Research map / Image generation / Autoregressive and latent generation
Holistic Semantic Representation for Navigational Trajectory Generation
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 HOSER, a destination-aware trajectory generator that combines multi-level road network encoding, multi-granularity trajectory encoding, and a destination-oriented navigator, generating road-segment trajectories via a min-heap search.
Problem
Navigational trajectory generation for human mobility lacks holistic semantic modeling across road segments, zones, partial trajectories, and origin-destination context, and prior methods largely ignore destination guidance.
Approach
Hierarchical road network representation (road-segment GATv2 plus zone-level GCN over traffic-flow graph), a Transformer trajectory encoder with spatio-temporal relative position encodings, and a navigator that conditions next-road prediction on destination zone and distance/angle-to-destination features, trained end-to-end with NLL plus time-interval MAE and decoded by heap-based optimal search.
Key results
- HOSER achieves best global and local scores across three cities, e.g., Beijing Hausdorff 0.5503 vs. next best 0.6086 (Markov*) and Distance 0.0024.
- On downstream DeepMove next-location prediction, HOSER-generated data reaches Acc@5 0.776 / 0.888 / 0.797 on Beijing / Porto / SF versus real data 0.804 / 0.929 / 0.847.
- In zero-shot transfer (trained on Beijing, generated in Porto), HOSER gets Distance 0.0052 and Hausdorff 0.5843, beating Dijkstra and DiffTraj.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the Road Network Encoder (ablation) | HOSER · Beijing, Porto, and San Francisco | Hausdorff: 0.5694 -> 0.5503; 0.5993 -> 0.5746; 0.5403 -> 0.5351 | Helped |
| Add the Trajectory Encoder (ablation) | HOSER · Beijing, Porto, and San Francisco | Hausdorff: 0.5650 -> 0.5503; 0.5956 -> 0.5746; 0.5381 -> 0.5351 | Helped |
| Add the Destination-Oriented Navigator (ablation) | HOSER · Beijing, Porto, and San Francisco | Hausdorff: 0.5704 -> 0.5503; 0.6263 -> 0.5746; 0.5661 -> 0.5351 | Helped |
| Switch Markov, SeqGAN, MoveSim, or STEGA to a search-based paradigm | Markov, SeqGAN, MoveSim, and STEGA · Beijing, Porto, and San Francisco | Distance | Mixed |
| Train HOSER on fewer trajectories | HOSER · Few-shot trajectory generation | trajectory generation quality | Hurt |
| Use HOSER-generated trajectories instead of real trajectories | DeepMove · Beijing, Porto, and San Francisco location prediction | Acc@5: 0.776 / 0.804; 0.888 / 0.929; 0.797 / 0.847 | Hurt |
Surprising findings
- Dijkstra's shortest-path algorithm outperforms most deep learning baselines on global and local metrics.
- Adding a search-based decoding paradigm alone improves several baselines (e.g., Markov*, SeqGAN*), but they still lag HOSER.
Limitations
- Relies on map matching of GPS trajectories to road segments, which may introduce errors; baselines producing GPS points were excluded from utility evaluation for this reason.
- FlashBack utility shows slight deviations from real data due to reliance on timestamps, indicating room for improvement.
- Evaluated only on three taxi-based datasets and road-segment level generation.
Experimental setup
Datasets: Beijing taxi trajectories (provided by Bayou Tech) Porto taxi trajectories (PKDD-15) San Francisco taxi trajectories (CRAWDAD epfl/mobility) OpenStreetMap road networks
Benchmarks: Beijing Porto San Francisco trajectory datasets
Metrics: JSD Distance JSD Radius JSD Duration Hausdorff distance DTW EDR Acc@5 MRR
Baselines: Markov Dijkstra SeqGAN SVAE MoveSim TSG TrajGen DiffTraj STEGA TS-TrajGen
Builds on: HRNR GATv2 TS-TrajGen DiffTraj
Closest papers by meaning
- TrajMamba: An Efficient and Semantic-rich Vehicle Trajectory Pre-training Model (NeurIPS 2025 · similarity 0.89)
- GTR: A General, Multi-View, and Dynamic Framework for Trajectory Representation Learning (ICML 2025 · similarity 0.86)
- Generative Human Trajectory Recovery via Embedding-Space Conditional Diffusion (ICML 2025 · similarity 0.86)
- Scalable Trajectory-User Linking with Dual-Stream Representation Networks (AAAI 2025 · similarity 0.85)
- RHYTHM: Reasoning with Hierarchical Temporal Tokenization for Human Mobility (NeurIPS 2025 · similarity 0.84)
- UniTR: A Unified Framework for Joint Representation Learning of Trajectories and Road Networks (AAAI 2025 · similarity 0.84)
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.