- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper addresses the observation that semantic priors from pretrained language models are not always beneficial for tabular imputation and proposes SemImputer to incorporate such information selectively. SemImputer consists of Structure-Semantic Alignment (SSA), a Missing Pattern Encoder, and Counterfactual Utility-Guided Fusion (CUF). In particular, CUF trains a utility gate using the difference in prediction errors with and without semantic injection. Across 10 datasets under MCAR, MAR, MNAR, and MIXED missingness, the paper reports average improvements of 5.4% in RMSE and 16.1% in over existing methods.
Critical strengths
- Meaningful problem formulation: The paper recognizes that semantic priors are not universally useful and proposes selectively incorporating semantic information depending on the prediction target.
- Strong empirical performance: The paper reports consistent average improvements over existing methods on 10 datasets. Appendix C.1 also reports statistically significant dataset-level improvements over all baselines for RMSE and .
- Extensive additional experiments: The paper provides component ablations and evaluations under unseen missing rates and held-out missingness mechanisms, as well as comparisons against fixed-view
Critical weaknesses
- Insufficient justification for utility-guided fusion: Equation (4) constructs binary utility labels, whereas Equation (6) uses the predicted probability of these labels as a continuous fusion weight. The paper lacks theoretical or empirical evidence that the probability of semantic benefit determines an appropriate injection strength.
- Unclear contribution of counterfactual supervision: In Table 3, AUROC improves substantially from 0.543 to 0.910, while RMSE improves by only approximately 1.3%. Moreover, the ablation also changes the gradient flow and training configuration, making it difficult to isolate the independent contribution of counterfactual supervision.
- Limited validation on genuinely incomplete data: All main experiments apply artificial missingness to complete datasets. It remains unverified whether the proposed utility supervision is effective when ground-truth values at naturally missing training entries are unavailable.
Questions for authors
- Q1. Utility probability and injection strength: What justifies using the predicted probability of a binary counterfactual utility label as a continuous fusion weight? Are there comparisons with a fixed injection strength or a continuous utility target?
- Q2. Independent effect of counterfactual supervision: The ablation in Table 3 changes not only the utility loss but also the gradient flow. Can the authors provide a comparison that controls for other training conditions? Also, what exactly distinguishes “w/o CUF” in Figure 4 from “w/o Counterfactual Supervision” in Table 3?
- Q3. Validity of field-level alignment: Equation (1) applies alignment loss to both observed and masked fields. Is it appropriate to align identical semantic descriptors with row-specific structural representations? Are comparisons with observed-only or missing-only alignment available?
- Q4. Applicability to genuinely incomplete data: How are utility labels constructed when the training data itself contains missing values? Does the improvement persist when the model is trained by additionally masking observed entries?
4 comments
Response to Reviewer TYgG (Part 2)
Thank you for your careful assessment. The alignment and native-missingness studies use seeds 42/84/126 and test rates of 30/40/50%; in SemImputer’s native training, 10/20/30/40% of observed entries are masked. Results are averaged over settings and seeds, with datasets weighted equally. Native-missingness results are reported as mean ± standard deviation across the three seeds.
Q3
Observed-cell texts include values; masked descriptors express field meaning shared across rows. Structural tokens encode both field identity and row context, so the descriptor serves as a shared semantic prior. Eq. (1) aligns projected tokens through an auxiliary cosine loss with a decaying weight, while reconstruction and row-level InfoNCE provide row-specific supervision. We compare all-field, observed-only, masked-only and no alignment on four datasets under MCAR/MAR/MNAR/MIXED, normalizing by selected fields and keeping other settings fixed.
| Field-alignment scope | RMSE ↓ | MAE ↓ | R² ↑ | WD ↓ | Acc (%) ↑ |
|---|---|---|---|---|---|
| All fields | 0.7231 | 0.3664 | 0.4641 | 0.1953 | 80.05 |
| Masked only | 0.7428 | 0.3893 | 0.4605 | 0.2033 | 79.09 |
| Observed only | 0.7293 | 0.3749 | 0.4527 | 0.2047 | 79.11 |
| No field alignment | 0.7391 | 0.3851 | 0.4591 | 0.2182 | 78.25 |
Joint alignment leads all five means: RMSE falls 2.16% and Acc rises 1.80 percentage points versus no field alignment. Adding masked-field alignment to observed-only alignment lowers RMSE from 0.7293 to 0.7231, while masked-only RMSE is higher than without field alignment. These results support shared-descriptor alignment within the joint observed/masked objective.
W3 & Q4
Our benchmark follows CACTI and UnIMP (Section 4.1). We evaluated all rows of Cervical Cancer [1] (858 rows, 32 inputs, 13.19% missing) and Mammographic Mass [2] (961 rows, four inputs, 4.16% missing), excluding outcome/screening labels.
Following your suggestion, we retain native missingness and additionally mask observed values. Zero/capped predictions share inputs and parameters; their loss differences on known masked targets define Eq. (4)'s labels. Only these targets supervise reconstruction and utility. Methods share stratified splits and test masks, with training-only preprocessing. Metrics score known masked entries. CtrTab is omitted: its evaluated formulation needs complete targets.
Cervical Cancer
| Method | RMSE ↓ | MAE ↓ | R² ↑ | WD ↓ | Cat. Acc (%) ↑ |
|---|---|---|---|---|---|
| ReMasker | 0.9456 ± 0.3155 | 0.4954 ± 0.0799 | 0.2803 ± 0.1225 | 0.3587 ± 0.0653 | 95.88 ± 0.91 |
| KPI | 1.1077 ± 0.2703 | 0.5514 ± 0.0793 | 0.0448 ± 0.0264 | 0.3481 ± 0.0657 | 95.66 ± 0.41 |
| DiffPuter | 0.9671 ± 0.3040 | 0.4657 ± 0.0782 | 0.2612 ± 0.1067 | 0.2982 ± 0.0644 | 96.51 ± 0.16 |
| UnIMP | 1.0283 ± 0.3097 | 0.5349 ± 0.0911 | 0.1619 ± 0.0997 | 0.4170 ± 0.1048 | 95.30 ± 0.55 |
| CACTI | 1.0232 ± 0.5096 | 0.5777 ± 0.2841 | 0.2712 ± 0.2347 | 0.4634 ± 0.2960 | 90.73 ± 12.75 |
| RefiDiff | 1.0890 ± 0.2701 | 0.6726 ± 0.0569 | 0.1009 ± 0.0443 | 0.4590 ± 0.0688 | 94.79 ± 0.70 |
| SemImputer | 0.9664 ± 0.2793 | 0.5025 ± 0.0631 | 0.2406 ± 0.0662 | 0.3803 ± 0.0333 | 95.59 ± 0.22 |
Mammographic Mass
| Method | RMSE ↓ | MAE ↓ | R² ↑ | WD ↓ | Cat. Acc (%) ↑ |
|---|---|---|---|---|---|
| ReMasker | 0.9407 ± 0.0987 | 0.7682 ± 0.1140 | 0.1170 ± 0.0397 | 0.4908 ± 0.0531 | 63.90 ± 2.04 |
| KPI | 0.9510 ± 0.1291 | 0.7778 ± 0.1374 | 0.1073 ± 0.0625 | 0.4639 ± 0.0602 | 60.83 ± 1.27 |
| DiffPuter | 1.0307 ± 0.0889 | 0.8378 ± 0.0998 | 0.0658 ± 0.0513 | 0.4611 ± 0.1128 | 60.91 ± 2.37 |
| UnIMP | 0.9252 ± 0.1056 | 0.7448 ± 0.1176 | 0.1352 ± 0.0517 | 0.5015 ± 0.1648 | 62.31 ± 5.61 |
| CACTI | 0.9173 ± 0.0983 | 0.7445 ± 0.1027 | 0.1408 ± 0.0548 | 0.5251 ± 0.0973 | 64.82 ± 4.46 |
| RefiDiff | 0.9682 ± 0.1166 | 0.7747 ± 0.1193 | 0.1037 ± 0.0545 | 0.3566 ± 0.0703 | 61.49 ± 3.12 |
| SemImputer | 0.9271 ± 0.0979 | 0.7584 ± 0.1164 | 0.1272 ± 0.0357 | 0.4772 ± 0.0657 | 63.17 ± 2.78 |
SemImputer's mean Cervical RMSE ranks second, 5.55% below CACTI and 6.02% below UnIMP. On Mammographic Mass, RMSE is 1.45% below ReMasker's and WD is below both semantic baselines. With cached semantics, inference needs one gated prediction; Fig. 6 on Default reports the second-lowest online latency, 0.02 GFLOPs/row and peak GPU memory below 100 MB.
[2] Matthias Elter. Mammographic Mass. UCI Machine Learning Repository, 2007.
The authors’ responses and additional experiments have partially addressed the earlier questions. However, concerns about the novelty of the proposed model and the stability of the experimental results remain.
Additionally, the procedure for selecting the weights of the six loss terms in Equation (7) is unclear. Although the schedule for changing the weights during training and their final values are provided, it is difficult to determine from the main text, appendix, or provided code what criteria and search procedure were used to select those final values. Please explain the search range, selection criteria, and tuning budget for each weight. Please also clarify how much performance varies with the choice of the main weights and whether a comparable tuning budget was applied to the baseline methods.
Thank you very much for your feedback!
Thank you for your follow-up comments. We address your remaining concerns below.
Q1 What is the core novelty of the model?
Existing semantic imputers leave a key question implicit: does semantics help predict each missing target? Our core novelty is to explicitly learn whether semantic information improves the prediction of each missing target given the available observations. Our method constructs counterfactual supervision by comparing the prediction losses of the same model with and without semantic injection. The predicted probability of benefit then regulates injection to improve imputation while limiting interference from unhelpful semantics.
Q2 How stable are the experimental results?
Our ablation and rebuttal comparisons include three random seeds (42/84/126). Appendix C.1 additionally reports dataset-level paired comparisons: RMSE and R² improvements over all seven baselines remain significant after Holm correction. Appendix C.2 evaluates 60% and 70% missingness, beyond the training range, and Appendix C.3 evaluates missingness mechanisms unseen during training. Overall, we consider the experimental results of our method to be reasonably stable.
Q3 How were the six loss weights selected?
Appendix B.7 specifies the resulting schedule and coefficient values. The initial weights were set empirically. Checkpoints are selected by minimizing the unweighted sum of numerical and categorical validation reconstruction losses.
Q4 How sensitive is performance to the main loss weights?
Figure 5 varies the final utility-loss weight after warm-up (0.1, 0.2 or 0.4). Each setting keeps the weight at zero during alignment, increases it linearly over the first 10 transition epochs, and holds it constant thereafter.
Q5 Were comparable tuning budgets used for the baselines?
Our evaluation follows the benchmark datasets used by CACTI and UnIMP. We adopt the baselines’ official default hyperparameters, with benchmark adaptations and training budgets documented in Appendix B.8. No additional hyperparameter search was performed for the baselines.
Response to Reviewer TYgG (Part 1)
Thank you for your constructive suggestions. The following experiments use test rates of 30/40/50% under MCAR/MAR/MNAR/MIXED on the paper’s ten datasets. Native tests additionally mask observed entries and retain original missingness. Fixed-strength controls use seed-42 checkpoints. Full-model retraining uses seeds 42/84/126. Means average available settings and seeds within each dataset; macro means weight datasets equally.
CUF trains a sigmoid gate with BCE to predict whether capped injection improves a target's loss over zero injection beyond the prescribed margins. The predicted probability of semantic benefit scales a type-specific cap learned through reconstruction, giving a continuous interpolation between zero and capped injection. Identifying the loss-minimizing intermediate strength additionally requires losses at intermediate strengths.
Our fixed-strength controls change only the injection coefficient on the frozen checkpoints: zero, half-cap, full-cap or the original CUF gate. All other parameters and inputs remain fixed.
Separately, we freeze a common reference predictor and its caps and train only new gates and context projections. Binary utility uses BCE; strength regression uses MSE toward the lowest-loss fraction among 0, 0.05, 0.1, 0.25, 0.5, 0.75 and 1 for each known training target, selected using numerical Huber loss or categorical cross-entropy. Initialization, masks, optimizer, budget and validation selection are matched.
CUF leads fixed policies on RMSE, MAE, WD and Acc; capped injection gives the highest R². Strength regression improves RMSE and R², while binary utility gives better MAE, WD and Acc. We will combine endpoint-benefit supervision with intermediate-strength loss differences to better capture beneficial partial injection.
CUF's core contribution is explicit counterfactual supervision of target-specific semantic benefit. Section 3.3 states that CUF “explicitly supervises this predictive gain, which reconstruction alone leaves implicit.”
Figure 4 removes the entire CUF module before training and independently trains the remaining model. Table 3 retains the gate and semantic fusion but removes utility supervision, allowing reconstruction to train the gate. Gate-weighted row pooling also supplies row-alignment gradients. Initialization and checkpoint settings are documented in Appendix D.1. This compares explicit utility learning with end-to-end gating.
Our independent reruns match initial ρ=0.5, unweighted row pooling, detached aligned gate inputs, data, masks, architecture, optimizer, curriculum and checkpoint rules within each study. Full uses BCE and stops reconstruction gradients at the gate. Recon removes BCE and enables those gradients. Fixed half removes BCE and keeps them stopped, fixing ρ=0.5 while other parameters and caps train. Full–Fixed half tests utility supervision under the same reconstruction-gradient rule.
Full leads all five macro means on the ten datasets, with RMSE reductions of 0.04% over Recon and 0.11% over Fixed half. On both native datasets, Full improves mean RMSE, R² and WD over both controls; Recon gives lower MAE and higher categorical Acc.