A fine-grained graph-based spatiotemporal network for bike flow prediction in bike-sharing systems

Abstract

Since bike-sharing systems have been widely adopted for urban traveling, the system operators need to provide a fine-grained management to ensure a balanced distribution of bikes in the city. Predicting the number of riding bikes from one station to another can help to grasp the bike transition pattern and then optimize the repository of bikes. However, such a station-to-station bike prediction task is challenging due to multiple spatial and temporal factors. To manage this problem, in this work we propose a fine-grained graph-based spatiotemporal network called FGST model to predict station-to-station bike usages. Specifically, the FGST model consists of two main processes: the graph generation and the spatiotemporal embedding. At first, we propose an algorithm to construct the bike flow correlation graph, which is able to express the influence relationship of bike flows. Then, based on the generated graph, two blocks are designed for spatiotemporal dependences modeling: the spatial embedding block which uses the graph convolutional network learns the topological structure of graph for spatial dependence modeling, and the temporal embedding block which uses recurrent neural network captures the dynamic change of bike data for temporal dependence modeling. Experiments conducted on three real-world bike-sharing datasets demonstrate that our FGST model can effectively deal with the spatiotemporal dependences of bike data, and outperforms the existing neural network-based baselines.