Closing the Loop: Co-Evolving EM for Irregular Time Series Generation in Lifted Representations
Organizations: Department of Computer Science, Ben-Gurion University of The Negev
Abstract
Regular time-series generation from irregular time series is typically reduced to one-shot imputation: complete each partial sequence once, freeze the completions, and train a generator on the resulting surrogate dataset. We show that this open-loop strategy is structurally flawed. Point imputers collapse ambiguous missing regions to conditional averages, while stochastic imputers become stale as the generator evolves; the binding constraint is therefore iteration, not imputer design. We introduce a co-evolving Monte Carlo EM framework that closes the imputer-generator loop. Each E-step samples missing values from the posterior of the current generator, and each M-step retrains the generator on these completions, preserving unconditional generator training. Since our diffusion prior operates in a lifted 2D representation while observations live in time-series space, we further introduce Posterior Sampling for Lifted Representations (PSLR), which enforces operator-consistent conditioning, uncertainty preservation, and manifold consistency during the E-step. Across ten datasets and four corruption regimes, our method achieves state-of-the-art performance, improving 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 our 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.
Figures & tables
| Metric | Setting | TimeGAN- | GT-GAN | KoVAE | ImagenI2R | Ours |
|---|---|---|---|---|---|---|
| Disc. | Timestep (Standard) | 0.496 | 0.387 | 0.228 | 0.067 | |
| Blocking | 0.499 | 0.432 | 0.321 | 0.366 | ||
| Cell | 0.494 | 0.447 | 0.341 | 0.304 | ||
| Continuous | 0.498 | 0.427 | 0.326 | 0.274 | ||
| Pred. | Timestep (Standard) | 0.317 | 0.168 | 0.150 | 0.132 | |
| Blocking | 0.377 | 0.139 | 0.131 | 0.155 |
| Method | Closed-loop | Imputer | PSLR | Energy | Weather | Stock | Avg. |
| Point imputer (NCDE) + ImagenTime | ✗ | Deterministic | ✗ | 0.225 | 0.281 | 0.149 | 0.218 |
| CSDI + ImagenTime | ✗ | Stochastic | ✗ | 0.120 | 0.095 | 0.070 | 0.095 |
| Ours (1 iter, extended training) | ✗ | Posterior | ✓ | 0.140 | 0.120 | 0.085 | 0.115 |
| Lifted-mask MMPS-EM | ✓ | Posterior | ✗ | 0.095 | 0.098 | 0.096 | 0.087 |
| w/o C1 | ✓ | Posterior | partial | 0.065 | 0.035 | 0.014 | 0.038 |
| w/o C2 | ✓ | Posterior | partial | 0.072 | 0.048 | 0.022 | 0.047 |
Appendix figures & tables29 assets
Supplementary material from the paper’s appendix.
Appendix
| Weather | Energy | |||
|---|---|---|---|---|
| Initialization | Disc. | Iters. | Disc. | Iters. |
| Random noise | 0.057 | 0.045 | 10 | |
| Linear interpolation | 0.026 | 4 | 0.157 | |
| Kalman smoothing | 0.021 | 3 | 0.040 | 5 |
| STL (default) | 0.020 | 3 | 0.042 | 5 |
| Sampler | Posterior covariance | Linear solver | Dual-space failure mode | Ablation cell |
|---|---|---|---|---|
| DPS [ 9 ] | point estimate (delta) | – | MAP-like, collapses posterior diversity | — |
| GDM [ 34 ] | isotropic diagonal | – | no inter-pixel correlation, drops lift redundancy | — |
| DiffPIR [ 47 ] | none (proximal) | inner solver collapses | no closed form once is composed with | — |
| TMPD [ 5 ] | row-sum diagonal | per-step closed form | still drops block correlation; converges-then-diverges | 0.112 |
| MMPS [ 32 ] | full Jacobian | CG (image-space) | wrong operator , not ; biased fixed point | |
| (ours) | full Jacobian | CG (obs-space) | none (regime a) |
| Failure mode | Fix | Disc. without fix |
|---|---|---|
| M1: conditions through an image-space mask instead of | C1 | 0.038 |
| M2: posterior uncertainty collapses, unstable linear solve | C2 | 0.047 |
| M3: samples drift off across iterations | C3 | 0.056 |
| Method | imp. MSE | imp. MAE | imp. ES | disc |
|---|---|---|---|---|
| Point regression | ||||
| Point imputer (NCDE) | ||||
| EM Iter 1 | ||||
| Frozen stochastic | ||||
| Co-evolving EM (Ours) | ||||
| Full-data generator | — | — |
| Claim | Holds in |
|---|---|
| Co-evolving EM Frozen stochastic | |
| Frozen stochastic 1 EM iter | |
| Co-evolving EM Point imputer (NCDE) | |
| Co-evolving EM Point regression | |
| Co-evolving EM within disc of Full-data generator |
| Controlled method | Main-ablation analog | Prediction |
|---|---|---|
| Point regression | ImagenTime, no masked-loss patch | worst disc / FID |
| Point imputer (NCDE) | GT-GAN / deterministic NCDE imputation step | deterministic ceiling, above Point regression |
| Stochastic ( iter) | Ours, EM iteration | captures stochastic gain, misses iteration gain |
| Frozen stochastic | CSDI-impute EDM (commit-and-forget) | best frozen baseline; bounded by capacity gap |
| Co-evolving EM (Ours) | Full pslr -EM ( , iters) | oracle-matching; wins in |
| Sine | Stocks | MuJoCo | Energy | Weather | ETTh1 | ETTh2 | ETTm1 | ETTm2 | Electricity | |
|---|---|---|---|---|---|---|---|---|---|---|
| Image size | ||||||||||
| Delay | ||||||||||
| Skip | ||||||||||
| U-Net base channels | ||||||||||
| Batch size | ||||||||||
| Learning rate |
| Sine | Stocks | Energy | Weather | ETTh1 | ETTh2 | ETTm1 | ETTm2 | |
|---|---|---|---|---|---|---|---|---|
| Image size | ||||||||
| Delay | ||||||||
| Skip | ||||||||
| U-Net base channels | ||||||||
| Batch size | ||||||||
| Learning rate |
| Sine | Stocks | Energy | Weather | ETTh1 | ETTh2 | ETTm1 | ETTm2 | |
|---|---|---|---|---|---|---|---|---|
| Image size | ||||||||
| Delay | ||||||||
| Skip | ||||||||
| U-Net base channels | ||||||||
| Batch size | ||||||||
| Learning rate |
| Drop | Drop | Drop | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Metric | Model | ETTh1 | ETTm2 | Energy | Weather | ETTh2 | ETTm1 | Weather | Energy | Stocks |
| t-SNE | ImagenI2R | |||||||||
| Ours | ||||||||||
| KDE ( ) | ImagenI2R | |||||||||
| Ours | ||||||||||
| Dataset | Method | 30% | 50% | 70% |
|---|---|---|---|---|
| Energy | L-STFT (Ours) | 0.042 | 0.075 | |
| Delay Emb. (Ours) | 0.038 | |||
| Weather | L-STFT (Ours) | 0.019 | 0.030 | |
| Delay Emb. (Ours) | 0.020 |
| Rate | KoVAE | ImagenI2R | Ours | Avg. EM iters |
|---|---|---|---|---|
| 30% | 9.89 | 3.34 | 2.89 | |
| 50% | 7.86 | 3.76 | 6.56 | |
| 70% | 8.25 | 3.68 | 5.22 | |
| Avg. | 8.67 | 3.59 | 4.89 |
| Dataset | Method | 30% | 50% | 70% | Avg. |
|---|---|---|---|---|---|
| Energy | Masked TMPD | 0.105 | 0.112 | 0.118 | 0.112 |
| Ours (L - tmpd ) | |||||
| Weather | Masked TMPD | 0.098 | 0.104 | 0.110 | 0.104 |
| Ours (L - tmpd ) | |||||
| Clean-data Oracle | 0.009 | 0.003 | 0.006 | 0.006 | |
| (A) Noise scale | |||||
|---|---|---|---|---|---|
| Energy | 0.484 (unstable) | 0.043 | 0.042 | 0.049 | 0.065 |
| Weather | 0.091 (unstable) | 0.019 | 0.017 | 0.021 | 0.034 |
| Training data | CT | SAD |
|---|---|---|
| GT-GAN | 0.379 | 0.468 |
| KoVAE | 0.653 | 0.660 |
| ImagenI2R | 0.782 | 0.817 |
| Ours | 0.938 | 0.935 |
| Oracle (real, clean) | 0.990 | 0.979 |
| MCAR | MNAR-value | MNAR-state | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Dataset | Method | |||||||||
| Weather | KoVAE | 0.229 | 0.498 | 0.269 | 0.301 | 0.499 | 0.412 | 0.287 | 0.495 | 0.403 |
| ImagenI2R | 0.029 | 0.035 | 0.106 | 0.074 | 0.112 | 0.231 | 0.089 | 0.148 | 0.257 | |
| Ours | 0.017 | 0.020 | 0.029 | 0.028 | 0.041 | 0.066 | 0.024 | 0.031 | 0.058 | |
| Energy | KoVAE | 0.280 | 0.298 | 0.392 | 0.269 | 0.281 | 0.405 | 0.266 | 0.279 | 0.398 |
| ImagenI2R | 0.048 | 0.065 | 0.128 | 0.117 | 0.171 | 0.284 | 0.095 | 0.129 | 0.246 | |
| MCAR | MNAR-latent | ||||||
|---|---|---|---|---|---|---|---|
| Dataset | Method | ||||||
| Weather | KoVAE | 0.229 | 0.498 | 0.269 | 0.312 | 0.499 | 0.437 |
| ImagenI2R | 0.029 | 0.035 | 0.106 | 0.096 | 0.184 | 0.293 | |
| Ours | 0.017 | 0.020 | 0.029 | 0.031 | 0.052 | 0.089 | |
| Energy | KoVAE | 0.280 | 0.298 | 0.392 | 0.274 | 0.288 | 0.411 |
| ImagenI2R | 0.048 | 0.065 | 0.128 | 0.124 | 0.203 | 0.318 | |
| Run | Diagnostic | start | iter 1 | iter 3 | iter 5 | iter 10 |
|---|---|---|---|---|---|---|
| ETTh1 | M-step gap | – | ||||
| Held-out CRPS | 0.0431 | 0.0385 | 0.0333 | 0.0317 | 0.0298 | |
| Likelihood bound | 0.9177 | 0.7903 | 0.6841 | 0.6353 | 0.5925 | |
| Weather | Likelihood bound | – | 0.5441 | 0.4819 | 0.4130 | 0.3741 |
| Metric | Model | Etth1 | Etth2 | Ettm1 | Ettm2 | Weather | Electricity | Energy | Sine | Stock | Mujoco |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 30% Drop | |||||||||||
| Disc. | TimeGAN- | 0.499 | 0.499 | 0.499 | 0.499 | 0.493 | 0.498 | 0.448 | 0.494 | 0.463 | 0.471 |
| GT-GAN | 0.473 | 0.371 | 0.420 | 0.369 | 0.472 | 0.409 | 0.333 | 0.363 | 0.251 | 0.249 | |
| KoVAE | 0.208 | 0.075 | 0.045 | 0.077 | 0.229 | 0.497 | 0.280 | 0.035 | 0.162 | 0.123 | |
| ImagenI2R | 0.020 | 0.009 | 0.029 | 0.399 | 0.048 | 0.013 | 0.007 | 0.009 | |||
| Ours | 0.015 | 0.015 | |||||||||
| Metric | Model | Etth1 | Etth2 | Ettm1 | Ettm2 | Weather | Energy | Sine | Stock |
|---|---|---|---|---|---|---|---|---|---|
| 30% Drop | |||||||||
| Disc. | TimeGAN- | 0.500 | 0.500 | 0.500 | 0.500 | 0.498 | 0.461 | 0.499 | 0.479 |
| GT-GAN | 0.485 | 0.384 | 0.432 | 0.381 | 0.489 | 0.347 | 0.379 | 0.268 | |
| KoVAE | 0.255 | 0.096 | 0.264 | 0.075 | 0.290 | 0.416 | 0.244 | 0.114 | |
| ImagenI2R | 0.037 | 0.036 | 0.033 | 0.028 | 0.130 | 0.004 | 0.011 | ||
| Ours | 0.208 | ||||||||
| Metric | Model | Etth1 | Etth2 | Ettm1 | Ettm2 | Weather | Energy | Sine | Stock |
|---|---|---|---|---|---|---|---|---|---|
| 30% Drop | |||||||||
| Disc. | TimeGAN- | 0.500 | 0.500 | 0.500 | 0.500 | 0.500 | 0.473 | 0.500 | 0.489 |
| GT-GAN | 0.492 | 0.395 | 0.441 | 0.394 | 0.497 | 0.358 | 0.387 | 0.281 | |
| KoVAE | 0.239 | 0.237 | 0.282 | 0.160 | 0.411 | 0.371 | 0.284 | 0.289 | |
| ImagenI2R | 0.045 | 0.032 | 0.067 | 0.047 | 0.093 | 0.145 | 0.005 | 0.025 | |
| Ours | |||||||||
| Metric | Model | Etth1 | Etth2 | Ettm1 | Ettm2 | Weather | Energy | Sine | Stock |
|---|---|---|---|---|---|---|---|---|---|
| 30% Drop | |||||||||
| Disc. | TimeGAN- | 0.500 | 0.500 | 0.500 | 0.500 | 0.499 | 0.490 | 0.499 | 0.485 |
| GT-GAN | 0.492 | 0.440 | 0.465 | 0.435 | 0.490 | 0.405 | 0.430 | 0.315 | |
| KoVAE | 0.206 | 0.210 | 0.215 | 0.212 | 0.433 | 0.457 | 0.180 | 0.249 | |
| ImagenI2R | 0.238 | 0.245 | 0.250 | 0.247 | 0.432 | 0.364 | 0.200 | 0.343 | |
| Ours | |||||||||
| Metric | Model | Etth1 | Etth2 | Ettm1 | Ettm2 | Weather | Energy | Sine | Stock |
|---|---|---|---|---|---|---|---|---|---|
| 30% Drop | |||||||||
| Disc. | TimeGAN- | 0.500 | 0.500 | 0.500 | 0.500 | 0.499 | 0.490 | 0.499 | 0.485 |
| GT-GAN | 0.492 | 0.440 | 0.465 | 0.435 | 0.490 | 0.405 | 0.430 | 0.315 | |
| KoVAE | 0.346 | 0.350 | 0.360 | 0.355 | 0.484 | 0.320 | 0.388 | ||
| ImagenI2R | 0.483 | 0.490 | 0.492 | 0.491 | 0.499 | 0.499 | 0.470 | 0.251 | |
| Ours | 0.486 | ||||||||
| Metric | Model | Etth1 | Etth2 | Ettm1 | Ettm2 | Weather | Energy | Sine | Stock |
|---|---|---|---|---|---|---|---|---|---|
| 30% Drop | |||||||||
| Disc. | TimeGAN- | 0.500 | 0.500 | 0.500 | 0.500 | 0.499 | 0.490 | 0.499 | 0.485 |
| GT-GAN | 0.492 | 0.440 | 0.465 | 0.435 | 0.490 | 0.405 | 0.430 | 0.315 | |
| KoVAE | 0.222 | 0.225 | 0.230 | 0.228 | 0.441 | 0.460 | 0.190 | 0.207 | |
| ImagenI2R | 0.133 | 0.140 | 0.145 | 0.142 | 0.262 | 0.233 | 0.120 | 0.121 | |
| Ours | |||||||||
| 30% Drop | 50% Drop | 70% Drop | ||||||||||||||
| Metric | Model | Etth1 | Ettm1 | Weather | Energy | Stock | Etth1 | Ettm1 | Weather | Energy | Stock | Etth1 | Ettm1 | Weather | Energy | Stock |
| Disc. | TimeGAN- | 0.500 | 0.492 | 0.500 | 0.500 | 0.495 | 0.500 | 0.500 | 0.480 | 0.490 | 0.460 | 0.500 | 0.500 | 0.500 | 0.500 | 0.490 |
| GT-GAN | 0.500 | 0.485 | 0.500 | 0.415 | 0.325 | 0.470 | 0.485 | 0.500 | 0.410 | 0.355 | 0.485 | 0.485 | 0.495 | 0.485 | 0.315 | |
| KoVAE | 0.242 | 0.255 | 0.452 | 0.495 | 0.187 | 0.165 | 0.215 | 0.463 | 0.478 | 0.206 | 0.217 | 0.495 | 0.470 | 0.495 | 0.273 | |
| ImagenI2R | 0.143 | 0.155 | 0.282 | 0.253 | 0.131 | 0.088 | 0.125 | 0.498 | 0.500 | 0.388 | 0.187 | 0.475 | 0.500 | 0.500 | 0.332 | |
| Ours | ||||||||||||||||
| Metric | Model | Etth1 | Weather | Energy | Stock | Etth1 | Weather | Energy | Stock |
|---|---|---|---|---|---|---|---|---|---|
| Disc. | TimeGAN- | 0.500 | 0.499 | 0.500 | 0.485 | 0.500 | 0.500 | 0.500 | 0.500 |
| GT-GAN | 0.492 | 0.490 | 0.405 | 0.315 | 0.490 | 0.500 | 0.390 | 0.335 | |
| KoVAE | 0.222 | 0.441 | 0.460 | 0.207 | 0.193 | 0.433 | 0.448 | 0.206 | |
| ImagenI2R | 0.133 | 0.262 | 0.233 | 0.121 | 0.078 | 0.488 | 0.499 | 0.378 | |
| Ours | |||||||||
Time Series Generation
Time series generation creates synthetic sequential data that preserve temporal patterns, cross-variable structure, and controllable conditions.
See all 100 papers