Co-evolving EM with PSLR generates time series from irregular data
The paper tackles generating regular time series from irregular ones. The usual approach reduces this to one-shot imputation: complete each partial sequence once, freeze the completions, and train a generator on the resulting surrogate dataset. The authors show that this open-loop strategy is structurally flawed. In their account, point imputers collapse ambiguous missing regions to conditional averages, while stochastic imputers become stale as the generator evolves. From this they conclude that the binding constraint is iteration, not imputer design.
Their fix is a co-evolving Monte Carlo EM framework that closes the loop between imputer and generator. Each E-step samples missing values from the posterior of the current generator. Each M-step retrains the generator on those completions, which preserves unconditional generator training.
There is a technical complication. The diffusion prior works in a lifted 2D representation, while the observations live in time-series space. To handle this, the authors introduce Posterior Sampling for Lifted Representations (PSLR). It enforces operator-consistent conditioning, uncertainty preservation, and manifold consistency during the E-step.
Across ten datasets and four corruption regimes, the authors report state-of-the-art performance. They say the method improves over the strongest baselines by 59% in discriminative score, 15% in predictive score, 73% in Context-FID, and 71% in feature-correlation error. Ablations show that the approach reduces average discriminative score by 73% over single-step baselines, closing most of the remaining gap to the clean-data oracle. Code is available at https://github.com/azencot-group/PSLR.
Key facts
- The common pipeline (impute once, freeze the completions, train a generator) is called structurally flawed: point imputers average ambiguous gaps away, and stochastic imputers go stale as the generator changes.
- The proposed co-evolving Monte Carlo EM framework samples missing values from the current generator's posterior in each E-step and retrains the generator on them in each M-step.
- PSLR (Posterior Sampling for Lifted Representations) handles the mismatch between a diffusion prior in a lifted 2D representation and observations in time-series space.
- Reported gains over the strongest baselines: 59% in discriminative score, 15% in predictive score, 73% in Context-FID, 71% in feature-correlation error, across ten datasets and four corruption regimes.
- In ablations, the method cuts average discriminative score by 73% against single-step baselines and closes most of the remaining gap to the clean-data oracle. Code is on GitHub at azencot-group/PSLR.
Why it matters
Training a generator on irregular data is often handled by filling the gaps once and moving on. This paper's claim is that the weakness lies in that one-pass design, not in the choice of imputer: the authors say the binding constraint is iteration. If that holds, a better imputer alone will not fix the problem, and the imputer and generator need to improve together.
Who it affects
Researchers and engineers building generative models for time series whose training data has gaps or irregular sampling. The paper frames the task as generating regular time series from irregular ones, so it speaks to anyone currently using an impute-then-train pipeline.
How to use it
Code is available at https://github.com/azencot-group/PSLR. The practical shift the paper proposes is to stop freezing imputed data. Instead, alternate between sampling missing values from the current generator's posterior (E-step) and retraining the generator on those completions (M-step). For a diffusion prior in a lifted 2D representation, PSLR is the proposed way to do the E-step sampling.
How solid is it
The results are the authors' own, reported in the abstract: state-of-the-art performance across ten datasets and four corruption regimes, with relative gains of 59%, 15%, 73% and 71% over the strongest baselines on four metrics. The ablation figure of a 73% reduction in average discriminative score versus single-step baselines comes from the same abstract. Code is public, so the claims can be checked by anyone who runs it.
Risks and caveats
The abstract does not name the datasets, baselines or corruption regimes, and it does not give missingness rates. It does not say whether the percentage gains are averages across datasets or best cases. No absolute metric values are given, and no training cost, compute or runtime figures. The abstract names no authors or institutions, and no venue or peer-review status is stated. All performance numbers are self-reported.
“the binding constraint is therefore iteration, not imputer design”
— From the paper's abstract