Understanding the spatial networks formed by the trajectories of mobile userscan be beneficial to applications ranging from epidemiology to local search.Despite the potential for impact in a number of fields, several aspects ofhuman mobility networks remain largely unexplored due to the lack oflarge-scale data at a fine spatiotemporal resolution. Using a longitudinaldataset from the location-based service Foursquare, we perform an empiricalanalysis of the topological properties of place networks and note theirresemblance to online social networks in terms of heavy-tailed degreedistributions, triadic closure mechanisms and the small world property. Unlikesocial networks however, place networks present a mixture of connectivitytrends in terms of assortativity that are surprisingly similar to those of theweb graph. We take advantage of additional semantic information to interprethow nodes that take on functional roles such as `travel hub', or `food spot'behave in these networks. Finally, motivated by the large volume of new linksappearing in place networks over time, we formulate the classic link predictionproblem in this new domain. We propose a novel variant of gravity models thatbrings together three essential elements of inter-place connectivity in urbanenvironments: network-level interactions, human mobility dynamics, andgeographic distance. We evaluate this model and find it outperforms a number ofbaseline predictors and supervised learning algorithms on a task of predictingnew links in a sample of one hundred popular cities.
展开▼
