Dynamic Multi-Scale Spatio-Temporal Graph ODE for Metro Ridership Prediction

Accurate metro ridership prediction is a very crucial task for the daily operation of metro system. However, complex spatio-temporal correlations of ridership makes it challenging. Existing studies widely use Graph Neural Networks (GNNs) to model spatial correlations, but traditional GNN models cannot expand the spatial receptive field by increasing the depth, and all node representations will converge to the same value with deeper layers. In this paper, we propose a novel Dynamic Multi-Scale Spatio-Temporal Graph Ordinary Differential Equation Networks (DMSTGODE). First, the temporal module captures dynamic multi-scale temporal dependencies through a temporal convolution module and a temporal attention module. Second, we design a tensor-based Ordinary Differential Equation (ODE) solver, which not only captures spatio-temporal features but also avoids over-smoothing problems. To understand the network more comprehensively, we consider both the geographic adjacency matrix and the semantic adjacency matrix. Extensive experiments on real-world datasets demonstrate that our proposed method significantly outperforms the state-of-the-art methods.