Dynamic Origin/Destination modeling with Gravity-Latent Dirichlet Allocation (G-LDA)
Résumé
The proposed approach follows and develops the research direction initiated almost 50 years ago, during which Origin/Destination (O/D) matrices were in a very simple way evaluated by means of Gravity Models (see Casey (1955)). In the classical framework, the population density of locations and the distance between two locations are often considered as relevant factors to estimate the number of trips from one zone to another one. Our novel model is similar in the subject since it considers compromises and covariates to get insight of the mobility. But it overcomes the barriers of Gravity Modelling which is mainly static and push it in a dynamic context. Following the traces of Mixed-Membership modeling (see Airoldi et al. (2014)), which can perform analyses of O/D matrices (as in Come et al. (2014)), we propose to merge a gravity model approach with Latent Dirichlet Allocation (LDA) in order to deal with non stationary O/D matrices. Our proposal called G-LDA which stands for Gravity LDA, allows us to uncover the hidden patterns underlying urban mobility by taking into account temporal effects on usages and the effects of contextual socio-economical variables. In the following, we describe the novel extension of the LDA model and give some results on a dataset collected on the Paris Bike Shared System (BSS).