Pre-Training Location Representations via Spatial-Temporal Trajectory Subgraph Contrastive Learning

Location representations from people’s location-based service data are important and beneficial for urban downstream tasks. Locations usually have complex spatial-temporal contextual semantics, meaning that the same location has variable functionalities in different trajectories. Existing methods are mostly based on sequence or graph, where the former captures accurate temporal but limited spatial information (one trajectory), while the latter obtains a global spatial perspective (upstream and downstream nodes) but ignores the temporal information. Furthermore, the frequencies of visited locations are long-tail distributed, which is disadvantageous for infrequently visited locations. To that end, we propose a spatial-temporal trajectory subgraph contrastive learning framework entitled ST-TGCL, integrating comprehensive spatial-temporal information and relieving the long-tail issue with contrastive learning. Specifically, we construct contrastive trajectory subgraph pairs to stably learn variable functionalities and increase training opportunities for infrequently visited locations. To capture spatial-temporal contextual semantics, we design a trajectory network that formulates trajectories and a trajectory graph convolution network, which has the strengths of both sequence-based and graph-based models. Finally, we apply the location representations for downstream tasks to demonstrate our framework’s effectiveness and generalization. ST-TGCL is evaluated over real-world datasets, and the results demonstrate that our framework significantly outperforms existing methods in location representation learning.