acceptodds
Withdrawn from ICLR 2027

Marginal Optimal Transport for Missing Value Imputation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Optimal-transport methods for missing value imputation assume that compared batches share a common joint support. Heterogeneous masks violate this assumption: samples can have no jointly observed coordinates and therefore represent different marginals. We formulate Marginal Optimal Transport (MOT), which matches one Wasserstein marginal per missingness pattern. Under pattern coverage and chordal Markov regularity, the objective identifies the true joint; we also derive a finite-sample rate of order for an importance-weighted estimator and develop a scalable Kantorovich-dual estimator. To handle non-random missingness, an inverse-propensity-weighted variant is introduced as an empirical heuristic. Experiments confirm imputation, distributional, covariance, source, and graph recovery across synthetic and tabular benchmarks. We further prove that unprojected joint-support matching violates metric axioms and leaves its Kantorovich potential unidentified on unobserved coordinates.

Primary area:
optimization
Keywords:
missing data imputation, Optimal Transport

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_HHhgOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates OT-based missing value imputation (MVI) problem. Specifically, it first introduces marginal optimal transport and derive loss functions based on modified Wasserstein distance. It also incorporates propensity weighting to mitigate sample selection bias. Experiments on synthetic and tabular benchmarks demonstrate the effectiveness of the proposed method.

Critical strengths

The problem of missing value imputation is clearly motivated and well formulated, and the introduction of optimal transport to address this problem is reasonable and feasible.

The proposed algorithms are clearly described in the main text and appendix. Authors also provide code in the supplementary file to improve reproducibility.

Critical weaknesses
  1. OT-based imputation is a problem that attracts extensive investigation. Authors effectively discuss the original one (Muzellec et al., 2020) and one subsequent refinement (Jarrett et al., 2022), but ignore the existence of numerous literature in this field.

  2. The experiment would benefit from comparing with more comprehensive and recent baselines. The baselines selected in this paper are somewhat out-of-dated, as all of them except KPI were presented before 2024. Moreover, only one OT-based imputation method is selected as baseline (Muzellec et al., 2020). These two factors make it insufficient to demonstrate the effectiveness of the proposed approach, given the rapid progression of this field.

  3. The abstract noted that samples can have no jointly observed coordinates and therefore represent different marginals, which is claimed as a limitation of existing OT methods. However, existing OT methods can address it by employing a preliminary imputer to generate initial value of imputations. Therefore, if authors want to demonstrate that this limitation persists, please consider the existence of preliminary imputer. Moreover, according to figure 1, the proposed method also resorts to a preliminary imputer, which makes it less persuasive that the proposed approach can separately address the identified limitation without preliminary imputation.

  4. It is known that preliminary imputer highly impacts the imputation performance. Authors should elaborate more on the implementation of preliminary imputer. Specifically, it is necessary to perform sensitivity analysis, by using different models as the preliminary imputer (e.g., mean value, mode value, etc) and compare the resulting performance with that of baselines.

  5. The task to estimate propensity is truly difficult in practice. Moreover, if propensity is not accurately estimated, inverse propensity weighting induces bias. Authors should discuss the effectiveness of propensity score estimation accuracy in diverse scenarios, which is essential to back up the effectiveness of the method in practice.

  6. The repository provided is not complete. At least a concise readme file and requirement.txt should be included, to initiate the project. The hyperparameters should also be attached to improve reproducibility.

  7. Some theorems may be somewhat naive. For instance, Theorem 1 naturally holds, especially given the MCAR assumption.

  8. The experiments would need further supports and analysis, including but not limited to experiments with different missing ratios, confidence intervals, efficiency and computational cost comparison, isolated and detailed ablation study, sensitivity analysis to OT hyperparameters, sensitivity to preliminary imputers, case study and demos of imputed tables, etc.

Questions for authors

Please see the comments above.

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

The paper proposes Marginal Optimal Transport (MOT) for imputation with heterogeneous missingness patterns. Its population objective averages Wasserstein discrepancies between a model's projected distribution and the corresponding true marginal for each observed-coordinate pattern. Under MCAR and a shared chordal Markov structure whose maximal cliques coincide with the pattern support, the paper argues that matching these marginals recovers the joint distribution. Two estimators are proposed: MOT-IW, using conditional completions and importance weighting with Sinkhorn transport, and MOT-Var, using neural Kantorovich potentials. Experiments are done to validate the effectiveness of the proposed methods.

Critical strengths
  1. The problem is relevant: comparisons between incomplete samples require care when their observed coordinates differ. Matching distributions within a fixed observed subspace is a reasonable formulation.

  2. The attempt to evaluate distributional and structural properties beyond pointwise imputation error is appreciated.

  3. The population objective addresses a relevant observation problem. Under MCAR, matching each genuinely observed marginal avoids defining pairwise distances over inconsistent coordinate intersections. The manuscript correctly recognizes that marginal agreement alone does not generally identify an unrestricted joint distribution.

Critical weaknesses
  1. There are inconsistencies in the experiment results.
  • The displayed experimental results do not support several central statements. Under the legend currently shown in Figure 2, the covariance error of MOT-IW increases with sample size in subfigure (a), contrary to the statement that both MOT variants decay steadily. In subfigure (b), MOT-IW has an AUC of approximately 0.80 at dimension 50, despite the claim that both variants exceed 0.90 at every dimension. subfigure (c) does not show MOT-Var attaining the highest source correlation as claimed. subfigure (d) shows MissForest achieving the lowest RMSE, contrary to the claim in the final sentence on Page 8 that MOT variants occupy the lowest error envelope.

  • Figure 3 also displays identical numerical results across all four missingness mechanisms. Consequently, this figure does not demonstrate the mechanism-dependent baseline degradation described on Page 9. In addition, the MOT-Var/Iris RMSE is 0.570 in Figure 3 but 0.270 in Table 1.

  1. The stated cost and its metric proof are inconsistent. Equation (11) uses squared differences for numerical and ordinal features, whereas Appendix B.12 proves metricity for absolute differences. Squared differences do not generally satisfy the triangle inequality; for example, the squared distance between 0 and 2 exceeds the sum of those from 0 to 1 and 1 to 2. Thus, the appendix does not establish Lemma 3 for the cost as written.

  2. The model-side samples used in the training objective are not clearly defined. Algorithms 1–4 explicitly sample missing coordinates, but subsequently evaluate transport or critic losses on the observed coordinates of the generated sample. For example, Algorithm 2 samples \hat{X}{-\Lambda}, then compares the critic outputs on \hat{X}{\Lambda} and X_{\Lambda}. In addition, the construction of \hat{X}_{\Lambda} is not specified.

  3. The text attributes substantial degradation to the unprojected Full-DM ablation, but there is no result for MOT-VAR (Full-DM) in Table 1 or Figures 2–3. Because this comparison is used to support the proposed mechanism, it should be explicitly reported.

  4. The evaluation may not be comprehensive.

  • Comparison with more baselines such as ReMasker [1] may be needed.

[1] Du, Tianyu, Luca Melis, and Ting Wang. "Remasker: Imputing tabular data with masked autoencoding." In International Conference on Learning Representations, vol. 2024, pp. 9002-9024. 2024.

  • Only 6 datasets are tested, which may not be sufficient to show the improvement of the proposed scheme.
  1. There are quite a few confusing places.
  • In the last sentence of the introduction claims that MOT can improve "graph recovery." What is graph recovery and why is it related to missing value imputation?

  • In Section 2.2 Line 90, there is a question mark, probably a missing citation here.

  • In Lines 141 - 142, the mathematical notations and abbreviations such as MCAR and MNAR (Line 174) are not defined.

Questions for authors

Please kindly refer to the weaknesses.

Reviewer_Raz7OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Marginal Optimal Transport (MOT) for missing-value imputation under heterogeneous observation patterns. Its population-level objective matches model and data distributions on the observed subspace associated with each pattern. The paper presents an identifiability argument under MCAR and a graphical-model assumption, and introduces two estimators: MOT-IW and MOT-Var. It also claims finite-sample guarantees for the objective and imputation risk, and explores inverse-propensity weighting as an empirical extension to non-random missingness. Experiments cover tabular imputation, covariance estimation, independent-component recovery, and graphical-model recovery.

Critical strengths
  1. Relevant problem formulation Heterogeneous missingness raises an important question about which observed distributions can be meaningfully compared. Organizing distribution matching around observation patterns offers a useful perspective on this problem.

  2. Explicit attention to identifiability The paper recognizes that matching observed marginals does not generally identify the joint distribution. Introducing a graphical-model assumption makes this limitation explicit and provides a concrete setting for studying joint recovery.

  3. Broad evaluation scope The evaluation considers distributional and structural recovery alongside pointwise imputation error. These criteria help assess whether an imputation method preserves information needed for downstream analysis, although the reported results require clarification.

Critical weaknesses
  1. The training objective appears to lose the imputation signal.

    Algorithms 2 and 4 generate missing entries but compute the critic loss only on unchanged observed entries. Since

x^i,Λ=xi,Λ,\hat{x}_{i,\Lambda}=x_{i,\Lambda},

as the former one is obtained by only generating xi,−Λx_{i,-\Lambda}, the critic difference is zero and cannot train the imputer. Algorithm 3 has a similar issue after projecting and normalizing the completion weights.

  1. The cost definition and its proof are inconsistent.

    Equation (11) uses squared differences, while Appendix B.12 proves metricity for absolute differences. Squared differences violate the triangle inequality, so Lemma 3 does not hold for the stated cost. The implemented cost, Wasserstein order, and Lipschitz dual need to be clarified.

  2. Key theoretical claims are not adequately supported.

    Appendix B.11 replaces the model conditional distribution with an arbitrary proposal without the required correction, undermining Lemma 2. Appendix B.5 does not establish the claimed conversion from Wasserstein error to conditional-mean error. Appendix B.9 also incorrectly equates expected transport to a random point mass with transport to the population distribution. These issues affect the stated gradient, risk, and identification guarantees.

  3. Several experimental conclusions contradict the figures.

    Figure 2a shows increasing MOT-IW covariance error, Figure 2c shows lower source correlation for MOT-Var than the baselines, and Figure 2d shows lower RMSE for MissForest—contrary to the accompanying claims. Figure 3 repeats identical values across all four missingness mechanisms. MOT-Var’s Iris RMSE also differs between Table 1 (0.270) and Figure 3 (0.570). Corrected results and explanations are needed.

Questions for authors
  1. How do Algorithms 2–4 provide a nonzero learning signal for missing entries when the loss uses only unchanged observed coordinates?

  2. Which cost and Wasserstein order are implemented, and how do they agree with Lemma 3 and the Lipschitz dual?

  3. Can you clarify or correct the proposal substitution in Appendix B.11, the conditional-mean error bound in B.5, and the transport identity in B.9?

  4. Can you reconcile Figure 2 with the stated conclusions, explain the identical results across mechanisms in Figure 3, and resolve the Iris RMSE discrepancy between Table 1 and Figure 3?

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

Related papers

Discussion (0)

Sign in to comment.