Trajectory Completion via Context-Guided Neural Filtering and Encoding

Di Yao, Fangda Guo, Zhenyu Wen, Yu Guo, Peng Cheng, Yangyuan He, Jingping Bi

Abstract: Trajectories have been massively collected in a wide range of domains and play a critical role in data-driven task support. However, trajectories are often highly sparse and incomplete, which has become a key bottleneck that limits the applicability of trajectory analysis techniques. While many existing sequential models are seemingly applicable to the trajectory completion problem, they often suffer severely from data sparsity and irregularity and yield poor performance in practice. We propose an effective method, named TrajCom, for completing sparse and irregular trajectories. To address data sparsity, TrajCom leverages rich context information to filter a set of reference trajectories that correlate strongly with the target incomplete trajectory. Then, TrajCom learns time-aware encodings of these trajectories by a newly proposed time-aware recurrent unit. Moreover, a popularity-weighted attention mechanism is proposed to complete the missing locations. Extensive experiments on four datasets show that TrajCom outperforms competitive baselines with up to 25% relative improvements.