Research map / Label-efficient and robust learning / Recognition and long-tail

GTG: Generalizable Trajectory Generation Model for Urban Mobility

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Dense features and foundation priors · 5 citations (OpenAlex)

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 GTG, a cross-city trajectory generation model that combines Space Syntax road features, disentangled adversarial travel-cost prediction, and preference learning with shortest-path search to generate trajectories in cities without target trajectory data.

Problem

Trajectory generation models are data-driven and depend on a city's global road network, so they fail to transfer to new cities where no trajectory data is available.

Approach

Learn city-invariant road representations from Space Syntax topological features and disentangled adversarial domain adaptation to predict travel costs, then learn a preference combining observable and hidden costs and generate trajectories via shortest-path search on the target road network.

Key results

• On the new-city setting, GTG achieves the best Distance JSD on BJ (0.0006 vs next-best 0.0107 for TT) and the best DTW on BJ (4.89 vs 24.27 for TT) without target-city training.
• Using GTG-generated data to train DeepMove on XA yields ACC 0.82 (GTG 1) vs 0.46-0.78 for baselines, second only to real data at 0.88.
• Ablation on XA shows removing cost prediction raises Distance JSD from 4.039e-3 to 19.171e-3 (BJ to XA), the largest drop among components.

Measured effects reported in the paper

Surprising findings

• Fine-tuning on target-city trajectories does not monotonically improve all metrics; e.g., BJ-to-XA DTW and Distance vary non-monotonically with the fraction of target data.

Limitations

• Still requires the target city's road network to be available for inference and feature extraction
• Generated trajectories are conditioned on start and destination road segments from travel demand
• Fine-tuning results are non-monotonic across data fractions

Experimental setup

Datasets: BJ taxi GPS trajectories (Nov 2015) XA DiDi trajectories (Nov 2018) CD DiDi trajectories (Nov 2018) OpenStreetMap road networks

Benchmarks: BJ, XA, CD cross-city trajectory generation

Metrics: JSD on Distance JSD on Radius JSD on LocFreq Hausdorff DTW EDT EDR

Baselines: Random Walk Density-EPR Spatial-EPR TrajGen SeqGAN SVAE MoveSim TS-TrajGen DiffTraj VOLUNTEER

Builds on: GATv2 maximum entropy inverse reinforcement learning disentangled representation learning with adversarial domain adaptation METIS graph partitioning

Closest papers by meaning

• MobTA: Bus-Conditioned Zero-Shot Trajectory Generation via Task Arithmetic (NeurIPS 2026 · similarity 0.84)
• TransitLM: A Large-Scale Dataset and Benchmark for Map-Free Transit Route Generation (NeurIPS 2026 · similarity 0.83)
• TrajMamba: An Efficient and Semantic-rich Vehicle Trajectory Pre-training Model (NeurIPS 2025 · similarity 0.82)
• GTR: A General, Multi-View, and Dynamic Framework for Trajectory Representation Learning (ICML 2025 · similarity 0.82)
• Satellites Reveal Mobility: A Commuting Origin-destination Flow Generator for Global Cities (NeurIPS 2025 · similarity 0.82)
• TrajAgent: An LLM-Agent Framework for Trajectory Modeling via Large-and-Small Model Collaboration (NeurIPS 2025 · similarity 0.82)