Spatial modeling of bike-sharing trip data: A methodological comparison

Abstract

Usage data plays a major role in evaluating and planning sharing systems such as bike-sharing. Hence, effective methods are needed to analyze and model this kind of data. In this research, trip data of the bike-sharing system in Cologne, Germany is modeled. We compare two methods that can be applied in the modeling of spatial trip data, facing the requirements of spatial autocorrelation, zero-inflation and count data simultaneously. A generalized additive model (GAM) based on a Tweedie distribution is compared to a machine learning approach using the XGBoost algorithm. While the results of the GAM are easier to interpret and allow for the direct integration of spatial interdependencies in the model estimation, XGBoost leads to more precise predictions and can potentially be estimated in a shorter amount of time.