A traffic-flow forecasting network that models spatial and temporal correlation jointly through a spatio-temporal synchronous graph convolution kernel, capturing temporal correlation during the graph convolution itself rather than in a separate module. It reduces the graph computation from O(N^2) to O(KN) time and learns a static adaptive road-network graph from data that is then dynamically adjusted, achieving competitive accuracy at low computational and resource cost without self-attention, recurrent networks, or temporal convolutional networks.
Traffic-flow forecasting depends on the spatio-temporal correlation in traffic data, but most models capture spatial and temporal correlation separately and pay for accuracy with rising model complex…