Research map / Graph learning / Knowledge and temporal graphs
KGTS: Contrastive Trajectory Similarity Learning over Prompt Knowledge Graph Embedding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes KGTS, a trajectory similarity framework combining GRot knowledge-graph grid embedding, a prompt-based GCN+GRU trajectory encoder, and unsupervised contrastive learning with three positive-sample generation strategies.
Problem
Deep learning trajectory similarity computation needs costly supervised label preprocessing (quadratic in trajectory count and length), and learned models generalize poorly because training data lacks diverse highly-similar trajectory pairs.
Approach
Embeds map grids with a modified RotatE (GRot) so neighbouring grids get similar embeddings, bridges grid and trajectory embeddings with a learnable attentive prompt, and trains the encoder with contrastive learning using positive samples generated by copy/add/delete/move operations on trajectories.
Key results
- On GeoLife, KGTS reaches HR@1 0.5123, HR@5 0.579, HR@10 0.5942, above the best baseline GTS (0.4327 HR@1).
- On Porto, KGTS reaches HR@1 0.5244, HR@5 0.6277, HR@10 0.6542, above the best baseline T3S/NeuTraj (HR@10 0.6109 for T3S).
- Ablation on Porto shows removing the GRU order encoder drops HR@1 from 0.5244 to 0.152, the largest drop among variants.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use KGTS instead of the strongest baseline, GTS | Trajectory similarity computation · GeoLife | HR@1: 0.4327 -> 0.5123 | Helped |
| use KGTS instead of the strongest baseline, NeuTraj | Trajectory similarity computation · GeoLife | HR@5: 0.53 -> 0.579 | Helped |
| use KGTS instead of the strongest baseline, T3S | Trajectory similarity computation · GeoLife | HR@10: 0.5843 -> 0.5942 | Helped |
| use KGTS instead of the strongest baseline, t2vec | Trajectory similarity computation · Porto | HR@1: 0.4105 -> 0.5244 | Helped |
| use KGTS instead of the strongest baseline, T3S | Trajectory similarity computation · Porto | HR@5: 0.5506 -> 0.6277 | Helped |
| use KGTS instead of the strongest baseline, T3S | Trajectory similarity computation · Porto | HR@10: 0.6109 -> 0.6542 | Helped |
| add GRot grid embedding (ablation) | KGTS · Porto | HR@1: 0.3243 -> 0.5244 | Helped |
| use modified GRot rather than original RotatE (ablation) | KGTS · Porto | HR@1: 0.5067 -> 0.5244 | Helped |
| add attentive prompt for grid embedding (ablation) | KGTS · Porto | HR@1: 0.5096 -> 0.5244 | Helped |
| add GCN trajectory structure embedding (ablation) | KGTS · Porto | HR@1: 0.4853 -> 0.5244 | Helped |
| add GRU point-order embedding (ablation) | KGTS · Porto | HR@1: 0.152 -> 0.5244 | Helped |
Surprising findings
- Replacing the GRU order encoder causes the largest performance drop (HR@1 falls to 0.152), while removing GRot drops HR@1 to 0.3243.
- CL-TSim, another unsupervised contrastive method, performs worst among the baselines (HR@1 0.3233 on GeoLife).
Limitations
- Evaluation limited to two datasets (GeoLife and Porto).
- Positive sample generation relies on hand-designed rule-based operations rather than learned augmentation.
- Efficiency advantage over supervised baselines is asserted from preprocessing cost, not measured end-to-end.
Experimental setup
Datasets: GeoLife Porto
Benchmarks: GeoLife Porto
Metrics: HR@1 HR@5 HR@10
Baselines: SRN t2vec NeuTraj T3S GTS CL-TSim TMN
Builds on: RotatE SimCSE t2vec GCN infoNCE
Closest papers by meaning
- GTR: A General, Multi-View, and Dynamic Framework for Trajectory Representation Learning (ICML 2025 · similarity 0.85)
- Holistic Semantic Representation for Navigational Trajectory Generation (AAAI 2025 · similarity 0.83)
- TrajMamba: An Efficient and Semantic-rich Vehicle Trajectory Pre-training Model (NeurIPS 2025 · similarity 0.83)
- Scalable Trajectory-User Linking with Dual-Stream Representation Networks (AAAI 2025 · similarity 0.83)
- Road Network Representation Learning with the Third Law of Geography (NeurIPS 2024 · similarity 0.83)
- UniTR: A Unified Framework for Joint Representation Learning of Trajectories and Road Networks (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.