Research map / Image generation / Autoregressive and latent generation
Generative Human Trajectory Recovery via Embedding-Space Conditional Diffusion
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 DiffMove, a conditional diffusion model that denoises location embeddings (conditioned on current and historical trajectories) and decodes them to discrete POI/grid locations for human trajectory recovery.
Problem
Recovering missing locations in sparse, irregularly observed human trajectories, where locations are discrete IDs rather than continuous values that standard diffusion imputation models handle.
Approach
Runs the diffusion denoising process in a learned location-embedding space conditioned on observed current-day locations and historical trajectories (via gated GNN over transition graphs, cross-attention for periodicity, and a target condition block), then decodes embeddings to discrete locations by inner-product matching against the embedding table.
Key results
- On Geolife, DiffMove improves Recall@1 over the best baseline TRILL by 9.57% (0.5173 vs 0.4721).
- On Foursquare, DiffMove improves Recall@1 over TRILL by 11.56% (0.3600 vs 0.3227).
- Reducing generated samples to 1 lowers Recall@1 by 1.85% on Geolife and 1.7% on Foursquare, supporting probabilistic generation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the embedding-space conditional diffusion (DiffMove) versus best baseline TRILL | DiffMove · Geolife | Recall@1: +9.57% Recall (relative, as reported) | Helped |
| Add the embedding-space conditional diffusion (DiffMove) versus best baseline TRILL | DiffMove · Foursquare | Recall@1: +11.56% Recall (relative, as reported) | Helped |
| Add the embedding-space conditional diffusion (DiffMove) versus best baseline TRILL | DiffMove · Geolife | Distance@1: 19.66% lower distance (relative) | Helped |
| Use a single generated sample (M=1) instead of M=4 generations (ablation) | DiffMove · Geolife | Recall@1: 0.5173 -> 0.4988 (-1.85 points) | Hurt |
| Use a single generated sample (M=1) instead of M=4 generations (ablation) | DiffMove · Foursquare | Recall@1: 0.3600 -> 0.3430 (-1.7 points) | Hurt |
| Remove all conditioning (unconditional diffusion) (ablation) | DiffMove · Foursquare | Recall@1: 0.3600 -> 0.0416 (-88.44%) | Hurt |
| Add Spatial Conditional Block (ablation) | DiffMove · Foursquare | Recall@1: 0.3382 -> 0.3600 (-6.06% when removed) | Helped |
| Add Target Conditional Block (ablation) | DiffMove · Foursquare | Recall@1: 0.3493 -> 0.3600 (-2.97% when removed) | Helped |
| Add Temporal Transformer (ablation) | DiffMove · Foursquare | Recall@1: 0.3023 -> 0.3600 (-16.03% when removed) | Helped |
| Increase missing ratio of historical trajectories from 20% to 80% | DiffMove · Geolife (robustness) | Dist@10: DiffMove 80% missing 695 vs TRILL 20% 682 (not better); DiffMove 80% beats TRILL 40% (720) | Helped |
Surprising findings
- Removing the unconditional diffusion's conditioning drops Foursquare Recall to 0.0416, showing the bare diffusion model is nearly nonfunctional for this task.
Limitations
- Evaluated only under random masking; structured missingness (e.g., persistent device-failure gaps) is not studied.
- Only two datasets (Foursquare and Geolife) are used for evaluation.
Experimental setup
Datasets: Foursquare Geolife
Benchmarks: Foursquare Geolife
Metrics: Recall@K Distance@K MAP
Baselines: Top Markov PMF LSTM BiLSTM DeepMove AttnMove PeriodicMove TRILL
Builds on: DDPM CSDI AttnMove TRILL DiffWave
Closest papers by meaning
- ProDiff: Prototype-Guided Diffusion for Minimal Information Trajectory Imputation (ICML 2025 · similarity 0.90)
- Pedestrian Trajectory Prediction with Missing Data: Datasets, Imputation, and Benchmarking (NeurIPS 2024 · similarity 0.87)
- Holistic Semantic Representation for Navigational Trajectory Generation (AAAI 2025 · similarity 0.86)
- TrajCLIP: Pedestrian trajectory prediction method using contrastive learning and idempotent networks (NeurIPS 2024 · similarity 0.86)
- Rethinking Diffusion for Text-Driven Human Motion Generation: Redundant Representations, Evaluation, and Masked Autoregression (CVPR 2025 · similarity 0.86)
- SingularTrajectory: Universal Trajectory Predictor Using Diffusion Model (CVPR 2024 · similarity 0.85)
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.