Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic\n Forecasting

Spatiotemporal forecasting has various applications in neuroscience, climate\nand transportation domain. Traffic forecasting is one canonical example of such\nlearning task. The task is challenging due to (1) complex spatial dependency on\nroad networks, (2) non-linear temporal dynamics with changing road conditions\nand (3) inherent difficulty of long-term forecasting. To address these\nchallenges, we propose to model the traffic flow as a diffusion process on a\ndirected graph and introduce Diffusion Convolutional Recurrent Neural Network\n(DCRNN), a deep learning framework for traffic forecasting that incorporates\nboth spatial and temporal dependency in the traffic flow. Specifically, DCRNN\ncaptures the spatial dependency using bidirectional random walks on the graph,\nand the temporal dependency using the encoder-decoder architecture with\nscheduled sampling. We evaluate the framework on two real-world large scale\nroad network traffic datasets and observe consistent improvement of 12% - 15%\nover state-of-the-art baselines.\n