A machine learning approach to traffic flow prediction using CP tensor decomposition

This paper deals with the prediction of highway traffic flow based on historic data. The methodology is based on canonical polyadic (CP) tensor decompositions of traffic flow data. This step captures the regular elements of the traffic signal based on daily and weekly rhythms and typical geographical distributions of the traffic, while significantly reducing the amount of data required to describe these. The key factors are then extrapolated into the future, and the traffic data is reconstructed from the decomposition. Applied to traffic flow data from the M62 in the North of England in October 2019, this approach provides a surprisingly accurate prediction based on a very compact model, which is a distinct advantage compared to conventional machine learning approaches. Using 4 factors, the prediction captures 90% of the signal energy, which beats existing rolling average prediction techniques.