Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Detecting Entanglement in High-Spin Quantum Systems via a Stacking Ensemble of Machine Learning Models

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that a stacking ensemble of a neural network, XGBoost, and Extra Trees, combined by a CatBoost meta-learner, accurately predicts entanglement negativity in pure and Werner states across spin dimensions, and that its…

desk verdict Plausible regression result undercut by data-construction and benchmark-leakage issues; strong claims need revision. read the letter →

arxiv 2507.12775 v1 pith:ETN4W2BI submitted 2025-07-17 quant-ph

classification quant-ph
keywords quantumentanglementnegativitystackingensembleneuralnetworksXGBoostExtraTreeshigh-spinsystemsWernerstates
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that machine-learning ensembles can serve as a fast surrogate for a standard entanglement quantifier: the negativity of a bipartite state, computed from the partial transpose. It shows that a stacking ensemble, built from three base regressors and a CatBoost meta-learner, reproduces negativity for random pure states and mixed Werner states with spin $J = 0.5$, $1$, and $5$. The central claim is not just high aggregate accuracy, but that the ensemble's individual predictions scatter less around the true negativity than those of a standalone neural network, making it more trustworthy for scientific use. If true, this offers a way to characterize entanglement in high-dimensional systems without repeated partial-transpose computations, at the cost of generating training data.

What carries the argument

The load-bearing object is the negativity measure $N(\rho_{AB}) = (\|\rho_{AB}^{T_B}\|_1 - 1)/2$, the trace-norm distance of the partial transpose from positivity, which labels every training state. The predictor is a stacking ensemble: base learners NN, XGBoost, and Extra Trees are trained on standardized state coefficients for pure states or flattened density-matrix entries for Werner states, and their out-of-fold predictions feed a CatBoost meta-learner. Stacking with out-of-fold predictions is what lets the meta-learner learn the optimal combination of base outputs without leakage, and that combination is the mechanism claimed to cancel correlated errors and reduce prediction variance.

What would settle it

A concrete check: take random pure bipartite states with exactly two nonzero coefficients $C_{mn}$ whose row and column indices both differ, so the coefficient matrix has rank two; these states are entangled by construction. Compute their exact negativity and compare with the ensemble's predictions: if the model, trained on a class balance that assumes sparse coefficients are mostly separable, predicts near-zero negativity for them, the sparsity-based labeling premise fails.

Watch

Extended reading notes

Core claim

The central discovery, as the authors state it, is that stacking heterogeneous regressors yields a reliable estimator of negativity across a range of spin dimensionalities. For pure states, predictions on random states reach $R^2 = 0.9999$ for $J=0.5$, $0.9962$ for $J=1$, and $0.9717$ for $J=5$; for mixed Werner states, $R^2 = 0.9997$, $0.9977$, and $0.9928$ respectively. The ensemble also reproduces the exact negativity curve of the tunable state $\cos(\theta)|-J,-J\rangle + \sin(\theta)|J,J\rangle$ and the Werner-state separability threshold in $\alpha$. The paper's distinctive claim is that the stacked model outperforms each base learner in deviation and consistency even when aggregate metrics are close, attributing this to error cancellation and variance reduction.

Load-bearing premise

The load-bearing premise is that sparse coefficient vectors in Eq. (8) describe separable or nearly separable states, because 90% of the training set is labeled that way; but for pure bipartite states separability means the coefficient matrix factors as a single product, which generic sparse superpositions are not.

Editorial extensions

If this is right

  • Entanglement in a new high-spin bipartite state can be estimated directly from its coefficients or density-matrix entries, without recomputing the partial transpose and trace norm.
  • The ensemble tracks the exact negativity of the tunable pure state $\cos(\theta)|-J,-J\rangle+\sin(\theta)|J,J\rangle$ and of Werner states, including the separability threshold in $\alpha$, so it has learned physical structure rather than memorized averages.
  • Data requirements grow roughly exponentially in $J$; the empirical formula $\log_{10}(S)\approx 2.8+0.502J-3.042\,\mathrm{MSE}-8.012\,\mathrm{MAE}+1.012 R^2$ lets future studies estimate the sample size needed for a target accuracy.
  • Because ensemble predictions cluster more tightly around true negativity than the neural network's do, the method is claimed to be more reliable for individual states, not only on aggregate error.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the ensemble has learned the mapping from state coefficients to negativity rather than the specific training families, the same stacking recipe should transfer to other bipartite entanglement measures such as concurrence or log-negativity, since the input representation is generic.
  • A testable extension is to train on Haar-random complex coefficients instead of real Gaussian amplitudes; a sharp drop in accuracy would indicate the model exploits real-state symmetries rather than learning a general coefficient-to-negativity function.
  • The paper's scatter-plot consistency claim could be quantified by reporting per-state absolute-error distributions or prediction intervals for the ensemble versus the neural network, turning the visual observation into a statistical test.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a stacking ensemble regressor—composed of a neural network, XGBoost, and Extra Trees as base learners with CatBoost as the meta-learner—to predict the negativity of bipartite quantum states from state coefficients or density-matrix elements. The method is trained and tested on datasets of pure states and Werner states for J = 0.5, 1, and 5, reporting high R² values (up to 0.9999) and low error metrics. The authors also derive an empirical scaling law, Eq. (12), intended to estimate the number of training samples needed for a given spin dimension and target accuracy, and they benchmark the ensemble on the tunable pure family |ζ(θ)⟩ and on Werner states, with reported near-perfect agreement.

Significance. If the reported results hold under rigorous out-of-sample testing, the ensemble would provide a practical ML surrogate for negativity that avoids direct partial-transpose computation in high-spin systems, which is a useful contribution to quantum-information tooling. The paper makes its data and code openly available on GitHub and Zenodo, which is a strength that enables independent verification. However, the current manuscript has several load-bearing methodological issues—the construction of 'predominantly separable' training states, the incorporation of benchmark families before the train/test split, and the descriptive rather than predictive nature of the scaling law—that must be resolved before the central claims can be accepted.

major comments (4)
  1. [Section 2.1, Eq. (8)] The claim that states with a controlled number of nonzero Cmn amplitudes are 'predominantly separable or exhibited very low entanglement' is not correct. For a pure bipartite state, separability requires the coefficient matrix C to have rank 1; a random sparse superposition of several product basis states with two or more nonzero amplitudes generically has rank > 1 and is entangled. For example, a state of the form cos(θ)|i,j⟩ + sin(θ)|k,l⟩ with (i,j) ≠ (k,l) is entangled for all θ except the endpoints. Since the paper states that roughly 90% of the training data was generated under this sparsity assumption, the actual target distribution (and class balance) of the training set is mischaracterized, which directly affects the interpretation of every reported performance metric.
  2. [Sections 2.1 and 3.2/3.3] The two benchmark families, |ζ(θ)⟩ and Werner states, are explicitly 'incorporated' into the comprehensive dataset before the data is partitioned 80/20 for training and testing. The subsequent evaluations in Figures 4 and 5 on these same families therefore are not guaranteed to be out-of-sample. If any benchmark states were included in the training portion, the near-perfect agreement with exact curves could reflect memorization rather than generalization. The authors should either exclude all benchmark states from the training data or provide explicit proof that the 20% test partition contains these benchmark points, and then re-evaluate.
  3. [Section 3.1, Eq. (12)] The scaling law log10(S) ≈ 2.8 + 0.502J − 3.042MSE − 8.012MAE + 1.012R² is a linear regression fitted to the very same performance curves (Figures 2 and 3) that it is claimed to explain. There is no independent validation set for this formula, and the predictors MSE, MAE, and R² are mathematically interdependent, so the fitted coefficients do not establish a predictive relationship. In particular, the statement that the positive coefficient for J 'quantitatively confirms' an exponential growth of S with J is not supported, because S was chosen by the authors at a few discrete values and the regression merely describes the observed metrics at those chosen sizes.
  4. [Sections 3.2, 3.3, and 4] The central claim that the ensemble exhibits 'superior predictive consistency and lower deviation' compared to individual learners is based on visual inspection of scatter plots, not on a paired statistical test or a quantitative consistency metric. Since Tables 1 and 2 show that the aggregate MSE/MAE/R² of the ensemble and the standalone NN are often close (e.g., J = 5 pure states: MSE 0.0011 vs 0.0014), the claim that the ensemble is more reliable for individual predictions is not demonstrated. A paired test on residuals or an explicit measure of prediction scatter (e.g., standard deviation of errors across repeated runs) is needed.
minor comments (5)
  1. [Section 2, Eq. (7)] The text states that normalized negativity ranges from 0 to 1, but the definition N(ρ) = (‖ρ^TB‖₁ − 1)/2 does not by itself produce a dimension-independent maximum of 1. The normalization convention used for the target variable across different J values should be stated explicitly, since it affects the interpretation of MSE values.
  2. [Figures 2 and 3] The x-axis is described as 'number of random states' in the captions but the text refers to training sample size; it should be clarified whether the reported metrics are computed on the training or test partition, and at what sample sizes the final models for each J are evaluated.
  3. [Section 2.1] The description of the 10% 'demonstrably entangled' class is vague ('ensuring a sufficient number of non-zero, appropriately distributed Cmn amplitudes'); please specify the concrete algorithm used to generate these states and the resulting range of target negativity values.
  4. [Throughout] Several equations and inline symbols have rendering issues (e.g., Eq. (1) and the CatBoost target statistic formula), and some parameter names appear with typographical inconsistencies ('nestimators' vs 'n_estimators'). A careful proofreading pass is recommended.
  5. [References] Reference [4] is described as 'a first-hand account by alain aspect'; please verify that this is the intended citation and format it consistently with the journal's style.

Circularity Check

2 steps flagged · score 5.0 of 10

Benchmark families are inserted into the training pool before the 80/20 split, so the 'stringent' tests on |ζ(θ)⟩ and Werner states may be in-sample; Eq. (12) is a fit to the curves it claims to explain.

  1. other [Section 2.1 (dataset generation), evaluated in Sections 3.2 and 3.3 (Figures 4d-f and 5d-f)]
    "To stringently evaluate the model’s predictive capabilities on well-characterized quantum states, two specific families were incorporated: 1. A class of tunable pure entangled states |ζ(θ)⟩ ... 2. Werner states ... The comprehensive dataset was partitioned using a standard 80% for training the models and the remaining 20% for rigorous, unseen testing and performance evaluation."

    The lower panels of Figures 4 and 5 then report the model's 'near-perfect overlay' and 'exceptional fidelity' on exactly these two families. Because the families were placed in the comprehensive dataset before the random 80/20 split, nothing ensures that the displayed benchmark points are out of sample; any that fell in the training 80% are memorized. The claimed 'stringent' validation is therefore not independent of the training data by construction: the test family is a subset of the same data pool used for training, so the reported agreement can be a restatement of fitted points rather than evidence of generalization.

  2. fitted input called prediction [Section 3.1, Eq. (12)]
    "Leveraging these observed performance trends, we sought to derive a phenomenological scaling law to estimate the number of training samples (S) required. A linear regression model was applied to the aggregated data from Figures 2 and 3, using performance metrics (MSE, MAE, R2) and the spin quantum number J as input features, with log10(S) as the target variable. This analysis yielded the following empirical relationship: log10(S) ≈ 2.8 + (0.502) J − (3.042) MSE − (8.012) MAE + (1.012) R2 (12)"

    The paper calls this a 'derived' scaling law and uses the coefficient +0.502 on J to 'quantitatively confirm' that sample size grows exponentially with spin. But Eq. (12) is a linear fit to the same Figures 2 and 3 performance curves, so the coefficient is fitted from the data it is then said to confirm. The formula is a re-expression of the training curves, not an independent prediction, and the dataset sizes it 'guides' (10^4, 2×10^4, 10^5) are the very sizes plotted in those figures; evaluating at those sizes is a self-consistency check rather than an out-of-sample validation.

full rationale

The ensemble regression of negativity from state coefficients is, in principle, a legitimate supervised-learning construction: the target is computed directly from the input state, and the random-state test metrics are based on an 80/20 split, so the core R2 values (0.9999, 0.9962, 0.9717 for pure states; 0.9997, 0.9977, 0.9928 for Werner states) are not circular by themselves. There is no load-bearing self-citation chain: the ML methods and negativity definition are standard, and no prior work of the authors is invoked to forbid alternatives. However, two elements undermine the claim of independent validation. First, the two benchmark families used for the 'stringent' lower-panel tests in Figures 4 and 5 are stated to have been incorporated into the comprehensive dataset before the 80/20 partition, so the displayed agreement with exact curves is not guaranteed to be out of sample; the test family is drawn from the same pool as the training data, making the validation partially circular and the claim of 'unseen' testing unsupported for those figures. Second, Eq. (12) is presented as a derived scaling law but is a linear regression fit to the same Figures 2 and 3 it is used to explain; the positive J coefficient is an estimate from those curves, so the 'confirmation' of exponential growth is a restatement of the fit. These are localized circularities: the random-state regression content remains independent, but the paper's most stringent tests and its empirical formula are weakened by construction. The sparsity-based class-balance assumption in Section 2.1 is a physical correctness risk, not a circular-derivation issue, and does not affect this score.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper's central results rest on standard supervised learning plus several choices: an artificial class balance, an unspecified negativity normalization, and a linear-regression formula fitted to its own curves. The most fragile assumption is that sparse random superpositions are mostly separable, which is not generally true. No new physical entities are introduced.

free parameters (5)
  • Eq. (12) regression coefficients = intercept 2.8, J 0.502, MSE -3.042, MAE -8.012, R2 1.012
    All five constants are fitted by linear regression on the performance data in Figures 2 and 3; they are neither derived nor validated on independent data.
  • Entangled-to-separable training mix = 10% entangled, 90% low/separable
    Chosen by hand in Section 2.1 to balance classes; the actual distribution of random pure states is not this mix.
  • Sparsity thresholds for non-zero Cmn amplitudes = not specified
    Section 2.1 sets a controlled number of non-zero amplitudes 'ranging from a single non-zero amplitude up to a threshold that typically maintains separability or minimal entanglement', but no explicit threshold is given.
  • ML hyperparameters = NN 128/64/32; XGB 300 trees, lr 0.05, depth 10; ET 300 trees, depth 15; CB up to 1000 with early stopping
    Chosen by hand in Section 2.1; no tuning or sensitivity analysis is reported.
  • Negativity normalization convention = not specified
    The paper says normalized negativity ranges from 0 to 1 but never states the normalization factor; the target labels may differ from standard negativity by a dimension-dependent factor.
assumptions (5)
  • domain assumption Negativity as defined in Eq. (7) is an adequate and unambiguous entanglement measure for the states tested.
    The paper never states the normalization used to make negativity range from 0 to 1, and zero negativity does not certify separability in general dimensions; the ML target labels inherit this convention.
  • standard math Sampling Gaussian random coefficients and normalizing approximates a uniform (Haar) distribution over real pure states.
    Section 2.1 relies on this for diverse state generation; it is standard for real states but the paper does not sample over complex phases.
  • ad hoc to paper Sparse coefficient vectors with few non-zero Cmn amplitudes are predominantly separable or weakly entangled.
    Section 2.1 uses sparsity to construct 90% low-entanglement training examples; this is false in general because a superposition of product states can be entangled even with few terms.
  • domain assumption A random 80/20 split of the full dataset, including the benchmark families, yields an unbiased held-out test.
    The tunable pure and Werner families are incorporated before splitting, so test curves for the same families may have near-duplicates in training; the paper does not describe an explicit seen/unseen protocol.
  • ad hoc to paper A linear function of J, MSE, MAE, and R2 is an appropriate model for the data-requirement scaling law.
    Section 3.1 fits Eq. (12) to the aggregated performance curves with no independent validation or residual analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Detecting Entanglement in High-Spin Quantum Systems via a Stacking Ensemble of Machine Learning Models." pith.science (2026). https://pith.science/paper/ETN4W2BI

@misc{pith2026250712775,
  author       = {Pith},
  title        = {Pith review of: Detecting Entanglement in High-Spin Quantum Systems via a Stacking Ensemble of Machine Learning Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ETN4W2BI}},
  note         = {Machine review of arXiv:2507.12775}
}
read the original abstract

Reliable detection and quantification of quantum entanglement, particularly in high-spin or many-body systems, present significant computational challenges for traditional methods. This study examines the effectiveness of ensemble machine learning models as a reliable and scalable approach for estimating entanglement, measured by negativity, in quantum systems. We construct an ensemble regressor integrating Neural Networks (NNs), XGBoost (XGB), and Extra Trees (ET), trained on datasets of pure states and mixed Werner states for various spin dimensions. The ensemble model with stacking meta-learner demonstrates robust performance by CatBoost (CB), accurately predicting negativity across different dimensionalities and state types. Crucially, visual analysis of prediction scatter plots reveals that the ensemble model exhibits superior predictive consistency and lower deviation from true entanglement values compared to individual strong learners like NNs, even when aggregate metrics are comparable. This enhanced reliability, attributed to error cancellation and variance reduction inherent in ensembling, underscores the potential of this approach to bypass computational bottlenecks and provide a trustworthy tool for characterizing entanglement in high-dimensional quantum physics. An empirical formula for estimating data requirements based on system dimensionality and desired accuracy is also derived.

Figures

Figures reproduced from arXiv: 2507.12775 by the authors.

Figure 1
Figure 1. The dataset is split into 80% for training and 20% for testing. An ensemble model using stacking is constructed with three base learners( Neural Networks, Extra Trees, and XGBoost) each trained inde￾pendently. Their predictions are used as input features for the meta-learner (CatBoost), which learns to optimally combine these predictions to produce the final, more accurate output systems characterized by high-dimens… view at source ↗
Figure 2
Figure 2. Pure state metrics with respect to the number of random states in the case of pure states [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Werner states metrics with respect to the number of random states in the case of Werner states. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Upper plates: Comparison of actual and predicted values for pure state (a) [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Upper plates: Comparison of actual and predicted values for Werner state (a) [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Scalable Entanglement Detection in Quantum Systems via Fisher Linear Discriminant Analysis

    quant-ph 2025-09 reject novelty 3.0 of 10

    Fisher Linear Discriminant Analysis, trained on Pauli expectation values from simulated two-, three-, and four-qubit states, classifies entangled versus separable states with reported accuracies between 33% and 100% d...

Reference graph

Works this paper leans on

28 extracted references · 28 canonical work pages · cited by 1 Pith paper

  1. [1]

    Entanglement-based quantum information technology: a tutorial

    Zheshen Zhang, Chenglong You, Omar S Maga˜na-Loaiza, Robert Fickler, Roberto de J Le´on-Montiel, Juan P Torres, Travis S Humble, Shuai Liu, Yi Xia, and Quntao Zhuang. Entanglement-based quantum information technology: a tutorial. Adv. Opt. Photonics, 16(1):60–162, 2024

  2. [2]

    Modern Quantum Theory: From Quantum Mechanics to En- tanglement and Quantum Information

    Reinhold Bertlmann and Nicolai Friis. Modern Quantum Theory: From Quantum Mechanics to En- tanglement and Quantum Information . Oxford University Press, 2023

  3. [3]

    Experimental tests of realistic local theories via bell’s theorem

    Alain Aspect, Philippe Grangier, and G ´erard Roger. Experimental tests of realistic local theories via bell’s theorem. Phys. Rev. Lett., 47(7):460, 1981

  4. [4]

    Experimental tests of bell’s inequalities: a first-hand account by alain aspect

    William D Phillips and Jean Dalibard. Experimental tests of bell’s inequalities: a first-hand account by alain aspect. Eur . Phys. J. D, 77(1):8, 2023

  5. [5]

    A secure and privacy-preserving signature protocol using quantum teleportation in metaverse environment

    Pankaj Kumar, Vivek Bharmaik, Sunil Prajapat, Garima Thakur, Ashok Kumar Das, Sachin Shetty, and Joel JPC Rodrigues. A secure and privacy-preserving signature protocol using quantum teleportation in metaverse environment. IEEE Access, 2024

  6. [6]

    Securing blockchain transactions using quantum teleportation and quantum digital signature

    Sheetal Singh, Nikhil Kumar Rajput, Vipin Kumar Rathi, Hari Mohan Pandey, Amit Kumar Jaiswal, and Prayag Tiwari. Securing blockchain transactions using quantum teleportation and quantum digital signature. Neural Process. Lett., pages 1–16, 2023

  7. [7]

    Navigating complex multiclass classification in high-dimensional spaces: A hybrid approach

    Praneeth Nemani, Venkata Surya Sundar Vadali, Prathistith Raj Medi, Ashish Marisetty, and Satya- narayana V ollala. Navigating complex multiclass classification in high-dimensional spaces: A hybrid approach. In 2023 7th International Conference on Computation System and Information Technology for Sustainable Solutions (CSITSS) , pages 1–6. IEEE, 2023. 18

  8. [8]

    The future of adaptive machine learning algorithms in high-dimensional data processing

    Anne Wilson and Muhammad Rehan Anwar. The future of adaptive machine learning algorithms in high-dimensional data processing. Int. Trans. Artif. Intell., 3(1):97–107, 2024

Show all 28 references
  1. [9]

    Scalable neural network algorithms for high dimensional data

    Mukesh Soni and Marwan Ali Shnan. Scalable neural network algorithms for high dimensional data. Mesopotamian J. Big Data, 2023:1–11, 2023

  2. [10]

    Nonlinear approximation and (deep) relu networks

    Ingrid Daubechies, Ronald DeV ore, Simon Foucart, Boris Hanin, and Guergana Petrova. Nonlinear approximation and (deep) relu networks. Constr . Approx., 55(1):127–172, 2022

  3. [11]

    Adaptive neural network control for a class of nonlinear systems with function constraints on states

    Yan-Jun Liu, Wei Zhao, Lei Liu, Dapeng Li, Shaocheng Tong, and CL Philip Chen. Adaptive neural network control for a class of nonlinear systems with function constraints on states. IEEE Trans. Neural Netw. Learn. Syst., 34(6):2732–2741, 2021

  4. [12]

    Convxgb: A new deep learning model for classification problems based on cnn and xgboost

    Setthanun Thongsuwan, Saichon Jaiyen, Anantachai Padcharoen, and Praveen Agarwal. Convxgb: A new deep learning model for classification problems based on cnn and xgboost. Nucl. Eng. Technol., 53(2):522–531, 2021

  5. [13]

    Tree boosting with xgboost-why does xgboost win” every” machine learning compe- tition? Master’s thesis, NTNU, 2016

    Didrik Nielsen. Tree boosting with xgboost-why does xgboost win” every” machine learning compe- tition? Master’s thesis, NTNU, 2016

  6. [14]

    Extremely randomized trees

    Pierre Geurts, Damien Ernst, and Louis Wehenkel. Extremely randomized trees. Mach. Learn., 63:3– 42, 2006

  7. [15]

    Online extra trees regressor

    Saulo Martiello Mastelini, Felipe Kenji Nakano, Celine Vens, Andr ´e Carlos Ponce de Leon Ferreira, et al. Online extra trees regressor. IEEE Trans. Neural Netw. Learn. Syst., 34(10):6755–6767, 2022

  8. [16]

    Multistrategy ensemble learning: Reducing error by combining ensemble learning techniques

    Geoffrey I Webb and Zijian Zheng. Multistrategy ensemble learning: Reducing error by combining ensemble learning techniques. IEEE Trans. Knowl. Data Eng., 16(8):980–991, 2004

  9. [17]

    A survey of ensemble learning: Concepts, algorithms, applications, and prospects

    Ibomoiye Domor Mienye and Yanxia Sun. A survey of ensemble learning: Concepts, algorithms, applications, and prospects. Ieee Access, 10:99129–99149, 2022

  10. [18]

    An improved ensemble learning method for classifying high-dimensional and imbalanced biomedicine data

    Hualong Yu and Jun Ni. An improved ensemble learning method for classifying high-dimensional and imbalanced biomedicine data. IEEE/ACM Trans. Comput. Biol. Bioinform., 11(4):657–666, 2014

  11. [19]

    High-dimensional ensemble learning classification: An ensemble learn- ing classification algorithm based on high-dimensional feature space reconstruction

    Miao Zhao and Ning Ye. High-dimensional ensemble learning classification: An ensemble learn- ing classification algorithm based on high-dimensional feature space reconstruction. Appl. Sci. , 14(5):1956, 2024

  12. [20]

    Mathematics of neural networks: models, algorithms and applications, volume 8

    Stephen W Ellacott, John C Mason, and Iain J Anderson. Mathematics of neural networks: models, algorithms and applications, volume 8. Springer Science & Business Media, 2012

  13. [21]

    Learning of neural networks based on weighted mean squares error function

    Yang Sai, Ren Jinxia, and Li Zhongxia. Learning of neural networks based on weighted mean squares error function. In 2009 Second International Symposium on Computational Intelligence and Design , volume 1, pages 241–244. IEEE, 2009

  14. [22]

    Multi-loss regularized deep neural network

    Chunyan Xu, Canyi Lu, Xiaodan Liang, Junbin Gao, Wei Zheng, Tianjiang Wang, and Shuicheng Yan. Multi-loss regularized deep neural network. IEEE Trans. Circuits Syst. Video Technol., 26(12):2273– 2283, 2015

  15. [23]

    Novel feature-based difficulty prediction method for math- ematics items using xgboost-based shap model

    Xifan Yi, Jianing Sun, and Xiaopeng Wu. Novel feature-based difficulty prediction method for math- ematics items using xgboost-based shap model. Mathematics, 12(10):1455, 2024. 19

  16. [24]

    Catboost for big data: an interdisciplinary review

    John T Hancock and Taghi M Khoshgoftaar. Catboost for big data: an interdisciplinary review. J. Big Data, 7(1):94, 2020

  17. [25]

    Comparison of the catboost classifier with other machine learning methods

    Abdullahi A Ibrahim, Raheem L Ridwan, Muhammed M Muhammed, Rabiat O Abdulaziz, and Ganiyu A Saheed. Comparison of the catboost classifier with other machine learning methods. Int. J. Adv. Comput. Sci. Appl., 11(11):738–748, 2020

  18. [26]

    Computable measure of entanglement

    Guifr ´e Vidal and Reinhard F Werner. Computable measure of entanglement. Phys. Rev. A , 65(3):032314, 2002

  19. [27]

    Entanglement negativity in quantum field theory

    Pasquale Calabrese, John Cardy, and Erik Tonni. Entanglement negativity in quantum field theory. Phys. Rev. Lett., 109(13):130502, 2012

  20. [28]

    Measures and applications of quantum correlations

    Gerardo Adesso, Thomas R Bromley, and Marco Cianciaruso. Measures and applications of quantum correlations. J. Phys. A:Math. Theor ., 49(47):473001, 2016. 20

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.