Highlights • A graph wavelet gated recurrent neural network (GWGR) is proposed. • GWGR incorporating graph wavelet units as gates learns traffic networks as graphs. • GWGR achieves better prediction performance with fewer weight parameters. • The sparsity of learned weights in GWGR can enhance model interpretability. • The learned weights in GWGR can help to identify the key roadway links.

Abstract Network-wide traffic forecasting is a critical component of modern intelligent transportation systems for urban traffic management and control. With the rise of artificial intelligence, many recent studies attempted to use deep neural networks to extract comprehensive features from traffic networks to enhance prediction performance, given the volume and variety of traffic data has been greatly increased. Considering that traffic status on a road segment is highly influenced by the upstream/downstream segments and nearby bottlenecks in the traffic network, extracting well-localized features from these neighboring segments is essential for a traffic prediction model. Although the convolution neural network or graph convolution neural network has been adopted to learn localized features from the complex geometric or topological structure of traffic networks, the lack of flexibility in the local-feature extraction process is still a big issue. Classical wavelet transform can detect sudden changes and peaks in temporal signals. Analogously, when extending to the graph/spectral domain, graph wavelet can concentrate more on key vertices in the graph and discriminatively extract localized features. In this study, to capture the complex spatial-temporal dependencies in network-wide traffic data, we learn the traffic network as a graph and propose a graph wavelet gated recurrent (GWGR) neural network. The graph wavelet is incorporated as a key component for extracting spatial features in the proposed model. A gated recurrent structure is employed to learn temporal dependencies in the sequence data. Comparing to baseline models, the proposed model can achieve state-of-the-art prediction performance and training efficiency on two real-world datasets. In addition, experiments show that the sparsity of graph wavelet weight matrices greatly increases the interpretability of GWGR.