Cluster Modeling of Operating Speed for Improved Transferability on Large Road Networks

Abstract

There is consensus among researchers that the transfer of operating speed models is limited in geographical scope. This paper advances the argument that, even within a single region, the transferability of operating speed models can be problematic when applied to dissimilar road segments of a large road network. To address this issue, a novel methodology is introduced aimed at enhancing the transferability. This paper involves identifying similar road segments using landmark-based spectral clustering, based on attributes that prominently factor into speed models. Models are developed using data from select segments within a cluster, and their transferability to other segments is subsequently assessed on segments situated at a distance but belonging to the same cluster. To gauge this transferability, a multifaceted approach is employed, including conventional cross-validation methods, a transferability index, and covariance analysis. The results indicate an enhancement in the transferability of cluster-calibrated models. In addition, the nontransferability of the models from one cluster to an entirely different cluster within the same network is discussed.