Deep spatial-temporal networks for crowd flows prediction by dilated convolutions and region-shifting attention mechanism.