Research map / Probabilistic and scientific modeling / Causal discovery
Causal Deciphering and Inpainting in Spatio-Temporal Dynamics via Diffusion Model
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 CaPaint, a plug-in that uses ViT attention scores to split spatio-temporal data into causal and environmental patches, then inpaints the environmental patches with a fine-tuned DDPM to augment training data under front-door adjustment.
Problem
Spatio-temporal prediction suffers from scarce, imbalanced sensor data and models lacking causal grounding and generalizability; prior causal ST augmentation (NuwaDynamics) has exponential training complexity and can generate non-existent distributions.
Approach
Identify causal vs. environmental patches via self-supervised ViT reconstruction attention scores, then apply diffusion inpainting only to environmental patches, justified by front-door adjustment, so augmentation cost is roughly linear in patches rather than exponential over environment samples.
Key results
- Integrating CaPaint improves MAE/MSE across seven backbones on five datasets, with relative gains of 3.7% to 77.3% (e.g., ViT on DRS MAE 13.59 to 7.52).
- On TaxiBJ+, CaPaint augmentation yields MAE 12.87 versus 15.11 for NuwaDynamics mixup augmentation and 15.94 for cropping.
- Gains are largest in low-data regimes; e.g., SSIM improvement exceeds 50% for ViT on TaxiBJ+ with 10% of training data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CaPaint to ViT | ViT · TaxiBJ+ (10→10) | MAE: 16.59 -> 14.54 | Helped |
| Add CaPaint to ViT | ViT · TaxiBJ+ (10→10) | MSE: 11.40 -> 8.89 | Helped |
| Add CaPaint to ViT | ViT · KTH (10→10) | MAE: 32.03 -> 29.52 | Helped |
| Add CaPaint to ViT | ViT · KTH (10→10) | MSE: 36.11 -> 32.79 | Helped |
| Add CaPaint to ViT | ViT · SEVIR (10→10) | MAE: 18.69 -> 17.56 | Helped |
| Add CaPaint to ViT | ViT · SEVIR (10→10) | MSE: 9.93 -> 9.16 | Helped |
| Add CaPaint to ViT | ViT · DRS (10→10) | MAE: 13.59 -> 7.52 | Helped |
| Add CaPaint to ViT | ViT · DRS (10→10) | MSE: 6.21 -> 1.41 | Helped |
| Add CaPaint to ViT | ViT · FireSys (10→10) | MAE: 17.32 -> 15.97 | Helped |
| Add CaPaint to ViT | ViT · FireSys (10→10) | MSE: 23.40 -> 21.06 | Helped |
| Use CaPaint rather than NuWa augmentation | ST forecasting models · DRS | MAE: 2.02_{± 0.09} -> 1.57_{± 0.14} | Helped |
| Use CaPaint rather than NuWa augmentation | ST forecasting models · KTH | MAE: 22.32_{± 0.94} -> 20.56_{± 1.02} | Helped |
Surprising findings
- Traditional augmentations (flip, rotate, crop) often yield results on par with or slightly worse than no augmentation, suggesting they disrupt ST structure.
Limitations
- Gains are most pronounced under data scarcity and smaller on abundant, uniform data.
- Only single-channel generated copies per sequence were evaluated in main results; arbitrary-channel generation left to future work.
- Causal patch selection is heuristic (attention column sums) rather than verified causal identification.
Experimental setup
Datasets: TaxiBJ+ KTH SEVIR DRS FireSys
Benchmarks: TaxiBJ+ KTH SEVIR DRS FireSys
Metrics: MAE MSE SSIM
Baselines: NuwaDynamics Flip Rotate Crop ConvLSTM PredRNN-V2 MAU SimVP MMVP Earthfarsser ViT
Builds on: NuwaDynamics DDPM Vision Transformer RePaint-style diffusion inpainting
Closest papers by meaning
- Causal Spatio-Temporal Prediction: An Effective and Efficient Multi-Modal Approach (NeurIPS 2025 · similarity 0.90)
- Discovering Latent Causal Graphs from Spatiotemporal Data (ICML 2025 · similarity 0.89)
- RDPI: A Refine Diffusion Probability Generation Method for Spatiotemporal Data Imputation (AAAI 2025 · similarity 0.88)
- SPACETIME: Causal Discovery from Non-Stationary Time Series (AAAI 2025 · similarity 0.88)
- Seahorse: A Unified Benchmarking Framework for Spatiotemporal Event Modeling (NeurIPS 2026 · similarity 0.87)
- Learning from Highly Sparse Spatio-temporal Data (NeurIPS 2024 · similarity 0.87)
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.