Holistic Prediction for Public Transport Crowd Flows: A Spatio Dynamic Graph Network Approach

This paper targets at predicting public transport in-out crowd flows of different regions together with transit flows between them in a city. The main challenge is the complex dynamic spatial correlation of crowd flows of different regions and origin-destination (OD) paths. Different from road traffic flows whose spatial correlations mainly depend on geographical distance, public transport crowd flows significantly relate to the region’s functionality and connectivity in the public transport network. Furthermore, influenced by commuters’ time-varying travel patterns, the spatial correlations change over time. Though there exist many works focusing on either predicting in-out flows or OD transit flows of different regions separately, they ignore the intimate connection between the two tasks, and hence lose efficacy. To solve these limitations in the literature, we propose a Graph spAtio dynamIc Network (GAIN) to describe the dynamic non-geographical spatial correlation structures of crowd flows, and achieve holistic prediction for in-out flows of each region together with OD transit flow matrix between different regions. In particular, for spatial correlations, we construct a dynamic graph convolutional network for the in-out flow prediction. Its graph structures are dynamically learned from the prediction of OD transit flow matrix, whose spatial correlations are further captured via a multi-head graph attention network. For temporal correlations, we leverage three blocks of gated recurrent units, which capture minute-level, daily-level and weekly-level temporal correlations of crowd flows separately. Experiments on real-world datasets are used to demonstrate the efficacy and efficiency of GAIN.