Diff-MN: Diffusion Parameterized MoE-NCDE for Continuous Time Series Generation with Irregular Observations
Source
- Raw Markdown: paper_diff-mn-2026.md
- PDF: paper_diff-mn-2026.pdf
- Preprint: arXiv 2601.13534
- Official code: microsoft/TimeCraft/Diff-MN
Status And Credibility
Diff-MN was first posted to arXiv on 2026-01-20 and the inspected arXiv metadata reports version 2, updated 2026-01-30. It is referenced by the official Microsoft TimeCraft repository as MN-TSG and implemented under Diff-MN. No peer-reviewed venue is listed in the arXiv metadata snapshot.
Core Claim
Diff-MN targets continuous time-series generation from irregular observations. It combines a mixture-of-experts Neural Controlled Differential Equation with diffusion-generated sample-specific dynamics weights, so generated samples can be evaluated at arbitrary time points.
Key Contributions
- Replaces a single NCDE dynamics function with dense MoE dynamics.
- Uses a decoupled autoencoder-backed optimization design so training focuses on dynamics.
- Learns the joint distribution of regularized time-series observations and MoE weights through diffusion.
- Evaluates irregular-to-regular and irregular-to-continuous generation under 30%, 50%, and 70% dropped observations.
Evidence And Results
The paper evaluates ten public and synthetic datasets, including Sines, Stocks, Energy, MuJoCo, ECG datasets, and synthetic function datasets. It reports improvements over strong baselines on discriminative score, MDD, KL, and membership-inference risk, and it includes continuous-generation visualizations.
Limitations
- Irregularity is simulated by random observation dropping for several datasets, so missingness mechanisms may be simpler than real clinical or operational sampling.
- The method is evaluated as generation/reconstruction from irregular observations, not as action-conditioned dynamics.
- Continuous interpolation quality does not by itself prove useful latent state for decisions.
- Diffusion over MoE weights adds model complexity and may make serving cost part of the contract.
Foundation TSFM Relevance
| Agenda slot | Verdict | Evidence | Missing pieces |
|---|---|---|---|
| Irregular time and continuous state | partially closes | Generates continuous trajectories from irregular observations through MoE-NCDE dynamics. | Needs real irregular domains and streaming update tests. |
| Time-series generation and editing | partially closes | Supports irregular-to-regular and irregular-to-continuous generation. | No explicit edit instructions or action-conditioned rollouts. |
| Dense numeric fidelity | adjacent | Reports DS, MDD, KL, MIR, and visualizations. | Needs calibrated uncertainty and downstream utility tests. |