REVIEW 5 major objections 5 minor 22 references
EAPCR: A Universal Feature Extractor for Scientific Data without Explicit Feature Relation Patterns
T0 review · 5 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read 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…
desk verdict 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. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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.
What would settle it
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.
Extended reading notes
Core claim
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%.
Load-bearing premise
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.
Editorial extensions
If this is right
- 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.
Reading between the lines
- 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.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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.
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 (5)
- [Sec. 3.2, App. A, Eq. (2), App. G] 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.
- [App. F, Props. F.1-F.3] 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.
- [Secs. 3.1, 3.3, Tabs. 1-3] 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.
- [Sec. 2.2, Eqs. (1)-(2)] 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.
- [Sec. 3.2, Fig. 3, App. C] 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.
minor comments (5)
- [Sec. 2.1] 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.
- [Tab. 5] The table layout is garbled and contains a typo 'EACPR' instead of 'EAPCR'; the CNN architecture details are difficult to parse.
- [Sec. 3.2, Fig. 3] 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.
- [App. G] The recall rates are reported without defining the binarization threshold or the matching procedure; without precision and baselines these numbers are hard to interpret.
- [App. H] 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.
Circularity Check
The synthetic robustness benchmark is circular: EAPCR's permuted CNN uses the same designed permutation that generates the 'FRP-free' data, so the claimed recovery of hidden FRPs is wired in by construction.
-
self definitional
[Sec. 3.2 (synthesized data); Sec. 2.2 Eq. (2); App. A; App. G]
"However, our approach differs by using a carefully designed permutation, as outlined in Sec. 2.2, which strategically moves adjacent matrix elements further apart and brings distant elements closer (Fig. 3-a). ... Although both the permuted CNN in EAPCR and the synthesized data use a permutation matrix that disrupts the original spatial relationships, they are not identical. For instance, the synthesized data matrix is [28, 28], while the model's matrix is [784, 784] ... We apply a designed permutation matrix M on A giving a new matrix P, defined as: P ≜ M AM⊤."
App. A defines one permutation M (arrange 784 elements into an R×L matrix, transpose, reshape). The synthetic 'FRP-free' MNIST is generated with exactly this designed permutation, and EAPCR's permuted CNN applies the same M via Eq. (2). Both act on the same 784 flattened feature indices; displaying the image as [28,28] while M is [784,784] is only a reshape, not a different permutation. For the square 28×28 case, the App. A permutation is transposition and hence self-inverse. If A learns the correlation matrix of the permuted data, A ≈ M C M^T, then P = M A M^T ≈ C, the original correlation matrix. Thus the headline synthetic result and App.
full rationale
The real-data benchmarks (heart disease, TiO2 catalysis, sensor failure, lung/breast cancer, thermocatalysis, centrifugal pumps) are external datasets with fixed feature sets and are not circular: superiority there is an empirical claim that does not reduce to EAPCR's own construction. The central weakness is the synthetic robustness experiment in Sec. 3.2, which is the paper's load-bearing evidence for recovering hidden FRPs and for 'universal' robustness. That experiment is circular because the same designed permutation from App. A both destroys the FRPs in the data and is used as M in the permuted CNN, making the apparent recovery of the original correlation matrix a definitional consequence of Eq. (2) rather than evidence of discovering arbitrary feature relations. The self-citations in the paper (Zhou et al., Gao et al., Liu et al.) appear in peripheral discussions of data scarcity and are not load-bearing, so no self-citation chain circularity is present. Appendix F's claim that independent features cannot interact is mathematically incorrect (e.g., XOR with independent fair bits), but that is a correctness flaw rather than a circularity; it does not affect this score. Because the circularity affects a central robustness demonstration while the real-data results remain independent, a partial-circularity score of 6 is appropriate.
Assumptions & free parameters
free parameters (4)
- Discretization thresholds for numerical features =
not specified in paper
- Embedding size Es =
64, 128, or 256 depending on dataset
- Permutation matrix factorization R x L =
e.g., 28 x 28 for N=784
- CNN architecture hyperparameters =
kernel sizes 3x3 or 4x4, channels 4-16, layer counts as in Tabs. 5-6
assumptions (4)
- domain assumption Pairwise products of learned embeddings in A=Tanh(E E^T) capture all feature interactions that matter for prediction.
- ad hoc to paper Important interacting features are correlated, so sampling correlated pairs is sufficient.
- domain assumption Discretizing numeric features into categorical strings preserves the information needed for prediction.
- ad hoc to paper Applying a designed permutation M to A improves sampling of distant feature relations without harming local ones.
Cite this review
Pith. "Pith review of EAPCR: A Universal Feature Extractor for Scientific Data without Explicit Feature Relation Patterns." pith.science (2026). https://pith.science/paper/5WWVJUDY
@misc{pith2026241108164,
author = {Pith},
title = {Pith review of: EAPCR: A Universal Feature Extractor for Scientific Data without Explicit Feature Relation Patterns},
year = {2026},
howpublished = {\url{https://pith.science/paper/5WWVJUDY}},
note = {Machine review of arXiv:2411.08164}
}
read the original abstract
Conventional methods, including Decision Tree (DT)-based methods, have been effective in scientific tasks, such as non-image medical diagnostics, system anomaly detection, and inorganic catalysis efficiency prediction. However, most deep-learning techniques have struggled to surpass or even match this level of success as traditional machine-learning methods. The primary reason is that these applications involve multi-source, heterogeneous data where features lack explicit relationships. This contrasts with image data, where pixels exhibit spatial relationships; textual data, where words have sequential dependencies; and graph data, where nodes are connected through established associations. The absence of explicit Feature Relation Patterns (FRPs) presents a significant challenge for deep learning techniques in scientific applications that are not image, text, and graph-based. In this paper, we introduce EAPCR, a universal feature extractor designed for data without explicit FRPs. Tested across various scientific tasks, EAPCR consistently outperforms traditional methods and bridges the gap where deep learning models fall short. To further demonstrate its robustness, we synthesize a dataset without explicit FRPs. While Kolmogorov-Arnold Network (KAN) and feature extractors like Convolutional Neural Networks (CNNs), Graph Convolutional Networks (GCNs), and Transformers struggle, EAPCR excels, demonstrating its robustness and superior performance in scientific tasks without FRPs.
Figures
Figures from the paper (4 more)
Reference graph
Works this paper leans on
-
[9]
Semi-supervised classification with graph convolutional networks
Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907,
-
[10]
toronto.edu/~kriz/learning-features-2009-TR.pdf
URL https://www.cs. toronto.edu/~kriz/learning-features-2009-TR.pdf . Dana Lahat, Tülay Adali, and Christian Jutten. Multimodal data fusion: an overview of methods, challenges, and prospects. Proceedings of the IEEE, 103(9):1449–1477,
work page 2009
-
[12]
The specific experimental results are shown in Tab
Table 11: Comparison of our method with others in the diagnosis of Lung cancer dataset Method Accuracy Precision Recall F1 Score AUC Bagging 89.76% 91.88% 89.35% 90.00% 95.30% AdaBoost 90.70% 90.70% 90.70% 90.70% 97.62% LightGBM 92.56% 93.93% 92.10% 92.71% 92.71% XGBoost 94.42% 95.66% 94.46% 94.74% 98.14% EAPCR 96.30% 96.49% 96.49% 96.49% 98.61% Table 12:...
work page 2023
-
[13]
Umap: Uniform manifold approximation and projection for dimension reduction
Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426,
-
[14]
Efficient estimation of word representations in vector space
Tomas Mikolov. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781,
-
[15]
Automated flower classification over a large number of classes
Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pp. 722–729. IEEE,
2008
-
[17]
URLOST: Unsupervised Representation Learning without Stationarity or Topology
Zeyu Yun, Juexiao Zhang, Bruno Olshausen, Yann LeCun, and Yubei Chen. Urlost: Unsupervised representation learning without stationarity or topology. arXiv preprint arXiv:2310.04496,
-
[18]
Here we test GCN only on data without FRPs
Linear1(128,10) Linear1(784,128) Linear2(128,10) relationships between nodes. Here we test GCN only on data without FRPs. When GCN with adjacency matrices (AM) given by the correlation matrix recovered from simple statistics on the data, its performance slightly improved, achieving an accuracy of 91.2%, as shown in Fig. 3-c-4. However, the classification ...
work page 2017
Show all 22 references
-
[21]
The dataset contains 5,118 rows of measurements from two centrifugal pumps from the same manufacturer
More system anomaly detection: Centrifugal pumps dataset (Mallioris et al., 2024). The dataset contains 5,118 rows of measurements from two centrifugal pumps from the same manufacturer. These measurements include key features: value_ISO, value_DEM, value_ACC, value_P2P, value_...
2024
-
[22]
/ 0.073 / 0.927 EAPCR 0.131 ± 0.004 0.054 ± 0.001 0.233 ± 0.003 0.940 ± 0.002 Table 14: Comparison of our method with others in the Inorganic catalysis data set. Model MAE MSE RMSE R2 RF / / 3.40 0.89 EAPCR 1.27 ± 0.02 2.88 ± 0.09 1.69 ± 0.02 0.97 ± 0.00 Table 15: Comparison o...
2022
-
[48]
Fe 0 500 MB 100 7 0 545 60 28.1Fe 0 500 MB 100 7 0 545 120 44Fe 0 500 MB 100 7 0 545 180 52.8Fe 0 500 MB 100 7 0 545 300 69 Table 10: Sensor measurements dataset footfall tempMode AQ USS CS VOC RP IP Temperature fail 0 7 7 1 6 6 36 3 1 1 190 1 3 3 5 1 20 4 1 0 5 5 3 3 6 1 24 6...
2022
-
[1986]
Modeling relational data with graph convolutional networks
12 Michael Schlichtkrull, Thomas N Kipf, Peter Bloem, Rianne Van Den Berg, Ivan Titov, and Max Welling. Modeling relational data with graph convolutional networks. In The semantic web: 15th international conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018, proceedi...
2018
-
[1989]
11 Jin-Hwa Kim, Jaehyun Jun, and Byoung-Tak Zhang
DOI: https://doi.org/10.24432/C52P4X. 11 Jin-Hwa Kim, Jaehyun Jun, and Byoung-Tak Zhang. Bilinear attention networks. Advances in neural information processing systems, 31,
-
[1998]
Pre-trained cnns as visual feature extractors: A broad evaluation
Andrew Holliday and Gregory Dudek. Pre-trained cnns as visual feature extractors: A broad evaluation. In 2020 17th conference on computer and robot vision (CRV), pp. 78–84. IEEE,
2020
-
[2009]
Bert: Pre-training of deep bidirectional transformers for language understanding
Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805,
-
[2013]
Feature selection via mutual information: New theoretical insights
Mario Beraha, Alberto Maria Metelli, Matteo Papini, Andrea Tirinzoni, and Marcello Restelli. Feature selection via mutual information: New theoretical insights. In 2019 international joint conference on neural networks (IJCNN), pp. 1–9. IEEE,
2019
-
[2017]
doi: 10.1007/s11222-016-9646-1
ISSN 1573-1375. doi: 10.1007/s11222-016-9646-1. Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30,
-
[2018]
An image is worth 16x16 words: Transformers for image recognition at scale
Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929,
2010 arXiv
-
[2020]
Multimodal compact bilinear pooling for visual question answering and visual grounding
Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach. Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847,
-
[2022]
Imagenet: A large-scale hierarchical image database
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp. 248–255. Ieee,
2009
-
[2023]
Kan 2.0: Kolmogorov-arnold networks meet science
Ziming Liu, Pingchuan Ma, Yixuan Wang, Wojciech Matusik, and Max Tegmark. Kan 2.0: Kolmogorov-arnold networks meet science. arXiv preprint arXiv:2408.10205, 2024a. Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Solja ˇci´c, Thomas Y . Hou, and Ma...
-
[2024]
Lung cancer prediction model using ensemble learning techniques and a systematic review analysis
Muntasir Mamun, Afia Farjana, Miraz Al Mamun, and Md Salim Ahammed. Lung cancer prediction model using ensemble learning techniques and a systematic review analysis. In 2022 IEEE World AI IoT Congress (AIIoT), pp. 187–193. IEEE,
2022
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.