Research map / Time series / Foundation models for time series
TimeDP: Learning to Generate Multi-Domain Time Series with Domain Prompts
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 TimeDP, a label-free, text-free multi-domain time series diffusion model that conditions generation on domain prompts built from learned prototype assignments, enabling in-domain and few-shot unseen-domain generation.
Problem
Generating time series across multiple domains with one model is hard because domain patterns diverge widely; existing generators are single-domain and domain-label or text conditioning is limited for unseen or unlabeled domains.
Approach
A set of learnable time series prototypes (basis "words") is shared across domains; a prototype assignment module maps each sequence to non-negative weights over prototypes, which serve as cross-attention conditions for a U-Net diffusion model. At sampling, domain prompts are formed from few-shot target-domain samples' assignments, enabling generation without labels or text.
Key results
- In-domain generation at sequence length 168: TimeDP achieves best MMD on 11 of 12 datasets, e.g., Electricity MMD 0.001 vs 0.002 for TimeVQVAE-C and Exchange MMD 0.151 vs 0.233.
- Unseen-domain generation (Stock, Web) with 3-100 shots: DP-Diff (TimeDP) gets best MMD in most settings, e.g., Stock 100-shot MMD 0.014 vs 0.096 for TimeVQVAE unconditional.
- Ablation: removing PAM raises average MMD from 0.071 to 0.281; removing the prompt raises it to 0.143 (16 prototypes is default).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TimeDP with learned domain prompts rather than multi-domain baselines | Time series generation across 12 in-domain datasets · In-domain datasets, sequence length 168 | MMD | Helped |
| Use domain prompts for few-shot unseen-domain generation | TimeDP · Stock and Web, 3, 10, and 100 shots | MMD: Stock: 0.048_{\pm 0.014}, 0.050_{\pm 0.012}, 0.014_{\pm 0.005}; Web: 0.068_{\pm 0.035}, 0.065_{\pm 0.014}, 0.021_{\pm 0. | Helped |
| Increase few-shot prompt samples | TimeDP unseen-domain generation · Stock and Web | MMD: Stock: 0.048_{\pm 0.014} -> 0.014_{\pm 0.005}; Web: 0.068_{\pm 0.035} -> 0.021_{\pm 0.004} | Helped |
| Add the prototype assignment module (ablation) | TimeDP · In-domain datasets, length 168 | Average MMD: 0.281 -> 0.074 | Helped |
| Add prompt conditioning (ablation) | TimeDP · In-domain datasets, length 168 | Average MMD: 0.143 -> 0.074 | Helped |
| Vary prototype count among 4, 8, 16, 32, and 64 (ablation) | TimeDP · In-domain datasets, length 168 | Average MMD: 0.083, 0.071, 0.074, 0.073, 0.072 | Mixed |
| Fine-tune baseline models with few-shot data | Time-series generation baselines · Stock and Web unseen domains | MMD | No clear effect |
Surprising findings
- Removing the prompt conditioning gives better average KL (0.193) than full model, and removing PAM gives the best average KL (0.181), suggesting MMD and KL disagree on which component matters.
- Fine-tuning baselines on few-shot unseen-domain data does not consistently improve over unconditional generation.
Limitations
- Evaluated only on univariate series with channel-independent treatment.
- Prototypes are frozen after random orthogonal initialization, which may limit expressiveness.
- Unseen-domain evaluation uses only two extra datasets (Stock, Web).
Experimental setup
Datasets: Electricity Solar Wind Traffic Taxi Pedestrian Air Quality Temperature Rain NN5 Fred-MD Exchange
Benchmarks: Electricity Solar Wind Traffic Taxi Pedestrian Air Quality Temperature Rain NN5 Fred-MD Exchange
Metrics: Maximum Mean Discrepancy Kullback-Leibler divergence Marginal Distribution Difference
Baselines: TimeGAN GT-GAN TimeVAE TimeVQVAE TimeVQVAE-C
Builds on: DDPM Diffusion-TS TimeVQVAE
Closest papers by meaning
- Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach (NeurIPS 2025 · similarity 0.89)
- Large Pre-trained time series models for cross-domain Time series analysis tasks (NeurIPS 2024 · similarity 0.87)
- Towards a General Time Series Forecasting Model with Unified Representation and Adaptive Transfer (ICML 2025 · similarity 0.87)
- TEMPO: Prompt-based Generative Pre-trained Transformer for Time Series Forecasting (ICLR 2024 · similarity 0.87)
- Population Aware Diffusion for Time Series Generation (AAAI 2025 · similarity 0.86)
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.86)
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.