{"id":"f9944c16-06dd-46f1-b215-e71fd7a48d6c","arxiv_id":"2502.07181","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"Bar-image encoding with tailored image augmentation and ImageNet transfer learning achieves the best average F1 on six small UCI datasets, though the comparison is confounded by asymmetric augmentation and pretraining.","lead":"Tab2Visual turns each row of a spreadsheet into a bar-chart image and feeds the images to a pretrained image classifier. Its augmented version wins the average F1 comparison on the six smallest UCI datasets tested, but the win relies on advantages the baseline methods did not receive.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed superiority is not established: the winning Tab2Visual variant (A4) is selected on the same test folds, baselines receive no augmentation or transfer learning, and the average-F1 margin has no significance test.","rationale":"The reader correctly identifies the augmentation component as central, but the most load-bearing flaw is not whether elastic distortions preserve class semantics; it is that the experimental design cannot support the causal claim 'Tab2Visual outperforms other methods.' Three specific problems are decisive. First, the comparison is confounded: only Tab2Visual receives augmentation and ImageNet transfer learning, so the observed gains could come from having more effective training data and a strong pretrained feature extractor, not from the bar-image representation itself. Second, the best augmentation level A4 is highlighted after evaluating all levels A0-A4 on the same test folds, which is a form of test-set selection; without a nested validation loop, the reported 74.6% F1 is optimistically biased. Third, the paper reports no confidence intervals, error bars, or significance tests, and the most favorable metric (average F1) still leaves Tab2Visual-A4 only about 2-4 points above several baselines while being tied or slightly behind on average AUC. The paper itself admits in Section 4.1 that no single method consistently outperforms all others across datasets, which undercuts the abstract and conclusion. The label-preservation assumption flagged by the reader would matter if the paper claimed the augmented images are faithful synthetic samples, but the central claim is about classification performance; even label-corrupting augmentations could improve test accuracy through regularization. Therefore the load-bearing concern is the validity of the empirical comparison, not the semantic interpretation of the augmentation. A concrete nested-CV rerun with matched augmentation for baselines would settle whether the claimed advantage survives a fair test. Since the reported evidence does not support the strong claim as stated, the reader's REJECT verdict remains appropriate.","tokens_in":18571,"tokens_out":4377,"duration_ms":38809,"concrete_test":"Run a nested cross-validation protocol on the six small datasets: for each outer fold, tune K (and the augmentation hyperparameters) on the inner training folds only, then evaluate the chosen Tab2Visual variant on the outer test fold. In the same protocol, give the strongest baselines (CatBoost, MLP, TabPFN) an equivalent augmentation budget (e.g., SMOTE or noise-based augmentation at matched multiplicity) and identical Optuna tuning. If Tab2Visual's average F1 and AUC are no longer best, or if the difference falls within a bootstrap or paired-test confidence interval crossing zero, the claimed superiority is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that Tab2Visual outperforms all compared methods on smaller datasets rests on a comparison that does not isolate the method's contribution and selects its best variant after seeing the results. In Table 4 / Fig. 6, the only consistent win is average macro-F1 for Tab2Visual-A4 (74.6% vs 72.6% for CatBoost/MLP/TabPFN), while CatBoost ties or slightly beats it on average AUC (89.2% vs 88.7%). The advantage appears only after augmentation: Tab2Visual-A0 averages roughly 73.2% F1 and 86.6% AUC, which is not top among methods. Yet no baseline receives any form of augmentation or transfer learning, so the comparison conflates the visual encoding with the extra training data and ImageNet pretraining that only Tab2Visual receives. Moreover, the augmentation scale K is not chosen by validation: the paper reports A0-A4 for each small dataset and then highlights A4, and Figures 6 and 8 select the best-performing K on the same folds used for evaluation. No confidence intervals or significance tests are reported, so the 2-4% F1 margin over three baselines could easily be noise. The label-preservation question in Algorithm 2 is secondary: even if elastic and morphological transforms generate out-of-distribution images, they could still act as a regularizer; the load-bearing failure is that the evidence does not establish a fair, statistically supported superiority.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Tab2Visual, a method that encodes each tabular sample as an image of colored bars whose widths are proportional to normalized feature values, then trains or fine-tunes an ImageNet-pretrained EfficientNetV2 (or EfficientViT) on these images, with optional image augmentations based on elastic distortion and morphological operations. The authors evaluate the method on ten UCI datasets, comparing it against classical classifiers, tree ensembles, TabNet, and TabPFN, using 5-fold cross-validation with macro-F1 and AUC. The central claim, stated in the abstract and conclusions, is that Tab2Visual outperforms all compared methods on datasets with at most 1000 samples.","tokens_in":18767,"tokens_out":5133,"duration_ms":44697,"significance":"If the claim were substantiated, Tab2Visual would be a practically useful contribution: it is conceptually simple, enables transfer learning and augmentation for tabular data, and the paper includes informative ablations on transfer learning, backbone choice, and feature arrangement. The algorithmic descriptions are clear and the benchmark covers ten datasets and a broad set of baselines. However, the evidence presented does not establish the claimed superiority. The winning Tab2Visual variant receives both heavy data augmentation and ImageNet pretraining that no baseline receives, the augmentation scale is selected after seeing the test results, and no statistical significance is reported. As a result, the paper's main conclusion is currently unsupported.","major_comments":[{"comment":"The comparison is confounded. Tab2Visual uses ImageNet-pretrained EfficientNetV2-B0 and image augmentation, whereas none of the baselines receive transfer learning or any data augmentation. The reported advantage of Tab2Visual-A4 over CatBoost, MLP, and TabPFN on small datasets therefore cannot be attributed to the visual encoding; Table 5 shows that transfer learning alone yields a 7.5% average AUC gain on the small datasets. To support the central claim, the authors should compare against baselines that receive equivalent augmentation and transfer learning (e.g., SMOTE-augmented or otherwise pre-trained variants), or ablate Tab2Visual by removing augmentation and pretraining to isolate the contribution of the visual representation.","section":"Section 3.3 and Table 4 / Figure 6"},{"comment":"The augmentation scale K is selected post hoc. The paper reports results for A0 through A4 on each small dataset and then highlights A4 in the abstract and conclusion, but K is not chosen by a validation set or nested cross-validation; it is the best-performing variant on the same test folds used for evaluation. This selection-on-test procedure inflates the reported performance and should be replaced with an independent validation-based selection of K or a pre-specified augmentation level.","section":"Section 4.2 and Table 4"},{"comment":"No confidence intervals, standard deviations, or significance tests are reported for any method. The margins that drive the conclusion are small (e.g., 74.6% vs 72.6% average F1 in Figure 6), and on average AUC CatBoost beats Tab2Visual-A4 (89.2% vs 88.7%). Without repeated runs or statistical testing, the claimed superiority of Tab2Visual on small datasets is not established.","section":"Section 3.4, Table 4, and Figures 6-10"},{"comment":"The label-preserving property of the proposed augmentations is not validated. Elastic distortions and morphological dilation/erosion on bar-width images can alter feature values in ways that do not correspond to valid tabular samples; the claim that they generate 'synthetic samples that closely resemble the original data distribution' requires empirical support, for example by reconstructing feature values from augmented images and comparing their distributions. This issue is secondary to the confound above, but it matters because augmentation is the main source of the reported performance gain.","section":"Algorithm 2 and Section 2"}],"minor_comments":[{"comment":"Algorithm 1 line 7 contains a typo: 'h ←= H/r' should be 'h ← H/r'. In Section 2, 'hight' should be 'height'.","section":"Algorithm 1 and Section 2"},{"comment":"The column headers contain spacing artifacts ('HR T', 'SA T', 'JU', 'GL', 'PE') and the table is difficult to parse; consider a layout that separates F1 and AUC rows more clearly.","section":"Table 4"},{"comment":"The sentence 'data is first divided into non-overlapping training and testing partitions' is ambiguous in the context of 5-fold cross-validation; please specify how augmentation is applied within each fold's training partition.","section":"Section 3.4"},{"comment":"The paper reports AUC for multiclass datasets (Glass, Satellite) without specifying the averaging scheme; please clarify whether this is macro one-vs-rest AUC.","section":"Section 3.4"},{"comment":"Figure 1 is presented without a description of the OpenML query or the date of access; please provide these details for reproducibility.","section":"Figure 1"},{"comment":"The claim of a 'consistent improvement' from augmentation is based on averages over six datasets without any measure of variance; showing per-dataset trajectories or error bars would be more informative.","section":"Section 4.2 and Figure 10"}],"recommendation":"reject","confidential_remarks":"The core issue is experimental design rather than presentation. The main claim is not supported because the comparison does not give baselines access to augmentation or transfer learning, the augmentation level is chosen after seeing test results, and no statistical tests are reported. If the authors can rerun the comparison with equivalent baselines, select K by validation, and report confidence intervals, the manuscript might be reconsidered; in the current form, I would not recommend publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before reading: the paper is a solid, clearly written empirical study, and its main claim is overblown. Tab2Visual is a bar-image encoding of tabular rows, built on the authors' prior work, plus a tailored augmentation scheme (elastic distortion and morphology) and ImageNet-pretrained EfficientNetV2. The experiments are extensive: ten UCI datasets, 5-fold CV, Optuna hyperparameter tuning, and ablations over augmentation level, backbone, transfer learning, and feature arrangement. Those ablations are genuinely useful, especially Table 5 showing that transfer learning helps substantially on small datasets. The time analysis is also a nice touch, though Tab2Visual is the slowest method to train by a large margin.\n\nThe soft spot is the central claim. The abstract says Tab2Visual outperforms other methods on limited tabular data; the conclusion says it outperformed all compared methods on datasets ≤1000 samples. Looking at Table 4 and Figure 6, the only consistent win is average macro-F1 for the A4 variant: 74.6% vs 72.6% for CatBoost/MLP/TabPFN. CatBoost matches or slightly beats it on average AUC (89.2% vs 88.7%). That gap could easily be noise; no confidence intervals or significance tests are reported. More importantly, the comparison is confounded: Tab2Visual receives ImageNet pretraining and up to 4x data augmentation, while none of the baselines get either. And A4 is selected after seeing the results on the same folds used for evaluation, which is post-hoc selection, not validation. The label-preservation question about the morphological augmentations is real but secondary; even if they produce out-of-distribution images, they could still help as regularizers. The bigger issue is that the evidence doesn't establish a fair, statistically supported superiority.\n\nThe paper also builds directly on the authors' earlier encoding [29], and it says so, which is honest. The new bits are the augmentation set and the systematic comparison, and those are worth publishing in some form. But the conclusion needs to be toned down, the baselines need a fairer setup (e.g., augmented data for tree-based methods, or at least reporting per-fold variance), and the augmentation level should be chosen by validation, not after inspecting test-fold results.\n\nWho is this for? Anyone working on tabular-to-image conversion or small-sample tabular classification will find the ablations and the augmentation details useful. It deserves a serious referee: the method is concrete, the write-up is clear, and the empirical scope is decent. But a reviewer should push hard on the evaluation protocol and the claims.","headline":"A clean empirical study of a tabular-to-image method with useful ablations, but the headline claim of superiority on small datasets is not supported by the evidence as presented.","tokens_in":19430,"tokens_out":1798,"would_cite":false,"duration_ms":17993,"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":"Tab2Visual claims that converting tabular rows into proportional-width bar images lets a pretrained vision model beat all compared classifiers, including TabNet and TabPFN, on small datasets.","keywords":["tabular data","limited data","deep learning","visual representation","data augmentation","transfer learning","image classification","small datasets"],"falsifier":"A reader could apply Algorithm 2 to a set of Tab2Visual images, measure the bar widths back into feature vectors, and check whether the recovered values stay within the original class's observed feature ranges; if distorted images routinely turn near-zero features into visible bars or invert the ranking of two features, the augmentation is creating label noise rather than valid new samples.","tokens_in":18250,"feed_emoji":"📊","tokens_out":8489,"duration_ms":70630,"temperature":0.7,"pith_summary":"Tab2Visual addresses a practical problem: deep learning models routinely lose to tree ensembles on small tabular datasets, yet many real-world datasets, especially in healthcare, have only a few hundred rows. The paper's proposal is to convert each tabular row into an image in which every feature is a colored vertical bar whose width is proportional to the normalized feature value, then fine-tune a pretrained EfficientNetV2 on those images. To stretch limited data, it adds elastic distortions and morphological operations designed to alter bar boundaries and widths while keeping the sample's class. Across the small datasets in the study, the augmented Tab2Visual variants achieved the best average F1-score and best average rank among all compared classifiers, including TabNet and TabPFN. If this holds, deep learning becomes usable on small tabular data without specialized tabular architectures.","feed_headline":"Bar-chart images beat specialized deep nets on small tables","feed_subtitle":"The method draws rows as proportional bars, then fine-tunes a pretrained vision model.","key_machinery":"The central object is the Tab2Visual image: a canvas divided into rows and columns, one bar per feature, with bar width equal to the normalized feature value times a maximum width, and each feature assigned a distinct color. The mechanism is a two-step pipeline: first, tabular rows are normalized, one-hot encoded where needed, and drawn as bars; second, each image is optionally augmented by elastic distortion followed by random dilation, erosion, opening, or closing using randomly sized structuring elements. The augmented images then fine-tune a pretrained vision model whose feature weights are frozen except for the final classification layer. The bar layout lets a CNN read relative feature magnitudes as spatial extent.","core_discovery":"The central claim is that the visual format is the enabler: once tabular features are laid out as proportional-width bars in a compact image, a vision model pretrained on natural images can transfer its feature knowledge to the tabular task, and image-style augmentation can generate new training samples that stay close to the original distribution. The paper reports that on datasets with no more than about a thousand samples, Tab2Visual outperformed every compared method, with the version that augments each image four times posting the highest average F1-score and the best average rank, while CatBoost tied it on average AUC. On larger datasets, tree ensembles remained the best classifiers, so the claimed advantage is specifically for the limited-data regime.","pith_inferences":["The reported gains likely combine pretrained feature reuse, strong regularization, and augmentation, and the paper does not isolate these effects; a controlled comparison against the same backbone fine-tuned on the original tabular features would separate the visual encoding from the transfer-learning benefit.","The row-arrangement finding suggests the CNN is reading bars as a length-ranking code rather than as a spatial pattern; this predicts that deterministic orderings by feature correlation would behave differently from random orderings, which the paper leaves untested.","The method's dependence on pretrained natural-image features could be probed by comparing against a randomly initialized small CNN on larger augmented sets; the paper's from-scratch results on small data are much weaker, indicating transfer is doing real work."],"forward_implications":["Small clinical and survey datasets with only a few hundred rows can be tackled with off-the-shelf vision models by rendering features as bar images, without building a new tabular architecture.","Transfer learning from natural-image models becomes a viable strategy for tabular data, at least when the input is a compact bar layout.","Image augmentation operations such as elastic distortion and morphological transformations can serve as tabular augmentation, expanding effective training size without collecting new records.","On larger tabular datasets, the method does not displace tree ensembles; its claimed advantage is specific to the limited-data regime.","Feature arrangement matters: one or two rows of bars outperforms taller layouts, so compact bar layouts should be the default configuration for this method."],"supporting_citations":[{"why":"Introduces TabPFN, the transformer-based small-tabular classifier that serves as a key deep-learning baseline in the small-data comparison.","marker":"[12]"},{"why":"Extends TabPFN's small-data applicability and is cited as the basis for limiting TabPFN results to datasets under 1,000 samples.","marker":"[13]"},{"why":"TabNet, a transformer-based tabular architecture used as a state-of-the-art deep-learning baseline across all datasets.","marker":"[10]"},{"why":"EfficientNetV2, the pretrained CNN backbone that provides the transfer-learning mechanism central to Tab2Visual.","marker":"[41]"},{"why":"The image-augmentation library used to implement the elastic-distortion and morphological operations in Algorithm 2.","marker":"[35]"},{"why":"The survey that documents tree ensembles' dominance and the limited-data challenge that Tab2Visual targets.","marker":"[7]"},{"why":"DeepInsight, the prior tabular-to-image method whose feature-rich-data limitation Tab2Visual claims to overcome.","marker":"[16]"},{"why":"The authors' earlier prostate-cancer study that motivated the transfer-learning approach and provided the preliminary idea.","marker":"[29]"}],"fun_headline_variants":["Visual tables beat deep nets on small datasets","Draw tables as bars, let vision AI take over","Bars turn tiny tables into transfer learning wins","Tab2Visual: images make small tabular data learnable","Pretrained vision models ace small tabular tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that elastic distortion and morphological operations leave the class meaning of each bar image intact, so each augmented picture is a valid new sample of the same class.","fun_headline_variants_meta":{"raw":{"variants":["Visual tables beat deep nets on small datasets","Draw tables as bars, let vision AI take over","Bars turn tiny tables into transfer learning wins","Tab2Visual: images make small tabular data learnable","Pretrained vision models ace small tabular tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000161,"raw_usage":{"total_tokens":1168,"prompt_tokens":808,"completion_tokens":360,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":424,"completion_tokens_details":{"reasoning_tokens":286}},"tokens_in":424,"tokens_out":360,"duration_ms":3998,"temperature":1.0,"reasoning_tokens":286,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T13:32:55.194357+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could apply Algorithm 2 to a set of Tab2Visual images, measure the bar widths back into feature vectors, and check whether the recovered values stay within the original class's observed feature ranges; if distorted images routinely turn near-zero features into visible bars or invert the ranking of two features, the augmentation is creating label noise rather than valid new samples.","supporting_citations":[{"cited_title":"Hollmann, S","cited_arxiv_id":null,"evidence_quote":"Introduces TabPFN, the transformer-based small-tabular classifier that serves as a key deep-learning baseline in the small-data comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"TabNet, a transformer-based tabular architecture used as a state-of-the-art deep-learning baseline across all datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"EfficientNetV2, the pretrained CNN backbone that provides the transfer-learning mechanism central to Tab2Visual."},{"cited_title":"Buslaev, V","cited_arxiv_id":null,"evidence_quote":"The image-augmentation library used to implement the elastic-distortion and morphological operations in Algorithm 2."},{"cited_title":"Borisov, T","cited_arxiv_id":null,"evidence_quote":"The survey that documents tree ensembles' dominance and the limited-data challenge that Tab2Visual targets."},{"cited_title":"Sharma, E","cited_arxiv_id":null,"evidence_quote":"DeepInsight, the prior tabular-to-image method whose feature-rich-data limitation Tab2Visual claims to overcome."},{"cited_title":"El-Melegy, A","cited_arxiv_id":null,"evidence_quote":"The authors' earlier prostate-cancer study that motivated the transfer-learning approach and provided the preliminary idea."}],"review_version":1}