Road-Network Trajectory Similarity Learning with Landmark Distance Encoding and Matching Knowledge Transfer

Representation-based trajectory similarity learning (TSL) uses deep neural encoders to map each trajectory independently to a reusable vector, which enables efficient road-network trajectory retrieval. However, existing methods do not fully exploit the information that defines road-network similarity. Quantitative road-network distances are rarely organized into reusable segment-level descriptions, and the position-level comparisons performed by the target measure are usually collapsed into trajectory-level supervision. To preserve this information, we propose two complementary designs, one for representation and one for supervision. Landmark distance encoding describes each visited road segment by its shortest-path distances to a shared set of landmark road segments, which gives every segment a common quantitative reference for road-network distance. Matching knowledge transfer converts the position-level comparison rule of the target measure into soft position targets and uses them to supervise contextual position representations during training. Both designs are integrated into a single shared Transformer encoder. At inference, each trajectory is still encoded independently, and similarity is evaluated by a vector distance alone. Experiments on two real-world road-network trajectory datasets under two representative target measures show that the proposed framework consistently outperforms representative TSL methods while retaining efficient, vector-based retrieval.