WGCN: A Novel Wavelet Graph Neural Network for Metro Ridership Prediction

Junjie Tang, Junhao Zhang, Juncheng Jin, Zehui Qu

Abstract: Metro ridership prediction is a significant and difficult task in intelligent urban transportation systems due to the unique traffic pattern of each station, the implicit transfer of ridership, and the sparse connectivity of the metro network. Existing methods mostly use auxiliary information for modeling, which increases the difficulty of data collection and the complexity of the model. Given the power of the wavelet transform in capturing complex information, we incorporate wavelet analysis into deep learning networks and propose a gated learnable wavelet filter module, it can fully excavate the spatio-temporal correlations without the graph. Meanwhile, a flow-aware K-order graph convolution network is designed to mine the connections of stations in topology by mixing weights and a K-order adjacency matrix. Based on these two modules, we propose a novel wavelet graph neural network: WGCN, which stacks them to capture spatio-temporal correlations from multiple scales. A large number of experiments on two real-world datasets demonstrate the superiority of the proposed model, which outperforms the state-of-the-art models on multiple metrics.