{"id":"97df1212-1b30-4c74-a90b-d5015113f543","arxiv_id":"2501.10910","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"DeepIFSAC combines feature and sample attention with CutMix augmentation and contrastive learning to impute missing tabular values, ranking first on average against 11 baselines in extensive benchmarks.","lead":"DeepIFSAC imputes missing values in tabular data by learning attention between features and attention between samples inside a contrastive learning framework. In benchmarks across 12 datasets, missing rates up to 90%, and three missingness types, it achieves the best average rank among 11 baselines, though the margin over its own sample-attention baseline is small.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significance tests: average-rank margins over the closest baseline are within one standard deviation, so the central superiority claim is not established.","rationale":"The paper presents a serious, extensive empirical study: 12 datasets, three missing types, five missing rates, 11 baselines, an ablation, and a downstream EHR task. The method is well-motivated, and the ablation provides some internal support for CutMix and contrastive learning. Credit is also due for the authors' self-declared limitations in Section 6.6, which delimit the method's scope. My concern is not about fraud or internal inconsistency; it is that the central claim of superiority is not quantitatively supported. The reader's stated weakest assumption was the CutMix surrogate for real missingness, especially MNAR. That is a valid concern about the method's mechanism, and I partially agree with it: training on uniform random corruption does not obviously model self-censoring MNAR. However, the most load-bearing weakness for the paper's headline claim is statistical: even if CutMix were a perfect surrogate, the reported rank differences between DeepIFSAC and the closest attention baseline are within the reported standard deviations, and no significance testing is performed. The abstract asserts 'superiority' on the basis of average ranks that may reflect noise, and the epoch-count discrepancy for generative baselines further weakens the comparison. A concrete Friedman/Nemenyi test, plus a rerun of GAIN and Diffputer with their usual training budgets, would settle whether the claimed superiority survives. Because the reader already recommended a CONDITIONAL verdict and my analysis supports that same conclusion, I leave the verdict unchanged, but I would emphasize the significance-testing requirement as the primary condition for acceptance.","tokens_in":29325,"tokens_out":5289,"duration_ms":62966,"concrete_test":"Run a Friedman rank test followed by Nemenyi post-hoc analysis on the 12 datasets for each missing type and missing rate, comparing DeepIFSAC against Between-sample, missForest, MICE, and KNN. Report the critical difference and adjusted p-values. Additionally, retrain GAIN and Diffputer with their recommended 10,000-epoch schedules and recompute the rank comparison. If DeepIFSAC's rank advantage over Between-sample is not significant (p < 0.05) or is smaller than the critical difference, the abstract should be revised to state that DeepIFSAC is competitive rather than superior.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract claims that DeepIFSAC's average performance rank 'demonstrates its superiority' over 11 methods across missing rates 10%–90% and three missing types, especially MNAR. The evidence is Table 2: mean ranks over 12 datasets. DeepIFSAC's closest competitor is the authors' own Between-sample attention baseline: 3.53 vs. 3.72 (MCAR), 3.50 vs. 3.83 (MAR), and 2.75 vs. 2.98 (MNAR). The reported standard deviations of these ranks are roughly 2.5–3.5, so the differences are far below one standard deviation. No significance test, confidence interval, or effect size is reported anywhere. The paper also trains all deep baselines for 1,000 epochs, while GAIN and Diffputer are typically trained for 10,000 epochs (a point the authors acknowledge in Section 6.1), so the comparison may understate generative baselines. Since the central claim is a comparative superiority claim, the absence of any statistical support for the comparison means the claim is not established. The novel contribution—contrastive learning added to the authors' own Between-sample method—changes the average rank by roughly 0.2–0.3, a magnitude that could easily arise from run-to-run and dataset-to-dataset noise.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DeepIFSAC, a transformer-based missing value imputation method that combines between-feature (column) attention and between-sample (row) attention, trained with CutMix data augmentation and a contrastive learning objective. The method is benchmarked against 11 baselines on 12 OpenML tabular datasets under MCAR, MAR, and MNAR missingness at rates from 10% to 90%, using held-out test folds and NRMSE as the primary metric; a secondary evaluation imputes missing values in an All of Us EHR dataset and measures downstream heart-failure classification F1. The central claim is that DeepIFSAC's average performance rank across datasets demonstrates superiority over state-of-the-art methods, especially under MNAR.","tokens_in":29552,"tokens_out":6025,"duration_ms":62823,"significance":"If the central claim were established, the paper would provide a valuable practical result: a relatively simple attention-based imputer that outperforms GAN-, diffusion-, and autoencoder-based methods across a wide range of missing-data scenarios, with a large and careful benchmark design. The evaluation has notable strengths: set-aside test folds avoid circular fitting, the missingness regimes cover three mechanisms, the missing rates extend to 90%, and a real EHR downstream task is included. However, the statistical support for the superiority claim is currently insufficient: the margins over the closest baseline are small relative to the reported variability, no significance tests are provided, and the acknowledged shorter training budget for generative baselines hampers fair comparison. The novelty of the contrastive component is also undermined by the ablation results, which show essentially no improvement over the same architecture without contrastive learning.","major_comments":[{"comment":"The abstract's claim of superiority for missing rates between 10% and 90% is not supported by the reported average ranks. The gap between DeepIFSAC and the closest baseline (Between-sample [40]) is 0.19 (MCAR), 0.33 (MAR), and 0.23 (MNAR) rank points, while the standard deviations of the ranks are roughly 2.5–3.5; no significance test, confidence interval, or effect size is reported. Moreover, at MCAR 70% (4.33 vs. 3.75) and MAR 70% (4.67 vs. 4.33) the Between-sample baseline has a better average rank than DeepIFSAC, directly contradicting the phrase 'for missing rates between 10% and 90%.' The authors should either provide paired tests across datasets (e.g., Wilcoxon signed-rank) or substantially temper the superiority claim.","section":"Section 5.1, Table 2"},{"comment":"The contrastive learning component is a central novel contribution, but the ablation shows that it provides at most 0.01 NRMSE improvement over the same model without the contrastive loss (e.g., dataset 37 at 30%: 0.30 vs. 0.29; dataset 1497 at 10%: 0.18 vs. 0.18). Given the standard deviations reported in the same table (on the order of 0.01–0.06), this difference is within noise. The authors should demonstrate that the contrastive contribution is statistically significant, or reposition contrastive learning as a minor auxiliary loss rather than a central novelty.","section":"Section 3.4, Table 8"},{"comment":"The deep baselines GAIN and Diffputer are trained for 1,000 epochs, while the authors themselves acknowledge in Section 6.1 that these methods are typically trained for 10,000 epochs. This comparison likely understates the performance of the generative baselines and undermines the claim that DeepIFSAC is superior to state-of-the-art deep methods. To support the central claim, the authors should retrain these baselines to convergence (or at least match their recommended training budgets) or provide evidence that 1,000 epochs is sufficient for them.","section":"Section 4.4 and Section 6.1"},{"comment":"The training procedure uses CutMix corruption (Equation 4) as the sole surrogate for missingness, including for MNAR where the missingness probability depends on the missing value itself. The paper does not justify that CutMix corruptions are a faithful surrogate for the target missingness mechanisms; the empirical NRMSE is the evidence offered, but that is the very claim under evaluation. The authors should either provide a diagnostic separating the effect of the augmentation distribution from the true missingness model or explicitly discuss this limitation as a threat to generalization.","section":"Section 3.2, Algorithm 1"}],"minor_comments":[{"comment":"The downstream F1 difference between DeepIFSAC and Between-sample is only 0.25 points (87.49 vs. 87.24) with overlapping standard deviations (1.35 vs. 1.24); no significance test is reported for this claim either.","section":"Table 6"},{"comment":"The text states the source code has been publicly shared, but the footnote says the link would be shared in the final version for anonymity; this inconsistency should be resolved and the code link provided.","section":"Footnote 1"},{"comment":"References [43] and [44] appear to duplicate the same paper; please check and remove the duplicate.","section":"References"},{"comment":"The NRMSE formula uses Var(X_j), but when X_j has missing values it is unclear whether the variance is computed on observed entries only or after some initialization; please clarify.","section":"Equation 14"},{"comment":"The two attention baselines 'Between-feature' and 'Between-sample' are from the authors' own prior work [40]; this should be explicitly acknowledged as a limitation of the independence of the comparison.","section":"Section 4.3"},{"comment":"The average-rank rows would be easier to interpret if accompanied by confidence intervals or error bars, since the raw standard deviations are large and the visual figures in the appendix do not display them.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's extensive benchmark and held-out evaluation are strengths, but the central novelty—contrastive learning—shows negligible effect in the ablation, and the superiority claim rests on average ranks without significance testing. The authors also use their own prior work as the closest baseline, and the GAIN/Diffputer training budgets are not matched to their recommended settings. I recommend major revision: the authors should add significance testing, retrain or justify the baselines, and recalibrate the claims to match the evidence. If the significance tests show no reliable difference, the paper could still be valuable as a benchmark study, but the novelty framing would need to be changed accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear X,\n\nThe short version: DeepIFSAC is a sensible combination of known components—feature-wise and sample-wise multi-head self-attention, CutMix corruption, and contrastive learning—for tabular imputation, and the empirical study is serious. But the headline claim that its average rank demonstrates superiority is not supported by the statistics actually reported.\n\nWhat is genuinely new: applying joint row-column attention to imputation. The building blocks all appear in the cited literature, but the specific combination for missing value imputation is, as far as I can tell, not previously published. The paper also does a broad, clean benchmark: 12 OpenML sets, three missing mechanisms, 10–90% missingness, held-out test folds, and a downstream EHR classification task. The ablations (Table 8) suggest CutMix helps relative to zero-filling, and contrastive learning adds a small consistent gain. Section 6.6 states real limitations—numerical features only, weak under ~200 samples, memory limits—which matches what the results show.\n\nThe soft spots are all in the strength of the comparative claim. In Table 2, DeepIFSAC's average rank beats the authors' own between-sample attention baseline by about 0.2–0.3, with standard deviations in the 2.5–3.5 range. No significance test or confidence interval is reported anywhere, so 'demonstrates its superiority' is overclaimed. The gap could easily be noise. This is fixable with a proper multi-dataset comparison (e.g., Friedman + Nemenyi or pairwise Wilcoxon). Second, the deep baselines are all run for 1,000 epochs; GAIN and Diffputer are commonly trained for 10,000. The authors acknowledge this but do not test sensitivity. Third, the CutMix corruption used during training is a surrogate for the missingness mechanisms; there is no diagnostic showing it behaves like MAR or MNAR, so the strong MNAR gains could partly reflect an artifact of the training corruption rather than a true mechanism-matched advantage.\n\nNone of this is disqualifying. The method is reasonable, the evaluation is otherwise careful, and the limitations are honestly stated. I'd send it to peer review, but with a request to add significance tests and soften the abstract to claim 'best mean rank in these experiments' rather than 'superiority.' For me, the paper is worth a reading-group discussion on how to interpret average-rank comparisons, but I wouldn't cite it for the SOTA claim as it stands.\n\nBest,","headline":"DeepIFSAC is a solid incremental study whose central superiority claim is not backed by significance tests; the claim should be softened.","tokens_in":30143,"tokens_out":3189,"would_cite":false,"duration_ms":34586,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A joint feature-sample attention imputer claims top rank among 11 missing-value methods.","keywords":["missing value imputation","tabular data","attention mechanism","contrastive learning","CutMix data augmentation","MNAR","electronic health records","NRMSE"],"falsifier":"Take a dataset with an MNAR mechanism whose missingness propensity can be estimated, and train two versions of DeepIFSAC: one with standard CutMix corruption and one with corruption masks drawn from the true propensity. If the true-propensity model does not achieve lower NRMSE on held-out missing entries, then the CutMix surrogate for non-random missingness is not doing the work the paper assigns it.","tokens_in":29083,"feed_emoji":"📊","tokens_out":7808,"duration_ms":73926,"temperature":0.7,"pith_summary":"Missing values in real-world tabular data are often neither few nor random, and standard imputation methods degrade as the missing rate climbs. DeepIFSAC is a deep imputation model that estimates a missing cell by attending both to other features in the same row and to similar samples in the same dataset, and it trains these two attention passes together with CutMix corruption and a contrastive loss. The paper's central claim is that this combined framework achieves the best average performance rank among eleven statistical, machine-learning, and deep baselines across twelve tabular data sets, for missing rates from 10% to 90% and for the three standard missingness mechanisms—MCAR (missing completely at random), MAR (missing at random), and MNAR (missing not at random)—with the clearest advantage under MNAR. A downstream test on real electronic health records reports the highest F1 for heart-failure classification when the features are imputed by DeepIFSAC. If the claim holds, the practical message is that between-sample attention and augmentation with real values can make deep imputation robust in exactly the high-rate, non-random regimes where classical methods struggle.","feed_headline":"An imputer that reads rows and columns beats 11 rivals on missing data","feed_subtitle":"Feature- and sample-attention with CutMix tops MICE, missForest, GAIN and diffusion, and boosts heart-failure F1.","key_machinery":"The central object is a two-stage multi-head self-attention pipeline over a per-feature embedding tensor. Stage one applies multi-head self-attention across features to capture column dependencies; stage two reshapes the embedding and applies multi-head self-attention across samples to capture row dependencies, after which an MLP head reconstructs the data values. Training corrupts mini-batches with CutMix, replacing a random subset of a sample's features with the corresponding values from another sample, and optimizes the sum of a reconstruction loss computed only on observed entries and an infoNCE contrastive loss that pulls corrupted and uncorrupted embeddings of the same sample together. The CutMix corruption, rather than zero-filling, is what lets the model learn to repair missing-like patterns from real value combinations, while contrastive learning sharpens the sample neighbourhoods that the between-sample attention then relies on.","core_discovery":"The paper's central claim is that a missing value, lying at the intersection of a row and a column, should be reconstructed by jointly learning between-feature attention and between-sample attention rather than by either alone. In the reported experiments, DeepIFSAC is compared with eleven baselines on twelve tabular data sets under three simulated missingness mechanisms and five missing rates, with methods ranked by normalized root mean squared error on held-out folds. The average rank places DeepIFSAC first overall for MCAR, MAR, and MNAR, and its advantage over the strongest alternatives is largest under MNAR, where missingness depends on the value itself and between-sample attention can supply information that between-feature attention cannot. On real electronic health records without ground truth, the imputed data support a heart-failure classifier with the highest mean F1 score among the compared methods. The paper additionally reports ablations showing that CutMix augmentation improves on zero-filling and that the contrastive loss improves on joint attention without it.","pith_inferences":["The paper leaves the CutMix-to-MNAR transfer unvalidated; a direct test would train on masks drawn from an estimated missingness propensity and compare NRMSE against CutMix-trained models, showing whether the surrogate assumption holds.","The dataset heterogeneity suggests the practical rule is conditional, not universal: attention-based imputation appears strongest with adequate sample sizes and comparable feature scales, while tree-based methods remain competitive on small or extreme-scale datasets; a formal decision rule could make this actionable.","Contrastive learning probably helps most when samples form meaningful clusters, so on weakly clustered datasets the between-sample attention could dilute rather than help; this is testable by varying clusterability.","Extending the reconstruction head to categorical features with a classifier loss is a natural next step, and the same attention weights could double as a diagnostic for which features and samples drive the imputation."],"forward_implications":["If the average-rank result holds, DeepIFSAC is the best default among the eleven tested methods for tabular data with missing rates from 10% to 90% across MCAR, MAR, and MNAR.","The MNAR result implies that row attention can partially compensate for the absence of feature dependencies, the regime where regression-based and generative baselines weaken.","The EHR result implies that lower imputation error translates into better predictive performance on real data where missing ground-truth values are unknowable.","The ablations imply that CutMix augmentation with real values and the contrastive objective are load-bearing parts of the method, not optional extras.","At high missing rates, between-sample attention is reported as the most resilient view, which supports attention-based imputation as the stronger deep-learning family for tabular missing values."],"supporting_citations":[{"why":"Supplies the CutMix augmentation that replaces masked features with real values from another sample, the paper's alternative to zero-filling.","marker":"[14]"},{"why":"Provides the row-attention and contrastive pre-training design for tabular data that DeepIFSAC adapts to imputation.","marker":"[30]"},{"why":"Reports that attention and contrastive learning outperform baselines on tabular classification, motivating their use for imputation.","marker":"[29]"},{"why":"Supplies the scaled dot-product multi-head attention blocks used in both feature and sample stages.","marker":"[33]"},{"why":"MICE is the leading statistical baseline whose average rank DeepIFSAC must beat across missing rates and types.","marker":"[6]"},{"why":"missForest is the strongest tree-based baseline and the most competitive classical method on hard datasets.","marker":"[7]"},{"why":"GAIN is the GAN-based deep baseline whose random-missingness assumptions the paper contrasts with its own.","marker":"[12]"},{"why":"Diffputer is the diffusion-based imputation baseline used to compare generative approaches with attention-based ones.","marker":"[39]"},{"why":"The authors' earlier attention-based imputation work provides the between-feature and between-sample components used as ablations.","marker":"[40]"}],"fun_headline_variants":["Joint attention to rows and columns tops 11 imputation rivals","Feature and sample attention imputer beats 11 baselines, excels on MNAR","CutMix and contrastive attention beat 11 imputers, best on non-random missingness","Attention to features and samples beats 11 imputers, excels with non-random gaps","Dual attention imputation beats 11 rivals, best for non-random missingness"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that CutMix corruption—replacing random features with values from another sample—is a faithful stand-in for real missingness, including MNAR where the chance of being missing depends on the missing value itself; if the two distributions diverge, the model could learn to repair CutMix patterns while producing biased imputations for true missing cells.","fun_headline_variants_meta":{"raw":{"variants":["Joint attention to rows and columns tops 11 imputation rivals","Feature and sample attention imputer beats 11 baselines, excels on MNAR","CutMix and contrastive attention beat 11 imputers, best on non-random missingness","Attention to features and samples beats 11 imputers, excels with non-random gaps","Dual attention imputation beats 11 rivals, best for non-random missingness"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000929,"raw_usage":{"total_tokens":3996,"prompt_tokens":981,"completion_tokens":3015,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":2909}},"tokens_in":597,"tokens_out":3015,"duration_ms":21351,"temperature":1.0,"reasoning_tokens":2909,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:50:30.251853+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a dataset with an MNAR mechanism whose missingness propensity can be estimated, and train two versions of DeepIFSAC: one with standard CutMix corruption and one with corruption masks drawn from the true propensity. If the true-propensity model does not achieve lower NRMSE on held-out missing entries, then the CutMix surrogate for non-random missingness is not doing the work the paper assigns it.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the CutMix augmentation that replaces masked features with real values from another sample, the paper's alternative to zero-filling."},{"cited_title":"Somepalli, A","cited_arxiv_id":null,"evidence_quote":"Provides the row-attention and contrastive pre-training design for tabular data that DeepIFSAC adapts to imputation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Reports that attention and contrastive learning outperform baselines on tabular classification, motivating their use for imputation."},{"cited_title":"Resche-Rigon, I","cited_arxiv_id":null,"evidence_quote":"MICE is the leading statistical baseline whose average rank DeepIFSAC must beat across missing rates and types."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"missForest is the strongest tree-based baseline and the most competitive classical method on hard datasets."},{"cited_title":"Zhang, L","cited_arxiv_id":null,"evidence_quote":"Diffputer is the diffusion-based imputation baseline used to compare generative approaches with attention-based ones."},{"cited_title":"Kowsar, S","cited_arxiv_id":null,"evidence_quote":"The authors' earlier attention-based imputation work provides the between-feature and between-sample components used as ablations."}],"review_version":1}