Data-driven investigation of traffic spatial correlation through functional data analysis

The article proposes a particular functional data analysis technique, named clustering-based functional regression (CBFR), to map pairwise relationships between raw within-day flow measurement profiles from traffic sensors spanning a large day-to-day time horizon. The proposed approach yields more effective traffic estimation, and its results are more interpretable in the sense of effectively disentangling different sources of correlations between traffic measurements. It also allows for the addressing of missing data, thus improving the data collection capabilities offered by movable sensors, such as drones and radars, which can be moved in different locations over time, yielding inherently incomplete datasets. A four-step methodology is developed, including an important functional clustering step, and applied to the real case study of the Turin network (Italy). The results show the practical applicability and effectiveness of CBFR in exploring spatial associations between traffic flow measurement profiles on different network links, highlighting, in particular, the ability to interpret better and disentangle the underlying traffic phenomena.