acceptodds
Under review as a conference paper at ICLR 2027

DoobImp: Diffusion Imputation without Artificial Masking

20 views

Abstract

Missing data are ubiquitous in real-world datasets, making reliable imputation essential for prediction and downstream analysis. Diffusion-based imputation methods typically learn conditional scores by artificially masking observed entries during training. This couples learning to a user-specified mask distribution, which may not match test-time missingness and can degrade imputation performance. We introduce DoobImp, a diffusion imputation framework that avoids artificial masking by decoupling test-time imputation from learning an unconditional diffusion backbone. Its two complementary components address conditioning and learning. DoobCond uses Doob's h-transform to characterize the ideal conditional reverse diffusion process and derives a practical approximation for conditioning on observed entries. DoobFit learns the unconditional backbone from incomplete training data by alternating between imputing missing entries with DoobCond and refitting the unconditional score. Across nine tabular datasets under three missingness mechanisms, DoobImp achieves the best overall performance over 17 baselines, attaining an average MAE rank of 1.19 and the highest aggregate categorical accuracy. Experiments on four time-series datasets further demonstrate robust performance under both pointwise and structured missingness. The resulting tabular imputations also support strong downstream classification performance. These results show that decoupling unconditional score learning from conditioning provides an effective and broadly applicable alternative to mask-based diffusion imputation.

Primary area:
generative models
Keywords:
Missing Data Imputation, Diffusion Models, Doob’s $h$-Transform, Inference-Time Conditioning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_NKAeOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a diffusion-based data imputation method called DoobImp that does not require artificial masking during training. It learns the unconditional distribution of the data and applies observational conditions at the inference stage using Doob's h-transform to recover missing data; meanwhile, through iterative imputation and retraining, the model is able to learn directly from incomplete data.

Critical strengths
  1. The ideal Doob transform has a clear probabilistic interpretation, whereas the practical algorithm uses Tweedie estimation, posterior precision approximation, and the Jacobian to convert the theory into a computable guidance term.
  2. The experiments are sufficient and, from the perspective of performance comparison, provide relatively strong support for the effectiveness of DoobImp.
Critical weaknesses
  1. The motivation may overstate the novelty of mask-free diffusion imputation. NewImp and DiffPuter have already explored imputation without additional artificial masking.
  2. There is no sufficient direct comparison with existing conditional sampling methods such as DPS and RePaint under the same unconditional backbone.
Questions for authors
  1. What specific limitations of existing mask-free imputation methods, particularly NewImp and DiffPuter, motivate the proposed framework? (Weakness 1)
  2. Can the authors demonstrate whether the advantage of DoobCond actually comes from the special properties of Doob's h-transform, or merely from adding an observation-consistency gradient to the reverse diffusion? (Weakness 2)
4 comments
AuthorsOpenReview ↗

Response to Reviewer NKAe (Part 1/2)

We thank you for the time you spent reviewing our paper and for recognizing our framework based on Doob's hh-transform and the strong empirical performance on a variety of imputation tasks. We also appreciate your detailed questions, which help us improve the completeness and clarity of the paper. We now address each question in turn.

Q1. The motivation may overstate the novelty of mask-free diffusion imputation. NewImp and DiffPuter have already explored imputation without additional artificial masking. What specific limitations of existing mask-free imputation methods, particularly NewImp and DiffPuter, motivate the proposed framework?

A1. Thank you for helping us clarify this point. In the first place, we do not claim to be the first to propose mask-free diffusion imputation models. We acknowledge prior work in the manuscript (lines 76–79; lines 101–104). Instead, our motivation starts from missingness distribution shift, which we address through DoobCond complemented by DoobFit. This is supported by the experiments in Appendix C.5 (Figure 8) and the real-data experiments below. We now explain in detail how our formulation based on Doob's hh-transform and its adaptation to imputation distinguish DoobImp from NewImp [1] and DiffPuter [2].

NewImp. Algorithm 1 of NewImp [1] alternates between fitting a joint score to the current completed matrix and updating its missing entries by a Wasserstein gradient flow, with a negative-entropy term that favors more concentrated imputations. In contrast, DoobCond provides reusable conditioning: given a frozen unconditional diffusion backbone, DoobCond takes a new sample's observed values and mask and draws its missing entries without updating the backbone (Algorithm 1). DoobFit calls the same sampler to complete its training samples before the next refit (Algorithm 2). NewImp's authors also note that their kernel computation scales roughly quadratically with the number of samples and that their real-valued support assumption limits direct application to categorical variables [1, Appendix G.1]. DoobCond computes its update from the current sample through the frozen denoiser (Equation (10)), without evaluating a kernel over pairs of samples. Our tabular implementation handles categorical variables by binary encoding and decodes the imputed values for evaluation (Appendix D.1; Table 1).

DiffPuter. Our main difference from DiffPuter [2] lies in the conditioning term. DiffPuter's Theorem 1 characterizes conditional sampling in the limit Δt→0\Delta t\to0. In the proof, Equation (11) of [2, Appendix B.1] makes the following approximation inside the reverse transition, where yy denotes the clean observed entries:

\approx p_\theta(x_{t-\Delta t}^{\mathrm{mis}}\mid x_t).$$ The right-hand side **drops the explicit dependence on $y$**. Instead, using Doob's $h$-transform, our Lemma 1 derives how $y$ changes the reverse drift through $\nabla_x\log h_\eta(x,t)$ (Equation (6)). DoobCond approximates its missing-coordinate component: the denoiser Jacobian maps the precision-weighted discrepancy between $y$ and its predicted value back to the missing entries (Eqs. (8)–(10)). Holding DiffPuter's own weights fixed, switching from its default RePaint-style sampler to DoobCond lowered both errors in 27/27 pairs. The table below shows this comparison on frozen DiffPuter weights. Cells give mean MAE / RMSE across MCAR, MAR, and MNAR at 30% missingness, with three paired seeds per mechanism. | Dataset | RePaint [2, 3] | DoobCond | |---|---:|---:| | Letter | 0.330 / 0.510 | 0.322 / 0.495 | | Gesture | 0.322 / 0.760 | 0.307 / 0.687 | | News | 0.299 / 0.784 | 0.278 / 0.573 | These results demonstrate **consistently lower imputation errors with DoobCond in the tested settings**. To reconcile reverse updates of the missing block with refreshed observations, DiffPuter's implementation runs 20 resampling rounds at each nonterminal noise level. DoobCond retains the explicit dependence on $y$ through its conditioning term derived from Doob's $h$-transform and applies this term before refinement. On Letter at 30% MCAR, our default sampler used five same-level refinement steps and achieved a lower measured cost: it took 0.555 ms per sample per posterior draw versus 2.171 ms for DiffPuter on the same RTX A6000. A complete training iteration averaged 9.618 min for DoobImp versus 16.436 min for DiffPuter (Table 42-43). The tabular benchmark in the manuscript gives a second comparison in which both methods reuse their weights as missingness changes. DoobImp had lower average MAE and RMSE in 9/9 training–test mechanism combinations (Appendix E.3). For models trained at 30% missingness and tested at 50%, Table 40 reports MAE / RMSE of 0.341 / 0.623 for DoobImp and 0.393 / 4.546 for DiffPuter. These results further demonstrate DoobImp's stronger imputation performance under the tested changes in missingness mechanism and rate.
AuthorsOpenReview ↗

Response to Reviewer NKAe (Part 2/2)

Q2. Can the authors demonstrate whether the advantage of DoobCond actually comes from the special properties of Doob's h-transform, or merely from adding an observation-consistency gradient to the reverse diffusion?

A2. Doob's hh-transform plays a key role in constructing DoobCond's conditioning term, beyond simply adding an observation-consistency gradient. Lemma 1 identifies the conditioning term required by Doob's hh-transform. In the Gaussian case, Equation (8) specifies how the posterior precision weights the observation residual and how the denoiser Jacobian maps it to the missing coordinates. Equations (9)–(10) provide the practical approximation used by DoobCond, giving the update a probabilistic basis. In Table 37 of the manuscript (lines 2453–2465), both samplers retain an observation-consistency gradient, allowing us to isolate the contribution of the precision weighting derived from Doob's hh-transform.Replacing temperature-only weighting (which already includes an observation-consistency gradient) with the diagonal precision derived from Doob's hh-transform reduced mean MAE from 0.291 to 0.272 and RMSE from 0.548 to 0.518. All three datasets improved.

For comparison with an established method, we additionally evaluated DoobCond against DPS [4], which also differentiates an observation residual through the denoiser. On the same frozen DoobFit checkpoints, DoobCond lowered both MAE and RMSE in 27/27 pairs. Each cell below gives mean MAE / RMSE across MCAR, MAR, and MNAR at 30% missingness, with three seeds per mechanism.

DatasetDoobCondDPS [4]
Letter0.284 / 0.4450.451 / 0.643
Gesture0.263 / 0.5630.358 / 0.759
News0.262 / 0.5330.423 / 0.768

The DPS comparison further demonstrates the advantage of DoobCond's full conditioning update over a sampler that already uses an observation-consistency gradient in these settings. A3 below provides the results by dataset and mechanism, including RePaint.

Q3. There is insufficient direct comparison with existing conditional sampling methods such as DPS and RePaint under the same unconditional backbone.

A3. We additionally evaluated DPS [4] and the RePaint-style sampler [2, 3] on exactly the same frozen DoobFit checkpoints as DoobCond. RePaint here is DiffPuter's inference implementation. We held the test samples, split, 30% mask, and normalization fixed and averaged 20 draws for each point prediction. Each cell reports mean MAE / RMSE over three seeds.

DatasetMechanismDoobCondDPS [4]RePaint-style [2, 3]
LetterMCAR0.270 / 0.4230.447 / 0.6350.280 / 0.441
LetterMAR0.297 / 0.4650.452 / 0.6440.301 / 0.474
LetterMNAR0.284 / 0.4470.454 / 0.6500.292 / 0.461
GestureMCAR0.255 / 0.5490.355 / 0.7510.274 / 0.808
GestureMAR0.263 / 0.5590.350 / 0.7440.271 / 0.574
GestureMNAR0.270 / 0.5810.370 / 0.7830.281 / 0.610
NewsMCAR0.263 / 0.5310.425 / 0.7630.281 / 0.551
NewsMAR0.261 / 0.5250.418 / 0.7590.276 / 0.660
NewsMNAR0.263 / 0.5440.425 / 0.7810.289 / 1.975

DoobCond achieved lower MAE and RMSE than DPS in 27/27 matched dataset–mechanism–seed comparisons and lower MAE and RMSE than RePaint in 27/27. This comparison isolates the contribution of the conditional sampling method under the same backbone, supporting DoobCond's superior performance in imputation tasks.

Thank you again for reviewing our paper. We hope these responses address your concerns, and we welcome further discussion.

References

[1] Zhichao Chen et al. Rethinking the Diffusion Models for Missing Data Imputation: A Gradient Flow Perspective. NeurIPS, 2024.

[2] Hengrui Zhang et al. DiffPuter: Empowering Diffusion Models for Missing Data Imputation. ICLR, 2025.

[3] Andreas Lugmayr et al. RePaint: Inpainting Using Denoising Diffusion Probabilistic Models. CVPR, 2022.

[4] Hyungjin Chung et al. Diffusion Posterior Sampling for General Noisy Inverse Problems. ICLR, 2023.

Thank you for the clarifications and additional experiments. The discussion of NewImp and DiffPuter better explains the method’s positioning, while the matched-backbone comparisons with DPS and RePaint strengthen the empirical support for DoobCond. These responses address my concerns to a reasonable extent. I maintain my original rating of 3 (Weak Acceptance).

Thank you for your positive feedback and for acknowledging our clarifications and additional experiments. We sincerely appreciate your thoughtful review and constructive comments throughout the discussion.

Reviewer_jM95OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents DoobImp, a diffusion-based imputation framework that separates unconditional score learning from inference-time conditioning to avoid artificial masking during training. It consists of DoobCond, which applies Doob's h-transform for conditional imputation, and DoobFit, which iteratively trains an unconditional diffusion model using incomplete data. The method is evaluated on tabular and time-series datasets under various missingness mechanisms and patterns, with results compared against existing imputation methods.

Critical strengths
  1. The paper is well written and clearly organized, making the motivation and proposed method easy to follow.
  2. The proposed framework is practical and flexible, as it avoids artificial masking during training and supports different missingness patterns at inference time.
  3. The paper addresses an important imputation problem and demonstrates strong empirical performance across a wide range of tabular and time-series benchmarks.
Critical weaknesses
  1. The practical guidance term in DoobCond is structurally very similar to existing posterior guidance methods. It computes the observation residual using a denoised estimate and propagates this residual via the Jacobian. Similar formulations have already been explored in the diffusion and flow-matching literature [1]. This raises questions about the novelty of the proposed conditioning mechanism.
  2. The exact Doob transform is itself an established result, while the iterative completion-and-refitting procedure used in DoobFit is closely related to DiffPuter [2] and diffusion-based EM methods [3]. Therefore, I see the main contribution of this work as an effective adaptation and integration of existing techniques for missing data imputation. The paper would be more convincing if the authors could clarify which aspects of the proposed conditioning approximation are genuinely new compared with prior work.
  3. Although the paper provides extensive comparisons with existing imputation baselines, it lacks direct comparisons with established posterior guidance methods using the same unconditional backbone. Based on the current results, it is difficult to determine whether the improvements come from the proposed conditioning term, precision weighting, observed-coordinate refresh, refinement steps, or the iterative training procedure.
  4. The projection-only and other ablation studies are useful, but comparisons with standard likelihood guidance or covariance-aware guidance would be more convincing. Such comparisons are important because the methodological contribution of the paper largely depends on whether DoobCond provides a meaningful improvement over existing guidance approaches. [1] On the Guidance of Flow Matching. ICML 2025 [2] DiffPuter: Empowering Diffusion Models for Missing Data Imputation. ICLR 2025 [3] An Expectation-Maximization Algorithm for Training Clean Diffusion Models from Corrupted Observations. NeurIPS 2024
Questions for authors
  1. How does the guidance term in DoobCond differ from existing posterior guidance methods, particularly those studied in [1]?
  2. What are the main methodological contributions of DoobImp beyond the established Doob transform and the iterative imputation frameworks in [2, 3]?
  3. Can the authors provide comparisons with existing posterior guidance methods using the same unconditional backbone to better isolate the contribution of DoobCond?
  4. How does DoobCond compare with standard likelihood guidance or covariance-aware guidance, and what specific advantages does it offer over these approaches?
6 comments
AuthorsOpenReview ↗

Response to Reviewer jM95 (Part 1/4)

We thank you for the time you spent reviewing our paper and for recognizing the clarity of our presentation, the flexibility of our framework, and the strong empirical performance. We also appreciate your detailed questions, which help us improve the completeness and clarity of the paper. We now address each question in turn.

Q1. DoobCond computes the observation residual using a denoised estimate and propagates this residual via the Jacobian. Similar formulations have already been explored in the diffusion and flow-matching literature [1]. This raises questions about the novelty of the proposed conditioning mechanism. How does the guidance term in DoobCond differ from existing posterior guidance methods, particularly those studied in [1]?

A1. Thank you for pointing us to the work of Feng et al. [1], which we will cite and discuss in the revised manuscript. Although the formulations are related, DoobCond differs from their Gaussian inverse-problem approximation (Section 3.4, Equation (9) of [1]) in two main respects.

1. Conditioning weights. Their approximation uses scalar schedules for guidance strength and covariance scale. Instead, starting from the endpoint law in Lemma 1, our Gaussian derivation weights the observed residual by the posterior precision (V0∣tobs,obs+η2IDobs)−1(V_{0\mid t}^{\rm obs,obs}+\eta^2 I_{D_{\rm obs}})^{-1} (Appendix B.2; Equation (8)). Equations (9)–(10) turn this conditioning term into a computable imputation update: the learned Tweedie estimate supplies the mean and Jacobian, while the scalar Gaussian posterior-variance formula supplies a diagonal precision for each coordinate and noise level. In Table 37 (Appendix E.2.3), replacing temperature-only weighting with this precision reduced average MAE from 0.291 to 0.272 and RMSE from 0.548 to 0.518.

2. Updating the missing coordinates. Their Equation (9) approximates the denoiser Jacobian by the identity. When the measurement operator selects the observed coordinates, this gives a conditioning vector with zero entries in the missing block. In contrast, DoobCond retains the Jacobian ∂x^θ,0obs(x,t)/∂xmis\partial\widehat x_{\theta,0}^{\rm obs}(x,t)/\partial x^{\rm mis} in Equation (10), which describes how the denoised observed block depends on the noisy missing block. Its transpose maps the precision-weighted observed residual to the missing-coordinate conditioning term ∇xmislog⁡h^η(x,t)\nabla_{x^{\rm mis}}\log\widehat h_\eta(x,t), which directly updates the missing entries in Equation (11). Thus, diagonal precision weighting still allows the update to use learned dependence between observed and missing coordinates.

To sum up, we build on Doob's hh-transform to develop an update rule tailored to imputation. Our empirical results demonstrate its effectiveness.

AuthorsOpenReview ↗

Response to Reviewer jM95 (Part 2/4)

Q2. The iterative completion-and-refitting procedure used in DoobFit is closely related to DiffPuter [2] and diffusion-based EM methods [3]. The main contribution appears to be an effective adaptation and integration of existing techniques for missing data imputation. What are the main methodological contributions of DoobImp beyond the established Doob transform and the iterative imputation frameworks in [2, 3]? Which aspects of the proposed conditioning approximation are genuinely new compared with prior work?

A2. First, we respectfully disagree that DoobImp is simply an integration of existing techniques for data imputation. Our core contribution is the specific conditioning approximation in DoobCond. As explained in A1, we build on Doob's hh-transform to derive the observed-residual weighting and the update of the missing coordinates, then make both computable on a frozen unconditional model. These techniques give a practical imputation sampler whose contribution is supported by the comparisons below.

Second, we agree that iterative completion and refitting are established ideas [2, 3], but our starting point is not the iterative procedure itself. On the one hand, DoobCond, our core conditional imputation method derived from Doob's hh-transform, operates on a frozen unconditional model and can be reused with backbones trained by different procedures. On the other hand, the overall DoobImp framework is motivated by the problem of missingness distribution shift, as discussed in the Introduction. To this end, DoobFit combines the observed-entry initialization objective in Equation (13) with DoobCond completions and score refitting (Section 4.3). This training procedure complements DoobCond in our experiments. Together, the two components allow learning directly from incomplete data and reusing the learned unconditional model as missingness patterns change.

We next explain how the conditioning term derived from Doob's hh-transform differs from DiffPuter's reverse update. Theorem 1 of [2] considers conditional sampling as Δt→0\Delta t\to0. In its proof, Equation (11) of [2, Appendix B.1] uses the following approximation inside the expectation over the noisy state xtx_t, where yy is the clean observed block:

\approx p_\theta(x_{t-\Delta t}^{\mathrm{mis}}\mid x_t).$$ The transition on the right uses the noisy observed coordinates in $x_t$; **it drops the explicit dependence on $y$**. Our Lemma 1 identifies how $y$ changes the reverse drift through $\nabla_x\log h_\eta(x,t)$ (Equation (6)). DoobCond approximates the missing-coordinate component by comparing $y$ with the denoiser's prediction of the observed block, weighting that residual by its approximate posterior precision, and propagating it to the missing coordinates through one vector–Jacobian product (Equations (8)–(11)). DoobFit initializes the score from observed entries and uses this same sampler to complete the training samples before each refit (Algorithms 1–2). Table 1(a) in the manuscript (lines 378–388) reports mean MAE ranks of 1.185 for DoobImp and 3.333 for DiffPuter across 27 dataset–mechanism settings. To separate training from inference, we additionally tested both samplers on each learned prior. Each cell gives mean MAE / RMSE over nine evaluations on one dataset: three mechanisms (MCAR, MAR, MNAR) at 30% missingness and three seeds per mechanism. | Learned prior | Inference sampler | Letter | Gesture | News | |---|---|---:|---:|---:| | DiffPuter | RePaint-style [2, 4] | 0.330 / 0.510 | 0.322 / 0.760 | 0.299 / 0.784 | | DiffPuter | DoobCond | 0.322 / 0.495 | 0.307 / 0.687 | 0.278 / 0.573 | | DoobFit | RePaint-style [2, 4] | 0.291 / 0.459 | 0.276 / 0.664 | 0.282 / 1.062 | | DoobFit | DoobCond | **0.284 / 0.445** | **0.263 / 0.563** | **0.262 / 0.533** | Changing only the sampler from RePaint to DoobCond **lowered both errors in 54/54 paired comparisons** across the two priors. Holding DoobCond fixed, changing the prior from DiffPuter to DoobFit **lowered both errors in 27/27 pairs**. The first comparison identifies what the sampler adds even to an existing DiffPuter model. The second identifies what training with DoobFit adds when both priors use the same sampler. These conditioning procedures also differ in computational cost. DiffPuter's implementation uses 20 local resampling rounds at each nonterminal noise level. DoobCond evaluates its conditioning term before same-level refinement. On Letter at 30% MCAR, the default DoobCond sampler ($J=5$) took 0.555 ms per sample per posterior draw, versus 2.171 ms for DiffPuter on the same RTX A6000 (Table 42 in the manuscript). Including completion and refitting, a training iteration averaged 9.618 min for DoobImp versus 16.436 min for DiffPuter (Table 43). These results show that DoobImp achieves lower MAE, with faster inference and training in the Letter timing study.
AuthorsOpenReview ↗

Response to Reviewer jM95 (Part 3/4)

Q3. Can the authors provide comparisons with existing posterior guidance methods using the same unconditional backbone to better isolate the contribution of DoobCond?

A3. Yes. We additionally compared DoobCond with DPS [5], pseudoinverse-guided diffusion (PiGDM) [6], and DiffPuter's RePaint-style sampler [2, 4] using identical frozen DoobFit weights within each comparison. We held the split, mask, normalization, and test samples fixed. Each method used its specified sampler on the same backbone and predicted each missing numerical entry by averaging 20 draws. We tested Letter, Gesture, and News at 30% missingness under MCAR, MAR, and MNAR, with three seeds per setting.

Each cell below gives mean MAE / RMSE across the three mechanisms and three seeds for that dataset.

DatasetDoobCondDPS [5]PiGDM [6]RePaint-style [2, 4]
Letter0.284 / 0.4450.451 / 0.6430.368 / 0.5490.291 / 0.459
Gesture0.263 / 0.5630.358 / 0.7590.272 / 0.5710.276 / 0.664
News0.262 / 0.5330.423 / 0.7680.311 / 1.0160.282 / 1.062

DoobCond lowered both mean errors in 9/9 settings against each comparator. Within matched seeds, DoobCond lowered MAE in 27/27 pairs against each of DPS, PiGDM, and RePaint. It lowered RMSE in 27/27 against DPS, 25/27 against PiGDM, and 27/27 against RePaint. Therefore, these additional experiments better isolate DoobCond's contribution and demonstrate its effectiveness in imputation tasks.

Q4. How does DoobCond compare with standard likelihood guidance or covariance-aware guidance, and what specific advantages does it offer over these approaches?

A4. We additionally compared DoobCond with standard likelihood guidance and covariance-aware guidance on the same frozen Letter checkpoints. We tested the following methods in these two categories:

  1. Standard likelihood guidance. DPS [5] differentiates the observation residual through the denoiser. PiGDM [6] uses the measurement pseudoinverse to map that residual back to the input coordinates.

  2. Covariance-aware guidance. MMPS [7] applies a covariance operator and uses one GMRES iteration per solve in our tabular comparison. TMPD [8] uses its measurement-diagonal approximation.

We also compared with TDS [9], a sequential Monte Carlo sampler that reweights and resamples particles. Each cell below reports mean MAE / RMSE over three seeds, using 20 draws per point imputation.

Inference methodMCARMAR
DoobCond0.270 / 0.4230.297 / 0.465
DPS0.447 / 0.6350.452 / 0.644
PiGDM0.361 / 0.5360.371 / 0.554
MMPS0.392 / 0.5720.402 / 0.590
TMPD>1 / >1>1 / >1
TDS0.332 / 0.4940.347 / 0.519

DoobCond achieved the lowest mean MAE and RMSE among all compared methods, demonstrating strong imputation performance.

Reviewer_DKZjOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes DoobImp, a diffusion-based imputation method that leverages unconditional prediction to characterize the conditional reverse diffusion process for imputation. Experimental validation shows DoobImp obtains excessive performance over a wide range of datasets.

Critical strengths
  • The paper is well written and easy to follow.
  • The performance of the proposed method is validated under a variety of tests across different imputation tasks.
Critical weaknesses
  • My main concern is that a solid theoretical illustration of the role of unconditional prediction is needed. A convergence analysis of how the generated imputed final results converge to the real underlying data distribution is beneficial.
  • What are the computational costs of DoobImp compared to the baseline method?
Questions for authors
  • see weakness
5 comments
AuthorsOpenReview ↗

Response to Reviewer DKZj (Part 1/2)

Thank you for your careful reading and for recognizing the clarity of our presentation and the strong performance across tabular and time-series tasks. We also appreciate your questions, which help us further clarify the paper, and we will incorporate these clarifications into the revised version as space permits. We now address each question in turn.

Q1. A solid theoretical illustration of the role of unconditional prediction is needed.

A1. The unconditional model learns the joint distribution of the observed and missing entries. Equipped with this model, DoobCond uses Doob's hh-transform to approximate the conditional law of the missing entries given a sample's observed values (Section 4.2.1, Lemma 1). At each reverse step, the unconditional model predicts the clean observed entries from the full noisy state, including the coordinates being imputed. For the Gaussian model of Appendix B.2, Equation (8) in Section 4.2.2 gives the conditioning term on the missing coordinates as

\left[\frac{\partial\mu_{0\mid t}^{\mathrm{obs}}(x)} {\partial x^{\mathrm{mis}}}\right]^{\top} \left(V_{0\mid t}^{\mathrm{obs,obs}}+\eta^2 I_{D_{\mathrm{obs}}}\right)^{-1} \times \left(x_0^{\mathrm{obs}}-\mu_{0\mid t}^{\mathrm{obs}}(x)\right).$$ The unconditional model enters both outer factors through its prediction. On the right, the residual compares the actual observation $x_0^{\rm obs}$ with its prediction $\mu_{0\mid t}^{\rm obs}(x)$. On the left, the Jacobian tells us how changing the noisy missing coordinates would change that same prediction. Its transpose maps the precision-weighted residual to an update of the missing coordinates. In practice, Equations (9)–(10) use the frozen unconditional model's Tweedie estimate in place of the Gaussian posterior mean in both factors.
AuthorsOpenReview ↗

Response to Reviewer DKZj (Part 2/2)

Q2. A convergence analysis of how the final imputations converge to the real underlying data distribution would be beneficial.

A2. For data imputation, we target the conditional law of the missing entries given the observed values x0obsx_0^{\rm obs}. To analyze convergence to this conditional law, we extend the endpoint identity and convergence result in Lemma 1 and Appendix B.1, including Equation (18), to a sequence of fitted unconditional models.

1. Doob endpoint convergence. Fix mm and x0obsx_0^{\rm obs}, and suppose the learned joint endpoint laws converge weakly to the data distribution, Pθn,0⇒PdataP_{\theta_n,0}\Rightarrow P_{\rm data} as n→∞n\to\infty. For any bounded continuous test function φ\varphi on the missing coordinates, Equation (18) gives

EPθn,0η[φ(X~0mis)]=Eθn[φ(X~0mis) wη(X~0)]Eθn[wη(X~0)].\mathbb E_{P_{\theta_n,0}^{\eta}}\left[\varphi(\widetilde X_0^{\mathrm{mis}})\right] = \frac{\mathbb E_{\theta_n}\left[\varphi(\widetilde X_0^{\mathrm{mis}})\thinspace w_\eta(\widetilde X_0)\right]}{\mathbb E_{\theta_n}\left[w_\eta(\widetilde X_0)\right]}.

At fixed η>0\eta > 0, both integrands are bounded and continuous, and the denominator has a positive limit. The ratio therefore converges to the same ratio under PdataP_{\rm data}. Now let η↓0\eta\downarrow0. If x0obsx_0^{\rm obs} lies in the support of the observed marginal and the data conditional law has a weakly continuous version there, the normalized Gaussian weights concentrate at x0obsx_0^{\rm obs}. The missing-coordinate law thus converges weakly to Pdata(⋅∣x0obs)P_{\rm data}(\cdot\mid x_0^{\rm obs}), the true conditional law for this sample.

2. Algorithm 1 convergence. Let εalg\varepsilon_{\rm alg} be the bounded-Lipschitz distance dBLd_{\rm BL} between the law of one Algorithm 1 draw X^0mis\widehat X_0^{\rm mis} and the missing-coordinate marginal of the exact Doob endpoint for the same unconditional model and η\eta. Let b(η)b(\eta) be the distance from the smoothed data conditional to the true conditional at x0obsx_0^{\rm obs}. The ratio identity and the triangle inequality give

dBL(L(X^0mis),Pdata(⋅∣x0obs))≤εalg+2(1+η−1) dBL(Pθn,0,Pdata)EPdata[wη(X0)]+b(η).d_{\mathrm{BL}}\left(\mathcal{L}(\widehat X_0^{\mathrm{mis}}), P_{\mathrm{data}}(\cdot\mid x_0^{\mathrm{obs}})\right) \le \varepsilon_{\mathrm{alg}} + \frac{2(1+\eta^{-1})\thinspace d_{\mathrm{BL}}(P_{\theta_n,0},P_{\mathrm{data}})}{\mathbb E_{P_{\mathrm{data}}}\left[w_\eta(X_0)\right]} + b(\eta).

To bound the middle term, center each unit bounded-Lipschitz test function at its mean under the exact model tilt and use 0<wη≤10 < w_\eta\le1 and Lip⁡(wη)≤1/η\operatorname{Lip}(w_\eta)\le1/\eta. Choose η=ηn↓0\eta=\eta_n\downarrow0 slowly enough that this term vanishes. The support and continuity assumptions give b(ηn)→0b(\eta_n)\to0. Thus Algorithm 1 draws converge weakly to the data conditional provided εalg→0\varepsilon_{\rm alg}\to0 along the same sequence.

We also have empirical evidence supporting this analysis. In Gaussian simulations (Appendix C), reducing η\eta from 0.1000.100 to 0.0100.010 lowered the analytic smoothing bias from 9.660×10−39.660\times10^{-3} to 9.830×10−59.830\times10^{-5} (Figure 9b). By L=200L=200, the exact-conditioning Euler sampler reached the error scale of direct draws from the relaxed target (Figure 5). With the same learned unconditional score, practical conditioning reduced the normalized moment-based Wasserstein error against the true conditional from 0.416 to 0.197 (Figure 4). On real data, Figure 15 (Appendix E.3.1) holds DoobCond fixed and compares the point predictions of DoobFit and an unconditional model trained on fully observed samples. Across Gesture, Letter, and News, both mean MAEs rounded to 0.209 at 10% missingness, indicating very similar average imputation performance.

Q3. What are the computational costs of DoobImp compared to the baseline method?

A3. On Letter at 30% MCAR, the default DoobImp sampler took 0.555 ms per sample per posterior draw, compared with 2.171 ms for DiffPuter [1], as reported in Table 42 (Appendix E.4). A complete training stage averaged 9.618 min for DoobImp and 16.436 min for DiffPuter (Table 43). These measurements were taken sequentially over five seeds on one RTX A6000.

We additionally summed the same recorded training times across all ten stages within each seed, yielding mean total training times of 1.603 h for DoobImp and 2.739 h for DiffPuter over the five seeds. We also counted denoiser evaluations in the same Letter configuration. DiffPuter's RePaint-style sampler uses 20 resampling rounds at each nonterminal noise level, for 1,961 denoiser evaluations per draw on its 50-level grid. This lower sampling cost also contributes to DoobImp's training efficiency.

Thank you again for reviewing our paper. We hope these responses address your concerns, and we welcome further discussion.

References

[1] Hengrui Zhang et al. DiffPuter: Empowering Diffusion Models for Missing Data Imputation. ICLR, 2025.

Dear Reviewer,

As the discussion period is approaching its end (in around 2 days), we would greatly appreciate any further feedback on our rebuttal. We would be happy to provide additional clarifications to help address any remaining concerns.

Thank you again for your time and consideration!

Authors

Comments on the paper

2 comments

Dear Reviewers,

This is a friendly reminder to check the authors' responses to your reviews and leave a follow-up comment if appropriate. Your engagement in the discussion is greatly appreciated.

Thank you!

AC

AuthorsOpenReview ↗

Rebuttal Summary

Dear AC and Reviewers,

Thank you to the AC for guiding the discussion and to the reviewers for their thoughtful feedback and participation.

All three reviewers found the method clear and well motivated and recognized its strong empirical performance. Across nine tabular datasets and three missingness mechanisms, DoobImp achieved the best overall performance among the 18 methods compared, with an average MAE rank of 1.19. Reviewer jM95 noted that the proposed method is practical and flexible and addresses an important problem. Reviewer NKAe also highlighted the clear probabilistic interpretation behind the method.

During the rebuttal, we clarified how our framework differs from previous posterior guidance methods. Our contribution is to use Doob’s hh-transform to model the dependence between observed and missing entries and to develop practical approximations that make this conditioning effective for imputation. We thank Reviewers jM95 and NKAe for suggesting direct comparisons with existing methods. We added comparisons with DPS, PiGDM, and RePaint-style sampling on the same frozen backbone. These experiments strengthen the evidence for DoobCond, which achieved the lowest mean MAE and RMSE in all nine dataset–mechanism settings (Section 5.3 and Appendix E.5 of the revised manuscript). Following the rebuttal, Reviewers DKZj and NKAe maintained their support for the paper and their Weak Accept ratings.

Thank you again!

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.