Leveraging data sampling patterns in data-driven traffic characteristic modeling

Data sampling pattern is inherently determined by the deployment of sensors and the recording intervals. The deployment of sensors is determined by the road characteristics. In the domain of short-term traffic flow prediction, existing approaches often indirectly learn the sampling pattern from the changes of observation or modeling the sampling pattern using static spatial definition of road networks. The existing approach may not be optimum in the newer datasets with multiple sampling patterns. We argue learning the sampling patterns directly from the road characteristics is more effective. We conducted a case study on spatial temporal traffic flow prediction model. Findings from experiments show 2% to 4% improvement over prediction metrics and some interesting prediction characteristics are found at different prediction horizons.