation that contain both Euclidean spatial features and non-Euclidean structural features, and capture the spatio-temporal dynamics. We evaluate the proposed model on two real-world large scale urban traffic datasets and the experimental studies demonstrate MSTIF-Net has achieved superior performance of urban ride-Hailing demand prediction compared with some traditional state-of-art baseline models.