{"id":"330f1dcd-9ceb-451b-b953-ac6e0959ee06","arxiv_id":"2608.13513","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"TabSOM is a tabular-to-image encoding that places features via SOM component planes and adds a relational edge channel; it is competitive with, but not consistently superior to, existing encodings.","lead":"TabSOM converts tabular data into images using self-organizing maps to place features and encode pairwise relationships, then trains CNNs on those images. The paper claims strong benchmark results, but its own tables show TabSOM ranks third on two of four datasets, not first or second.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that TabSOM 'ranks first or second on every dataset' is contradicted by the paper's own Table 3, where TabSOM is third on Parkinsons and QSAR.","rationale":"The reader's weakest assumption concerned whether a single fixed CNN architecture is fair across all encodings. My stress-test identifies a more direct and more decisive problem: the paper's own benchmark table contradicts its central claim. This is an internal inconsistency, not a dispute with external consensus, so it does not rely on contested baselines. The reader's rationale already mentions the contradiction between the abstract and Table 3, so there is substantial agreement on the overall rejection, but the specific weakest assumption I would emphasize is the factual mismatch between the reported table and the headline conclusions. Because the central claim is falsified by the manuscript's own evidence, the REJECT verdict stands; no change to the reader's verdict is needed. At the same time, I would not go further and allege any methodological dishonesty; the issue is a reporting and verification failure in the central claim.","tokens_in":12523,"tokens_out":2351,"duration_ms":21917,"concrete_test":"Rerun the ranking from Table 3: for each dataset, rank all 13 methods by mean AUROC and record TabSOM's position. Then compare the minimum standard deviation per dataset with TabSOM's standard deviation. Finally, count the datasets in Table 2 (four) and compare with the phrase 'across six binary datasets' in Section 4.3. If the printed numbers are used, TabSOM is third on Parkinsons and QSAR, BarGraph has lower variance on WDBC and QSAR, and the dataset count is inconsistent; the abstract and conclusion claims must be revised or the benchmark rerun.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract, Discussion, and Conclusions claim that TabSOM 'ranks first or second on every dataset' and 'achieves the lowest variance of any method evaluated.' These claims are contradicted by the paper's own results in Table 3. On Parkinsons, TabSOM (0.8852) ranks third behind Combination (0.9179) and DistanceMatrix (0.8993). On QSAR, TabSOM (0.9098) ranks third behind Combination (0.9179) and DistanceMatrix (0.9140). TabSOM is first only on Pima and WDBC. The variance claim also fails: BarGraph has a smaller standard deviation than TabSOM on WDBC (0.0055 vs 0.0082) and on QSAR (0.0162 vs 0.0180). In addition, Section 4.3 says the comparison is 'across six binary datasets,' while Section 4.1 and Table 2 list four datasets and Table 3 reports four columns. This is not a matter of interpretation or of a debatable baseline: the headline conclusion is directly falsified by the manuscript's own reported numbers. Because the paper's central contribution is the claimed state-of-the-art ranking, this internal inconsistency is load-bearing. The encoding idea may still be worth studying, but the manuscript as written substantially overstates its evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes TabSOM, a tabular-to-image encoding method that trains a self-organizing map (SOM), derives per-feature anchors from the SOM component planes, resolves anchor collisions with Hungarian assignment, constructs a relational graph from component-plane correlations, and renders a multi-channel image containing two Gaussian node channels and one edge-interaction channel. It also introduces two SOM-based interpretability tools: a class-separation importance score and a prototype-based partial dependence plot. The method is benchmarked against twelve existing tabular-to-image encodings on public binary-classification datasets, and the authors claim that TabSOM ranks first or second on every dataset, achieves the lowest variance among all evaluated methods, and provides interpretability consistent with Random Forest, XGBoost, and SHAP.","tokens_in":12834,"tokens_out":6214,"duration_ms":58216,"significance":"If the performance claims were accurate, TabSOM would be a meaningful contribution to tabular deep learning, since it combines a topology-preserving feature placement with an explicit pairwise-interaction channel and model-agnostic interpretability tools in a single SOM-based framework. The conceptual design is interesting: using component planes both for placement and for graph construction is a natural extension of existing tabular-to-image methods, and the comparison against Random Forest, XGBoost, and SHAP is a useful interpretability check. However, the headline empirical claims are contradicted by the paper's own reported numbers, and the experimental reporting contains inconsistencies. The significance of the contribution as stated cannot be accepted without substantial correction.","major_comments":[{"comment":"The statement that TabSOM \"ranks first or second on every dataset\" is directly contradicted by Table 3 of the manuscript. On Parkinsons, TabSOM (0.8852) ranks third behind Combination (0.9179) and DistanceMatrix (0.8993); on QSAR, TabSOM (0.9098) ranks third behind Combination (0.9179) and DistanceMatrix (0.9140). The same false claim appears in the Discussion and Conclusions sections. This is a load-bearing error because the abstract's central performance claim is falsified by the paper's own evidence.","section":"Abstract, Section 4.3, Table 3"},{"comment":"The variance claim that TabSOM \"achieves the lowest variance of any method evaluated\" is also contradicted by Table 3. On Parkinsons, BarGraph (0.0297) and Combination (0.0312) have lower standard deviations than TabSOM (0.0390); on WDBC, BarGraph (0.0055), Combination (0.0046), and several other methods have lower or comparable standard deviations than TabSOM (0.0082); on QSAR, DistanceMatrix (0.0136), BarGraph (0.0162), and Combination (0.0163) all have lower standard deviations than TabSOM (0.0180). Only on Pima is TabSOM's standard deviation the smallest among the reported methods.","section":"Abstract, Section 4.3, Table 3"},{"comment":"Section 4.3 states that the comparison is performed \"across six binary datasets,\" but Section 4.1 and Table 2 define only four datasets (PID, PAR, QSA, WBC) and Table 3 reports results for only four datasets. The two missing datasets are never identified, so the reported benchmark does not support the stated experimental scope. Either the experiments must be extended to six datasets or the text must be corrected.","section":"Section 4.3"},{"comment":"Several SOM hyperparameters are left unspecified: the initial and final learning rates (η0, η1), the initial and final neighborhood radii (σ0, σ1), the number of training steps or epochs, the initialization scheme, and the stopping criterion. Because the component planes, anchors, and relational graph all depend on the trained SOM, the method is not reproducible without these values, and the comparison against other encodings is under-specified.","section":"Sections 3.1.1 and 4.2"},{"comment":"The evaluation protocol is internally inconsistent: it first says each dataset is split into an 80% training subset and a 20% test subset, but then says all methods are evaluated under 5-fold stratified cross-validation. The text also reports \"mean and standard deviation across five random seeds\" without clarifying whether the five seeds refer to CNN initialization, SOM initialization, fold construction, or some combination. The exact protocol must be stated precisely.","section":"Section 4.2"},{"comment":"The fixed CNN architecture and hyperparameter set used for every encoding method are asserted but not justified. If the fixed image size, number of channels, or training schedule happens to favor the layouts produced by some encodings over others, the reported ranking may reflect tuning mismatch rather than intrinsic encoding quality. A sensitivity analysis or per-method hyperparameter selection under a fixed validation protocol is needed to support the ranking conclusions.","section":"Section 4.2"}],"minor_comments":[{"comment":"The label \"T abSOM\" contains a spurious space and should read \"TabSOM.\"","section":"Table 3"},{"comment":"The manuscript uses both \"AUROC\" and \"AUCROC\" for the same metric; please unify the abbreviation.","section":"Sections 4.2 and 4.3"},{"comment":"The phrase \"two first node channels\" is awkward; since the primary configuration uses S=2 node channels plus one edge channel, the text should say \"two node channels.\"","section":"Section 3.1.5"},{"comment":"The Discussion contains two consecutive sentences beginning \"Future work will explore...\"; these should be merged or the duplication removed.","section":"Section 5"},{"comment":"The reference list appears to include both a journal version and an arXiv preprint for the same work (Tab2Visual, refs [2] and [22]); if they are the same work, only the published version should be cited.","section":"References"},{"comment":"The caption refers to the composite RGB image as a \"fourth channel,\" but it is a display of the three actual channels rather than a separate image channel.","section":"Figure 1 caption"}],"recommendation":"reject","confidential_remarks":"The encoding idea and the interpretability framework are interesting, but the paper's central advertised result—ranking first or second on every dataset and achieving the lowest variance—is directly contradicted by its own Table 3, and the experimental reporting has additional inconsistencies (dataset count, evaluation protocol, missing hyperparameters). In my view, the current manuscript cannot be accepted without major corrections that go beyond presentation fixes. The editor may wish to ask the authors to correct or remove the false claims and to supply full experimental details if a resubmission is considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The TabSOM encoding idea is real: anchoring features via SOM component planes, then resolving collisions with Hungarian assignment, and adding a relational edge channel plus multi-scale node rendering, is a combination I haven't seen in the cited prior work. The interpretability tools (class-separation importance, prototype-based PDP) are a thoughtful addition, and the validation against RF/XGB/SHAP gives them some credibility. The benchmark across twelve methods and four datasets is a reasonable scope for a methods paper, and the writing is clear about the pipeline. So there is substance here.\n\nThe problem is that the paper's central claim does not survive its own results. The abstract, discussion, and conclusions all say TabSOM \"ranks first or second on every dataset\" and has the \"lowest variance of any method evaluated.\" Table 3 shows TabSOM third on Parkinsons and QSAR, and BarGraph has a smaller standard deviation on WDBC and QSAR. That is not a debatable interpretation; it is a direct factual contradiction in the reported numbers. The softer claim that TabSOM is consistently in the top two methods, with the second-best overall mean and average rank behind Combination, actually is supported. So the fix is to report accurately, not to retract the method. There are also smaller issues: Section 4.3 says \"across six binary datasets\" but only four appear, the prior SOM-based method [18] is discussed but never benchmarked against, and the SOM learning-rate and neighborhood-decay values are not reported. No code is provided, which limits reproducibility but is not decisive for a methods paper of this type.\n\nI disagree with the reader's suggestion that the premise is weak because the same CNN architecture is used for all encodings. That is a standard protocol in this literature and not obviously unfair; if anything, the paper could be clearer about how image sizes differ across methods, but the protocol itself is defensible. The stress-test concern, however, lands squarely: the abstract's ranking claim is falsified by the manuscript's own table.\n\nThis paper should go to peer review rather than be desk-rejected. The encoding pipeline is novel and plausible, and the errors are fixable in a revision. A serious referee would require corrected claims, a consistent dataset count, and ideally full hyperparameters or code. As written, it is not acceptable, but the underlying work deserves engagement. I'd bring it to a reading group to discuss exactly how a good idea can be undermined by overclaiming.","headline":"A genuinely new tabular-to-image encoding idea is undermined by a headline ranking claim that its own Table 3 contradicts; the method deserves a serious look but the manuscript needs major revision.","tokens_in":13371,"tokens_out":2002,"would_cite":false,"duration_ms":20707,"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":"TabSOM converts tabular data into a stable, multi-scale image whose layout is fixed by a self-organizing map, and the paper reports top rankings against twelve prior encodings.","keywords":["self-organizing maps","tabular-to-image encoding","component planes","feature placement","feature relationship graph","interpretability","convolutional neural networks","Hungarian assignment"],"falsifier":"Re-run the comparison on the same four datasets with per-method tuning of image size, channel count, and training schedule; if TabSOM no longer sits at the top of the ranking or its variance is no longer lowest, the fixed-architecture premise is the load-bearing part of the claim. As a minimal check, recompute the ranks from the mean AUCROC values printed in Table 3 to verify the 'first or second on every dataset' and 'lowest variance' wording.","tokens_in":12351,"feed_emoji":"🗺️","tokens_out":12858,"duration_ms":112847,"temperature":0.7,"pith_summary":"The paper proposes TabSOM, a tabular-to-image encoding that gives every column a fixed, meaningful position on an image canvas before a CNN ever sees the data. Existing encodings place features by embedding them with t-SNE, PCA, or UMAP, then render only each feature's marginal value, so pairwise relationships are lost. TabSOM trains a self-organizing map on the normalized table, reads each feature's component plane to choose an anchor, resolves collisions with the Hungarian algorithm, and renders two value channels plus one channel that draws active feature-pair links. The paper reports that TabSOM ranks first or second on every dataset among thirteen encodings and shows the lowest variance, and that its SOM-derived importance score and partial-dependence curves agree with established interpretability baselines on the top features. If these results hold, tabular data can use convolutional architectures without sacrificing a readable, inspectable representation.","feed_headline":"TabSOM ranks first or second across 13 tabular-to-image encodings","feed_subtitle":"Each column gets a fixed pixel position plus a feature-pair channel, so CNNs read stable, interpretable images.","key_machinery":"The carrying object is the SOM component plane $\\Phi_j$, the $j$-th coordinate of every node's prototype viewed as a spatial field over the SOM grid. It does three jobs at once: it fixes a preferred canvas position for feature $j$ (via centroid or mode anchor), it defines the pairwise relationship graph (correlation or cosine similarity between flattened planes), and it supplies the interpretability quantities (class-weighted plane averages and prototype-based dependence curves). A Hungarian linear-assignment step turns the anchors into a collision-free placement, so the final image is a fixed, sample-independent layout plus a sample-dependent value and edge rendering.","core_discovery":"TabSOM's central claim is that a self-organizing map, not a dimensionality-reduction embedding, should determine where features sit on an image canvas. After training an $H\\times W$ SOM on normalized tabular data, each feature $j$ is represented by its component plane $\\Phi_j$, a smooth field over the grid. The plane yields an anchor (intensity centroid or peak), and the Hungarian algorithm assigns all anchors to distinct grid cells at minimal total displacement. The rendered image has three channels: two node channels place Gaussian bumps of width $\\sigma=0.05$ and $\\sigma=0.08$ at each feature's cell, scaled by the feature's value; the third channel draws line segments between correlated features, with intensity equal to the product of the two active feature values weighted by their relation. Because the layout is fixed once the SOM is trained, the encoding is stable across folds and seeds, and the same component planes support class-separation importance and prototype-based partial dependence plots.","pith_inferences":["TabSOM's fixed-layout property suggests a direct ablation test the paper does not run: removing the edge channel should isolate how much of the gain comes from pairwise interactions versus from the stable SOM placement itself.","The class-separation importance score is computed from training labels and BMU hits; comparing it with permutation importance on held-out data would test whether it tracks generalization rather than just training-set separation.","The relational edge channel's interaction term $\\sqrt{\\max(x_i,0)\\max(x_j,0)}$ is specific to non-negative, binary-target scenarios; extending it to multiclass or regression would require a signed or class-conditional interaction term, which the paper leaves for future work."],"forward_implications":["If TabSOM's benchmark results hold, CNN pipelines for tabular data no longer need per-dataset tuning of an embedding method; one SOM-derived layout is fixed and stable across folds.","Because feature positions are derived from the training distribution and then frozen, the same layout can be reused for every row, making the encoding's channel decomposition inspectable as a feature map, a value map, and an interaction map.","The class-separation importance score ranks features without training a downstream model, so it can be used as a model-free screening tool for binary tables.","The prototype-based partial dependence plot gives per-feature effect curves that respect the data's populated regions, complementing conventional PDPs that average over synthetic grids.","A three-channel image (sharp node, mid node, edge) is the default configuration; the design allows more scales or a static graph channel, so the method can be adapted to other image resolutions."],"supporting_citations":[{"why":"Supplies the DeepInsight t-SNE-based feature placement method that TabSOM benchmarks against and improves on.","marker":"[4]"},{"why":"Provides the TINTO PCA/t-SNE placement and blurring baseline that TabSOM is directly compared with.","marker":"[5]"},{"why":"Introduces REFINED, a neighborhood-dependency image representation baseline for the benchmark.","marker":"[6]"},{"why":"Defines the self-organizing map, component planes, and topology-preserving properties that TabSOM builds on.","marker":"[14]"},{"why":"Presents the prior SOM-based tabular-to-image activation-map method that TabSOM extends and distinguishes itself from.","marker":"[18]"},{"why":"Provides the categorization and benchmark context for spatial encoding methods that frames the evaluation.","marker":"[3]"}],"fun_headline_variants":["Self-organizing maps deliver top-2 tabular-image encoding","TabSOM: SOM-based tabular encoding ranks top-2 on all datasets","Fixed canvas spots plus pairwise links: TabSOM's winning recipe","TabSOM rethinks tabular-to-image with interpretable SOM layouts","Stable feature maps turn tables into images that CNNs master"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark claim rests on one fixed CNN architecture and hyperparameters being equally fair for all thirteen encodings; if some encodings need different image sizes, channel counts, or training schedules, the ranking could reflect tuning mismatch rather than encoding quality.","fun_headline_variants_meta":{"raw":{"variants":["Self-organizing maps deliver top-2 tabular-image encoding","TabSOM: SOM-based tabular encoding ranks top-2 on all datasets","Fixed canvas spots plus pairwise links: TabSOM's winning recipe","TabSOM rethinks tabular-to-image with interpretable SOM layouts","Stable feature maps turn tables into images that CNNs master"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000698,"raw_usage":{"total_tokens":3206,"prompt_tokens":1052,"completion_tokens":2154,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":2060}},"tokens_in":668,"tokens_out":2154,"duration_ms":16858,"temperature":1.0,"reasoning_tokens":2060,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:19:05.680201+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the comparison on the same four datasets with per-method tuning of image size, channel count, and training schedule; if TabSOM no longer sits at the top of the ranking or its variance is no longer lowest, the fixed-architecture premise is the load-bearing part of the claim. As a minimal check, recompute the ranks from the mean AUCROC values printed in Table 3 to verify the 'first or second on every dataset' and 'lowest variance' wording.","supporting_citations":[{"cited_title":"DeepInsight: A methodology to transform a non-image data to an image for convolution neural network architecture","cited_arxiv_id":null,"evidence_quote":"Supplies the DeepInsight t-SNE-based feature placement method that TabSOM benchmarks against and improves on."},{"cited_title":"TINTO: converting tidy data into image for classifi- cation with 2-dimensional convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"Provides the TINTO PCA/t-SNE placement and blurring baseline that TabSOM is directly compared with."},{"cited_title":"Repre- sentation of features as images with neighborhood dependencies for compatibility with convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"Introduces REFINED, a neighborhood-dependency image representation baseline for the benchmark."},{"cited_title":"Engineering applications of the self-organizing map","cited_arxiv_id":null,"evidence_quote":"Defines the self-organizing map, component planes, and topology-preserving properties that TabSOM builds on."},{"cited_title":"Topological Activation Maps for Visual Representation Learning from Tabular Data","cited_arxiv_id":null,"evidence_quote":"Presents the prior SOM-based tabular-to-image activation-map method that TabSOM extends and distinguishes itself from."},{"cited_title":"A comprehensive bench- mark of spatial encoding methods for tabular data with deep neural networks","cited_arxiv_id":null,"evidence_quote":"Provides the categorization and benchmark context for spatial encoding methods that frames the evaluation."}],"review_version":1}