Spatio-Temporal Dynamic Multi-graph Attention Network for Ride-Hailing Demand Prediction

Accurate ride-hailing demand prediction is of great importance in traffic management and urban planning. Meanwhile, this is a challenging task due to the complicated Spatio-temporal correlations. Existing methods mainly focus on modeling the Euclidean correlations among spatially adjacent regions and modeling the non-Euclidean correlations among distant regions through the similarities of features such as points of interest (POI). However, due to these invariable regional characteristics, the spatial correlations obtained from them are static. These approaches all ignore the real-time dynamic correlations which change over time such as passenger flow between regions. Dynamic correlations can reflect the travel status of residents in real time and is the important factor for an accurate demand forecasting. In this paper, we propose Spatio-temporal dynamic multi-graph attention network (STDMG) to solve this problem. First, we encode the feature similarity and passenger flow between regions into multiple static and dynamic graphs at each time step. Then, the dynamic multi-graph fusion module is proposed to capture spatial dependencies by modeling these graphs. Finally, we design a temporal attention module which consisting of ConvLSTM layer and attention layer, to capture the influence of adjacent Spatio-temporal dependencies by combining the global context information. Experiments on three real-world datasets demonstrate the effectiveness of our approach over state-of-the-art methods.