Revisiting the Topology of Urban Road Networks with Graph Embeddings

Representing and depicting topological structures of road networks is beneficial for route planning, traffic management and network optimization. Even though road segments are basic management units in practice, the topological structures of road-segment based networks have not been well characterized due to the homogeneity of their network metrics (e.g., degree). Therefore, this study introduced graph-embedding technique and investigated its potential for characterizing topological structures of road-segment based networks. A popular graph-embedding model called Node2Vec is applied to learn embeddings of road segments from the networks. On this basis, we firstly explored the relationship between road-segment embeddings and centrality metrics by testing if the latter can be predicted by the former. Second, we investigated the clustering results of the road-segment embeddings and tested if the clustered road segments are modularized in topology. Experiments based on road networks of six Chinese cities proved that road-segment embeddings can not only well represent various centrality metrics (e.g., degree, closeness, and PageRank), but also detect community structures that road segments with similar embeddings tend to be densely connected with each other. Those findings revealed the characteristics of road embeddings and proved their potential for representing road-segment based network structures, thus we can better apply them to more related downstream tasks such as transport network optimization and traffic prediction.