{"id":"7a4fa82c-e451-4230-9daa-292e1510f179","arxiv_id":"2411.08164","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"An embedding plus bilinear attention plus permuted CNN architecture reports higher accuracy than the cited tree and deep baselines on several small scientific tabular datasets.","lead":"This paper proposes EAPCR, a deep-learning module that turns scientific tables into a learned pairwise feature-combination matrix and samples it with small CNNs. The authors report that EAPCR beats tree-based models on heart-disease, catalysis, and sensor-failure tasks, and on synthetic benchmarks with shuffled feature positions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The synthetic benchmark is circular: the data and EAPCR's permuted CNN use the same designed permutation, so the claimed recovery of hidden FRPs may be an artifact; the Appendix F independence proof is also false.","rationale":"The reader's stated weakest assumption is Appendix F's false claim that independent features cannot jointly determine the target. I agree that this proposition is mathematically wrong, and it undermines the paper's justification for focusing on the pairwise correlation matrix. However, the more directly load-bearing flaw is the circular synthetic benchmark: it is the only experiment that claims to demonstrate discovery of truly unknown FRPs, and it uses the same designed permutation in the data generator and in the model's permuted CNN. The model's P≈C derivation is exact for linear correlation matrices and approximately valid for learned embeddings, so the 'recovery' in App G may be pre-wired. The real-world tables (Tabs. 1-3, 11-15) are small, lack error bars on most metrics, and compare against a handful of baselines, so they cannot carry the universal-extractor claim by themselves. Because the central claim relies on this circular evidence, the rejection verdict stands; a non-circular permutation test or an XOR-style independent-feature interaction test would provide the needed support.","tokens_in":19309,"tokens_out":9712,"duration_ms":104736,"concrete_test":"Run the Sec. 3.2 MNIST experiment under four conditions: (i) data permuted with P_data and model M=P_data (the current design); (ii) data permuted with P_data but model M=random permutation; (iii) data permuted with a random permutation and model M=P_data; (iv) both data and model use independent random permutations. Compare accuracy and the App G correlation-recovery recall. If conditions (ii)-(iv) retain ~94.5% accuracy and ~84.6% recall, the circularity concern fails; if accuracy and recall drop toward the MLP/CNN baseline (~90%) whenever M is not the inverse of the data permutation, the synthetic result is an artifact of sharing the permutation construction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing evidence for the universal-extractor claim is the synthetic benchmark in Sec. 3.2, where FRPs are destroyed by applying a designed permutation to MNIST pixels, and App G reports that EAPCR recovers the original correlation matrix. But EAPCR's permuted CNN is built from the same construction: the M in Eq. (2) is generated (App A) by arranging N=784 elements into an R×L matrix, transposing, and reshaping, exactly the operation applied when synthesizing the data. For permuted features y_i=x_{π(i)}, the pairwise matrix is C'=P_π C P_π^T; applying the same P_π as M yields P=M A M^T≈C, so the permuted branch can see the original image correlations. The Sec. 3.2 disclaimer that the data matrix is [28,28] while the model matrix is [784,784] does not break this equivalence: both use the same permutation on the same 784 feature indices, merely displayed at different resolutions. Thus the headline result and App G recall may be an artifact of giving the model the inverse permutation, not evidence of discovering arbitrary hidden FRPs. The Appendix F proof that independent features cannot interact (Props. F.1-F.3) is also false (e.g., Y=A xor B with independent fair binary A,B), so it cannot justify restricting attention to correlated pairs.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EAPCR, a feature extractor for tabular scientific data that the authors claim lacks explicit feature relation patterns (FRPs). The method first converts each feature to a categorical index, learns an embedding matrix E of shape [N, Es], forms A = Tanh(E E^T), and then applies a lightweight CNN both to A and to a permuted version P = M A M^T, where M is a designed permutation matrix; an MLP residual branch is added. The authors report results on heart disease, TiO2 catalysis, and sensor-failure datasets, plus a synthetic benchmark in which MNIST images are permuted to remove spatial correlations. They claim EAPCR consistently outperforms trees, kernel methods, CNNs, GCNs, Transformers, and KANs, and that it recovers hidden FRPs.","tokens_in":19563,"tokens_out":6030,"duration_ms":63099,"significance":"If the central claim were true, EAPCR would be a useful and simple general-purpose tabular representation that could close the gap between deep learning and tree ensembles on scientific data. The paper deserves credit for stating a clearly motivated problem, for describing the permutation construction in an understandable way, and for including an ablation that removes the permuted CNN. However, the controlled synthetic experiment is constructed from the same permutation that EAPCR itself uses, the information-theoretic appendix intended to justify the design is mathematically false, and the real-data comparisons are too small and too weakly reported to support a universal-extractor claim. The strengths are the clarity of the problem statement and the cheap architecture; the load-bearing evidence for the headline claim is not present.","major_comments":[{"comment":"The synthetic no-FRP construction and the model's second CNN branch use the same permutation. For a feature vector x, the synthetic data is y_i = x_{pi(i)}; the pairwise covariance matrix becomes C' = P_pi C P_pi^T. Equation (2) then applies P = M A M^T with M = P_pi, so the permuted branch can reconstruct the original correlation matrix C. The defense in Sec. 3.2 that the data matrix is [28,28] while the model matrix is [784,784] does not break this equivalence: both operate on the same 784 features and the same index permutation. Appendix G's reported recall of 84.6% is therefore consistent with the model being given the inverse permutation rather than discovering arbitrary hidden FRPs. The central controlled demonstration is circular.","section":"Sec. 3.2, App. A, Eq. (2), App. G"},{"comment":"The appendix's justification that only correlated feature pairs need be sampled is invalid. Proposition F.1 assumes H(Y|A,B) = H(Y|A) + H(Y|B) - H(Y), an independence condition on conditional entropies that does not follow from marginal independence of A and B and is generally false. As a concrete counterexample, for independent fair binary A and B with Y = A XOR B, IG(Y,A) = IG(Y,B) = 0 while IG(Y,A,B) = 1; A and B are independent yet interact. Hence Props. F.2 and F.3, which are used to restrict the model to correlated pairs, do not hold. This matters because the matrix A in Eq. (1) represents only pairwise embedding products; the paper gives no valid argument that this captures all relevant higher-order interactions.","section":"App. F, Props. F.1-F.3"},{"comment":"The real-data evidence is too weak to support 'consistently outperforms' and 'universal'. The three main datasets have 303, 760, and 944 samples with 13, 9, and 9 features; Table 1 baselines are taken from a different paper with no shared split or protocol; Table 2 reports no variance for LightGBM; Table 3 omits precision, recall, and F1 values for most baselines and gives no error bars for any model. Appendix E's additional datasets use baselines from prior works rather than re-run comparisons. Without a common evaluation protocol, uncertainty intervals, and statistical tests, the claimed margins over tree ensembles are not established.","section":"Secs. 3.1, 3.3, Tabs. 1-3"},{"comment":"The paper states that A and P 'expose' and 'efficiently sample' all possible feature relations, but no argument or ablation establishes this. A is an N x N pairwise matrix, so any signal requiring third-order or structured interactions can be absent; P is a single fixed rearrangement of A, not a set of rearrangements or a mechanism for exhaustively sampling all combinations. The 'exposes all possible feature relations' claim is therefore not supported by the construction. This is load-bearing because the universal-extractor claim rests on this coverage.","section":"Sec. 2.2, Eqs. (1)-(2)"},{"comment":"The synthetic benchmark is performed on MNIST and other image datasets whose features are pixels with a destroyed spatial layout. This does not model the heterogeneous, differently scaled, categorical-and-integer scientific features that the paper motivates. Furthermore, the comparison of designed versus random permutation in App. C reports a single 1.3% accuracy difference with no variance or repetitions, which is insufficient to show that the designed permutation helps.","section":"Sec. 3.2, Fig. 3, App. C"}],"minor_comments":[{"comment":"Calling Eq. (1) 'bilinear attention' is misleading because no attention weights or query/key/value structure is defined; clarify the relation to bilinear attention or rename the operation.","section":"Sec. 2.1"},{"comment":"The table layout is garbled and contains a typo 'EACPR' instead of 'EAPCR'; the CNN architecture details are difficult to parse.","section":"Tab. 5"},{"comment":"The claim that EAPCR performs consistently across raw, randomly permuted, and designed-permutation data is stated but no quantitative results are shown for all three conditions.","section":"Sec. 3.2, Fig. 3"},{"comment":"The recall rates are reported without defining the binarization threshold or the matching procedure; without precision and baselines these numbers are hard to interpret.","section":"App. G"},{"comment":"The statement that code and data will be released after peer review means that the experiments currently cannot be reproduced; for a method paper this is a significant practical limitation.","section":"App. H"}],"recommendation":"reject","confidential_remarks":"The paper has a potentially salvageable idea, but the main controlled experiment is circular and the theoretical support is false. A direct fix would be to test the method on a synthetic benchmark where the hidden permutation is a held-out random permutation rather than the model's designed permutation, and to rerun all baselines under a common protocol with uncertainty estimates. As submitted, the central claims are not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes EAPCR, a combination of categorical embeddings, a bilinear attention matrix, a permuted CNN, and a residual MLP, aimed at tabular scientific data where feature relations are unknown. The problem is real, and the architecture is clearly presented. On the three real datasets (heart disease, TiO2 catalysis, sensor failure), EAPCR reports strong numbers against the baselines they chose. The designed permutation is a neat idea, and the ablation showing it beats a random permutation is a useful datapoint.\n\nBut the central evidence for the \"universal\" claim has a load-bearing flaw. The synthetic benchmark is built by applying a designed permutation to MNIST pixels, and EAPCR's permuted CNN uses the same designed permutation defined by the same row/column transposition. The paper's disclaimer that the data matrix is [28,28] while the model matrix is [784,784] does not break the equivalence: both operations act on the same 784 feature indices, just reshaped differently. So the model may simply be given the inverse permutation, which would explain its success on the \"no-FRP\" data and the 84.6% recall in App. G. That circularity makes the headline robustness result uninterpretable.\n\nThe theoretical support is also wrong. Appendix F claims that independent features cannot jointly inform Y and proves it by assuming H(Y|A,B)=H(Y|A)+H(Y|B)-H(Y), which is false in general. A simple XOR over two independent fair bits gives a counterexample. So the argument that only correlated pairs matter collapses, and the \"all possible FRPs\" framing is doubly misleading because A=Tanh(EE^T) only captures pairwise interactions anyway.\n\nThe empirical evaluation is thin: small datasets, most baselines borrowed from other papers with no shared splits, no variance reported for many comparisons, no modern tabular deep learners (FT-Transformer, SAINT, TabNet), and no code despite the promise in App. H. The real-world results may still be genuine, but they only show EAPCR beating a handful of baselines on a few small datasets.\n\nWho is this for? A reader working on tabular deep learning might take the permutation trick as a starting point, but should not rely on the universality claim. The paper deserves a serious referee because the core question matters and the flaw is instructive, but it needs major revision: release code, fix the synthetic benchmark so the model's permutation is independent of the data corruption, add proper baselines and statistical tests, and correct or remove App. F. I'd engage with it as a cautionary example, not as a method I'd adopt.","headline":"Interesting architecture, but the main synthetic evidence is circular (same permutation in data and model) and the theoretical appendix is mathematically wrong; the universal-extractor claim is not established.","tokens_in":20146,"tokens_out":2286,"would_cite":false,"duration_ms":26206,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","68T05"],"pacs":[],"model":"deepseek-v4-flash","headline":"EAPCR is a feature extractor for tabular scientific data without explicit feature relation patterns; it builds A=Tanh(E E^T), runs a lightweight CNN on A and on a hand-designed permutation of it, and reports consistent wins over…","keywords":["EAPCR","feature relation patterns","feature extraction","tabular scientific data","bilinear attention","permuted CNN","deep learning without explicit topology"],"falsifier":"Construct a synthetic dataset with, say, 8 binary features in which the label is the XOR of two mutually independent features plus small noise, while all other features are pure noise. If EAPCR cannot exceed chance on this dataset while a decision tree or MLP can, then pairwise-correlation sampling is not sufficient to capture all feature interactions and the central claim fails.","tokens_in":19050,"feed_emoji":"📊","tokens_out":6101,"duration_ms":52076,"temperature":0.7,"pith_summary":"The paper addresses a blind spot in deep learning: scientific tabular data whose features have no known spatial, sequential, or graph relationships. It proposes EAPCR, a feature extractor that converts every feature into an embedding, forms the pairwise correlation matrix A=Tanh(E E^T), and runs a lightweight CNN on both A and a hand-designed permutation of it. The claim is that this exposes all candidate feature interactions and samples them efficiently enough to beat decision-tree-based methods on non-image medical diagnosis, inorganic catalysis prediction, and sensor-failure detection. On a synthetic dataset built by permuting MNIST pixels, the paper reports that EAPCR keeps high accuracy while CNNs, GCNs, Transformers, and KANs degrade.","feed_headline":"Pairwise matrix lets deep nets beat trees on scientific data","feed_subtitle":"EAPCR's pairwise matrix and permuted CNN beat tree ensembles on medical, catalysis, and sensor tasks.","key_machinery":"The central object is the correlation matrix A=Tanh(E E^T), which turns all pairwise feature embeddings into a single N×N matrix that claims to expose every possible feature relation pattern. The designed permutation matrix M (built by arranging indices in an R×L grid, transposing, and reshaping) produces P=M A M^T, and a lightweight CNN applied to both A and P serves as an efficient sampler of local and non-local element combinations. A residual MLP over average-pooled embeddings keeps the embedding layer trainable.","core_discovery":"EAPCR is a universal feature extractor for data without explicit Feature Relation Patterns (FRPs). Given N features, it embeds each into a vector, builds A=Tanh(E E^T) so every entry encodes the relation between one pair of features, and applies a permuted CNN: a second matrix P=M A M^T rearranges A so originally distant matrix elements become local, letting a two-layer CNN sample both nearby and far-apart feature combinations. An MLP residual path on the pooled embeddings is added. The paper reports that this module outperforms SVM, logistic regression, naive Bayes, kNN, decision trees, random forests, XGBoost, LightGBM, CNN, GCN, Transformer, and KAN across several scientific tasks, and that on a synthetic FRP-free dataset derived from images it recovers the hidden pixel-correlation structure with reported recall around 84.6%.","pith_inferences":["The method's success implies that for many tabular scientific problems, explicit feature-relation priors are unnecessary; a learnable pairwise correlation map plus local convolution may be enough, which could extend to inspection of A as an interpretability tool.","If pairwise correlations suffice, then higher-order interactions (three or more features jointly determining the target without pairwise correlation) remain an open test; the paper's own Appendix F claims independence implies no interaction, which does not hold for patterns like XOR.","A testable extension: replace the hand-designed permutation M with a learned or data-dependent permutation, or with several permutations, to see whether sampling diversity rather than the specific construction is what matters.","The reported recall of hidden correlation recovery suggests A could serve as an inspectable diagnostic for which features the network believes interact, potentially aiding scientific hypothesis generation."],"forward_implications":["On non-image medical diagnostics (heart disease, lung cancer, breast cancer, maternal risk), EAPCR reports accuracy gains over the best classical baselines, e.g., 93% versus 85% SVM on the Cleveland heart-disease set.","On inorganic catalysis prediction, EAPCR reports R2 of 0.937 on the TiO2 dataset, ahead of LightGBM's 0.928, and higher R2 on two further catalysis sets.","On system anomaly detection, EAPCR reports the top F1 score (88.64%) on the machine-failure sensor set and perfect F1 on the centrifugal-pump set.","On synthetic FRP-free datasets derived from MNIST, Flower, ImageNet subsets, and CIFAR-10, EAPCR keeps high accuracy while ConvNeXt-V2 and other CNNs degrade, suggesting feature extraction no longer depends on a predefined topology.","The ablation EACR (no permutation) still beats MLP, but EAPCR with permutation reaches similar accuracy with fewer parameters, indicating permutation sampling rather than raw capacity drives the gain."],"supporting_citations":[{"why":"Supplies the embedding mechanism that converts each categorical feature into a dense vector used in A=Tanh(E E^T).","marker":"Mikolov, 2013"},{"why":"Is the bilinear attention construction the paper adapts to build the pairwise correlation matrix.","marker":"Kim et al., 2018"},{"why":"Provides the CNN machinery that samples local regions of A and P.","marker":"LeCun et al., 1998"},{"why":"Grounds the argument that CNNs, GCNs, and other deep models rely on explicit relation patterns such as spatial or graph structure.","marker":"Bronstein et al., 2017"},{"why":"Supplies the random-permutation baseline the paper contrasts with its designed permutation when synthesizing FRP-free data.","marker":"Yun et al., 2023"},{"why":"Is the TiO2 photocatalysis dataset and LightGBM baseline that EAPCR is benchmarked against.","marker":"Liu et al., 2022"},{"why":"Is the heart-disease study whose SVM/kNN/logistic baselines EAPCR is compared with.","marker":"Anderies et al., 2022"}],"fun_headline_variants":["EAPCR: deep learning that beats trees without feature relations","For data lacking feature relations, EAPCR outperforms all baselines","EAPCR: universal extractor for data without feature relations","EAPCR: a deep net beats trees on features without relations","Pairwise matrix and permuted CNN: EAPCR wins on unordered data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that any feature interaction that matters for the target appears as a pairwise correlation between features, so the matrix A=Tanh(E E^T) contains all the signal; if two independent features jointly determine the outcome (as in an XOR relation), the pairwise sampling can miss the relevant combination.","fun_headline_variants_meta":{"raw":{"variants":["EAPCR: deep learning that beats trees without feature relations","For data lacking feature relations, EAPCR outperforms all baselines","EAPCR: universal extractor for data without feature relations","EAPCR: a deep net beats trees on features without relations","Pairwise matrix and permuted CNN: EAPCR wins on unordered data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000927,"raw_usage":{"total_tokens":3984,"prompt_tokens":969,"completion_tokens":3015,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":2923}},"tokens_in":585,"tokens_out":3015,"duration_ms":19531,"temperature":1.0,"reasoning_tokens":2923,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:54:42.707460+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a synthetic dataset with, say, 8 binary features in which the label is the XOR of two mutually independent features plus small noise, while all other features are pure noise. If EAPCR cannot exceed chance on this dataset while a decision tree or MLP can, then pairwise-correlation sampling is not sufficient to capture all feature interactions and the central claim fails.","supporting_citations":[],"review_version":1}