Incremental Learning Models of Bike Counts at Bike Sharing Systems

Abstract

Bike sharing systems (BSSs) have become a convenient and environmentally friendly transportation mode, but may suffer from logistical issues such as bike shortages at stations. Predicting bike counts would help mitigate imbalances in the system. Research has focused on global prediction techniques but has neglected the role of user incentives. We adopted two computational techniques to capture BSS dynamics: mini-batch gradient descent for the linear regression (MBGDLR) and locally weighted regression (LWR). The two approaches used incremental learning based only on the previous status of the station with neither weather nor time information. The models were applied to a BSS data set for one year (2014–2015) in the San Francisco Bay Area for different prediction windows. Both models gave comparable results. LWR performed slightly better than MBGDLR for all prediction windows. The smallest prediction error for LWR was 0.31 bikes/station (4% prediction error) for a 15-minute prediction window and 0.32 bikes/station for MBGDLR. The 120-minute prediction window had the largest prediction error of 1.1 bikes/station and 1.2 bikes/station for LWR and MBGDLR, respectively. Computationally, MBGDLR was 55 times faster than LWR and proved to be faster than other machine learning and time series algorithms.