Spatio-Temporal Clustering of Urban Traffic Networks

• Activity Number: 517 - Issues in Transportation Statistics
• Type: Contributed
• Date/Time: Wednesday, August 2, 2017 : 10:30 AM to 12:20 PM
• Sponsor: Transportation Statistics Interest Group
• Author(s): Ashwini Venkatasubramaniam* and Ludger Evers and Konstantinos Ampountolas
• Companies: University of Glasgow and University of Glasgow and University of Glasgow
• Keywords: Bayesian ; network ; clustering ; spatial ; temporal ; congestion

Abstract

Heterogeneous urban traffic networks with regions of varying congestion levels have unique fundamental properties that require tailor-made clustering algorithms. We propose a novel Bayesian clustering technique for spatio-temporal network data which is based on an amalgamation of a distance-dependent Chinese restaurant process (DDCRP) and a spatio-temporal conditional auto-regressive model (CAR). Our method employs a modified version of the DDCRP to incorporate the geographic constraints of the network and determine the shape and number of clusters. We do not expect the DDCRP to fully capture the dependency structure of the data and thus a CAR model is used to account for the spatial dependency within a cluster. This method is able to identify spatially contiguous clusters that also incorporate changes in levels of occupancy over time. Inference is carried out using a Metropolis within Gibbs sampler and we apply this developed clustering model to an urban traffic network, using occupancy data aggregated at the junction level.