Context Aware Trajectory Imputation via Spatio-temporal Representation Learning

The wide application of location-acquisition devices spurs the research about trajectory data, which could benefit a variety of domains including vehicle navigation, route planning, and travel time estimation. However, the limitation of environment and devices in reality affects the quality of original data, leading to unreliable trajectories. Considering that the traffic systems are always organized as topology structure, this paper aims to improve trajectory quality on road networks. However, it is a non-trivial task due to the uncertainty of outliers, the variety of segment features and the complexity between traffic segments. To address the above issues, this paper proposes Spatio-Temporal Bert-based Imputation (ST-BerImp) method for trajectory imputation. To improve the robustness and evaluation of the method, this study proposes a data generation strategy to introduce outliers. Then, to extract abundant node features, a traffic segment representation module is proposed to capture the temporal features, the Euclidean features, the non-Euclidean features and the attributes simultaneously. To handle the complex relations between traffic segments, a module based on powerful transformer encoder is designed to capture global dependency. Extensive experiments are conducted on two real-world datasets and the experimental results demonstrate the superiority of the proposed method.