Sparse mobile crowd sensing saves perception cost by recruiting a small number of users to perceive data from a small number of sub-regions, and then inferring data from the remaining sub-regions. The data collected by different people on their respective trajectories have different values, and we can select participants who can collect high-value data based on their trajectory predictions. In this paper, we study two aspects of user trajectory prediction and user recruitment. First, we propose an STGCN-GRU user trajectory prediction algorithm, which uses the STGCN algorithm to extract features related to temporal and spatial information from the trajectory map, and then inputs the feature sequences into GRU for trajectory prediction, and this algorithm improves the accuracy of user trajectory prediction. Second, an ADQN (action DQN) user recruitment algorithm is proposed.The ADQN algorithm improves the objective function in DQN on the idea of reinforcement learning. The action with the maximum input value is found from the Q network, and then the output value of the objective function of the corresponding action Q network is found. This reduces the overestimation problem that occurs in Q networks and improves the accuracy of user recruitment. The experimental results show that the evaluation metrics FDE and ADE of the STGCN-GRU algorithm proposed in this paper are better than other representative algorithms. And the experiments on two real datasets verify the effectiveness of the ADQN user selection algorithm, which can effectively improve the accuracy of data inference under budget constraints.
Keywords: STGCN-GRU; reinforcement learning; sparse mobile crowd sensing; user selection.