Learning the dependency structure of highway networks for traffic forecast

Forecasting road traffic conditions requires an accurate knowledge of the spatio-temporal dependencies of traffic flow in transportation networks. In this article, a Bayesian network framework is introduced to model the correlation structure of highway networks in the context of traffic forecast. We formulate the dependency learning problem as an optimization problem and propose an efficient algorithm to identify the inclusion-optimal dependency structure of the network given historical observations. The optimal dependency structure learned by the proposed algorithm is evaluated on benchmark tests to show its robustness to measurement uncertainties and on field data from the Mobile Millennium traffic estimation system to show its applicability in an operational setting.