acceptodds
Under review as a conference paper at ICLR 2027

Hybrid Methods for Robust Tabular Data Imputation

Abstract

Missing data are a fundamental challenge in statistical analysis and machine learning, as the choice of imputation method substantially impacts downstream inference. In this work, we propose two hybrid imputation methods called NuclearForest and SoftForest, which combine nuclear-norm-based low-rank initialization using Singular Value Thresholding (SVT) and SoftImpute, respectively, with a non-iterative Random Forest refinement. For the SVT-based component, we further introduce an adaptive step-size rule, prove adaptive step-size bounds, and establish convergence for the corresponding zero-initialized iteration. The low-rank initialization provides a structured warm start that captures the global covariance patterns in the data, while the subsequent Random Forest step recovers residual nonlinear signals encoding local dependencies. We conduct an extensive benchmark on diverse datasets from different application domains, comparing the proposed methods with seven established imputation methods under the Missing Completely at Random (MCAR), Missing at Random (MAR), and Missing Not at Random (MNAR) mechanisms across varying missingness rates. Our results demonstrate that NuclearForest and SoftForest match or exceed the imputation fidelity of state-of-the-art iterative methods such as MissForest, while significantly reducing computational cost. In particular, they achieve speedups of approximately and over MissForest by replacing iterative cycles with a single refinement step. Our approach effectively exploits the low-rank structure of real-world tabular data and accommodates mixed-type variables, providing an efficient and robust solution for data imputation in bioinformatics, economics, and beyond.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
missing data imputation, low-rank learning, matrix completion, random forests, tabular machine learning, nuclear norm minimization, MCAR/MAR/MNAR, computational efficiency

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_TeEYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes two hybrid methods for tabular data imputation, NuclearForest and SoftForest, which combine low-rank matrix completion with a single Random Forest refinement pass. NuclearForest employs a modified Singular Value Thresholding (SVT) algorithm with column-mean warm starting and an adaptive step-size rule, while SoftForest uses the existing SoftImpute algorithm for initialization. The central motivation is that a structured low-rank initialization can reduce the need for iterative Random Forest refinement while maintaining competitive imputation accuracy. The authors evaluate the proposed methods against classical imputation baselines on two real-world datasets under different missingness settings, using reconstruction, structural, statistical, downstream predictive, and runtime metrics. The results demonstrate substantial computational savings relative to iterative MissForest in the evaluated settings, although the proposed methods do not consistently achieve the best imputation accuracy across all metrics and missingness mechanisms.

Critical strengths

1.The proposed methods substantially reduce computational runtime compared with iterative MissForest. 2.The paper provides a comprehensive evaluation of imputation quality using complementary metrics, component ablation studies, and supplementary runtime analyses.

Critical weaknesses
  1. The methodological novelty is limited, and the adaptive step-size mechanism lacks sufficient justification.
  2. Despite the authors' explanation, the lack of comparisons with neural-network-based imputation methods weakens the claim of broad robustness and competitiveness.
Questions for authors

1.Why do the authors change the conventional SVT step size from (1.2/p) to (1.2p)? What is the theoretical justification for this choice, and what advantages does it offer over other admissible initial step sizes? 2.How were the contraction and expansion factors of 0.9 and 1.05 determined? Could the authors provide justification for these choices or a corresponding sensitivity analysis? 3.The authors state that no step-size contractions were triggered during the experiments, making the adaptive update equivalent to a predetermined multiplicative growth schedule. What additional benefits does the residual-driven mechanism offer over a simpler deterministic schedule? 4.The convergence proof is established for a zero-initialized SVT iteration, whereas the implemented algorithm uses column-mean warm starting. Does the convergence guarantee also apply to the implemented warm-started algorithm?

5 comments

We appreciate the time and care the reviewer has devoted to our work. Your comments have helped us refine our ideas and improve the presentation. We have incorporated your suggestions into the revised manuscript, with all changes highlighted in green for ease of review.

Reviewer: Q1.Why do the authors change the conventional SVT step size from (1.2/p) to (1.2p)? What is the theoretical justification for this choice?

We thank the reviewer for this observation. The step size was changed to guarantee admissibility. We keep Cai et al.'s constant 1.2, replace 1/p1/p by pp, and cap the step at 2p2p, which keeps every step below 2. The results are insensitive to this constant, because initial steps from 0.5p0.5p to 1.9p1.9p change the NRMSE by at most 0.002 (see Q2).

Cai et al. chose 1.2/p1.2/p for large, randomly generated, low-rank matrices (up to 30,000 × 30,000) observed at low sampling ratios. Their heuristic justification is that, under random sampling and suitable incoherence assumptions, the squared Frobenius norm of the observed part of a fixed low-rank matrix is approximately pp times that of the full matrix. This motivates a step size proportional to 1/p1/p to accelerate convergence. However, Cai et al. explicitly acknowledge that this argument is heuristic.

In contrast, the real metabolomics data used in our paper are medium-sized and approximately low-rank (after standardization, the top 10 principal components explain 65% of the variance). In this setting, the 1/p1/p compensation is less suitable, and a large step can make SVT diverge, and even an admissible step close to 2 can make SVT fit the noise in the observed entries within a fixed iteration budget. We added a statement to the Appendix of the revised manuscript to clarify this distinction.

With respect to the theoretical justification, we make the following observations. Cai et al.'s convergence theorem guarantees convergence for step sizes below 2. The value 1.2/p1.2/p exceeds 2 as soon as more than 40% of the entries are missing (p<0.6p<0.6), so the guarantee does not hold there. Our steps start at 1.2p1.2p and are capped at 2p2p, which is always below 2. In fact, 2p2p is the largest step proportional to pp that is admissible for every pp.

Our experiments are consistent with this. With Cai et al.'s step 1.2/p1.2/p, SVT diverged whenever 50% or more of the data were missing in the MCAR/MAR settings (150 of 270 runs in three settings). The residual overflowed at 60–90% missingness (120 runs) and grew above 108810^{88} within 1000 iterations at 50% (30 runs). It converged whenever the step stayed below 2.

We have also added a statement to the Introduction and the Appendix of the revised manuscript to make this argument explicit.

Reviewer: What advantages does it offer over other admissible initial step sizes?

Within the admissible range, the choice of the initial step has essentially no effect. Initial steps from 0.5p0.5p to 1.9p1.9p change the NRMSE by at most 0.002 at any missing rate in all four settings. The advantage of the rule is therefore that it needs no tuning of the initial step. Details are given in our answer to Q2.

Reviewer: Q2. How were the contraction and expansion factors of 0.9 and 1.05 determined? Could the authors provide justification for these choices or a corresponding sensitivity analysis?

We thank the reviewer for this suggestion and have added a sensitivity analysis to the revised manuscript. The constants 0.9 and 1.05 are common multiplicative adaptation factors. They do not affect the convergence guarantee. The proof of Proposition 1 extends directly to any contraction factor c∈(0,1)c\in(0,1) and any expansion factor above 1, provided only finitely many contractions occur. Only the contraction factor enters the result, through the lower bound cBδ0c^B\delta_0. The sensitivity analysis covers all missing rates (10–90%), 10 repeats and four settings (metabolomics MCAR/MAR and MNAR, housing MCAR and MAR).

For initial step and expansion factor, we tested initial steps of 0.5p0.5p, 0.8p0.8p, 1.0p1.0p, 1.1p1.1p, 1.2p1.2p, 1.5p1.5p and 1.9p1.9p, and expansion factors of 1.01, 1.02, 1.05, 1.1 and 1.2. In both cases, the NRMSE changed by at most 0.002 at any missing rate in any setting. On metabolomics, the step grows geometrically to the cap 2p2p within 3–52 of the 1000 iterations (11 with the paper's factor 1.05), so all variants run with essentially the same step. On housing, SVT converges within about 11 iterations to the unique solution of the same convex problem, regardless of the step.

Table A. Sensitivity to the initial step size δ0=c0p\delta_0 = c_0 p (cap 2p, expansion factor 1.05, contraction factor 0.9). NRMSE averaged over 10 repeats.

Initial step10%20%30%40%50%60%70%80%90%
Metabolomics, MCAR/MAR
0.5p0.7360.8100.8740.9381.0271.1261.2501.3771.325
0.8p0.7360.8100.8740.9381.0271.1271.2511.3781.326
1.0p0.7360.8100.8740.9381.0271.1271.2511.3781.326
1.1p0.7360.8100.8740.9381.0271.1271.2511.3781.326
1.2p (paper)0.7360.8100.8740.9381.0271.1271.2511.3781.327
1.5p0.7360.8100.8740.9381.0271.1271.2511.3791.327
1.9p0.7360.8100.8740.9381.0271.1271.2511.3791.327
Max. difference0.0000.0000.0000.0000.0000.0010.0010.0020.002
Metabolomics, MNAR
0.5p2.2772.6872.6262.6332.8202.8652.9813.1763.333
0.8p2.2772.6872.6272.6332.8212.8652.9813.1763.334
1.0p2.2772.6872.6272.6332.8212.8652.9823.1763.334
1.1p2.2772.6872.6272.6322.8212.8652.9823.1763.334
1.2p (paper)2.2772.6872.6272.6322.8212.8652.9823.1763.334
1.5p2.2772.6872.6272.6322.8212.8652.9823.1763.334
1.9p2.2772.6872.6272.6322.8212.8652.9823.1763.334
Max. difference0.0010.0000.0010.0010.0010.0010.0010.0010.001
Housing, MCAR
0.5p1.0191.0131.0131.0111.0071.0051.0041.0031.007
0.8p1.0191.0131.0131.0111.0071.0061.0041.0031.007
1.0p1.0191.0131.0131.0111.0071.0061.0051.0041.008
1.1p1.0191.0131.0131.0111.0071.0061.0051.0041.008
1.2p (paper)1.0191.0131.0131.0111.0071.0071.0051.0041.008
1.5p1.0191.0131.0131.0111.0071.0071.0051.0041.008
1.9p1.0191.0131.0131.0111.0071.0071.0051.0051.008
Max. difference0.0000.0000.0000.0000.0000.0010.0020.0020.001
Housing, MAR
0.5p1.0361.0241.0211.0241.0271.0361.0581.0811.087
0.8p1.0361.0241.0211.0241.0271.0351.0581.0801.086
1.0p1.0361.0241.0211.0241.0271.0351.0571.0801.086
1.1p1.0361.0241.0211.0241.0271.0351.0571.0801.086
1.2p (paper)1.0361.0241.0211.0241.0271.0351.0571.0801.085
1.5p1.0361.0241.0211.0241.0271.0351.0571.0801.085
1.9p1.0361.0241.0211.0241.0271.0351.0571.0801.085
Max. difference0.0000.0000.0000.0000.0000.0000.0010.0010.002

We continue our response to Q2 here.

Table B. Sensitivity to the expansion factor (initial step 1.2p, cap 2p, contraction factor 0.9). NRMSE averaged over 10 repeats.

Expansion factor10%20%30%40%50%60%70%80%90%
Metabolomics, MCAR/MAR
1.010.7360.8100.8740.9381.0271.1261.2501.3771.325
1.020.7360.8100.8740.9381.0271.1271.2511.3781.326
1.05 (paper)0.7360.8100.8740.9381.0271.1271.2511.3781.327
1.10.7360.8100.8740.9381.0271.1271.2511.3791.327
1.20.7360.8100.8740.9381.0271.1271.2511.3791.327
Max. difference0.0000.0000.0000.0000.0000.0010.0010.0010.002
Metabolomics, MNAR
1.012.2772.6872.6262.6332.8202.8652.9813.1763.333
1.022.2772.6872.6272.6332.8212.8652.9823.1763.334
1.05 (paper)2.2772.6872.6272.6322.8212.8652.9823.1763.334
1.12.2772.6872.6272.6322.8212.8652.9823.1763.334
1.22.2782.6872.6272.6322.8212.8652.9823.1763.334
Max. difference0.0010.0000.0000.0010.0010.0010.0010.0000.001
Housing, MCAR
1.011.0191.0131.0131.0111.0071.0061.0051.0041.007
1.021.0191.0131.0131.0111.0071.0061.0051.0041.008
1.05 (paper)1.0191.0131.0131.0111.0071.0071.0051.0041.008
1.11.0191.0131.0131.0111.0071.0071.0051.0041.008
1.21.0191.0131.0131.0111.0071.0071.0051.0041.008
Max. difference0.0000.0000.0000.0000.0000.0000.0010.0010.000
Housing, MAR
1.011.0361.0241.0211.0241.0271.0351.0571.0801.086
1.021.0361.0241.0211.0241.0271.0351.0571.0801.086
1.05 (paper)1.0361.0241.0211.0241.0271.0351.0571.0801.085
1.11.0361.0241.0211.0241.0271.0351.0571.0801.085
1.21.0361.0241.0211.0241.0271.0351.0571.0801.085
Max. difference0.0000.0000.0000.0000.0000.0000.0000.0000.001

For the contraction factor, on the real datasets no contraction is triggered, so this factor does not influence the reported results. It acts as a safeguard when the step is too large. To test it, we started from Cai et al.'s step size 1.2/p1.2/p and tested contraction factors of 0.5, 0.7, 0.9 and 0.95. With 0.5, 0.7 and 0.9, every one of the 360 runs recovered and converged. With 0.95, 52 of 360 runs failed, presumably because the steps shrink too slowly. Among the factors that worked, 0.9 gave the lowest error on metabolomics MCAR with 90% missing data (NRMSE 1.638, compared with 1.659 for 0.7 and 1.679 for 0.5). We note that the differences are small, with 0.9 providing competitive performance among the evaluated settings.

Table C. Sensitivity to the contraction factor when the initial step is too large (SVT started from δ0=1.2/p\delta_0 = 1.2/p, 1000 iterations, all 360 masks). In the paper setting no contraction occurs, so the factor has no effect there. A run fails if its residual becomes non-finite or does not fall below 1.

Contraction factor0.50.70.9 (paper)0.95
Failed runs0/3600/3600/36052/360
NRMSE, metab. MCAR 90%1.6791.6591.638diverges

We added these three tables to the Appendix of the revised manuscript.

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

This paper targets missing-value imputation for tabular data, with the primary objective of reducing the computational cost of random-forest-based approaches such as MissForest (2012). It combines existing low-rank matrix-completion techniques with a random-forest refinement step to improve runtime while retaining competitive imputation quality.

Critical strengths

The paper is exceptionally well structured and clearly presented. This should not be understated as a strength: among the many papers I have read recently, this one stands out for the quality in this regard.

The empirical evaluation is a major strength. The benchmarking is carefully conducted, with excellent comparisons and visualizations/display of results. In particular, the runtime analysis makes one of the main practical motivations of the proposed approach clear and presents its computational advantages in a positive light. One can clearly see an order of magnitude improvement in runtime on different datasets. Overall, the experimental section is a strong aspect of the paper, with a sufficient number of datasets and experiments.

Critical weaknesses

The main concern is the limited methodological novelty and/or the framing/importance placed on this novelty. The approach builds on established techniques, including SVT, introduced for matrix completion in 2010, and combines them in an intuitive and potentially useful way. However, the paper does not yet make a sufficiently compelling case for the significance of this combination. If the theoretical analysis is intended to be a central contribution, the main body should explicitly state and discuss the key theorem(s), rather than merely referring the reader to the appendix. The relationship between these results and the algorithm as actually implemented should also be made clear.

Without a stronger theoretical contribution, the paper's main value rests on its empirical evidence. Promoting this contribution could work for the paper, but in that case, the comparison with existing methods should be more comprehensive to include more recent approaches, not only established baselines. Otherwise, it is difficult to assess the method's significance relative to the current state of the art, or to determine what it adds beyond another variant in an already well-established area of imputation.

Finally, the evaluation focuses primarily on imputation quality and computational cost. For practical use, imputation is often an intermediate step in a downstream analysis or prediction task, where the uncertainty associated with missing-value estimates can be important. Studying how the proposed method handles or represents this uncertainty, and how it affects downstream tasks, would strengthen the practical case for the approach.

Questions for authors

Can the authors clearly state what is new in the convergence analysis relative to existing SVT results, in a way that would fit in the main body of the paper.

What in terms of downstream analysis beyond faster point imputation, particularly regarding uncertainty in the imputed values and its effect on subsequent inference or prediction? Any difference wrt state of the art methods?

2 comments

We appreciate the time and care the reviewer has devoted to our work. Your comments have helped us refine our ideas and improve the presentation. We have incorporated your suggestions into the revised manuscript, with all changes highlighted in green for ease of review.

Reviewer: Can the authors clearly state what is new in the convergence analysis relative to existing SVT results, in a way that would fit in the main body of the paper.

Thank you for this suggestion. We have moved the convergence theorem to the main body of the revised manuscript (Theorem 1, Section 2) and added a paragraph stating what is new relative to existing SVT results. Cai et al. (2010) prove convergence of zero-initialized SVT for step sizes 0<inf⁡kδk≤sup⁡kδk<20<\inf_k\delta_k\le\sup_k\delta_k<2, and Lei and Zhou (2019) give convergence rates in this setting. Our contributions are twofold. First, the proposed step-size rule always satisfies δk≤2p<2\delta_k\le 2p<2, and it is bounded away from zero under the finite-contraction assumption (Proposition 1). Second, we extend the guarantee to the column-mean initialized iteration used in our experiments, including an O(1/T)O(1/T) rate for the best iterate. Since the updates never change the unobserved entries of the auxiliary variable, the warm start fixes them, and our proof adapts the Uzawa (dual-gradient) analysis of Cai et al. to the resulting shifted problem. The newly added material is highlighted in green. The detailed proof is in the appendix.

Reviewer: What in terms of downstream analysis beyond faster point imputation, particularly regarding uncertainty in the imputed values and its effect on subsequent inference or prediction? Any difference wrt state of the art methods?


We address downstream use and uncertainty separately in the following.

For downstream analysis, beyond the reported reconstruction error, our benchmark already evaluates downstream use. 
(i) To assess the preservation of statistical inference signals, we compare Welch t-tests on complete versus imputed data and report the Pearson correlation of the log p-values (metabolomics, Figure 2, and housing, Appendix Figure 7).


(ii) For prediction, we report the degradation of the R2R^2 of a ridge regression for housing prices (Figure 4 and Appendix).


(iii) We also report PCA/PLS Procrustes distances and Gower’s distance for structure preservation.

The results depend on the missingness level. Overall, the proposed methods offer downstream utility comparable to MissForest at substantially lower cost.

For uncertainty, all methods in our benchmark (NuclearForest, SoftForest, MissForest, kNN, SVT, SoftImpute, mean, median, and half-min) are implemented as single-imputation point estimators and do not provide uncertainty estimates for the imputed values. Single imputation does not account for uncertainty in the missing values and can lead to underestimated standard errors. A possible direction for future work is to estimate imputation uncertainty through the refinement step, for example using quantile regression forests or bootstrapped forests. We have added this limitation to the Conclusion of the revised manuscript.

Thank you to the authors for replying. I can see now that uncertainty analysis is beyond the scope of this paper, and a mention of this for possible future direction is sufficient. The theoretical insight is useful. And I still like very much how the paper is structured and presented. My enthusiasm is only tempered by that I don't easily see a significant impact from these contributions.

Reviewer_Nwn7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes NuclearForest and SoftForest for tabular data imputation. They combine low-rank completion using an adapted SVT algorithm or SoftImpute with a single random-forest refinement pass. The adapted SVT uses column-mean initialization and residual-based step-size adjustment. The theory establishes step-size bounds under finitely many contraction steps and derives convergence for the corresponding zero-initialized iteration. Experiments on metabolomics and housing data evaluate reconstruction, structure preservation, statistical signals, downstream prediction, and runtime across missingness settings. The hybrids achieve competitive quality in several MCAR/MAR settings with substantially lower runtime than MissForest, while their advantages are less consistent under left-censored MNAR.

Critical strengths

S1. A useful trade-off between imputation quality and runtime.

The proposed hybrids achieve competitive imputation quality in several evaluated settings while reducing runtime relative to iterative forest imputation. In the metabolomics experiment labeled MCAR/MAR, Table 1 reports speedups of about 5.81× for NuclearForest and 9.52× for SoftForest over MissForest. This efficiency benefit is useful when imputation must be repeated across experimental conditions.

S2. Evaluation beyond reconstruction error.

The paper examines the preservation of data structure and statistical signals, alongside reconstruction error. It also evaluates downstream prediction on housing data. Ten runs with independently regenerated missingness masks help assess sensitivity to the sampled masks. The SVT ablations separately examine column-mean initialization and step-size adaptation.

Critical weaknesses

W1. The rationale for a single forest pass needs direct evidence.

Section 2 states that further refinement passes provide only marginal gains. However, the paper does not report a direct comparison across refinement counts under the same initialization. The comparison with MissForest changes both initialization and iteration count, so it cannot separate their effects. A comparison of zero, one, and several passes with the same SVT or SoftImpute initialization would clarify this design choice. It should report both imputation quality and runtime under matched settings. This gap limits the explanation for the single-pass design, although the reported results still support a useful overall trade-off.

W2. The evidence supports a narrower scope than the broad robustness claims.

The benchmark includes one metabolomics dataset and one housing dataset. This limits the support for the abstract’s description of an extensive benchmark on diverse datasets. The metabolomics MCAR/MAR experiment uses uniform random masking. A separate MAR experiment is reported only for housing, while left-censored MNAR is evaluated only on metabolomics. The results therefore leave uncertainty about whether the observed trends hold across both datasets and mechanisms. The paper also does not systematically vary effective rank, correlation strength, or nonlinear relationships. These factors matter because low-rank initialization is central to the method. The weaker MNAR results further show that the benefits depend on the setting. The authors should define robustness and align the title and abstract with the demonstrated scope.

Questions for authors

Q1. With identical SVT or SoftImpute initialization, how do zero, one, and several forest passes compare in quality and runtime? Please keep the masks and forest settings fixed across these comparisons. Can the authors provide the results behind the statement that further passes yield only marginal gains? A corresponding comparison with simple initialization would also help separate the benefits of initialization and refinement count.

Q2. What specific property does “robust” refer to in this paper? Which results support that property, and under what conditions should users expect the hybrids to help? Please clarify these conditions and qualify the title and abstract where needed. Additional datasets and controlled structural experiments would strengthen the broader claims.

Q3. The Improved SVT paragraph in Section 2 states a cap of 2. Algorithm 2 and Appendix C, Table 4, specify min(2p, 2), which equals 2p when 0 < p < 1. Proposition 1 also uses the bound 2p. Please align the main text with the reported setting, or explain any difference in the experimental implementation.

3 comments

We are especially grateful for the reviewer's encouraging assessment of our work and the thoughtful suggestions for further improvement. We have incorporated your suggestions into the revised manuscript, with all changes highlighted in green for ease of review.

Reviewer: Q1. With identical SVT or SoftImpute initialization, how do zero, one, and several forest passes compare in quality and runtime? Please keep the masks and forest settings fixed across these comparisons. Can the authors provide the results behind the statement that further passes yield only marginal gains? A corresponding comparison with simple initialization would also help separate the benefits of initialization and refinement count.

We thank the reviewer for this suggestion. To answer this question we ran an ablation study with masks and forest settings held fixed as recommended. We used the same initialization followed by K = 0, 1, 2, 3, 5 and 10 forest passes, with median (metabolomics) or mean (housing) initialization as a simple reference. Values are mean NRMSE over 9 missing rates × 10 repeats (MCAR). The full tables, with standard deviations and the runtime of the reference initializations, are added to the Appendix of the revised manuscript. 

Metabolomics. NRMSE

InitializationK=0 (init only)K=1 (ours)K=2K=3K=5K=10
SVT1.0520.7880.7670.7680.7720.779
SoftImpute1.1250.8130.7860.7810.7820.787
Median (reference)1.0320.8220.7830.7720.7700.777

Housing. NRMSE

InitializationK=0 (init only)K=1 (ours)K=2K=3K=5K=10
SVT1.0100.9540.9851.0051.0191.029
SoftImpute1.7111.0401.0241.0241.0351.040
Mean (reference)1.0040.9550.9891.0101.0261.034

Runtime. Cumulative seconds per imputation (initialization + K passes)

Dataset / initializationK=0K=1K=2K=3K=5K=10
Metabolomics, SVT11.719.627.635.551.491.1
Metabolomics, SoftImpute0.88.916.824.740.680.2
Housing, SVT0.0030.651.281.913.176.32
Housing, SoftImpute0.0010.651.291.933.216.38

For number of passes, the first pass gives most of the gain. A second pass gives a further 0.02 to 0.03 for the low-rank initializations on metabolomics (0.04 for median) but increases runtime by about 8 s, and beyond K = 3 there is no further gain. On housing, additional passes increase the error, so K = 1 is indeed best there. We therefore describe the later passes as “small gains on metabolomics and negative on housing”. As can be seen from the reported measurements, runtime grows approximately linearly with K.

For initialization vs. refinement, most of the improvement over the initialization alone comes from the forest step. With one pass, low-rank initialization improves over the simple reference on metabolomics (SVT 0.788 vs. median 0.822 vs. SoftImpute 0.813), but not on housing (SVT 0.954 vs. mean 0.955 vs. SoftImpute 1.040). With more passes, the metabolomics gap closes (when K = 5, median 0.770 vs. SVT 0.772). Low-rank initialization is thus most useful when only a single pass is affordable and the data are approximately low-rank. Its main practical role is to make a single pass sufficient. We have revised the wording in the abstract and the Methodology section to make this argument more clear.

Reviewer: Q2. What specific property does “robust” refer to in this paper? Which results support that property, and under what conditions should users expect the hybrids to help? Please clarify these conditions and qualify the title and abstract where needed. 

Thank you for pointing this out. We acknowledge that the term "robust" may be misleading in this context, as in statistics it is commonly associated with resistance to outliers, contamination, or deviations from modeling assumptions, none of which we study in this paper. We therefore follow the reviewer's recommendation and have changed the title to "Hybrid Low-Rank and Random Forest Methods for Efficient Tabular Data Imputation", and removed the term from the abstract and introduction. In addition, we added explicit statements to the Conclusion about the conditions under which the hybrids are expected to help. MCAR/MAR missingness at low to moderate missing rates, approximately low-rank data with nonlinear feature relationships, and larger datasets where the runtime savings matter.

Reviewer: Additional datasets and controlled structural experiments would strengthen the broader claims.

As suggested, we evaluated all methods on a larger blood-plasma metabolomics dataset from the NIH Metabolomics Workbench (Project PR000950, Study ST001386, Analysis AN002314). After preprocessing, the dataset comprised 9,952 samples and 144 metabolites. The results are consistent with our observations that the runtime advantage persists and that accuracy remains competitive. We have included this new experiment in the revised paper.



We agree that further datasets and controlled structural experiments (e.g., varying rank and nonlinearity) would give further insights. Due to the limited time of the rebuttal phase we have to perform these studies in future work and thank the reviewer for this suggestion.

Reviewer: Q3. The Improved SVT paragraph in Section 2 states a cap of 2. Algorithm 2 and Appendix C, Table 4, specify min(2p, 2), which equals 2p when 0 < p < 1. Proposition 1 also uses the bound 2p. Please align the main text with the reported setting, or explain any difference in the experimental implementation.

We would like to thank the reviewer for carefully checking our manuscript. This is indeed a typo in the Improved SVT paragraph. The cap is min(2p,2), i.e. 2p for 0<p<1, as in Algorithm 2, Table 4 and Proposition 1. We have corrected this statement in the revised paper.

We use 2p because it reduces the step-size cap as missingness increases. Within our fixed budget of 1000 iterations, this may act as implicit early stopping, since our NRMSE increases as SVT runs longer. At 90% missingness, 2p gives a much lower NRMSE than a constant cap of 1.99 (1.33 vs. 1.65), with comparable errors at low missingness (see the table below for a numerical experiment on metabolomics data with MCAR).

 SVT NRMSE (metabolomics, MCAR, 1000 iterations): step-size cap 2p vs. a constant cap of 1.99

Step-size cap10%20%30%40%50%60%70%80%90%
2p0.7360.8100.8740.9381.0271.1271.2511.3781.327
Constant 1.990.7380.8110.8780.9491.0511.1671.3411.5631.650

For the sake of clarity we have added the table to the revised manuscript in the Appendix.

We thank the reviewer for these valuable suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.