Contextual spatio-temporal graph representation learning for reinforced human mobility mining

Abstract

The rapid development of location-based services spurred a large number of user-centric applications. Particularly, an interesting topic has attracted the attention of researchers that is to link trajectories to users (TUL). Despite the significant progress made by recent deep learning-based human mobility learning models, tackling TUL problem is still challenging. In this paper, we propose a novel reinforced trajectory learning approach called GraphTUL that implements an adversarial network with the policy gradient to improve the identification ability and leverages both labeled and unlabeled trajectories to address the insufficient label issue in a semi-supervised manner. Besides, some critical factors related to personal context and indispensable elements in current mobility learning models are still missing. Thus, we propose a novel graph-based human motion representation model (CGE) to exploit the contextual information from users’ trajectories for alleviating data sparsity and contextual constraint issues. CGE builds a unified graph with historical check-ins to reflect users’ geographical preferences and visiting intentions. It allows us to sample synthetic but realistic trajectories for augmenting data and enhancing contextual check-in embedding. We also successfully apply it to next check-in prediction task. The experimental results conducted on several real-world datasets demonstrate that our proposed method achieves significantly better performance than the state-of-the-art baselines.