Learning transferable structural representations and reasoning for cross-domain sparse trajectory recovery

Accurate trajectory recovery is fundamental to transportation intelligence applications such as traffic prediction, path planning, and navigation services. However, real-world trajectories are often highly sparse, and existing methods struggle with limited reasoning capability and poor generalization under structural domain shifts across cities. To address these challenges, we formulate trajectory recovery as a joint problem of cross-domain structural generalization and reasoning under sparsity, and propose CROSSER, a unified framework for cross-domain sparse trajectory recovery. From a structural perspective, we introduce a cross-domain graph augmentation mechanism that enables explicit knowledge sharing across multiple transition graphs and facilitates the learning of transferable mobility patterns. From a reasoning perspective, we develop an LLM-enhanced recovery framework, where structural embeddings are fused with continuous spatio-temporal representations and injected into a LoRA-adapted LLM backbone for global trajectory inference. In addition, we provide a theoretical analysis showing that the target-domain recovery error is bounded by the source-domain error and a cross-domain structural divergence term. Extensive experiments on three large-scale real-world datasets demonstrate that CROSSER consistently outperforms state-of-the-art baselines, with significant improvements under cross-domain settings.