Modeling the interaction coupling of multi-view spatiotemporal contexts for destination prediction

Abstract

Bike-Sharing Systems (BSSs) are being introduced to more and more cities recently, and therefore they have generated huge amounts of data. Mobike is a station-less BSS which is suffering from the chaotic parking problem. To solve this problem, it is necessary to pre-dict where the bikes are going. Traditional works deal-ing with destination prediction mainly focus on station-based BSSs, and they merely leverages context-aware in-formation technically. Thus it is naturally promising to investigate how to improve the destination prediction of station-less bikes by context information. To that end, in this paper, we develop a multi-view machine (MVM) method, by incorporating the context information from Point of Interest (POI) data and human mobility data into destination prediction. Specifically, we first de-scribe three different views, namely start position, start time and destination by features extracted from POI data and human mobility data. Then, we capture the relationship between these three views' interactions and the trip's possibility by a multi-view machine. Finally, since multi-view machine contains too many parame-ters to be optimized, we leverage tensor factorization (TF) to reduce the computation costs. The experimen-tal results show that the model can effectively capture the potential relationship of three views with trip's pos-sibility and the approach is thus much more effective than traditional prediction methods for destination.