Research map / Image generation / Fast diffusion sampling
ProDiff: Prototype-Guided Diffusion for Minimal Information Trajectory Imputation
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 ProDiff, a prototype-guided conditional diffusion model that imputes trajectory points from only two endpoints, jointly trained with prototype clustering and contrastive losses.
Problem
Trajectory imputation under minimal information: reconstructing missing intermediate points given only the two endpoints of a fixed-length window, without relying on dense sparse trajectories or road graphs.
Approach
Combine a diffusion model (1D-UNet denoiser) with a prototype condition extractor that embeds unlabeled trajectories into learned prototypes capturing macro-level movement patterns, trained jointly with the diffusion loss plus K-means-consistency and prototype contrastive losses.
Key results
- On WuXi with k=4, ProDiff reaches 0.7155 TC@2k versus 0.6958 for DiffTraj (the best baseline).
- On FourSquare with k=8, ProDiff TC@2k drops by 11.59% from k=4 while DiffTraj drops by 21.02%.
- Ablation on WuXi k=8: removing the prototype condition extractor lowers TC@2k from 0.5752 to 0.4486.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ProDiff instead of the strongest baseline | trajectory imputation · WuXi, k=4 | TC@2k: 0.6958 -> 0.7155 | Helped |
| use ProDiff instead of the strongest baseline | trajectory imputation · FourSquare, k=4 | TC@10k: 0.8928 -> 0.8971 | Helped |
| add the prototype condition extractor (ablation) | ProDiff · WuXi, k=8 | TC@2k: 0.4486 -> 0.5752 | Helped |
| add classification consistency loss L_C1 (ablation) | ProDiff · WuXi, k=8 | TC@2k: 0.5395 -> 0.5752 | Helped |
| add contrastive loss L_C2 (ablation) | ProDiff · WuXi, k=8 | TC@2k: 0.4888 -> 0.5752 | Helped |
| increase prototype count from 15 to 20 (ablation) | ProDiff · WuXi, k=6 | TC@10k: 0.9237 -> 0.9285 | Helped |
| increase prototype count from 20 to 25 (ablation) | ProDiff · WuXi, k=8 | TC@10k: 0.8945 -> 0.9017 | Helped |
| increase diffusion steps from 100 to 300 (ablation) | ProDiff · WuXi | TC@10k: 0.9187 -> 0.9300 | Helped |
| increase diffusion steps from 300 to 700 (ablation) | ProDiff · WuXi | TC@10k: 0.9300 -> 0.9216 | Hurt |
| replace MASK with prototype conditioning (ablation) | cVAE · WuXi | TC@10k: 0.4513 -> 0.5009 | Helped |
| replace MASK with prototype conditioning (ablation) | cGAN · WuXi | TC@10k: 0.4896 -> 0.5883 | Helped |
| provide 4 fixed trajectory points instead of 2 (ablation) | ProDiff · WuXi, length 10 | TC@2k: 0.4996 -> 0.6820 | Helped |
Surprising findings
- Trajectory coverage with random (rather than fixed) supplementary points is notably lower, diverging from the cited finding that four points suffice for 95% of trajectories.
Limitations
- Diffusion models incur higher computational cost than alternatives.
- Performance drops markedly when supplementary points are randomly selected rather than fixed.
- Evaluation uses only two datasets, and FourSquare is limited to Tokyo data.
Experimental setup
Datasets: WuXi mobile signal data Foursquare check-ins (Tokyo)
Benchmarks: WuXi FourSquare
Metrics: TC@2k TC@4k TC@6k TC@8k TC@10k Density Distance Segment Distance Radius MAE
Baselines: VAR SAITS TimesNet Diffusion-TS DiffTraj RNTrajRec TS-TrajGen MM-STGED AttnMove
Builds on: DDPM Diffusion-TS DiffTraj SAITS
Closest papers by meaning
- Generative Human Trajectory Recovery via Embedding-Space Conditional Diffusion (ICML 2025 · similarity 0.90)
- Pedestrian Trajectory Prediction with Missing Data: Datasets, Imputation, and Benchmarking (NeurIPS 2024 · similarity 0.88)
- MS-TIP: Imputation Aware Pedestrian Trajectory Prediction (ICML 2024 · similarity 0.86)
- SingularTrajectory: Universal Trajectory Predictor Using Diffusion Model (CVPR 2024 · similarity 0.86)
- GeoSPRINT: Geometric Redundancy-Aware Step Pruning for Inference in Diffusion Trajectories (NeurIPS 2026 · similarity 0.85)
- PLMTrajRec: A Scalable and Generalizable Trajectory Recovery Method with Pre-trained Language Models (NeurIPS 2025 · 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.