acceptodds
Under review as a conference paper at ICLR 2027

Recoverability-Aware Imputation with Conditional Bayes Risk Estimation for Multivariate Time Series

Abstract

The problem of missing values in time series is usually framed as a prediction problem where, given the contextual information, the model predicts every missing value. Though conceptually correct this approach overlooks a crucial aspect: the missing regions may differ in how recoverable they are given the available information. In this paper, we propose a recoverability-aware perspective of time series imputation that takes into account not only what to impute but also the uncertainty in this imputation. In our framework, the unrecoverability of the missing region is quantified via the conditional Bayes risk of the corresponding mask-aware missing region through a Contextual UnRecoverability Estimator (CURE) that learns this quantity from incomplete sequences. We then use the standard Bayes-risk decomposition to combine CURE with an imputer-specific residual-risk correction for selective imputation. Our experiments on synthetic dynamical systems demonstrate that regions with same size can have different levels of oracle unrecoverability for different generating configurations, and the imputers' errors correlate highly with the normalized oracle risk. Across five real-world datasets, seven imputers, MCAR and MAR missingness, and point and block corruption, CURE achieves a mean error-ranking correlation of 0.372, which increases to 0.507 after incorporating the imputer-specific correction. Total-risk ranking reduces AURC on average by 58.7% relative to random selection, compared with 53.8% for the strong imputer–CURE disagreement baseline, and yields lower AURC than disagreement in 56.3% of matched evaluations. For 50% coverage, selective RMSE is reduced by 38.4–62.7% for all five datasets relative to full coverage. These results support recoverability as a measurable basis for identifying which imputations can be trusted.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-Series Imputation, Recoverability Estimation, Conditional Bayes Risk, Selective prediction, Uncertainty Quantification

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6h7wOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper estimates unrecoverability and model-specific error to select reliable time-series imputations.

Critical strengths

The conditional Bayes-risk formulation clearly separates contextual unrecoverability from imputer-specific error, providing a useful conceptual basis for assessing when imputations can be trusted.

Critical weaknesses
  1. The experimental evaluation is limited to the ETT series and one air-quality dataset, and the baselines are relatively outdated.

  2. Figure 2, the main framework diagram, contains noticeable color inconsistencies across modules, leading me to suspect that generative AI may have been used to create it. However, the AI use statement does not explain in detail whether or how AI was used in preparing this figure.

  3. The paper has substantial formatting and presentation issues, including large blank spaces around Section 5 and poor readability in parts of the appendix.

  4. The paper emphasizes distinguishing insufficient contextual information from limitations of the imputer, yet the final selection is based on total risk. Regardless of the source of error, the action remains the same: retain or reject the imputation. The motivation for estimating these components separately is therefore insufficiently justified.

Questions for authors

See Weakness。

2 comments

Thank you for the critical review. We address the four weaknesses in turn, starting with the conceptual one, which we agree is central.

Why estimate the two components separately if the final action is retain or reject?

We give 2 answers both of which are supported by our experiments.

1. The components point to different remedies. For every test region we computed the error of the best of the seven imputers ("oracle switching") and the share of each imputer's error that switching would remove.

  • Across U^G\hat{U}_G deciles, the best-of-seven error rises about 27-fold from the lowest to the highest decile (Spearman with U^G\hat{U}_G: 0.375). Regions CURE marks as unrecoverable stay hard whichever model is used: the remedy there is more or better observations, not a different imputer.
  • The share of error removable by switching falls from 0.91 to 0.76 as U^G\hat{U}_G rises.
  • The excess term tracks the removable error on the ETT datasets (Spearman 0.22–0.41), but only weakly on Beijing AQ (0.01) and PhysioNet (0.09), which we report as a limitation.
  • U^G\hat{U}_G is computed once and shared by all imputers, which is what makes this diagnosis possible; a single total-risk score cannot provide it.

2. The decomposition also yields a better selection rule. Proposition 1 implies a training-free plug-in estimate of total risk: U^G\hat{U}_G plus the deviation DG,fD_{G,f} of the imputer from CURE's conditional mean. It improves AURC over disagreement alone by 18.9% (95% CI 16.5–21.2), on every one of the six datasets, and over a learned direct predictor of total error with identical inputs by 18.1%. So even when the action is retain or reject, the decomposition matters because it tells us how to estimate the imputer-specific part.

Datasets and baselines.

  • A sixth dataset: PhysioNet 2012 (set-a, 4,000 ICU patients, about 80% natively missing), a clinical domain with heavy native missingness, run with 1 training seed × 3 mask seeds because of compute limits. Results are consistent with the other datasets: U^G\hat{U}_G's Spearman with error is 0.42, and U^G+DG,f\hat{U}_G+D_{G,f} reduces AURC relative to random by 61.6% and beats disagreement by 28.9%.
  • What we could not add: we attempted Electricity and PEMS, but they exceeded our compute budget; we will state this limitation explicitly in the paper.
  • The imputers include recent architectures (TimesNet, 2023; ImputeFormer, 2024) alongside CSDI, SAITS, BRITS and GP-VAE.
  • New scoring baselines: we added the baselines that compete directly with our risk score: learned direct total-risk predictors (with and without CURE inputs), the native predictive variance of CSDI and GP-VAE, and the plug-in U^G+DG,f\hat{U}_G+D_{G,f}. Results are in the general response and our reply to Reviewer 2.

Figure 2 and the AI-use statement. You are right. Figure 2 was generated with a generative AI tool, from a design and content specified by the authors, who checked that it depicts the method correctly. Our AI-use statement should have said so, and we apologize for the omission. In the revision we will declare this explicitly, fix the color inconsistencies, add a legend for every module, and extend the statement to cover any AI assistance with writing and code.

Formatting. In the revision we will remove the blank space around Section 5 and reorganize the appendix, with larger fonts in tables and figures and one appendix section per experiment.

Thank you for the response; however, my main concerns remain unresolved, so I will maintain my original rating.

Reviewer_hhGZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper characterizes the unrecoverability of missing time-series regions through Bayes risk conditional on the observed context. It proposes CURE to estimate this imputer-independent risk component, then learns an imputer-specific residual-risk correction to estimate total risk for selective imputation. Synthetic experiments evaluate the estimates against oracle risk, while experiments on five real-world datasets and seven imputers assess error ranking and risk–coverage performance under MCAR and MAR, with point and block missingness.

Critical strengths

S1. The paper defines recoverability relative to the observation mask, target region, and available context, providing a more precise target than gap length or uncertainty reported by a particular imputer. Although the Bayes-risk decomposition is not itself novel, applying it to rank missing regions gives selective imputation a clear analytical basis. This is the main reason I lean toward supporting the paper.

S2. The experiments examine both the validity of the risk estimates and whether the resulting rankings improve selective decisions. Synthetic systems provide an oracle reference, while real-data experiments cover multiple datasets, imputers, and missingness settings. The broadly positive error correlations across imputers on shared regions also support the presence of a difficulty component shared across models.

Critical weaknesses

W1. Baselines that directly predicts total imputation error is missing. Equation (6) shows that the residual correction can also absorb CURE estimation error, so improved ranking after adding the correction does not establish that estimating recoverability first is better than learning total risk directly. This limits the evidence for the additional value of the two-stage design.

W2. The equivalence between Eqs. (3) and (4) requires clarification when native missing values are present. The context (C_G) does not retain the native observation mask within the target region, but the loss depends on the evaluable positions (G^{\mathrm{eval}}) and their number. If these remain uncertain given (C_G), the conditional-variance average in Eq. (4) requires additional conditioning or derivation.

W3. The advantage of total-risk ranking over the strong imputer–CURE disagreement baseline is not yet sufficiently established. Total risk achieves lower AURC in only 56.3% of matched comparisons, while disagreement gives lower RMSE on ETTh2 and ETTm2 at 50% coverage. Although multiple training seeds and test masks are used, uncertainty estimates for paired differences are needed to assess the reliability of the gains.

Questions for authors

Q1. Could the authors train a model that directly predicts total imputation error on the same calibration split, using identical or comparable input features, model capacity, and training budget? Please compare its error-ranking correlation and AURC with the two-stage method to establish the additional benefit of explicitly estimating recoverability.

Q2. When a target block contains native missing values, does Eq. (4) implicitly condition on (G^{\mathrm{eval}})? If not, how does Eq. (4) follow from Eq. (3) when the evaluable positions and their number are not determined by (C_G)? Alternatively, is the equality restricted to target regions whose entries were all originally observed?

Q3. Could the authors report paired AURC differences between total-risk ranking and disagreement, with confidence intervals and results broken down by dataset? Please account for dependence arising from shared training seeds and test masks to assess whether the overall benefit is consistent across datasets.

1 comment

Thank you for the supportive and technically sharp review. Your Q1 changed how we present the method.

W1/Q1: Direct total-risk baseline.

We trained direct predictors of realized error with the same calibration units, calibration-window split, MLP (64→32→1), loss, seed and training budget as the excess model, on all six datasets (Beijing AQ: one training seed, re-evaluated on the original frozen masks and units).

  • CURE-feature Risk (same inputs zG,fz_{G,f} as the excess model): two-stage total risk has lower AURC by a median of 8.7% (95% CI 7.4–10.2; lower in 75.3% of 4,704 paired evaluations), significant on Beijing AQ, ETTh1, ETTm1, ETTm2 and PhysioNet, a tie on ETTh2. With identical information, fitting the residual on top of U^G\hat{U}_G beats learning total error directly.
  • Context-Risk (local context and the imputer's prediction, no CURE input): strong and dataset-dependent; total risk is better on Beijing AQ, ETTh1, ETTm1 and PhysioNet and worse on ETTh2 and ETTm2 (pooled +6.2%).
  • Your point that the correction can absorb CURE's error (Eq. 6) led us to the plug-in estimator U^G+DG,f\hat{U}_G+D_{G,f}, which follows directly from Proposition 1 and needs no training beyond CURE. It beats CURE-feature Risk by 19.0% (CI 17.3–21.0), our learned total risk by 8.6%, and Context-Risk overall by 4.0% (CI 0.2–7.8): better on Beijing AQ, ETTh1, ETTm1 and PhysioNet, worse on ETTh2 and ETTm2.

Our view is that estimating recoverability first does help, but the benefit comes through the decomposition (U^G\hat{U}_G plus a deviation from the conditional mean) rather than a learned correction. On two ETT datasets, hand-crafted interpolation context carries signal that CURE's features miss; combining them is a natural next step, and we will say so in the paper.

W2/Q2: Eqs. (3)–(4) with native missing values.

You are right. When the target region contains native missing values, CGC_G does not determine which positions are evaluable, so Eq. (4) must also condition on GevalG_{\mathrm{eval}}. We now define UG∗=1∣Geval∣,tr,Cov(XGeval∣CG,Geval)U_G^{\ast} = \frac{1}{|G_{\mathrm{eval}}|}\\,\mathrm{tr}\\,\mathrm{Cov}(X_{G_{\mathrm{eval}}} \mid C_G, G_{\mathrm{eval}}), and the risk in Eq. (3) likewise conditions on (CG,Geval)(C_G, G_{\mathrm{eval}}). GevalG_{\mathrm{eval}} is fixed by the native observation pattern, which is known whenever the error can be evaluated, so this is the quantity evaluation actually measures. The estimator already matches this definition: CURE's loss and the region-level average in Eq. (18) run over GevalG_{\mathrm{eval}} only. For fully observed target regions nothing changes.

W3/Q3: Paired AURC differences with uncertainty.

We report paired relative AURC differences with 95% CIs from a clustered bootstrap over (dataset, mechanism, train seed, mask seed) clusters, 10,000 draws, so that all imputers and conditions sharing a trained model and a mask realization fall in one cluster. Median paired difference [95% CI], negative = first score better:

DatasetTotal risk vs. DG,fD_{G,f}U^G+DG,f\hat{U}_G+D_{G,f} vs. DG,fD_{G,f}U^G+DG,f\hat{U}_G+D_{G,f} vs. Context-RiskTotal risk vs. CURE-feature Risk
All−11.1 [−15.1, −7.7]−18.9 [−21.2, −16.5]−4.0 [−7.8, −0.2]−8.7 [−10.2, −7.4]
Beijing AQ−39.4 [−42.2, −36.1]−42.3 [−44.6, −40.0]−34.8 [−44.5, −24.7]−30.9 [−32.9, −29.4]
ETTh1−10.1 [−12.8, −5.4]−16.5 [−18.9, −14.4]−15.6 [−18.0, −13.5]−10.1 [−11.5, −8.9]
ETTh2−2.0 [−8.4, +3.6]−11.0 [−15.1, −4.6]+31.7 [+20.6, +59.9]−4.2 [−8.0, +3.6]
ETTm1+0.1 [−0.6, +0.6]−10.9 [−11.5, −9.9]−17.1 [−19.6, −14.7]−7.4 [−8.4, −6.4]
ETTm2−0.3 [−4.0, +2.8]−15.5 [−16.7, −13.0]+11.5 [+9.1, +14.3]−10.0 [−12.6, −8.4]
PhysioNet−27.0 [−31.2, −23.7]−28.9 [−32.9, −25.6]−14.5 [−16.4, −11.9]−5.1 [−5.6, −4.5]

The gain of total risk over disagreement is clear on Beijing AQ, PhysioNet and ETTh1, and a tie on ETTh2, ETTm1 and ETTm2; we will state this heterogeneity explicitly. U^G+DG,f\hat{U}_G+D_{G,f} beats disagreement on every dataset. PhysioNet, and the Beijing AQ direct-baseline comparisons, use one training seed, so their intervals rest on 3 clusters per mechanism and are optimistic.

Reviewer_UmjAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies multivariate time-series imputation from the perspective of recoverability. The authors argue that two missing regions of the same size can differ in how much information the observed context provides, and that this difference should be separated from the error introduced by a particular imputer. The authors decompose the conditional prediction risk into a conditional Bayes-risk term and an imputer-specific excess-risk term. They define the Bayes-risk term as contextual unrecoverability, which is measured relative to the observed mask, the target region, and the available context. They propose CURE, a Transformer with mean and variance heads trained on artificial holdouts, to estimate this quantity from incomplete sequences. A lightweight MLP then estimates an imputer-specific residual correction from the CURE representation, the imputer reconstruction, and context descriptors. The sum of the two estimates ranks missing regions for selective imputation. The authors evaluate the framework on four synthetic dynamical families with oracle Bayes risk, and on five real datasets with seven imputers under MCAR and MAR point and block masking. They report error-ranking correlations, AURC, and selective RMSE.

Critical strengths

1.The paper separates two questions that are often conflated. The first is whether the observed context leaves a missing region ambiguous. The second is whether a particular imputer adds error. Figure 1 illustrates this distinction concretely. The question is practically relevant for domains where imputation outputs feed downstream decisions. 2.The evaluation covers five real datasets, seven heterogeneous imputers, two missingness mechanisms, and both point and block corruption. Test masks are shared across imputers, and the calibration split is disjoint from imputer training, validation, and testing. The authors aggregate 560 conditions and 7,728 replicate-level evaluations while explicitly noting that these are not independent repetitions.

Critical weaknesses

1.What is the calibration error of X relative to Oracle U * on synthetic data? Please report ECE, predicted NLL, or reliability chart. If CURE is a Bayesian risk estimator, why only validate rank correlation? 2.How well do the predicted variances of CSDI and GP-VAE rank missing units? Please report the Spearman correlation and AURC of these scores. 3.What specific constraints does Bayes risk decomposition impose on the algorithm? Can you prove that the uncertainty score is closer to U * than usual? 4.Why does Table 1 use full coverage instead of random retention as the benchmark for decline? 5. Please unify the key numbers in the main text. The average of the abstract is 58.7%, Section 6.5 is written as median 58.7%, and Figure 6a shows 58.6%, 53.7%, and 43.6%.

Questions for authors

see weakness

1 comment

Thank you for the precise questions which have helped us improve our work. Please also see the general response.

W1: Calibration of U^G\hat{U}_G against the oracle UG∗U_G^{\ast}. On the four synthetic families, where UG∗U_G^{\ast} is known, we now report ECE (10 equal-mass bins of U^G\hat{U}_G) and the predicted NLL: the Gaussian NLL of the Bayes predictor's realised gap error under U^G\hat{U}_G minus the same NLL under UG∗U_G^{\ast}. This is a proper scoring rule minimised in expectation at UG∗U_G^{\ast}, so the gap is 0 when U^G\hat{U}_G is calibrated. The variance ratio (realised error / U^G\hat{U}_G) is 1 when calibrated; with UG∗U_G^{\ast} itself it is 0.99–1.00, which validates the check. Reliability diagrams will be added to the paper.

FamilySpearmanSlopeECE (relative)ΔNLL vs. oracle, nats [95% CI]Variance ratio
Gaussian VAR0.931.090.165 (18%)0.032 [0.029, 0.035]0.77
Periodic0.931.350.019 (43%)0.066 [0.063, 0.069]0.65
Regime switching0.791.080.076 (46%)0.079 [0.072, 0.087]0.78
Nonlinear0.801.020.096 (12%)0.019 [0.018, 0.021]0.82

Ranking is strong and close to linear, but CURE is conservative: U^G\hat{U}_G overestimates UG∗U_G^{\ast} by 22–53% on average (variance ratio 0.65–0.82), most where UG∗U_G^{\ast} is small (1.4–3.4 times the oracle in the lowest tercile vs. 1.1–1.4 in the highest). The cost in predicted NLL is small (0.02–0.08 nats per gap). We therefore no longer describe CURE as a calibrated risk estimator: it is a conservative estimator of relative recoverability, and we use it for ranking, which is what AURC and selective RMSE evaluate.

W2: Ranking by the native uncertainty of CSDI and GP-VAE.

We used the variance across each model's own predictive samples (the same samples it uses to impute) as a ranking score. Pooled, it improves AURC over U^G\hat{U}_G alone by a median of 14.8% (95% CI 10.4–18.3), but this is dataset-dependent: −13.3% (ETTh1), −38.7% (ETTh2), −18.4% (ETTm1), +9.4% (ETTm2), +8.4% (PhysioNet). U^G+DG,f\hat{U}_G+D_{G,f} beats it by a median of 25.9% (CI 17.7–33.1). Native variance also exists only for probabilistic imputers, whereas U^G\hat{U}_G and U^G+DG,f\hat{U}_G+D_{G,f} apply to all seven.

W3: What the decomposition implies; is the score closer to UG∗U_G^{\ast}?

Proposition 1 is an identity, not a dominance result, and we do not claim CURE is provably closer to UG∗U_G^{\ast} than other scores; we will soften the text accordingly. What we add theoretically is a lemma: the population minimizer of the heteroscedastic Gaussian NLL over (μ,v)(\mu,v) is the conditional mean and conditional variance for any conditional distribution with finite second moments (proper scoring rules; Gneiting & Raftery, 2007), so the objective needs no Gaussian assumption. In practice, the decomposition implies the estimator U^G+DG,f\hat{U}_G+D_{G,f}, which beats disagreement on every dataset (pooled −18.9%) and a learned direct-risk model with identical inputs (−18.1%). Closeness to UG∗U_G^{\ast} itself is addressed empirically in W1.

W4: Table 1 against random retention. We agree with this suggestion. In the revised Table 1, random retention at the same coverage becomes the reference column (full coverage kept as secondary). Selective RMSE at 50% coverage (lower is better; best per dataset in bold):

DatasetFullRandomU^G\hat{U}_GDG,fD_{G,f}U^G+DG,f\hat{U}_G+D_{G,f}Context-RiskCURE-feature RiskTotal risk
Beijing AQ0.8650.8670.4850.3830.2510.4240.3570.280
ETTh10.6640.6610.6320.4350.3960.4640.4350.413
ETTh21.0751.0200.5820.4500.4230.3060.4420.440
ETTm10.5290.5260.5050.2750.2590.3260.2800.269
ETTm20.5250.5230.2850.2330.2150.2090.2620.251
PhysioNet0.7920.7830.5380.5750.4640.5240.4950.472

W5: Inconsistent numbers.

This is true and it led us to the audit in the general response. The discrepancies (58.7 vs. 58.6, 43.8 vs. 43.6) came from eight partial runs on an excluded dataset (Electricity) that our figure script read; removing them reproduces the text values exactly (58.71, 43.77). All numbers in these replies come from one script, which will generate every number in the revision, and we will state the statistic (mean or median) everywhere. On the corrected six-dataset results, total risk reduces AURC relative to random by a median of 60.6%.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.