LAUSR

Abstract With the increasing concerns about the environmental impact of motor vehicles, more cities are investing in Bike-Sharing Systems (BSSs) as an alternative mode of transport for their citizens. In such systems, predicting the fine-grained station-level bike-flow improves the operation and reliability. Some recent studies have employed graph-based approaches to model BSSs, however, considering spatial closeness and communities in the flow prediction has not been fully addressed yet. In this paper, we propose a Matrix Factorization model with Local and Global consistency (MFLOG) to be used for flow prediction in BSSs. MFLOG captures the dynamics and underlying structure of a BSS and models spatial closeness, temporal variations, and communities in a BSS. We also investigate the relationship between spatial closeness and bike-flow and explore the stability of communities in the BSS. The proposed method is evaluated on the Divvy Trips data set in the City of Chicago. The results show that the MFLOG model improves the accuracy of single and multiple-step bike-flow and check-in/out predictions over the baseline models.