Learning urban form through unsupervised graph-convolutional neural networks

Graph theory has long provided the basis for the computa-tional  modelling  of  urban  flows  and  networks  and,  thus,  for  the  studyof urban form. The development of graph-convolutional neural networksoffers the opportunity to explore new applications of deep learning ap-proaches  in  urban  studies.  In  this  paper,  we  propose  an  unsupervisedgraph representation learning framework for analysing urban street net-works. Our results illustrate how a model trained on a 1% random sampleof street junctions in the UK can be used to explore the urban form of thecity of Leicester, generating embeddings which are similar but distinctfrom  classic  metrics  and  able  to  capture  key  aspects  such  as  the  shiftfrom urban to suburban structures. We conclude by outlining the cur-rent limitations and potential of the proposed framework for the studyof urban form and function.