{"id":"b9ac193d-9039-4f6d-a0fc-59b51120febc","arxiv_id":"2412.06265","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Table2Image pairs each tabular row with a random same-class MNIST or Fashion-MNIST image, trains an autoencoder plus small CNN, and reports competitive classification with a VIF initialization and a dual SHAP explanation layer.","lead":"This paper turns spreadsheet rows into small synthetic images borrowed from MNIST and Fashion-MNIST, then trains a small convolutional network to classify the generated images. It reports accuracy close to gradient-boosted trees and adds a collinearity-based initialization plus a dual SHAP explanation method, but the gains lack error bars and the image mapping injects class labels into the training target.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The random same-class image mapping injects label information into the reconstruction target, so the CNN may be classifying class-mean images rather than tabular features; the reported gains and DualSHAP explanations are then not about the tabular data.","rationale":"The central claim is that transforming tabular data into realistic images enables competitive classification and interpretability. For this to be true, the learned image representation must carry information from the tabular feature vector x, not merely from the image class labels. The architecture in Sec. 3.1 makes the second possibility almost unavoidable: the reconstruction target M(x) is a random image of the true class, so the optimal reconstructor is the class-mean image. The paper's own Appendix H confirms this behavior. With such templates, the CNN classifier is essentially recognizing FashionMNIST/MNIST class means, a trivially separable task. The reported performance then does not demonstrate a tabular representation, and DualSHAP attributes importance to a signal whose origin is the image label rather than the tabular features. This is the most load-bearing concern because it attacks both the performance claim and the interpretability claim at once. The reader's weakest assumption pointed to the same mechanism; my analysis sharpens it by noting the reconstruction-minimizer is the class-conditional mean and that the paper explicitly observes averaged outputs. A label-shuffle control experiment would settle whether the method has any tabular-feature content. Since the reader already conditioned acceptance on exactly this kind of ablation, the verdict remains CONDITIONAL (UNCHANGED).","tokens_in":17013,"tokens_out":5041,"duration_ms":49786,"concrete_test":"Run a label-shuffle control on a representative subset of OpenML-CC18 (e.g., 10 datasets spanning 2-10 classes): randomly permute the class labels of the target images so each x is paired with an image from a random, possibly different class, while keeping everything else fixed. If accuracy drops to near chance, the original performance is due to the injected class signal in the reconstruction target. Additionally, train a variant whose reconstruction target is the class-mean image for the true y (no random sampling); if its accuracy matches Table 1, the random mapping and autoencoder are not providing any tabular-feature information.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The mapping schema (Sec. 3.1) pairs each tabular instance x with a randomly selected image ix of the same label y. The reconstruction loss L_recon = ||AE(x,r) - M(x)||^2 drives the decoder to reproduce ix from inputs (x, r). Because ix is drawn uniformly from the class-y image pool and is otherwise independent of x and r, the minimum expected reconstruction error is achieved by outputting the class-conditional mean image E[ix | y]. Indeed, Appendix H reports that the autoencoder 'tends to generate averaged images that represent the typical images within each class.' Thus AE(x,r) is approximately a per-class template, and the downstream two-layer CNN classifies these templates by their visual class. The only tabular information used is the inferred label y; the feature vectors themselves contribute no additional signal beyond what a simple label classifier would provide. Consequently, the headline accuracy and AUC numbers in Table 1 (Table2Image-VIF Avg ACC 0.8787 vs XGBoost 0.8675) do not establish that the method learns a useful tabular representation. DualSHAP (Sec. 3.3) is then explaining an image-class template signal rather than feature-level relationships in the original table. The claimed interpretability and classification advantages are therefore not supported by the current experiments.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Table2Image, a framework that maps tabular instances to randomly selected images of the same class, uses an autoencoder to reconstruct those images from tabular features plus noise, and then trains a two-layer CNN on the reconstructions for classification. A VIF-based weight initialization variant (Table2Image-VIF) is introduced to address multicollinearity, and DualSHAP is proposed as an interpretability framework that aligns SHAP values from the tabular and image branches. Experiments on OpenML-CC18 and TabZilla report competitive accuracy and AUC relative to GBDTs and deep tabular baselines, and the authors claim interpretability advantages.","tokens_in":17261,"tokens_out":4508,"duration_ms":39465,"significance":"If the label-injection confound were absent, the framework would offer a lightweight deep-learning alternative to GBDTs with a dual-view interpretability mechanism, and the public code release is a strength. However, the current experimental design cannot separate the contribution of tabular features from the class signal injected by the image-mapping schema, and the interpretability evaluation is circular. With appropriate controls, the method could interest the tabular deep learning community, but the present evidence does not support the headline claims.","major_comments":[{"comment":"The reconstruction target M(x) is, by construction, a randomly selected image of the same class y as x, and is otherwise independent of x and r. Under L_recon = ||AE(x,r) - M(x)||^2, the optimal reconstruction is the class-conditional mean image E[I|y], and the authors' own Appendix H states that the autoencoder 'tends to generate averaged images that represent the typical images within each class.' Thus the downstream CNN is effectively classifying per-class templates, and the only tabular information that can influence the classifier is the inferred label y, not the feature vector x. Consequently, the accuracy and AUC numbers in Table 1 (e.g., Table2Image-VIF Avg ACC 0.8787 vs XGBoost 0.8675) do not establish that the method learns a useful tabular representation. The authors should add a control experiment in which image targets are assigned independently of the label (or features are permuted within each class) and show that the classifier's performance collapses to chance if the tabular features carry no signal.","section":"Section 3.1, Appendix H"},{"comment":"The reported gains over GBDTs are small (e.g., Avg ACC 0.8787 vs 0.8675, Avg AUC 0.9219 vs 0.8758 for XGBoost), yet no standard deviations, confidence intervals, or significance tests are reported, despite Appendix G stating that each experiment is repeated three times. Without measures of variance, the 'best performance' and 'wins' claims are not statistically grounded. This is especially important given the label-injection confound described above.","section":"Section 4.1, Table 1"},{"comment":"DualSHAP is defined by minimizing L_DualSHAP = L_MSE + L_KLD + L_MMD between P and Q, and the paper then presents the low values of these losses (Table 5) as evidence of interpretability. This is circular: the objective is being minimized, so its attainment is not external validation. The paper also derives the P=Q equality from Bayes' theorem while explicitly treating X, I, and F 'as outputs of the model or mechanism, rather than as probability distributions,' which leaves the normative basis of the equality unclear. The authors need an external validation of the explanations (e.g., faithfulness metrics, comparison with true feature relevance on synthetic data, or ablation of features with low importance).","section":"Section 3.3, Appendix A, Appendix C"},{"comment":"The claimed benefit of the VIF initialization is not demonstrated. Table 6 shows that Table2Image-VIF improves Avg ACC from 0.8766 to 0.8787 relative to Table2Image, a marginal difference, and no analysis is provided on datasets with known multicollinearity or on how the initialization changes learned weights. The paper should evaluate the method on a suite of collinear datasets and compare against standard initialization to support the claim of improved robustness.","section":"Section 3.2, Appendix D, Table 6"}],"minor_comments":[{"comment":"The statement that Table2Image 'does not undergo any hyperparameter tuning' is imprecise: the MLP expansion dimension (N+4), the noise size, and the choice of FashionMNIST/MNIST mapping are architectural hyperparameters, even if not tuned per dataset. The claim should be qualified.","section":"Appendix G"},{"comment":"The dimension handling in the DualSHAP equations (e.g., S × φ_img / X_recon) is not clear from the main text; Appendix B explains pixel unshuffle, but the exact shapes of S, T, and the products are never specified. A concrete tensor-shape diagram would improve reproducibility.","section":"Section 3.3"},{"comment":"The standard deviations of feature importance scores for class 9 are substantially larger than for other classes (e.g., P: 2.1893 vs typically ~0.3), but this is not discussed. The authors should explain this outlier.","section":"Appendix C, Table 4"},{"comment":"The term 'realistic and diverse image representations' is overstated given that Appendix H shows the autoencoder outputs class-typical averaged images rather than instance-specific realistic images; consider rewording to 'class-typical prototypes'.","section":"Abstract / Section 1"}],"recommendation":"reject","confidential_remarks":"The label-injection confound is a fundamental flaw in the experimental design: the image targets carry the class label independently of the tabular features, so the reported classification performance and the DualSHAP explanations cannot be attributed to the tabular data. The circularity of the interpretability evaluation and the lack of statistical significance testing compound this problem. I recommend rejection of the current version, though the authors could potentially address the confound with additional controlled experiments and resubmit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Candid take: Table2Image is a clean engineering recipe - random class-conditioned mapping to MNIST/Fashion-MNIST, an MLP autoencoder that produces class-typical images, a two-layer CNN, and a VIF-based initialization. The code is public, the benchmark sweep is broad, and the comparison against DeepInsight, IGTD, and HACNet is fair and informative. The VIF initialization is a reasonable, testable idea. The paper deserves credit for being concrete and reproducible.\n\nThe problem is the load-bearing claim: that a CNN on these generated images learns about tabular features. It probably doesn't, at least not in the way claimed. Each tabular instance is paired with a random same-class image, and the reconstruction loss trains the autoencoder to reproduce that image. As the authors themselves note in Appendix H, the autoencoder tends to output per-class averaged images. That means the CNN is essentially classifying class-mean image templates. The only tabular signal that survives is a class label indirectly inferred through the reconstruction target. The reported accuracy deltas over XGBoost are small (0.8787 vs 0.8675), and there are no error bars or significance tests. Without an ablation that removes label information from the reconstruction target - e.g., mapping to images independent of the class - the experiments cannot separate learning from tabular structure and learning from the injected image class.\n\nThe interpretability claim has the same circularity. DualSHAP minimizes a distributional discrepancy between P and Q, then presents the low discrepancy as evidence of interpretability. Low MSE/KLD/MMD means the optimization objective was achieved; it does not mean the feature importances match ground truth or human judgment. The shuffle-based stability test is weak because it also uses the same self-referential pipeline.\n\nThe limitations section is honest about regression and the Gaussian assumption, but it does not flag the class-injection problem. This is a serious omission, not a minor one.\n\nSo: the paper is a solid engineering contribution and deserves a serious referee - the idea is interesting and the fix (a proper control mapping, variance reporting, and validation against known feature importance) is plausible. But as it stands, the headline accuracy and interpretability claims are not supported. I would not cite it yet. I would bring it to a reading group to discuss what counts as learning from tabular data in this pipeline.","headline":"The pipeline is coherent and the code is real, but the label-conditioned image mapping lets the CNN read class-mean images rather than tabular features, so the headline accuracy and interpretability claims are not supported as reported.","tokens_in":738,"tokens_out":1979,"would_cite":false,"duration_ms":28545,"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":"This paper claims that converting each tabular row into a realistic MNIST-style image lets a small CNN match or beat gradient-boosted trees, with a dual SHAP mechanism for interpretability.","keywords":["tabular data classification","tabular-to-image transformation","autoencoder","CNN classifier","variance inflation factor initialization","DualSHAP","OpenML-CC18 benchmark","TabZilla benchmark"],"falsifier":"Pair each training row with a fixed image per class (or with images from the wrong class) and retrain; if accuracy stays roughly the same, or if the autoencoder output is nearly unchanged when the tabular input is zeroed out, then the label-selection signal, not the table's features, is carrying the result.","tokens_in":16731,"feed_emoji":"🖼️","tokens_out":10014,"duration_ms":89171,"temperature":0.7,"pith_summary":"This paper tries to show that gradient-boosted decision trees are not the only practical way to classify tabular data: a table row can be turned into a realistic MNIST/Fashion-MNIST-style image, and a small CNN can read that image and match or beat XGBoost. The transformation is not a fixed pixel layout but an autoencoder that takes the row plus random noise and reconstructs a randomly chosen same-class image, so the row's features and the image class both shape the output. A variance-inflation-factor initialization is added to dampen correlated features, and a dual SHAP framework derives feature importance from both the table and the image. If the claims hold, tabular deep learning becomes lighter and more interpretable, and tabular data can enter multimodal pipelines that assume image inputs.","feed_headline":"Spreadsheet rows turned into MNIST-style images outrank boosted trees","feed_subtitle":"A small autoencoder plus CNN tops XGBoost on 67 datasets while explaining decisions via both table and image.","key_machinery":"The load-bearing machinery is the label-conditioned random image mapping together with the autoencoder $AE(x,r)$: each row $x$ is paired with a random image $M(x)$ whose class equals the row's label, and the reconstruction loss $\\|AE(x,r)-M(x)\\|^2$ forces the generated image to depend on both $x$ and $r$; the CNN then classifies $AE(x,r)$. The VIF initialization sets first-layer weights to $w_{ij}=1/\\mathrm{VIF}_i$, so features with high collinearity enter with smaller initial influence, and the VIF embedding is concatenated with the ordinary embedding before entering the autoencoder. DualSHAP ties the two modalities together by minimizing MSE, KL divergence, and MMD between rescaled tabular SHAP values and image SHAP values, so the final importance scores are derived from both views.","core_discovery":"The paper's central claim is that a tabular row can be classified by first converting it into a realistic 28x28 image through an autoencoder conditioned jointly on the row and on random noise, where the target image is a randomly drawn FashionMNIST or MNIST image whose class matches the row's label. The reconstruction loss trains the autoencoder to blend row information with the chosen image; a two-layer CNN then classifies the generated image. The authors report that with this design, Table2Image and its VIF-initialized variant score higher average accuracy and AUC than XGBoost, LightGBM, CatBoost, and a set of recent deep tabular models on 67 OpenML-CC18 datasets and 34 TabZilla datasets. The VIF variant weights the first layer by the inverse variance inflation factor of each feature and concatenates that embedding, which they argue stabilizes learning under multicollinearity. The same image pathway also supports DualSHAP, an interpretability method that aligns SHAP values computed on the original table with Deep SHAP values computed on the generated image.","pith_inferences":["The mapping itself injects the label into the autoencoder's target image, so the accuracy and explanation results could partly reflect class-prior reconstruction; a reader could test this by using fixed or wrong-class images and measuring how much accuracy changes.","The VIF initialization is stated in terms of a generic first-layer weight scaling, so it could be applied to ordinary MLP or transformer tabular models even without the image transformation.","Because the generated images are class-typical averages, they could serve as a visual diagnostic for label leakage or class overlap, a use the paper does not develop."],"forward_implications":["On the reported benchmarks, a CNN reading generated images matches or beats XGBoost, LightGBM, CatBoost, and recent deep tabular models, so tabular data does not obviously require tree-based models or large transformers.","The VIF-initialized variant ranks first in average accuracy on both OpenML-CC18 and TabZilla, indicating that explicitly down-weighting collinear features helps learning stability.","Multi-image random mapping outperforms the single-image mapping used by HACNet in the paper's three-class experiments, so the diversity of the image pool is part of the method's behavior.","DualSHAP gives per-row feature importance from both the table and the generated image, with reported stability under column shuffling, so interpretability is available without switching back to trees."],"supporting_citations":[{"why":"OpenML-CC18 benchmark suite providing the 67 datasets used for the headline comparison.","marker":"(Bischl et al., 2021)"},{"why":"TabZilla benchmark suite providing the second set of 34 datasets used for evaluation.","marker":"(McElfresh et al., 2024)"},{"why":"FashionMNIST image pool used to map tabular classes 0-9 to realistic images.","marker":"(Xiao et al., 2017)"},{"why":"MNIST image pool used to extend the mapping to datasets with up to 20 classes.","marker":"(Deng, 2012)"},{"why":"Presents HACNet, the single-image-per-class table-to-image baseline that Table2Image compares against.","marker":"(Matsuda et al., 2024)"},{"why":"Presents DeepInsight, an earlier tabular-to-image method used as a comparison baseline.","marker":"(Sharma et al., 2019)"},{"why":"Presents IGTD, another tabular-to-image conversion method used in the comparison experiments.","marker":"(Zhu et al., 2021)"},{"why":"Provides SHAP, the feature-attribution mechanism that DualSHAP builds on for both tabular and image inputs.","marker":"(Lundberg & Lee, 2017)"},{"why":"Provides the maximum mean discrepancy used in the DualSHAP objective to align the two importance distributions.","marker":"(Gretton et al., 2012)"},{"why":"Supplies the XGBoost baseline that the proposed method is compared against in the main benchmark tables.","marker":"(Chen & Guestrin, 2016)"}],"fun_headline_variants":["Table rows become MNIST images, beat XGBoost on 67 datasets","Autoencoder draws row as FashionMNIST, CNN classifies better","VIF-initialized autoencoder images outrank boosted trees","Turn tabular rows into realistic images for interpretable AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the autoencoder's reconstructions actually encode the row's feature values, rather than just reproducing the class information that was already used to pick each random image; if the image class does most of the work, the reported accuracy and the DualSHAP explanations would not be about the tabular data.","fun_headline_variants_meta":{"raw":{"variants":["Table rows become MNIST images, beat XGBoost on 67 datasets","Autoencoder draws row as FashionMNIST, CNN classifies better","VIF-initialized autoencoder images outrank boosted trees","Turn tabular rows into realistic images for interpretable AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000705,"raw_usage":{"total_tokens":3166,"prompt_tokens":921,"completion_tokens":2245,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":2179}},"tokens_in":537,"tokens_out":2245,"duration_ms":14704,"temperature":1.0,"reasoning_tokens":2179,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:52:10.231449+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Pair each training row with a fixed image per class (or with images from the wrong class) and retrain; if accuracy stays roughly the same, or if the autoencoder output is nearly unchanged when the tabular input is zeroed out, then the label-selection signal, not the table's features, is carrying the result.","supporting_citations":[{"cited_title":"G., van Rijn, J","cited_arxiv_id":null,"evidence_quote":"OpenML-CC18 benchmark suite providing the 67 datasets used for the headline comparison."},{"cited_title":"Hacnet: End-to-end learning of interpretable table-to-image converter and convolutional neural network","cited_arxiv_id":null,"evidence_quote":"Presents HACNet, the single-image-per-class table-to-image baseline that Table2Image compares against."},{"cited_title":"A., and Tsunoda, T","cited_arxiv_id":null,"evidence_quote":"Presents DeepInsight, an earlier tabular-to-image method used as a comparison baseline."},{"cited_title":"A., Doroshow, J","cited_arxiv_id":null,"evidence_quote":"Presents IGTD, another tabular-to-image conversion method used in the comparison experiments."}],"review_version":1}