{"id":"8a38d7e7-f0d1-4230-9c8c-c4e30255e4a2","arxiv_id":"2411.11148","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"TabDeco pairs SAINT-style attention with SwitchTab-style feature decoupling and six contrastive losses, but its claim of consistently beating gradient boosting is contradicted by its own results.","lead":"TabDeco is a contrastive learning framework for tabular data that combines row and column attention with feature decoupling and six contrastive losses. The paper claims it beats XGBoost, CatBoost, and LightGBM across benchmarks, but its own tables show that claim is not consistent.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Even taking Table 2 at face value, TabDeco loses to the best GBDT on Bank, Income, and Spambase; with 14 loss combinations × 3 variants and no stated selection rule, 'consistent superiority' is unsupported.","rationale":"I agree with the reader that the paper should be rejected, but I would sharpen the load-bearing concern. The reader's weakest assumption focuses on per-dataset selection of the best loss combination from 14 options and 3 variants with access to test results. That is a serious protocol concern, and Section 4.3 indeed reports the best performance for each combination across variants without stating how the main-table choices were made. However, an even more direct problem is that the paper's own Table 2 already falsifies the abstract's 'consistently surpasses' claim under any reasonable reading: even if one is allowed to pick the best TabDeco variant per dataset, TabDeco still loses to the best GBDT on Bank, Income, and Spambase. These are not tiny differences hidden in noise, and Table 5 shows standard errors of 0.04 to 0.27, so the three losses are material. Thus the headline claim fails on the reported numbers before any protocol criticism is added. The unspecified loss/variant selection compounds this by making it impossible for a reader to know what fixed method Table 2 actually evaluates, or to reproduce the reported wins. I do not see a mathematical step in Equations (2)-(4) that needs to be checked before this evaluation-level problem is resolved; the missing selection protocol and the contradiction with Table 2 are sufficient. The reader's verdict of REJECT is therefore unchanged.","tokens_in":13701,"tokens_out":7585,"duration_ms":66188,"concrete_test":"Obtain the authors' code and rerun Table 2 with a single fixed configuration: full TabDeco with Lgg (the best single column in Table 3), hyperparameters tuned only on the 15% validation split, and the same 65/15/20 data split. Count how many of the 11 datasets this fixed configuration beats the best of XGBoost/CatBoost/LightGBM over 10 seeds using a paired bootstrap. If the fixed configuration does not win on at least 9 of 11 datasets, or if per-dataset oracle selection was needed to reach the 8/11 win count, the abstract's 'consistently surpasses' claim is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract promises that TabDeco 'consistently surpasses' XGBoost, CatBoost, and LightGBM. The paper's own Table 2 contradicts this at face value: comparing the best TabDeco variant per dataset against the best GBDT, TabDeco loses on Bank (93.34 vs LightGBM 93.39), Income (91.26 vs LightGBM 92.57), and Spambase (97.97 vs LightGBM 99.01). That is 8/11 wins, not consistent superiority, and with the standard errors in Table 5 most margins are within about 2 SE. The protocol behind the reported numbers is also underspecified. Section 4.3 reports the best performance for each of 14 contrastive-loss combinations across 3 variants, but the main results never state which loss combination and which architectural variant produced each Table 2 entry, nor whether the choice was made on validation or test. Selecting the best of 14 × 3 configurations per dataset on the test set would inflate results, and no code is provided to rule this out. Both the reported evidence and the reproducibility of the claim are therefore load-bearing and unsatisfied.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TabDeco, a supervised contrastive learning framework for tabular data that combines row- and column-level attention (in the style of SAINT) with global-local feature decoupling (inspired by SwitchTab) and a menu of six contrastive losses that can be combined into fourteen loss configurations. The authors evaluate TabDeco and two architectural variants on eleven binary and multiclass tabular datasets, reporting AUROC or accuracy against logistic regression, random forests, XGBoost, CatBoost, LightGBM, MLP, VIME, TabNet, TabTransformer, SwitchTab, and three SAINT variants. The abstract and conclusion claim that TabDeco 'consistently surpasses' leading gradient boosting algorithms across benchmark tasks, and Section 4.2 reports that one TabDeco variant ranks best or second-best on all eleven datasets.","tokens_in":13946,"tokens_out":3914,"duration_ms":36983,"significance":"If the central claim were supported, TabDeco would be a practically valuable contribution to tabular representation learning, since the combination of feature decoupling and multi-level contrastive losses is a plausible and interesting design. The paper also has some strengths: it builds on credible prior components (SAINT, SwitchTab), reports means and standard deviations over ten trials, and provides a relatively broad baseline comparison. However, the empirical evidence presented in the manuscript does not support the headline claim of consistent superiority over gradient boosting, and the evaluation protocol is insufficiently specified to allow the results to be reproduced or trusted. The conceptual contribution is therefore currently obscured by an overclaimed and under-specified evaluation.","major_comments":[{"comment":"The claim that TabDeco 'consistently surpasses' XGBoost, CatBoost, and LightGBM is directly contradicted by Table 2. The best TabDeco variant is worse than the best GBDT on Bank (93.34 vs 93.39 by LightGBM), Income (91.26 vs 92.57 by LightGBM), and Spambase (97.97 vs 99.01 by LightGBM). This is 8 of 11 datasets, not consistent superiority. The abstract and conclusion must be revised to reflect the actual results, or the experiments must be redone with a protocol that supports the stronger claim.","section":"Abstract; Section 5; Table 2"},{"comment":"The ablation study reports the best performance for each of 14 contrastive-loss combinations across the three TabDeco variants, but the manuscript never states how the final model used in Table 2 was selected. If the per-dataset best loss combination was chosen after inspecting test-set results in Table 3, the reported Table 2 numbers are optimistically biased by selection on the test set, and the usual safeguards (validation-based selection, multiple-testing correction) are not described. The authors must either disclose the exact selection rule or report results for a single fixed loss configuration chosen before seeing test labels.","section":"Section 4.3; Table 3"},{"comment":"The paper describes TabDeco as having 'significant predicting power,' but Table 5 shows that most differences from the strongest GBDT baselines are within about two standard deviations (e.g., Bank TabDeco 93.34 ± 0.15 vs LightGBM 93.39 ± 0.21; Income 91.26 ± 0.07 vs LightGBM 92.57 ± 0.13). No statistical significance test (paired test, confidence interval, or multiple-comparison correction) is reported. Without such an analysis, the claim of consistent superiority is not supported by the evidence in the tables.","section":"Section 4.2; Tables 2 and 5"},{"comment":"The training details specify the optimizer, learning rate, batch size, and temperature, but the contrastive loss weight α and the specific loss combination used for the main results in Table 2 are not given. Algorithm 1 leaves both as free choices (steps 7 and 8). This omission makes the main results non-reproducible and prevents a reader from determining whether the reported performance comes from a fixed model or from per-dataset tuning. The manuscript must specify the exact configuration used to produce each Table 2 entry.","section":"Section 4.1.3; Algorithm 1"}],"minor_comments":[{"comment":"There are numerous typos and misspellings, including 'ehance' in Section 1, 'simutaneously' in Section 3.1, 'demoted' before Equation (2), 'constrastive' in Algorithm 1, 'Table‘3' in Section 4.3, 'TabTransormer' in Table 2, and inconsistent spacing in 'V olkert'.","section":"Throughout"},{"comment":"The baseline list refers to TabTransformer but cites reference [49], which is the TranTab paper; either the citation or the baseline name is mismatched.","section":"Section 4.1.5; Reference [49]"},{"comment":"The use of a negative sign inside the exponential for Llocal and Lcross is unusual for InfoNCE-style losses; please clarify whether this is intentional (to repel similar features) or a typo, and relate the formulation to standard contrastive losses.","section":"Equations (2)-(4)"},{"comment":"The statement that TabDeco is 'the first attempt' to explicitly facilitate structured embeddings through contrasting for tabular data is an overclaim given existing contrastive tabular methods (e.g., SCARF, Recontab); the claim should be softened or restricted to the specific multi-level decoupling setup.","section":"Section 1; Section 4.2"},{"comment":"The paper states that baseline results are 'quoted from original papers when possible and reproduced otherwise,' but it does not say which baselines were quoted and which were reproduced, nor whether the same data splits and hyperparameters were used; this should be stated to rule out discrepancies in the comparison.","section":"Section 4.1.5; Table 2"}],"recommendation":"reject","confidential_remarks":"The central claim of consistent superiority over gradient boosting is contradicted by the paper's own Table 2, and the evaluation protocol is not disclosed in enough detail to verify the results. The manuscript would need a substantial redesign of the evaluation (fixed model selection, statistical testing, and full disclosure of configurations) before its claims could be taken seriously. As it stands, the paper does not meet the bar for a journal publication, though the core idea of combining decoupled global/local representations with multi-level contrastive losses may be worth pursuing in a future, more rigorous study."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"TabDeco is a combination of known components—SAINT-style row/column attention, SwitchTab's global/local decoupling, and a family of InfoNCE-style contrastive losses—organized into a six-way taxonomy. That taxonomy and the ablation across 14 loss combinations are the most useful parts of the paper. The authors test on 11 public datasets and compare against a reasonable set of baselines, which is more than many tabular-learning papers do.\n\nThe problem is the headline claim. The abstract says TabDeco 'consistently surpasses' XGBoost, CatBoost, and LightGBM. Table 2 contradicts that: TabDeco's best variant loses to LightGBM on Bank (93.34 vs 93.39), Income (91.26 vs 92.57), and Spambase (97.97 vs 99.01). That is 8/11 wins, not consistent superiority. More worrying, the main results seem to come from per-dataset selection of the best loss combination among the 14 tried (plus three architectural variants). Section 4.3 says they report the best performance for each combination, but the paper never says which combination produced each Table 2 entry, nor whether the choice was made on validation or test. If it was made on test, the numbers are inflated and the 'consistent' claim collapses. No code is provided to rule this out, and the contrastive loss weight alpha appears in Algorithm 1 but is never given a value.\n\nThe ablation itself shows no fixed configuration wins everywhere; some losses actually hurt on specific datasets. So the framework is not a single method with a defined recipe, but a menu with per-dataset tuning. That is a load-bearing flaw, not a cosmetic one.\n\nFor a workshop paper, this might be acceptable as a position piece. But as a claim of state-of-the-art performance, it needs a fixed, pre-registered configuration and validation-based selection. I would send it to peer review because the loss taxonomy and ablation are useful, and the flaws are fixable, but I would expect major revision and a much more careful evaluation.\n\nWho would benefit? Researchers working on tabular contrastive learning who want a structured look at which loss granularities help. But do not cite the headline claim; cite the taxonomy if anything.","headline":"A useful taxonomy of contrastive losses for tabular data, but the headline claim of consistent superiority over GBDTs is contradicted by the paper's own tables and the evaluation protocol is not trustworthy.","tokens_in":14501,"tokens_out":3812,"would_cite":false,"duration_ms":33287,"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":"TabDeco claims that multi-level contrastive losses on decoupled local and global features let a tabular transformer beat XGBoost, CatBoost, and LightGBM on most benchmark datasets.","keywords":["contrastive learning","tabular data","representation learning","feature decoupling","attention mechanism","self-supervised learning","gradient boosting","AUROC"],"falsifier":"Re-run Table 2 with the loss combination and architecture for each dataset fixed before test labels are examined, choosing them only on the 15% validation split, and count how often TabDeco beats XGBoost, CatBoost, and LightGBM; if the average AUROC edge disappears, the paper's central 'consistent superiority' claim is not supported.","tokens_in":13494,"feed_emoji":"📊","tokens_out":10355,"duration_ms":150361,"temperature":0.7,"pith_summary":"TabDeco sets out to show that contrastive learning can succeed on tabular data when positive and negative pairs are constructed at feature, instance, and batch levels rather than only at the instance level. It splits each encoded sample into a local representation and a global representation, then trains with a supervised loss plus combinations of six contrastive losses. The paper reports that on 11 classification benchmarks the TabDeco variants take the best or second-best score on every dataset, and that they surpass leading gradient-boosting algorithms such as XGBoost, CatBoost, and LightGBM on most of them. A sympathetic reader should care because tabular data dominate applied machine learning, and deep representation learners that beat tuned boosting would change the default tool of choice.","feed_headline":"TabDeco tops XGBoost, CatBoost, and LightGBM on most tabular datasets","feed_subtitle":"Row-and-column attention plus local-global contrast losses lands best or second-best scores on all 11 benchmarks.","key_machinery":"The load-bearing device is global-local feature decoupling. After feature-level (column) and instance-level (row) attention blocks produce an embedding, a local projector $L(\\cdot)$ captures instance-specific and feature-specific detail while a global projector $G(\\cdot)$ captures dataset-wide and cross-batch patterns. On these two views the paper defines six contrastive losses—$L_{\\mathrm{all}}$, $L_{\\mathrm{gg}}$, $L_f$, $L_s$, $L_{\\mathrm{fs}}$, and $L_{\\mathrm{sf}}$—each built from a global term, a local term, and a cross term, and each comparing similarity matrices at a different granularity (feature, instance, batch, or combinations). This machinery is what lets the paper vary only the pairing structure while keeping the architecture fixed, so observed gains are attributed to the decoupled contrastive objective rather than to added capacity.","core_discovery":"The central claim is that the weakness of prior contrastive tabular methods lies in how they choose sample pairs, not in the network architecture. TabDeco encodes data through attention over both columns and rows, projects the result into local and global feature vectors, and then applies six contrastive losses that contrast features, instances, and batches in different combinations, with the total loss being the supervised loss plus a weighted contrastive term. The paper argues that this multi-level decoupling produces well-separated, interpretable embeddings, and it reports that the best TabDeco variant is the top performer on 7 of 11 datasets and second-best on the remaining 4, including wins against XGBoost, CatBoost, and LightGBM.","pith_inferences":["Beyond the paper's protocol, the per-dataset choice of the best of 14 loss combinations and 3 architecture variants after seeing test-set results may inflate the reported edge over boosting; a validation-based selection protocol would test whether the advantage survives.","Beyond the paper, because gains are attributed to pairing structure, a natural extension is to replace the fixed loss zoo with a learned or validation-driven loss selector, which could reduce the 14-combination search to a single training run.","Beyond the paper, the same local-global projectors could likely be applied in semi-supervised or transfer settings where labels are scarce, even though the paper only demonstrates supervised training.","Beyond the paper, a cheap practical rule suggested by Table 3 is to skip $L_f$ and $L_s$ entirely, since they rarely beat the baseline, and to focus on the fine-grained and cross-batch losses."],"forward_implications":["A deep tabular model trained with the right contrastive loss can match or exceed tuned gradient boosting on the 11 benchmarks, so tree ensembles are not automatically the safe default for tabular classification.","Fine-grained contrastive losses such as $L_{\\mathrm{fs}}$, $L_{\\mathrm{sf}}$, and $L_{\\mathrm{gg}}$ improve on the no-contrastive baseline on 6 to 11 of the 11 datasets, while the simple feature-level and instance-level losses $L_f$ and $L_s$ help on only 0 and 3 datasets.","The comparison against the SAINT variants isolates the contribution of the decoupling and contrastive modules, since TabDeco without those modules reduces to the SAINT architecture.","TabDeco's reported standard deviations are often smaller than those of the boosting baselines, which the paper reads as evidence of more stable predictions.","The framework also reaches the best accuracy on the two multi-class datasets (Volkert and MNIST), indicating the decoupled contrastive objective transfers beyond binary classification."],"supporting_citations":[{"why":"Introduces the feature-decoupling idea that TabDeco extends and is a direct comparator in the experiments.","marker":"[51]"},{"why":"Supplies the row- and column-attention architecture and the SAINT variants used as both the backbone and the baselines.","marker":"[43]"},{"why":"XGBoost is one of the leading gradient-boosting baselines whose reported performance TabDeco claims to surpass.","marker":"[14]"},{"why":"LightGBM is a leading gradient-boosting baseline in the main comparison table.","marker":"[23]"},{"why":"CatBoost is a leading gradient-boosting baseline in the main comparison table.","marker":"[39]"},{"why":"VIME is the primary prior contrastive self-supervised tabular baseline.","marker":"[65]"},{"why":"TabNet is a transformer-based tabular deep learning baseline used in the comparison.","marker":"[1]"},{"why":"FT-Transformer is the column-attention encoder design referenced for feature-level attention.","marker":"[17]"}],"fun_headline_variants":["TabDeco beats XGBoost, CatBoost, LightGBM on tabular data","TabDeco: contrastive method outshines gradient boosting on benchmarks","Row-column attention lifts TabDeco past LightGBM on 11 benchmarks","TabDeco rethinks contrastive pairs, tops boosting on tabular sets","Decoupled representations: TabDeco wins majority of tabular tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim of consistent superiority assumes that selecting, for each dataset, the best of 14 loss combinations and 3 architecture variants after seeing test-set numbers does not inflate the reported performance; were that selection fixed in advance or made only on validation data, the margin over gradient boosting would likely shrink.","fun_headline_variants_meta":{"raw":{"variants":["TabDeco beats XGBoost, CatBoost, LightGBM on tabular data","TabDeco: contrastive method outshines gradient boosting on benchmarks","Row-column attention lifts TabDeco past LightGBM on 11 benchmarks","TabDeco rethinks contrastive pairs, tops boosting on tabular sets","Decoupled representations: TabDeco wins majority of tabular tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1246,"prompt_tokens":869,"completion_tokens":377,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":274}},"tokens_in":485,"tokens_out":377,"duration_ms":3870,"temperature":1.0,"reasoning_tokens":274,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:51:47.666483+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run Table 2 with the loss combination and architecture for each dataset fixed before test labels are examined, choosing them only on the 15% validation split, and count how often TabDeco beats XGBoost, CatBoost, and LightGBM; if the average AUROC edge disappears, the paper's central 'consistent superiority' claim is not supported.","supporting_citations":[{"cited_title":"Switchtab: Switched autoencoders are effective tabular learners","cited_arxiv_id":null,"evidence_quote":"Introduces the feature-decoupling idea that TabDeco extends and is a direct comparator in the experiments."},{"cited_title":"Catboost: unbiased boosting with categorical features","cited_arxiv_id":null,"evidence_quote":"CatBoost is a leading gradient-boosting baseline in the main comparison table."},{"cited_title":"Vime: Extending the success of self-and semi-supervised learning to tabular domain","cited_arxiv_id":null,"evidence_quote":"VIME is the primary prior contrastive self-supervised tabular baseline."},{"cited_title":"Revisiting deep learning models for tabular data.Advances in Neural Information Processing Systems, 34:18932– 18943, 2021","cited_arxiv_id":null,"evidence_quote":"FT-Transformer is the column-attention encoder design referenced for feature-level attention."}],"review_version":1}