Pith. sign in

REVIEW 4 major objections 5 minor 36 references

Quantum Inspired Encoding Strategies for Machine Learning Models: Proposing and Evaluating Instance Level, Global Discrete, and Class Conditional Representations

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper proposes three quantum-inspired encoding strategies—row-wise, global unique-value, and class-conditional—and claims they cut encoding time by 40-60% while keeping classifier accuracy within about 1-2 percentage points.

desk verdict The 40–60% encoding-time claim is contradicted by the paper's own Table 1, and the encoding pipeline is too underspecified to support any accuracy comparison. read the letter →

arxiv 2507.00019 v1 pith:H4MPUKJG submitted 2025-06-15 cs.LG cs.AIquant-ph

Minati Rath , Hema Date This is my paper · ORCID
classification cs.LGcs.AIquant-ph
keywords quantum-inspiredencodinginstance-levelglobaldiscreteclass-conditionaltimereductionclassificationaccuracyquantumdatarepresentationcustomerchurn
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 the expensive step of turning classical data into quantum-style representations can be made far cheaper without sacrificing classification quality. It proposes three strategies—encode every row separately, encode each unique value once across the whole dataset, or encode unique values separately for each class—and applies them with six quantum-inspired embedding methods to a customer-churn classification task. The paper reports encoding-time reductions of roughly 40-60% compared with direct row-by-row encoding, with classifier accuracy staying within about ±1-2%. A reader should care because encoding time is often the practical bottleneck in quantum-inspired machine learning, and a strategy that preserves accuracy while cutting that cost would make such pipelines usable on larger data.

What carries the argument

The central object is the encoding map $E$ from a classical feature value to a quantum state $|\psi\rangle$. The three strategies differ only in the argument set fed to $E$: one row at a time (ILS), the set of unique values over the whole dataset (GDS), or unique values restricted by class (CC-GDS) and rows within a class (CC-ILS). The load-bearing move is deduplication: the complexity stated by the paper falls from $O(n\cdot d\cdot C_{\mathrm{embed}})$ to roughly $O(m)$ or $O(c\cdot m)$, where $m$ counts unique values rather than rows. Within the experiments, six embedding implementations supply concrete instances of $E$, and identical rows are reused to reduce ILS cost.

What would settle it

Rerun the six encodings with a written-down $E$ and numeric feature extraction: if QAOA ILS still exceeds QAOA DE in time, or if Displacement or Squeezing GDS still exceeds its corresponding direct baseline, the paper's blanket 40-60% time-reduction claim is false. Likewise, any concrete implementation that changes the accuracies in Table 2 by more than the claimed ±1-2% would falsify the accuracy-preservation claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that value-deduplication can replace brute-force direct encoding. Formally, it proposes replacing per-row maps $E(x_i)\to|\psi_i\rangle$ with per-value maps $E(u_k)\to|\psi_k\rangle$ (GDS) or per-class-per-value maps $E_y(u_k^y)\to|\psi_k^y\rangle$ (CC-GDS), plus a row-level class-conditioned variant; because repeated values are encoded once, the embedding workload drops from $O(n\cdot d)$ embedding calls to roughly $O(m)$ or $O(c\cdot m)$ calls. The authors assert that across Basis, Angle, IQP, QAOA, Displacement, and Squeezing embeddings, this drop translates into 40-60% less encoding time, and that the accuracy of classical classifiers trained on the reconstructed data varies by only ±1-2%. They also single out CC-ILS with Squeezing embedding as the best accuracy configuration and GDS as the fastest.

Load-bearing premise

The load-bearing premise is that the unshown encoding function $E$ and the unshown conversion from $|\psi_k\rangle$ to a numeric feature vector work as intended; the paper gives no definition of either, and every reported time and accuracy depends on those absent details.

Editorial extensions

If this is right

  • If the reported trade-off holds, quantum-inspired embedding ceases to be the dominant cost in classical machine learning pipelines, and larger tabular datasets become feasible to encode.
  • Practitioners can choose GDS when encoding time is the binding constraint and accept that value-sharing may cost fidelity on complex data.
  • Class-conditional instance encoding, especially with squeezing-style embeddings, offers the paper's best accuracy-runtime balance.
  • The proposed strategies are preprocessing schemes, so they compose with any downstream classical classifier a workflow already uses.

Reading between the lines

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

  • Editorial: the paper's own Table 1 contains outliers that contradict the blanket 40-60% reduction claim—QAOA ILS takes 67,171.59 seconds versus 63,505.59 seconds for QAOA direct encoding, and Displacement and Squeezing GDS both take longer than their direct baselines.
  • Editorial: with $E$ and the state-to-feature conversion unspecified, the accuracy results in Table 2 are not reproducible; a concrete choice of implementation could shift the numbers well beyond the claimed ±1-2%.
  • Editorial: CC-GDS is defined but never tested, so the experiments cannot cleanly separate the benefit of class-aware encoding from the benefit of per-class deduplication; testing it would close that gap.
  • Editorial: a natural scaling law to check is whether time savings grow with value redundancy, i.e., with the ratio $m/(n\cdot d)$; if so, GDS-style strategies would be most attractive for categorical or low-cardinality datasets.
Share X Bluesky LinkedIn Reddit HN

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 three quantum-inspired data encoding strategies — Instance Level Strategy (ILS), Global Discrete Strategy (GDS), and Class Conditional Value Strategy (CCVS, presented as CC-ILS and CC-GDS) — for transforming classical tabular data into symbolic quantum-state representations intended for use with purely classical machine learning classifiers. The authors evaluate the strategies on a telecom churn dataset with six embedding types (Basis, Angle, IQP, QAOA, Displacement, Squeezing) and report encoding times and classification metrics for several classical models. The central claim is that ILS, GDS, and CC-ILS consistently reduce encoding time by 40–60% relative to Direct Encoding (DE) while keeping accuracy within ±1–2%.

Significance. If the 40–60% time reduction at near-constant accuracy held, the paper would offer a practically useful low-cost preprocessing recipe for quantum-inspired encoding workflows. The paper also provides a structured taxonomy of three encoding strategies with illustrative examples, which may be pedagogically useful. However, the study does not provide code, machine-checked proofs, or parameter-free derivations; the complexity expressions O(n·d), O(m), and O(c·m) merely restate the definitions of the strategies. The main empirical claims are unsupported by the paper's own tables, and the encoding function is never specified, so the results are not reproducible. The significance of the contribution as presented is therefore low.

major comments (4)
  1. [Results and Discussion; Table 1] The headline claim that 'ILS, GDS, and CCILS consistently reduce the encoding time by approximately 40 to 60% compared to Direct Encoding (DE) across all six quantum inspired embedding methods' is directly contradicted by Table 1. For QAOA, ILS takes 67171.5900 s versus DE's 63505.5914 s, a 5.8% increase. For Displacement, GDS takes 63.3668 s versus DE's 41.4430 s, a 52.9% increase; for Squeezing, GDS takes 66.9308 s versus DE's 52.3153 s, a 27.9% increase. Several CC-ILS reductions are far below 40%: Basis 26.3%, Angle 6.0%, IQP 7.2%, QAOA 1.1%. Only a minority of rows fall in the stated 40–60% band, and the direction of the effect is reversed in critical cases. The stated conclusion is not supported by the paper's own measurements.
  2. [Sections 3.2 and 4] The encoding function E is never defined. Equations (1)–(4) introduce the notation E(x_i) → |ψ_i⟩ and E(u_k) → |ψ_k⟩, and Section 4 says each value is 'replaced with its quantum representation,' but the paper does not specify how a classical value is mapped to a concrete quantum state, how that state is represented numerically, or how the symbolic states are converted into feature vectors for the purely classical classifiers. Without this specification, the classification pipeline and every accuracy number in Table 2 are unreproducible, and the results could change arbitrarily under a different implementation of E.
  3. [Table 2] Table 2 contains multiple internal inconsistencies that undermine the reliability of the accuracy results. For example, several IQP DE rows report accuracy values greater than 1 (49.3316 for Logistic Regression and SVM Linear, 62.0321 for KNN, 67.2460 for AdaBoost), and the SVM Linear Angle DE row shows '66.8449' in the accuracy column, which appears to be a percentage-like value rather than a proportion. Additionally, the DE rows for Basis and Angle are identical across every classifier (accuracy 0.6684, precision 0.6513, etc.), which is implausible unless the two embeddings produce the same features or the table has been mis-copied. These data-integrity issues mean Table 2 cannot be used to support the claimed ±1–2% accuracy invariance.
  4. [Table 2; Results and Discussion] The claim that accuracy 'remains within a small margin of variation, typically ±1 to 2%,' is also contradicted by the paper's own numbers. For instance, Logistic Regression with Basis encoding shows DE accuracy 0.6684 versus ILS accuracy 0.7433, a 7.5-percentage-point change; KNN with Basis encoding shows DE 0.5067 versus ILS 0.6941, an 18.7-percentage-point change. These differences are an order of magnitude larger than the stated margin, further invalidating the paper's central accuracy-efficiency trade-off claim.
minor comments (5)
  1. [Throughout] There are pervasive typos and nomenclature inconsistencies: 'Sqeezing' in Table 2, 'keggal' in the Data Availability Statement, and the abstract uses 'CCVS' while the main text uses 'CC-ILS' and 'CC-GDS'; please align the abbreviations.
  2. [Section 4] The text says the classifiers used were Logistic Regression, KNN, SVM, and ensemble methods such as Random Forest, LightGBM, AdaBoost, and CatBoost, but Table 2 also reports Decision Tree, Extra Trees, Gradient Boosting, and XGBoost; clarify which models were actually run and why the list differs.
  3. [Table 1] The GDS Basis entry reports 2 unique values and an encoding time of 0.0000 seconds; since the PCA-transformed features are continuous, it is unclear how only two unique values arise, and the zero time should be explained or rounded transparently.
  4. [Figure 2 caption] The caption for Figure 2 reads 'Flow Chart for Instance level strategy' but the figure illustrates the Global Discrete Strategy; correct the caption.
  5. [Sections 3.1–3.5] The complexity expressions O(n·d·C_embed), O(m), and O(c·m) are presented as the paper's own results, but they are restatements of the definitions and omit the per-value embedding cost for GDS and CC-GDS; please specify these costs explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the encoding strategies are defined by explicit maps and the runtime and accuracy results come from direct measurement; the only self-citation is background, and the headline 40-60% time-saving claim is internally contradicted by Table 1, which is a correctness issue, not circularity.

full rationale

The paper's proposed strategies are specified as explicit transformations: ILS maps each instance via E(x_i)->|psi_i>, GDS maps each unique value via E(u_k)->|psi_k>, and CC-GDS/CC-ILS map class-conditioned values via E_y(u_k^y)->|psi_k^y> (Eqs. 1-4). The encoding-time and accuracy numbers in Tables 1 and 2 are measured outcomes of an implementation (PennyLane on a Telco churn dataset), not quantities derived from the definitions by construction. The complexity statements O(n*d), O(m), and O(c*m) are immediate counting restatements of the procedural definitions, but the paper does not use these as predictions or as the evidence for its empirical time-saving claim; Table 1 is the stated evidence. The one self-citation, Rath and Date (2024), is used only as background for quantum data encoding and carries no load-bearing argument; no uniqueness theorem or fitted parameter is invoked. The central claim that ILS/GDS/CC-ILS consistently reduce encoding time by approximately 40-60% is contradicted by the paper's own Table 1 (e.g., QAOA ILS is 67,171.59 s vs. DE 63,505.59 s, and Displacement GDS is 63.37 s vs. DE 41.44 s), but this is an internal-consistency and empirical-validity problem, not circular reasoning. Therefore no circularity is found.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper introduces no physical entities. Its free parameters and assumptions are mostly modeling choices and unstated implementation details. The PCA dimension is the only explicit fitted number. The largest gaps are the undefined encoding function and the unverified conversion of symbolic states to classifier features, which make the entire benchmark uninterpretable.

free parameters (1)
  • Number of PCA components = 23
    Chosen from an elbow point in explained variance on the training data. This value determines the feature dimensionality that is later encoded and thus affects all reported timings and accuracies.
assumptions (4)
  • ad hoc to paper The encoding function E maps each value to a distinct symbolic state |ψk⟩ with no specified numerical representation.
    Sections 3.1 to 3.5 define E only through its outputs. The actual circuits or amplitudes are never stated, and the conversion to classifier features is never given.
  • domain assumption Replacing feature values with symbolic states preserves enough information for the classical classifiers to learn.
    The paper assumes that the relabeled data carry the same or similar predictive content as the original numeric PCA features. No experiment compares against the raw features without encoding.
  • domain assumption The PennyLane embedding implementations (Basis, Angle, IQP, QAOA, Displacement, Squeezing) are used correctly and measure the intended computational cost.
    Section 4 states that embeddings were implemented in PennyLane, but no circuit diagrams, parameters, or verification results are provided. The identical DE results for Basis and Angle suggest a possible implementation flaw.
  • domain assumption The GDS efficiency argument assumes the number of unique values m is much smaller than the number of row entries n·d.
    Section 3.2 argues GDS is efficient because only unique values are encoded. For the continuous PCA features in the experiment, Table 1 shows m ≈ 76,000 while n·d ≈ 86,000, so the reduction is modest and the timing data contradict the assumption for several embedding types.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantum Inspired Encoding Strategies for Machine Learning Models: Proposing and Evaluating Instance Level, Global Discrete, and Class Conditional Representations." pith.science (2026). https://pith.science/paper/H4MPUKJG

@misc{pith2026250700019,
  author       = {Pith},
  title        = {Pith review of: Quantum Inspired Encoding Strategies for Machine Learning Models: Proposing and Evaluating Instance Level, Global Discrete, and Class Conditional Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H4MPUKJG}},
  note         = {Machine review of arXiv:2507.00019}
}
read the original abstract

In this study, we propose, evaluate and compare three quantum inspired data encoding strategies, Instance Level Strategy (ILS), Global Discrete Strategy (GDS) and Class Conditional Value Strategy (CCVS), for transforming classical data into quantum data for use in pure classical machine learning models. The primary objective is to reduce high encoding time while ensuring correct encoding values and analyzing their impact on classification performance. The Instance Level Strategy treats each row of dataset independently; mimics local quantum states. Global Discrete Value Based encoding strategy maps all unique feature values across the full dataset to quantum states uniformly. In contrast, the Class conditional Value based encoding strategy encodes unique values separately for each class, preserving class dependent information. We apply these encoding strategies to a classification task and assess their impact on en-coding efficiency, correctness, model accuracy, and computational cost. By analyzing the trade offs between encoding time, precision, and predictive performance, this study provides insights into optimizing quantum inspired data transformations for classical machine learning workflows.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 25 canonical work pages

  1. [1]

    Machine learning & artificial intelligence in the quantum domain: a review of recent progress.Reports on Progress in Physics, 81(7):074001, 2018

    Vedran Dunjko and Hans J Briegel. Machine learning & artificial intelligence in the quantum domain: a review of recent progress.Reports on Progress in Physics, 81(7):074001, 2018

    2018

  2. [2]

    Quantum machine learning in feature hilbert spaces

    Maria Schuld and Nathan Killoran. Quantum machine learning in feature hilbert spaces. Physical review letters, 122(4):040504, 2019

    2019

  3. [3]

    Machine learning: Quantum vs classical.IEEE Access, 8:219275–219294, 2020

    Tariq M Khan and Antonio Robles-Kelly. Machine learning: Quantum vs classical.IEEE Access, 8:219275–219294, 2020

    work page 2020

  4. [4]

    Sok: quantum computing methods for machine learning optimization.Quan- tum Machine Intelligence, 6(2):47, 2024

    Hamza Baniata. Sok: quantum computing methods for machine learning optimization.Quan- tum Machine Intelligence, 6(2):47, 2024

    work page 2024

  5. [5]

    Parameterized quantum circuits as machine learning models.Quantum science and technology, 4(4):043001, 2019

    Marcello Benedetti, Erika Lloyd, Stefan Sack, and Mattia Fiorentini. Parameterized quantum circuits as machine learning models.Quantum science and technology, 4(4):043001, 2019

    2019

  6. [6]

    Quantum data encoding: A comparative analysis of classical-to- quantum mapping techniques and their impact on machine learning accuracy.EPJ Quantum Technology, 11(1):72, 2024

    Minati Rath and Hema Date. Quantum data encoding: A comparative analysis of classical-to- quantum mapping techniques and their impact on machine learning accuracy.EPJ Quantum Technology, 11(1):72, 2024

    work page 2024

  7. [7]

    Implementing a distance-based classifier with a quantum interference circuit.Europhysics Letters, 119(6):60002, 2017

    Maria Schuld, Mark Fingerhuth, and Francesco Petruccione. Implementing a distance-based classifier with a quantum interference circuit.Europhysics Letters, 119(6):60002, 2017

    work page 2017

  8. [8]

    Exponential data encoding for quan- tum supervised learning.Physical Review A, 107(1):012422, 2023

    Seongwook Shin, Yong-Siah Teo, and Hyunseok Jeong. Exponential data encoding for quan- tum supervised learning.Physical Review A, 107(1):012422, 2023

    work page 2023

Show all 36 references
  1. [9]

    Robust data encodings for quantum classifiers.Physical Review A, 102(3):032420, 2020

    Ryan LaRose and Brian Coyle. Robust data encodings for quantum classifiers.Physical Review A, 102(3):032420, 2020

    2020

  2. [10]

    Quantum computation over continuous variables.Phys- ical Review Letters, 82(8):1784, 1999

    Seth Lloyd and Samuel L Braunstein. Quantum computation over continuous variables.Phys- ical Review Letters, 82(8):1784, 1999

    1999

  3. [11]

    Quantum computing with continuous-variable clusters.Physical Review A—Atomic, Molecu- lar, and Optical Physics, 79(6):062318, 2009

    Mile Gu, Christian Weedbrook, Nicolas C Menicucci, Timothy C Ralph, and Peter van Loock. Quantum computing with continuous-variable clusters.Physical Review A—Atomic, Molecu- lar, and Optical Physics, 79(6):062318, 2009

    2009

  4. [12]

    Hybrid discrete-and continuous-variable quantum information.Nature Physics, 11(9):713–719, 2015

    Ulrik L Andersen, Jonas S Neergaard-Nielsen, Peter Van Loock, and Akira Furusawa. Hybrid discrete-and continuous-variable quantum information.Nature Physics, 11(9):713–719, 2015

    2015

  5. [13]

    Quantum computing overview: discrete vs

    Sophie Choe. Quantum computing overview: discrete vs. continuous variable models.arXiv preprint arXiv:2206.07246, 2022. 23

    2022 arXiv

  6. [14]

    Circuit-centric quantum classifiers.Physical Review A, 101(3):032308, 2020

    Maria Schuld, Alex Bocharov, Krysta M Svore, and Nathan Wiebe. Circuit-centric quantum classifiers.Physical Review A, 101(3):032308, 2020

    2020

  7. [15]

    Data re- uploading for a universal quantum classifier.Quantum, 4:226, 2020

    Adri´ an P´ erez-Salinas, Alba Cervera-Lierta, Elies Gil-Fuster, and Jos´ e I Latorre. Data re- uploading for a universal quantum classifier.Quantum, 4:226, 2020

    2020

  8. [16]

    Supervised learning with quantum-enhanced feature spaces.Nature, 567(7747):209–212, 2019

    Vojtˇ ech Havl ´ ıˇ cek, Andr´ es M C´ ardenas-Valencia, and Al´ an Aspuru-Guzik. Supervised learning with quantum-enhanced feature spaces.Nature, 567(7747):209–212, 2019

    2019

  9. [17]

    Quantum circuit learning.Physical Review A, 98(3):032309, 2018

    Kosuke Mitarai, Makoto Negoro, Masahiro Kitagawa, and Keisuke Fujii. Quantum circuit learning.Physical Review A, 98(3):032309, 2018

    2018

  10. [18]

    The quest for a quantum neural network.Quantum Information Processing, 13(11):2567–2586, 2018

    Maria Schuld, Ilya Sinayskiy, and Francesco Petruccione. The quest for a quantum neural network.Quantum Information Processing, 13(11):2567–2586, 2018

    2018

  11. [19]

    Braunstein and Peter Van Loock

    Samuel L. Braunstein and Peter Van Loock. Quantum information with continuous variables. Reviews of Modern Physics, 77(2):513, 2005

    2005

  12. [20]

    Quantum embeddings for machine learning.Physical Review A, 101(3):032305, 2020

    Seth Lloyd, Maria Schuld, Asif Ijaz, Josh Izaac, and Nathan Killoran. Quantum embeddings for machine learning.Physical Review A, 101(3):032305, 2020

    2020

  13. [21]

    Efficient measurement-based quantum computing with continuous-variable systems.Physical Review A—Atomic, Molecular, and Optical Physics, 85(6):062318, 2012

    Matthias Ohliger and Jens Eisert. Efficient measurement-based quantum computing with continuous-variable systems.Physical Review A—Atomic, Molecular, and Optical Physics, 85(6):062318, 2012

    2012

  14. [22]

    Quantum-inspired machine learning: En- coding strategies and interpretability.Quantum, 8:117, 2024

    Christoph Zoufal, Ryan Sweke, and Madalin Guta. Quantum-inspired machine learning: En- coding strategies and interpretability.Quantum, 8:117, 2024

    2024

  15. [23]

    Continuous-variable quantum neural networks.Physical Review Research, 1(3):033063, 2019

    Nathan Killoran, Thomas R Bromley, Juan Miguel Arrazola, Maria Schuld, Nicol´ as Quesada, and Seth Lloyd. Continuous-variable quantum neural networks.Physical Review Research, 1(3):033063, 2019

    2019

  16. [24]

    Measuring analytic gradients of general quantum evolution with the stochastic parameter shift rule.Quantum, 5:386, 2021

    Leonardo Banchi and Gavin E Crooks. Measuring analytic gradients of general quantum evolution with the stochastic parameter shift rule.Quantum, 5:386, 2021

    2021

  17. [25]

    Encoding strategies for efficient quantum data representation.Quantum Science and Technology, 8(1):015005, 2023

    Edward Grant, Marcello Benedetti, Seth Lloyd, and Leonard Wossnig. Encoding strategies for efficient quantum data representation.Quantum Science and Technology, 8(1):015005, 2023

    2023

  18. [26]

    On fundamental aspects of quantum extreme learning machines

    Weijie Xiong, Giorgio Facelli, Mehrad Sahebi, Owen Agnel, Thiparat Chotibut, Supanut Thanasilp, and Zo¨ e Holmes. On fundamental aspects of quantum extreme learning machines. Quantum Machine Intelligence, 7(1):20, 2025

    2025

  19. [27]

    A quantum approximate optimization algorithm, 2014

    Edward Farhi, Jeffrey Goldstone, and Sam Gutmann. A quantum approximate optimization algorithm, 2014

    2014

  20. [28]

    Quantum approximate optimization is computationally universal, 2018

    Seth Lloyd. Quantum approximate optimization is computationally universal, 2018

    2018

  21. [29]

    Gaussian quantum information.Reviews of Modern Physics, 84(2):621–669, 2012

    Christian Weedbrook, Stefano Pirandola, Ra´ ul Garc ´ ıa-Patr´ on, Nicolas J Cerf, Timothy C Ralph, Jeffrey H Shapiro, and Seth Lloyd. Gaussian quantum information.Reviews of Modern Physics, 84(2):621–669, 2012

    2012

  22. [30]

    John Wiley & Sons, 2013

    David W Hosmer Jr, Stanley Lemeshow, and Rodney X Sturdivant.Applied logistic regression. John Wiley & Sons, 2013

    2013

  23. [31]

    Nearest neighbor pattern classification.IEEE transactions on information theory, 13(1):21–27, 1967

    Thomas Cover and Peter Hart. Nearest neighbor pattern classification.IEEE transactions on information theory, 13(1):21–27, 1967. 24

    1967

  24. [32]

    Support-vector networks.Machine learning, 20:273–297, 1995

    Corinna Cortes and Vladimir Vapnik. Support-vector networks.Machine learning, 20:273–297, 1995

    1995

  25. [33]

    Random forests.Machine learning, 45:5–32, 2001

    Leo Breiman. Random forests.Machine learning, 45:5–32, 2001

    2001

  26. [34]

    Lightgbm: A highly efficient gradient boosting decision tree.Advances in neural information processing systems, 30, 2017

    Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. Lightgbm: A highly efficient gradient boosting decision tree.Advances in neural information processing systems, 30, 2017

    2017

  27. [35]

    A decision-theoretic generalization of on-line learning and an application to boosting.Journal of computer and system sciences, 55(1):119–139, 1997

    Yoav Freund and Robert E Schapire. A decision-theoretic generalization of on-line learning and an application to boosting.Journal of computer and system sciences, 55(1):119–139, 1997

    1997

  28. [36]

    Catboost: unbiased boosting with categorical features.Advances in neural information processing systems, 31, 2018

    Liudmila Prokhorenkova, Gleb Gusev, Aleksandr Vorobev, Anna Veronika Dorogush, and Andrey Gulin. Catboost: unbiased boosting with categorical features.Advances in neural information processing systems, 31, 2018. Data Availability Statement The datasets analysed during the curr...

    2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.