Fusing Local and Global Mobility Patterns for Trajectory Recovery

Abstract

The prevalence of various mobile devices translates into a proliferation of trajectory data that enables a broad range of applications such as POI recommendation, tour recommendation, urban transportation planning, and the next location prediction. However, trajectory data in real-world is often sparse, noisy and incomplete. To make the subsequent analysis more reliable, trajectory recovery is introduced as a pre-processing step and attracts increasing attention recently. In this paper, we propose a neural attention model based on graph convolutional networks, called TRILL, to enhance the accuracy of trajectory recovery. In particular, to capture global mobility patterns that reflect inherent spatio-temporal regularity in human mobilities, we construct a directed global location transition graph and model the mobility patterns of all the trajectories at point level using graph convolutional networks. Then, a self-attention layer and a window-based cross-attention layer are sequentially adopted to refine the representations of missing locations by considering intra-trajectory and inter-trajectory information, respectively. Meanwhile, an information aggregation layer is designed to leverage all the historical information. We conduct extensive experiments using real trajectory data, which verifies the superior performance of the proposed model.