Physics-Informed Spatio-Temporal Model for Human Mobility Prediction

Human mobility prediction is the fundamental problem in studying human social behaviors. However, current approaches overlook the dynamic physics processes inherent in the movement of human, especially in an ever-changing temporal environments. In this work, we propose a physics-informed spatial-temporal prediction network (Physics-ST) for human mobility prediction. Specifically, we employ the assumption that human mobility is driven by a potential energy filed and establish a physics-informed equation to describe the dynamics of human mobility in the city. Then, we propose a multi-head node attention graph convolutional network to express the process of potential energy transfer within human mobility between regions. To model the impact of temporal information on mobility, we introduce correction terms in the physics-informed equation and propose a history-future information reasoning module which can extrapolate the human mobility energy under the future scenarios. Experiments conducted on real-world datasets demonstrate that our model exhibits superior performance, particularly in its understanding of physics mechanism.