Bike share travel time modeling: San Francisco bay area case study

Abstract

Bike Share Systems (BSSs) are emerging in many US cities as a new sustainable transportation mode that provides a last-mile solution for short-distance transfers between different private and public transportation modes. In order to encourage the increased use of bikes as a mode of transportation, tools, measures, and planning techniques similar to those used for other transportation modes need to be developed. With precise information on the trip travel time, route planner systems can suggest optimal alternative routes, and manage and control traffic congestion. Although there is a growing body of literature dealing with BSSs, bike travel time has been studied sparingly up to this point. In this paper, we addressed this issue by developing different bike travel time models using random forest (RF), least square boosting (LSBoost) and artificial neural network (ANN) techniques. We studied 33 different predictors affecting bike travel time, including such predictors as travel distance, biker experience, time-of-day, and weather conditions. The RF model produced a reasonable prediction rate with a mean absolute error (MAE) of 84.01 sec and a mean absolute percentage error (MAPE) of 16.92%. We further improved the bike travel time prediction model by using RF and forward stepwise regression to select the best subset of predictors to explain the bike travel time variability. The resulting model, with only seven predictors, reduced the MAE to 82.04 sec and the MAPE to 16.2%.