Cluster analysis of day-to-day traffic data in networks

Day-to-day traffic data has been widely used in transportation planning and management. However, with the emerging of new technologies, one conventional assumption, on which many models rely, that all the day-to-day observations on the network follow a single pattern appears to be questionable. To better understand network flow patterns and their respective similarities, cluster analysis that partitions the day-to-day data into groups is an effective solution, but directly applying generic clustering algorithms may not always be appropriate identifying and interpreting day-to-day pattern changes due to the ignorance of the transportation network characteristics. In view of this practical issue, we propose a new clustering method that integrates network flow models, namely a statistical traffic assignment model and a probabilistic OD travel demand estimation model, into generic clustering algorithms. It essentially examines the probabilistic characteristics of traffic data by projecting those onto the dimensions of OD demands. For this reason, it can deal with traffic data where observations on some days and locations may be missing, or observing locations may change from day to day. The proposed algorithm embeds the domain knowledge of the transportation network, and is tested on two toy networks and one real-world network. Numerical experiments show the new clustering algorithm can effectively identify and interpret patterns that are hard to see by generic clustering algorithms otherwise, even with missing values or day-varying sensing locations.