TimeDP: Learning to Generate Multi-Domain Time Series with Domain Prompts

Source

Status And Credibility

TimeDP was posted to arXiv on 2025-01-09, and the arXiv metadata comment states AAAI 2025. It is an official Microsoft Research/TimeCraft paper with paper-linked code under YukhoY/TimeDP, a Microsoft TimeCraft implementation, and a Microsoft Research blog published on 2025-05-13.

Core Claim

TimeDP is a multi-domain diffusion model for time-series generation. It learns semantic prototype vectors and a Prototype Assignment Module so a few examples from a target domain can be converted into a domain prompt for generating synthetic samples.

Key Contributions

  • Learns time-series prototypes as reusable basis elements across domains.
  • Uses prototype assignment weights as domain prompts for diffusion generation.
  • Supports few-shot unseen-domain generation without labels or text descriptions.
  • Evaluates in-domain and unseen-domain generation across energy, transportation, weather, and finance datasets.

Evidence And Results

The paper evaluates MMD, KL, and marginal distribution distance on 12 datasets. It reports state-of-the-art in-domain generation quality and strong unseen-domain generation compared with GAN, VAE, and diffusion baselines. The Microsoft blog repeats the headline result that TimeDP improves MMD and KL over baselines.

Limitations

  • TimeDP is domain-prompted synthetic generation, not forecasting from observed history.
  • The prototype vocabulary is interpretable only to the extent that learned assignments match meaningful temporal factors.
  • It does not expose actions, control inputs, interventions, or counterfactual alternatives.
  • Generation quality is mostly judged by distributional similarity, not downstream decision utility.
  • The paper motivates privacy-preserving synthetic data, but it does not report duplicate, nearest-neighbor, subsequence, membership-inference, or checkpoint-age memorization probes; treat MMD, KL, and marginal distribution distance as fidelity evidence, not privacy evidence.

Foundation TSFM Relevance

Agenda slotVerdictEvidenceMissing pieces
Time-series generation and editingpartially closesLearns prototype-conditioned diffusion generation across multiple time-series domains.No editing interface and no action-conditioned rollout.
Context interfaceadjacentFew target-domain examples become domain prompts.Prompts encode style/domain, not rich context, exogenous variables, or actions.
Data diversitypartially closesTargets low-resource unseen-domain generation.Needs artifact and leakage audits when generated samples are used for pretraining or benchmarking.