{"id":"4ae6b302-b83e-4f7c-bce8-8bf017834bfb","arxiv_id":"2608.07132","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Converting tabular credit data into images with target-based feature allocation lets a 2D CNN match or beat XGBoost on large credit scoring benchmarks.","lead":"This paper describes a way to turn rows of credit data into small grayscale images, arranging the features so that related ones sit close together. A standard image-classifying neural network trained on these images then predicts loan defaults, and the authors report it beats XGBoost on two large credit datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The state-of-the-art claim is not capacity-controlled: ConvNeXt's gains over the 1D CNN/MLP baselines on HC/FM could come from model scale, not from Tabular Image; Table 7 controls capacity only for image-to-image comparisons, not for the headline claim.","rationale":"The paper has a clear method, a leakage-conscious WOE/IV protocol, and a same-architecture comparison against One-hot and DeepInsight in Table 7, which is real evidence that the encoding choice matters when architecture is fixed. The missing control is a capacity-matched tabular deep model; without it, the central state-of-the-art claim is underdetermined. This is exactly the reader's weakest assumption, and the proposed test would settle it. The reader's conditional verdict remains appropriate: the concern is addressable and does not by itself invalidate the method or the transformation comparison, but the headline claim should not be accepted as stated without the capacity control. No code release and unspecified preprocessing details are secondary reproducibility concerns, not the load-bearing issue.","tokens_in":27619,"tokens_out":4994,"duration_ms":46481,"concrete_test":"Run a capacity-matched control on HC and FM: train (a) a wide MLP (e.g., 4 hidden layers of 1024 units with dropout/batch-norm) and (b) a 1D version of ConvNeXt/ResNet on the same WOE-transformed, z-scored, oversampled tabular data, using the same five-fold CV and early stopping. If either matches ConvNeXt+Tabular Image within the paper's ROPE (0.5% AUC/H-measure, 0.005 KS), the state-of-the-art claim is not attributable to the image transformation; if ConvNeXt+Tabular Image still wins by the same margins, the concern fails and the conditional can be lifted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that Tabular Image lets a 2D CNN achieve state-of-the-art credit scoring—requires that the image representation, not the choice of a very large modern CNN, drives the improvement. The evidence in Tables 3-5 compares ConvNeXt on 32x32 Tabular Images with a LeNet-5-like 1D CNN and a 5-layer MLP capped at 100 hidden units. These baselines have orders of magnitude fewer parameters and no modern training tricks, so the large AUC/H-measure/KS margins on HC and FM are exactly what one would expect from model capacity and optimization even without any representational benefit from images. Table 7 does address this for the transformation comparison: One-hot and DeepInsight images fed to the same ConvNeXt are worse than Tabular Image, showing the encoding matters among image methods. But it does not answer whether the same ConvNeXt-scale capacity applied directly to tabular features (e.g., a wide modern MLP or 1D CNN on WOE-transformed data) would match or beat the reported numbers. Because the abstract and Section 5.2 make the state-of-the-art claim against XGBoost/GBDT and neural baselines, this missing control is load-bearing. A second-order issue is that the preprocessing details (oversampling ratio, WOE computation before/after oversampling, feature-arrangement computed per fold) are not fully specified, which limits reproducibility, but the capacity confound is the decisive one.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Tabular Image, a supervised preprocessing method that converts a tabular credit-scoring row into a 32x32 grayscale image. Categorical features are replaced by weight-of-evidence values, numerical features are discretized to compute information values, pixels are allocated to features in proportion to IV, and features are arranged in the image by maximizing local Spearman correlations. A ConvNeXt 2D CNN is then trained on these images. The method is evaluated on three credit datasets (Taiwan Credit, Home Credit, Fannie Mae) with five-fold cross-validation on AUC, H-measure, and KS, compared with nine tabular-data baselines and two alternative image transformations (One-hot and DeepInsight), supplemented by Bayesian correlated t-tests and robustness checks on image size, feature arrangement, and random oversampling.","tokens_in":27949,"tokens_out":3117,"duration_ms":30837,"significance":"The paper contains several genuinely useful elements: WOE and IV are computed on the training fold only, so the evaluation protocol avoids the most obvious form of test leakage; the comparison with alternative image encodings under the same ConvNeXt architecture in Table 7 is informative; the Bayesian correlated t-tests are a stronger form of comparison than simple mean-rank reporting; and the robustness checks on image size and feature arrangement address relevant practical questions. If the central claim were supported by capacity-controlled evidence, the method would be a practical contribution to applying 2D CNNs to credit data. However, the headline state-of-the-art claim rests on a comparison in which the proposed model is a large modern ConvNeXt while the neural baselines are a small LeNet-style 1D CNN and a shallow MLP with at most 100 hidden units. Because the image representation and model capacity are confounded, the reported large gains on Home Credit and Fannie Mae cannot currently be attributed to Tabular Image itself.","major_comments":[{"comment":"The central state-of-the-art claim is not capacity-controlled. The proposed method uses a deep modern ConvNeXt, while the neural tabular baselines are a LeNet-5-like 1D CNN and a 5-layer MLP with hidden units capped at 100. On the HC and FM datasets the reported margins over these baselines are exactly what one would expect from model scale and modern optimization even if the image representation contributed nothing. Table 7 controls capacity for image-to-image comparisons (all encodings are fed to the same ConvNeXt), but it does not answer whether a comparably large modern MLP or 1D CNN trained on the same WOE-transformed tabular features would match or exceed the reported numbers. Add that control; if a wide modern MLP or deep 1D CNN on tabular input reaches the same AUC/H/KS, the claim that the transformation itself drives the improvement collapses.","section":"§4.2, §5.2, Tables 3–5"},{"comment":"The abstract claims state-of-the-art predictive performance, but the TC results do not support that wording. In Table 3, ConvNeXt with Tabular Image achieves AUC 77.98% versus XGBoost 77.99%, H-measure 28.88% versus 29.14%, and KS 0.4281 versus 0.4307, and the Bayesian analysis in Section 5.2.2 reports practical equivalence with GBDT and XGBoost on TC. The claim should be restricted to the large datasets (HC and FM) or otherwise tempered. As written, the abstract overstates the evidence.","section":"Abstract, §5.2.1, Table 3"},{"comment":"The preprocessing details needed for reproducibility are incomplete. The paper states that random oversampling is applied to the training set but never gives the oversampling ratio; the interval is not part of the hyperparameter grid in Table 2, even though the oversampling ratio is a free parameter of the pipeline. It is also not stated explicitly whether WOE and IV are computed before or after oversampling, despite Section 5.6 comparing IV ranks before and after balancing. Please specify the exact ratio, the point in the pipeline at which WOE/IV are calculated, the number of quantile bins for numerical discretization, and the exact image size and block size choices for every dataset. If possible, release code or pseudocode with these settings.","section":"§4.1, §4.3, §5.6"},{"comment":"Algorithm 1 lacks a precise mathematical statement of the optimization problem. Line 12 says 'Apply S to obtain a feature names vector that can maximise the sum of the Spearman correlation coefficient in the current block', but the constraint structure—how the unknown positions in the block are filled, how already assigned features interact with the fixed first row, and how the per-feature pixel budgets N_pi enter as constraints—is not specified. Since the method is essentially an integer program, write out the objective and constraints in equations; otherwise the reported feature arrangement is not exactly reproducible.","section":"§3.3, Algorithm 1"}],"minor_comments":[{"comment":"The framework figure is not self-explanatory; each transformation stage (WOE, IV calculation, pixel allocation, arrangement, z-score normalization, image construction) should be labeled on the figure and referenced in the text.","section":"Fig. 1"},{"comment":"Table 2 reports search spaces but not the selected hyperparameter values for each model. Reporting the chosen values is important for reproducibility, especially since the text says the grid search yields optimal settings.","section":"Table 2"},{"comment":"Several references are malformed or incomplete, e.g., 'Anna Montoya i, KirillOdintsov MK (2018)' for the Home Credit Kaggle competition, and some dataset URLs lack access dates. Please provide a complete reference list with repository identifiers.","section":"References"},{"comment":"The definition of the region of practical equivalence is reported in text but it would be clearer to state the ROPE values in the figure captions or in a small table, since the probabilities in Figures 5–7 depend directly on those thresholds.","section":"§5.2.2"},{"comment":"The One-hot transformation is described as selecting features with IV larger than 0.1, but the number of resulting pixels and the exact resizing procedure from the original one-hot matrix to 32x32 are not specified. Please clarify how non-selected features are dropped and how the sparse matrix is resized.","section":"§4.1"},{"comment":"The explainability claim that each pixel directly corresponds to a feature is true for the filled cells before padding, but the paper does not demonstrate SHAP or any explanation method on the transformed images. The claim of seamless compatibility should be stated as a potential advantage, not as an experimentally supported result.","section":"§6"}],"recommendation":"major_revision","confidential_remarks":"The capacity confound is the decisive issue: the missing control is a modern, large 1D CNN or MLP on the WOE-transformed tabular features. That experiment is well within the scope of the manuscript, so a major revision rather than rejection seems appropriate. I would also ask the editor to encourage the authors to release code and dataset preprocessing details, because the transformation is the main artifact of the paper and current reporting is not sufficient for exact replication. The paper is also somewhat over-claimed in the abstract relative to the TC results; this is fixable in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is the encoding: IV-proportional pixel allocation plus target-aware Spearman feature arrangement. That combination is not in the prior work they survey (Hosaka, Zhu, DeepInsight, REFINE, IGTD, superTML), and it is a sensible fit for credit scoring. The method is described clearly enough to reimplement, and the cross-validation protocol is honest: WOE and IV are computed on training folds only, and the Bayesian correlated t-tests are a real plus over the usual \"our mean is higher\" reporting. The robustness checks on image size, feature arrangement, and oversampling stability are also useful. Credit where due: this is a solid engineering contribution to the tabular-data-to-image program, not a paradigm shift.\n\nThe soft spots are real but addressable. The largest is the capacity confound. The headline claim is that Tabular Image lets a 2D CNN beat XGBoost on large credit portfolios, but the neural baselines are a small LeNet-style 1D CNN and a 5-layer MLP capped at 100 hidden units, while the proposed method uses ConvNeXt. On HC and FM the gains over those baselines are exactly what you would expect from model scale and modern optimization, even if the image encoding added nothing. Table 7 does show Tabular Image beats One-hot and DeepInsight under the same ConvNeXt, so the encoding matters among image methods, but it does not answer whether a similarly sized 1D CNN or wide MLP on WOE-transformed tabular features would match the reported numbers. That control is load-bearing for the \"state-of-the-art\" wording in the abstract and Section 5.2. The abstract also overclaims: on TC the method ties with XGBoost and GBDT, which the paper itself acknowledges in the Bayesian section. Minor but real: the random oversampling ratio is never stated, pixel scaling from normalised values to greyscale is not fully pinned down, and no code or data is released. Those are fixable in revision, but they currently limit reproducibility.\n\nWho is this for? Researchers working on tabular deep learning, especially in credit scoring, and practitioners who want a concrete way to feed tabular data into 2D CNNs. The paper deserves a serious referee: the method is novel enough, the evaluation is mostly careful, and the flaws are fixable rather than fatal. My recommendation is to send it to review, with the explicit requirement that the authors add a capacity-matched neural baseline on tabular features and either release code or spell out the preprocessing details fully. If that control kills the advantage, the paper still stands as a competitive transformation method, just not a state-of-the-art claim.","headline":"A genuinely new tabular-to-image encoding with a plausible but not yet capacity-controlled empirical claim; worth refereeing with mandatory fixes.","tokens_in":28444,"tokens_out":819,"would_cite":true,"duration_ms":9729,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Tabular Image turns credit tables into images; a 2D CNN then beats XGBoost on large loan datasets.","keywords":["credit scoring","tabular data","data-to-image transformation","convolutional neural networks","weight of evidence","information value","default prediction","ConvNeXt"],"falsifier":"Train a comparably sized modern deep network, such as an MLP or 1D CNN with the same parameter count as ConvNeXt, directly on the same tabular features with the same five-fold protocol and tuning budget; if that model matches or exceeds Tabular Image's AUC and H-measure on Home Credit and Fannie Mae, the claimed benefit of the image representation is not the source of the improvement.","tokens_in":27424,"feed_emoji":"🖼️","tokens_out":5512,"duration_ms":47639,"temperature":0.7,"pith_summary":"Tabular Image is a proposed transformation that converts one row of tabular credit data into a compact grayscale image, encoding each feature with weight of evidence (WOE) and allocating image space by information value (IV). The authors claim that a standard 2D CNN, specifically ConvNeXt, trained on these images achieves state-of-the-art AUC, H-measure, and KS on three credit-scoring benchmarks, outperforming nine baselines including XGBoost and GBDT on the larger Home Credit and Fannie Mae datasets. If correct, this gives lenders a way to apply powerful image networks to the tabular data that dominates credit scoring, while preserving pixel-to-feature correspondence for interpretation.","feed_headline":"Pictures of credit data beat XGBoost on large loan sets","feed_subtitle":"Tabular Image turns each borrower row into a 32x32 image, and a ConvNeXt CNN then leads on Home Credit and Fannie Mae.","key_machinery":"The load-bearing object is the Tabular Image generation pipeline: categorical values are replaced by WOE, numerical features are discretized into ten bins and then also assigned WOE, each feature receives a pixel count proportional to its IV, and a feature-arrangement matrix places features with high absolute Spearman correlation into neighboring 3x3 blocks. The normalized feature values are then written into the matrix, with padding by the median pixel value. This arrangement gives the 2D CNN spatially meaningful neighborhoods to convolve over, and the WOE/IV encoding injects target-separation information directly into pixel intensities. The downstream model is ConvNeXt, a deep residual-style CNN with depthwise convolutions and global response normalization.","core_discovery":"The central claim is that tabular credit data can be converted into images whose pixel layout embeds credit-specific statistics, and that doing so lets deep 2D CNNs outperform traditional tabular models and shallower neural networks. In five-fold cross-validation, ConvNeXt with Tabular Image is reported to achieve the highest average test AUC, H-measure, and KS on Taiwan Credit, Home Credit, and Fannie Mae, with the advantage over GBDT and XGBoost growing as datasets become larger and more imbalanced. On Fannie Mae the reported AUC is 91.75% versus 90.12% for GBDT, and the H-measure rises from roughly 50% to 55%. Bayesian correlated t-tests are presented as evidence that the differences are not due to chance. The transformation produces compact 32x32 images even for datasets with 120 features, and each pixel can be traced back to an original feature, which the paper argues preserves information and supports explainability.","pith_inferences":["Beyond the paper's claims: the reported comparison does not control for model capacity, because the neural baselines are a small LeNet-like 1D CNN and a 5-layer MLP with at most 100 hidden units, while the proposed method uses a deep ConvNeXt; a comparably sized modern MLP or transformer on raw tabular data would be the decisive test of whether the image representation itself drives the gain.","Beyond the paper's claims: if each pixel is truly traceable to a feature, Tabular Image could enable pretraining a single 2D CNN on pooled credit data from multiple lenders and fine-tuning on small portfolios, a transfer-learning route that tree ensembles do not naturally support.","Beyond the paper's claims: the feature-arrangement matrix is learned once and can be frozen, so the marginal deployment cost of the method is close to the cost of a lookup-and-normalize pass per new applicant, making the approach practical for online lending."],"forward_implications":["On the Home Credit and Fannie Mae datasets, ConvNeXt with Tabular Image is reported to beat GBDT and XGBoost on AUC, H-measure, and KS, with the advantage increasing with data size and imbalance.","The method adapts a standard 2D CNN to tabular credit data with minimal changes, and the paper reports stable performance across 32x32 and 96x96 image sizes.","Because each pixel maps to one original feature, pixel-level explanation techniques such as SHAP or Grad-CAM can be applied to the downstream CNN without additional machinery.","The WOE/IV encoding and correlation-based arrangement are presented as a flexible framework that can be repurposed for other tabular domains by swapping the binning, importance, and similarity measures.","The conversion is reported to be fast: image generation is linear in the number of rows, and the feature-arrangement step is a one-time computation that can be reused on new data."],"supporting_citations":[{"why":"Supplies the tabular preprocessing protocol and the baseline claim that XGBoost is the strongest model on credit data, which Tabular Image must surpass.","marker":"(Gunnarsson et al, 2021)"},{"why":"Provides the ConvNeXt architecture used as the 2D CNN for the proposed transformation.","marker":"(Woo et al, 2023)"},{"why":"Defines the weight-of-evidence transformation used to encode categorical and discretized features.","marker":"(Siddiqi, 2012)"},{"why":"Defines information value as the feature-importance measure used to allocate pixels to features.","marker":"(Hand, 2005)"},{"why":"Defines the H-measure used as a cost-sensitive evaluation metric.","marker":"(Hand, 2009)"},{"why":"Provides the Bayesian correlated t-test used to assess whether performance differences are practically significant.","marker":"(Benavoli et al, 2017)"},{"why":"Supplies the one-hot image transformation baseline that Tabular Image is compared against.","marker":"(Dastile and Celik, 2021)"},{"why":"Supplies the DeepInsight transformation baseline that Tabular Image is compared against.","marker":"(Sharma et al, 2019)"},{"why":"Documents why tree-based models typically outperform deep learning on tabular data, motivating the transformation approach.","marker":"(Grinsztajn et al, 2022)"},{"why":"Establishes the credit-scoring benchmark context and the comparison against ensemble classifiers.","marker":"(Lessmann et al, 2015)"}],"fun_headline_variants":["Credit rows as images: CNN beats XGBoost on large loan sets","Tabular Image: CNN tops GBDT on big credit datasets","Pixelated credit data helps CNNs beat GBDT on big sets","Convert credit to 32x32 images, CNN outdoes XGBoost","CNN with Tabular Image leads on large imbalanced loans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that the performance gains come from the image representation itself rather than from the much larger model capacity of ConvNeXt compared with the small 1D CNN and 5-layer MLP baselines.","fun_headline_variants_meta":{"raw":{"variants":["Credit rows as images: CNN beats XGBoost on large loan sets","Tabular Image: CNN tops GBDT on big credit datasets","Pixelated credit data helps CNNs beat GBDT on big sets","Convert credit to 32x32 images, CNN outdoes XGBoost","CNN with Tabular Image leads on large imbalanced loans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000271,"raw_usage":{"total_tokens":1670,"prompt_tokens":1028,"completion_tokens":642,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":547}},"tokens_in":644,"tokens_out":642,"duration_ms":6093,"temperature":1.0,"reasoning_tokens":547,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:13:31.400317+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a comparably sized modern deep network, such as an MLP or 1D CNN with the same parameter count as ConvNeXt, directly on the same tabular features with the same five-fold protocol and tuning budget; if that model matches or exceeds Tabular Image's AUC and H-measure on Home Credit and Fannie Mae, the claimed benefit of the image representation is not the source of the improvement.","supporting_citations":[],"review_version":1}