Abstract
Origin-Destination (O-D) travel demand prediction is a fundamental challengein transportation. Recently, spatial-temporal deep learning models demonstratethe tremendous potential to enhance prediction accuracy. However, few studiestackled the uncertainty and sparsity issues in fine-grained O-D matrices. Thispresents a serious problem, because a vast number of zeros deviate from theGaussian assumption underlying the deterministic deep learning models. Toaddress this issue, we design a Spatial-Temporal Zero-Inflated NegativeBinomial Graph Neural Network (STZINB-GNN) to quantify the uncertainty of thesparse travel demand. It analyzes spatial and temporal correlations usingdiffusion and temporal convolution networks, which are then fused toparameterize the probabilistic distributions of travel demand. The STZINB-GNNis examined using two real-world datasets with various spatial and temporalresolutions. The results demonstrate the superiority of STZINB-GNN overbenchmark models, especially under high spatial-temporal resolutions, becauseof its high accuracy, tight confidence intervals, and interpretable parameters.The sparsity parameter of the STZINB-GNN has physical interpretation forvarious transportation applications.