Trajectories are sequences of timestamped location samples. In sparse trajectories, the locations are sampled infrequently; and while such trajectories are prevalent in real-world settings, they are challenging to use to enable high-quality transportation-related applications. Current methodologies either assume densely sampled and accurately map-matched trajectories, or they rely on two-stage schemes, yielding sub-optimal applications. To extend the utility of sparse trajectories, we propose a novel sparse trajectory learning framework, GenSTL. The framework is pre-trained to form connections between sparse trajectories and dense counterparts using auto-regressive generation of feature domains. GenSTL can subsequently be applied directly in downstream tasks, or it can be fine-tuned first. This way, GenSTL eliminates the reliance on the availability of large-scale dense and map-matched trajectory data. The inclusion of a well-crafted feature domain encoding layer and a hierarchical masked trajectory encoder enhances GenSTL's learning capabilities and adaptability. Experiments on two real-world trajectory datasets offer insight into the framework's ability to contend with sparse trajectories with different sampling intervals and its versatility across different downstream tasks, thus offering evidence of its practicality in real-world applications.

我们提出了一种新颖的稀疏轨迹学习框架GenSTL，通过自回归生成特征域的方法，将稀疏轨迹与密集轨迹建立联系，从而消除了对大规模密集和匹配的轨迹数据的依赖，实验结果表明GenSTL在处理不同采样间隔的稀疏轨迹和不同下游任务上的适用性和实用性。

GenSTL：通用稀疏轨迹学习通过特征域自回归生成