Title:Context-aware demand prediction in bike sharing systems: incorporating spatial, meteorological and calendrical context

Abstract:Bike sharing demand is increasing in large cities worldwide. The proper functioning of bike-sharing systems is, nevertheless, dependent on a balanced geographical distribution of bicycles throughout a day. In this context, understanding the spatiotemporal distribution of check-ins and check-outs is key for station balancing and bike relocation initiatives. Still, recent contributions from deep learning and distance-based predictors show limited success on forecasting bike sharing demand. This consistent observation is hypothesized to be driven by: i) the strong dependence between demand and the meteorological and situational context of stations; and ii) the absence of spatial awareness as most predictors are unable to model the effects of high-low station load on nearby stations.
This work proposes a comprehensive set of new principles to incorporate both historical and prospective sources of spatial, meteorological, situational and calendrical context in predictive models of station demand. To this end, a new recurrent neural network layering composed by serial long-short term memory (LSTM) components is proposed with two major contributions: i) the feeding of multivariate time series masks produced from historical context data at the input layer, and ii) the time-dependent regularization of the forecasted time series using prospective context data. This work further assesses the impact of incorporating different sources of context, showing the relevance of the proposed principles for the community even though not all improvements from the context-aware predictors yield statistical significance.