Analysis and visual exploration of prediction algorithms for public bicycle sharing systems

Abstract

Public bicycle sharing systems have become an increasingly popular means of transportation in many cities around the world. However, the information shown in mobile apps or websites is commonly limited to the system’s current status and is of little use for both citizens and responsible planning entities. The vast amount of data produced by these managing systems makes it feasible to elaborate and present predictive models that may help its users in the decision-making process. For example, if a user finds a station empty, the application could provide an estimation of when a new bicycle would be available. In this paper, we explore the suitability of several prediction algorithms applied to this case of bicycle availability, and we present a web-based tool to visually explore their prediction errors under different time frames. Even though a quick quantitative analysis may initially suggest that Random Forest yields a lower error, our visual exploration interface allows us to perform a more thorough analysis and detect subtle but relevant differences between algorithms depending on variables such as the station’s behavior, hourly intervals, days, or types of days (weekdays and weekends). This case illustrates the potential of visual representation together with quantitative metrics to compare prediction algorithms with a higher level of detail, which can, in turn, assist application designers and decision-makers to dynamically adjust the best model for their specific scenarios.