Modeling bike-sharing demand using a regression model with spatially varying coefficients

As an emerging mobility service, bike-sharing has become increasingly popular around the world. A critical question in planning and designing bike-sharing services is to know how different factors, such as land-use and built environment, affect bike-sharing demand. Most research investigated this problem from a holistic view using regression models, where assume the factor coefficients are spatially homogeneous. However, ignoring the local spatial effects of different factors is not tally with facts. Therefore, we develop a regression model with spatially varying coefficients to investigate how land use, social-demographic, and transportation infrastructure affect the bike-sharing demand at different stations to address this problem. Unlike existing geographically weighted models, we define station-specific regression and use a graph structure to encourage nearby stations to have similar coefficients. Using the bike-sharing data from the BIXI service in Montreal, we showcase the spatially varying patterns in the regression coefficients and highlight more sensitive areas to the marginal change of a specific factor. The proposed model also exhibits superior out-of-sample prediction power compared with traditional machine learning models and geostatistical models.