Skip to main content
Back to timeline
arXivSource publication:

Irregular time-series generation stops freezing its imputations: PSLR co-evolves filler and generator, cutting discriminative score 59% below the strongest baseline

Synopsis

Irregular time-series generation usually fills missing values once, freezes them, and then trains a generator; PSLR instead runs a Monte Carlo EM loop in which each round samples complete series from the current generator's posterior and retrains the generator on them. Because the diffusion prior lives in a lifted 2D representation while observations live in time-series space, PSLR enforces operator-consistent conditioning, uncertainty preservation and manifold consistency to connect posterior sampling across the two spaces. Across ten datasets and four corruption regimes, it improves discriminative score, predictive score, Context-FID and feature-correlation error by 59%, 15%, 73% and 71% over the strongest baselines.

Source-provided article image: Closing the Loop: Co-Evolving EM for Irregular Time Series Generation in Lifted Representations
arXiv

Interpretation

Regular generation from irregular time series is recast as a closed-loop training problem: filling missing values is no longer a one-time preprocessing step but is updated together with the generator. The prior approach was open-loop commit-and-forget: a pretrained imputer ran once over each partial sequence, its completions were frozen, and the generator was trained on them with no further feedback from the generator to the imputer. On ten datasets and four corruption regimes (standard timestamp masking, contiguous block gaps, per-cell asynchronous missingness and continuous-time fractional resampling), the method improves discriminative score, predictive score, Context-FID and feature-correlation error by 59%, 15%, 73% and 71% over the strongest baselines.

The loop alternates two steps: an E-step samples missing values from the current generator's posterior, and an M-step retrains the generator on those completed series with the standard diffusion denoising objective. The M-step is identical to training an image diffusion generator on complete data, with no masked loss, auxiliary imputation head or separate conditional generator; the only difference from standard training is that the completed training set is refreshed between rounds using the current generator. An ablation on Energy, Weather and Stock compares average discriminative score: the deterministic point-imputation baseline scores 0.218, swapping in the stochastic CSDI imputer gives 0.095, a single sampling step with extended training gives 0.115, the full loop gives 0.027, and the clean-data oracle gives 0.006.

To sample posteriors across the two spaces, PSLR conditions through the true observation operator, retains the full posterior covariance, and projects samples back onto the subspace of valid lifted representations. Off-the-shelf diffusion posterior samplers assume the observation operator acts in the model's own space; dropped into the lifted setting they use a wrong image-space mask, collapse multimodal posteriors to a mean, and let samples drift off the valid manifold. Removing any one of the three conditions worsens average discriminative score: closing the loop with a naive single-space sampler gives 0.087, dropping operator-consistent conditioning gives 0.038, dropping uncertainty preservation gives 0.047, dropping manifold consistency gives 0.056, and all three together give 0.027.

Perspective

The demonstrated scope is four corruption regimes on ten public time-series datasets, with training only on corrupted observations and evaluation comparing unconditionally generated samples against fully observed data; it targets research and engineering settings that need regular series generated from irregular observations.

When whole timesteps are dropped, the observations determine only marginals over short windows, so the model can match the observed data and still differ from the truth; EM converges only to a local optimum, so the final solution depends on initialization; wall-clock is roughly that of the fastest single-pass diffusion baseline, ImagenI2R, and is not in the cheapest compute tier.