Research map / Image generation / Autoregressive and latent generation
Spatio-Temporal Few-Shot Learning via Diffusive Neural Network 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 GPD, a Transformer-based conditional diffusion model that generates spatio-temporal prediction network parameters from region prompts (urban KG and self-supervised temporal embeddings) for few-shot transfer to data-scarce cities.
Problem
Spatio-temporal prediction (traffic speed, crowd flow) in data-scarce cities or regions, where transferring knowledge across cities with divergent data distributions is hard; existing fine-grained transfer methods rely on elaborate matching designs.
Approach
Recast spatio-temporal few-shot learning as generative pre-training in parameter space: train a prompt-conditioned Transformer diffusion model on the optimized weights of per-region prediction networks from source cities, then sample target-region network parameters from region prompts (KG embeddings plus temporal embeddings), with several conditioning strategies and a model-agnostic design.
Key results
- GPD reduces MAE versus the best baseline by 4.31% (Washington D.C.), 17.1% (Baltimore), 2.1% (LA) and 8.17% (Didi-Chengdu), averaging over six-step predictions.
- The paper reports an average improvement of 7.87% over the best baseline across four datasets.
- Using two source cities (N+B) instead of one improves six-step predictions by 16.4% and 12.0% for Washington D.C. (per the source-city ablation).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GPD instead of the strongest baseline | Spatio-temporal few-shot prediction · Washington D.C. | MAE: -4.31% | Helped |
| Use GPD instead of the strongest baseline | Spatio-temporal few-shot prediction · Baltimore | MAE: -17.1% | Helped |
| Use GPD instead of the strongest baseline | Spatio-temporal few-shot prediction · LA | MAE: -2.1% | Helped |
| Use GPD instead of the strongest baseline | Spatio-temporal few-shot prediction · Didi-Chengdu | MAE: -8.17% | Helped |
| Add spatial and temporal prompts rather than using either prompt alone (ablation) | GPD · Washington D.C. | prediction error | Helped |
| Add inductive bias to prompt conditioning (ablation) | GPD · Multiple datasets | prediction error | Helped |
| Include prompts within Transformer self-attention rather than conditioning without self-attention (ablation) | GPD · Multiple datasets | prediction error | Helped |
| Use two source cities instead of one (ablation) | GPD · Washington D.C., Baltimore, and NYC crowd-flow prediction | MAE: 8.1% and 3.2% for Washington; 1.5% and 11.4% for Baltimore; 17.5% and 20.3% for NYC (one-step); 16.4% and 12.0% for Wash | Helped |
Limitations
- Denoising with DDPM's step-by-step sampling makes training and generation slower than some baselines (about 3 hours).
- Region prompts depend on a pre-trained urban knowledge graph and temporal embeddings from three days of target data; prompt design is hand-crafted.
- Only evaluated on a handful of city datasets with a single target city per experiment.
Experimental setup
Datasets: NYC taxi flow Washington D.C. crowd flow Baltimore crowd flow METR-LA PEMS-BAY Didi-Chengdu Didi-Shenzhen
Benchmarks: Washington D.C. crowd flow Baltimore crowd flow LA traffic speed Didi-Chengdu traffic speed
Metrics: MAE RMSE
Baselines: HA ARIMA RegionTrans DASTNet AdaRNN MAML TPB ST-GFSL
Builds on: DDPM STGCN Graph WaveNet STID TuckER
Closest papers by meaning
- Diffusion Transformers as Open-World Spatiotemporal Foundation Models (NeurIPS 2025 · similarity 0.89)
- Drawing Informative Gradients from Sources: A One-stage Transfer Learning Framework for Cross-city Spatiotemporal Forecasting (AAAI 2025 · similarity 0.86)
- FlashST: A Simple and Universal Prompt-Tuning Framework for Traffic Prediction (ICML 2024 · similarity 0.85)
- Get Rid of Isolation: A Continuous Multi-task Spatio-Temporal Learning Framework (NeurIPS 2024 · similarity 0.85)
- GeoPro-Net: Learning Interpretable Spatiotemporal Prediction Models Through Statistically-Guided Geo-Prototyping (AAAI 2025 · similarity 0.85)
- Efficient Traffic Prediction Through Spatio-Temporal Distillation (AAAI 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.