Pith. sign in

REVIEW 5 major objections 5 minor 50 references

Boosting KNNClassifier Performance with Opposition-Based Data Transformation

T0 review · 5 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read Opposition-based mirror images boost KNN accuracy across 26 datasets.

desk verdict A simple OBL reflection trick for KNN augmentation, undermined by incomplete reporting, missing results, and a likely data-leakage problem. read the letter →

arxiv 2504.16268 v2 pith:EBB54Y4R submitted 2025-04-22 cs.LG cs.AI

classification cs.LGcs.AI
keywords Opposition-BasedLearningdataaugmentationK-nearestneighborsclassificationhigh-dimensionaltransformationclass-wiseOBLfeatureselection
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

The paper proposes using Opposition-Based Learning (OBL)—a reflection operation from optimization—to create synthetic 'opposite' training samples for classification. It tests three variants, with bounds taken from the whole dataset, from each class, or from local same-class neighborhoods, and combines them with K-nearest neighbors (KNN). Across 26 datasets, the reported accuracy and F1-scores for OBL-augmented KNN consistently exceed plain KNN, with class-wise OBL singled out as the strongest scheme. The motivation is that reflected points add geometry-aware, class-aligned coverage of sparse or high-dimensional feature spaces, at negligible computational cost.

What carries the argument

The load-bearing object is the opposition transformation $x^*_k = a_k + b_k - x_k$, a deterministic mirror reflection of a feature value about the midpoint of its range. Varying the source of the bounds ($a_k,b_k$)—global, class-specific, or local neighborhood—yields three augmentation schemes. Because the reflection is geometric rather than stochastic, the added points are claimed to expand the feature-space footprint of each class without injecting random noise.

What would settle it

Run the same 26-dataset protocol with a strict fold-wise pipeline in which Z-score normalization, mutual-information feature selection, and OBL min/max bounds are recomputed inside each training fold only, then compare OBL-KNN against plain KNN; if the accuracy advantage disappears or shrinks to statistical noise, the original gains came from leakage of test-set information into preprocessing.

Watch

Extended reading notes

Core claim

The central claim is that reflecting each training point through feature-wise boundaries, $x^*_k = a_k + b_k - x_k$, produces synthetic opposite samples that improve KNN classification when appended to the training set. The bounds $a_k,b_k$ are taken globally (Global OBL), per class (Class-Wise OBL), or from the $P$ nearest neighbors of the same class (Localized Class-Wise OBL). On 26 datasets, the paper reports that the OBL-augmented KNN variants beat both plain KNN and distance-weighted WKNN on accuracy and F1-score, and that Class-Wise OBL performs best. The authors attribute the gain to enriched decision boundaries and better coverage of underexplored regions, especially in high-dimensional and sparse settings.

Load-bearing premise

The reported gains assume all preprocessing statistics—Z-score parameters, chosen features, and OBL bounds—are computed from the training folds only, with the test fold never influencing them; the paper does not explicitly state that the train/test split occurs before these statistics are calculated.

Editorial extensions

If this is right

  • If the reported gains hold under clean evaluation, OBL augmentation is a parameter-light way to improve KNN on high-dimensional and sparse datasets, with runtime overhead close to that of plain KNN.
  • Class-Wise OBL being the best of the three schemes suggests that class-specific bounds preserve or sharpen between-class structure better than global or local bounds.
  • Weighted KNN variants (WOBLKNN) add small additional gains in the reported tables, indicating that distance weighting and OBL augmentation are complementary.
  • The paper lists SVM, logistic regression, and gradient boosting as intended beneficiaries, so testing OBL augmentation on those classifiers is a direct next step if the KNN results are confirmed.
  • Because the opposite set is generated deterministically from the training data, it provides a reproducible augmentation that does not depend on random seeds or generative models.

Reading between the lines

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

  • The class-wise variant effectively mirrors each class separately, which should increase the density of minority classes relative to majority classes; if so, its advantage should be largest on imbalanced datasets, a pattern that could be tested by stratifying the 26 datasets by imbalance ratio.
  • A natural extension is to treat the bounds as learnable or to apply OBL repeatedly after different preprocessing stages, which might further improve separation.
  • The localized variant's neighborhood size $P$ is a new hyperparameter; its optimal value may scale with dimensionality, so an adaptive $P$ could be a testable improvement.
  • Combining OBL reflection with interpolation-based augmentation could produce a two-stage scheme that both expands the convex hull and adds opposite-region coverage; this is an editorial suggestion, not a paper claim.
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

5 major / 5 minor

Summary. The paper proposes three opposition-based learning (OBL) transformations—Global OBL, Class-Wise OBL, and Localized Class-Wise OBL—as a data-augmentation preprocessing step for KNN and other classifiers. It claims that OBL-enhanced classifiers consistently outperform plain KNN across 26 datasets, with evaluation by 5-fold cross-validation averaged over 30 runs and reported accuracy, F1, and runtime. The manuscript describes the OBL schemes, provides a preprocessing pipeline, and presents tables comparing KNN, WKNN, OBLKNN, and WOBLKNN, plus Friedman test figures for accuracy, F1, and runtime.

Significance. If the central claim were sound, the contribution would be practically useful: a simple, geometry-aware reflection of training samples could improve KNN in high-dimensional or sparse settings with negligible overhead. The paper has several strengths in presentation: the OBL variants are clearly defined in Section 5, the dataset collection is sizable and includes many high-dimensional problems, and the authors report multiple metrics and average over 30 runs. However, the central claim is not supported by the paper's own tables, the evaluation protocol is under-specified, and the proposed class-wise and localized variants are absent from the reported experimental results. These issues are load-bearing for the paper's contribution, so the current manuscript does not establish its advertised finding.

major comments (5)
  1. [Section 5 and Section 6.1] The experimental protocol does not specify that all preprocessing statistics are computed on the training folds only. The pipeline in Section 5 lists Z-score normalization, mutual-information feature selection, OBL bounds (a_k, b_k, a_{c,k}, b_{c,k}, a_{i,k}, b_{i,k}), and post-augmentation normalization as a linear sequence ending in model training, with no mention of a train/test split. Section 6.1 states only that a 5-fold CV was used and results were averaged over 30 runs, not that the pipeline was refit inside each fold. If any of these statistics are computed on the full dataset before splitting, the accuracy and F1 values in Tables 2 and 3 are optimistically biased and the central claim is invalidated. This is the most serious issue because it affects every reported result.
  2. [Abstract and Table 2] The claim that OBL-enhanced classifiers 'consistently outperform' basic KNN is contradicted by the paper's own numbers. In Table 2, OBLKNN accuracy is lower than KNN accuracy on 10 of the 26 datasets, including dataset 1 (0.9250 vs. 0.9256), dataset 3 (0.9247 vs. 0.9300), dataset 4 (0.8494 vs. 0.8497), dataset 9 (0.7442 vs. 0.7448), dataset 13 (0.9804 vs. 0.9807), dataset 17 (0.9892 vs. 0.9901), dataset 19 (0.6950 vs. 0.6952), dataset 20 (0.9478 vs. 0.9486), dataset 21 (0.9252 vs. 0.9253), and dataset 25 (0.9591 vs. 0.9607). A pattern of mixed and often tiny differences does not support the abstract's 'consistently outperform' wording.
  3. [Section 7.1 and Tables 2-4] The paper's proposed class-wise and localized variants are central to the contribution, but no experimental results are reported for them. Section 6 states that 'Class-Wise Opposition-Based Learning is the most performed,' and Section 7.1 says the evaluation covers OBLKNN-CW, LOBLKNN-CW, WOBLKNN-CW, and WLOBLKNN-CW. However, Tables 2-4 contain only KNN, WKNN, OBLKNN, and WOBLKNN. The Friedman test figures (Figures 3-5) are referenced but their contents, p-values, and post-hoc comparisons are not reported in the text. The selection of a best variant is therefore unsupported by any presented evidence.
  4. [Section 5, pipeline bullets] There is an internal inconsistency about whether OBL generates additional samples or replaces the original data. The surrounding text says the scheme 'enriches the training set' and describes an 'OBL-augmented' dataset, but one pipeline bullet says the transformation produces 'oppositional samples that replace the original dataset,' and another bullet says training is done on the 'OBL-augmented and normalized dataset.' If opposite samples replace rather than augment, the method is not data augmentation as claimed, and the reported comparisons are not testing the stated approach. The authors need to state unambiguously which dataset is used for training.
  5. [Section 7.1 and Figures 3-5] No statistical uncertainty is reported for any of the accuracy or F1 values: there are no standard deviations, confidence intervals, p-values, or effect sizes. The manuscript claims Friedman tests were conducted, but the text does not report the test statistics or associated p-values, and the figures referenced as evidence are not reproduced in the manuscript body. Given that many pairwise differences in Table 2 are on the order of 0.001 or less, the claim of statistically validated improvement is not currently testable.
minor comments (5)
  1. [Introduction] The phrase 'adversarial-based learning (OBL)' near the end of Section 1 should read 'opposition-based learning (OBL)'.
  2. [Equation numbering] Equation numbers are not unique: Equation (1) appears for min-max scaling in Section 3.1 and again for generalized OBL in Section 4.1; similar duplicate numbering occurs for other equations. Please renumber equations sequentially throughout the manuscript.
  3. [References] The reference list contains inconsistencies and editorial notes, including duplicate entries for [43], a reference marked as a duplicate of [42], and a mismatched in-text citation for OBL: the text cites Tizhoosh's OBL paper as [7], but the reference list entry [7] is the ADASYN paper, with Tizhoosh at [8].
  4. [Tables 2-4] The tables do not include dataset names or a key row label; each row is identified only by a number. Adding abbreviated dataset names would make it much easier for readers to check individual results.
  5. [Table 1] For the Semeion dataset, the class count is listed as 2, but Semeion is usually a 10-class handwritten digit dataset; if the data were binarized, this should be stated explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: OBL augmentation is a fixed geometric preprocessing rule and the reported gains are empirical, not derived from the target.

full rationale

The paper contains no derivation chain in which a predicted quantity is algebraically or statistically equivalent to an input. The OBL transformations are explicit preprocessing rules (Eqs. 9–11): each opposite point is a reflection of an original point around global, class-wise, or local feature bounds; no parameter is fitted to the accuracy/F1 values that are later reported. The classifiers are standard algorithms, and the evaluation is an external benchmark (26 UCI/scikit-feature datasets, 5-fold CV, 30 runs) rather than a theorem derived from the method's assumptions. The author does not invoke a self-authored uniqueness theorem or prior work to justify the choice of OBL; the cited OBL literature (Tizhoosh 2005; Rahnamayan et al. 2008) is independent. A real methodological risk remains: Section 5 presents Z-score normalization, mutual-information feature selection, and OBL min/max bounds before describing any train/test split, so the reader cannot verify that preprocessing statistics are computed per training fold; if they were computed on full data, the cross-validation numbers would be leaky. But leakage is an experimental-validity concern, not a circularity in the paper's derivation chain, and the declared protocol (5-fold CV) is consistent with an honest evaluation.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the assumption that mirrored points preserve class semantics and that the evaluation is leakage-free. The only hand-set quantities are the KNN neighborhood, local neighborhood size, and per-dataset feature counts; none are cross-validated. No new theoretical entities are introduced.

free parameters (3)
  • KNN neighborhood k = 3
    Fixed for all algorithms with no sensitivity analysis, despite being a central hyperparameter for KNN.
  • Local neighborhood size P = 3 and 5
    Used for LOBL-CW variants, whose results are not reported in the paper.
  • Number of selected features per dataset = 100 for high-dimensional sets, smaller values for low-dimensional sets
    Table 1 lists per-dataset feature counts (100, 36, 20, etc.) without explaining how they were chosen; this choice directly affects reported accuracy.
assumptions (3)
  • ad hoc to paper Opposite samples computed from feature-wise bounds are informative and preserve their assigned class labels after re-normalization.
    Section 5 assumes reflection across min/max creates useful synthetic samples; no proof or ablation shows the reflected points carry the class label they are assigned.
  • domain assumption No data leakage: preprocessing statistics are computed on training folds only.
    The pipeline in Section 5 lists normalization, feature selection, and OBL transformation before model training and evaluation without stating when the train/test split occurs; if global bounds are fitted on all data, results are optimistic.
  • domain assumption Euclidean distance is the appropriate proximity measure for all 26 datasets.
    Section 2.1 defines KNN with Euclidean distance, but no distance metric sensitivity analysis is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Boosting KNNClassifier Performance with Opposition-Based Data Transformation." pith.science (2026). https://pith.science/paper/EBB54Y4R

@misc{pith2026250416268,
  author       = {Pith},
  title        = {Pith review of: Boosting KNNClassifier Performance with Opposition-Based Data Transformation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EBB54Y4R}},
  note         = {Machine review of arXiv:2504.16268}
}
read the original abstract

In this paper, we introduce a novel data transformation framework based on Opposition-Based Learning (OBL) to boost the performance of traditional classification algorithms. Originally developed to accelerate convergence in optimization tasks, OBL is leveraged here to generate synthetic opposite samples that enrich the training data and improve decision boundary formation. We explore three OBL variants Global OBL, Class-Wise OBL, and Localized Class-Wise OBL and integrate them with K-Nearest Neighbors (KNN). Extensive experiments conducted on 26 heterogeneous and high-dimensional datasets demonstrate that OBL-enhanced classifiers consistently outperform the basic KNN. These findings underscore the potential of OBL as a lightweight yet powerful data transformation strategy for enhancing classification performance, especially in complex or sparse learning environments.

Figures

Figures reproduced from arXiv: 2504.16268 by the authors.

Figure 1
Figure 1. 2D and 3D of OBL data augmentation [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. t-SNE Visualization: Original Points (◊) vs. OBL-transformed Points (X) To evaluate the effectiveness of Opposition-Based Learning (OBL) as a data augmentation technique for classification algorithms, a structured preprocessing pipeline is applied prior to model training. This pipeline begins with essential steps such as handling missing values, normalizing feature scales, and preparing the dataset for learning. For… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 41 canonical work pages

  1. [1]

    Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25

  2. [2]

    Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781

  3. [3]

    M., & Furht , B

    Shorten, C., Khoshgoftaar, T. M., & Furht , B. (2021). Data augmentation for machine learning: A survey. Journal of Big Data, 8(1), 1-48

  4. [8]

    Tizhoosh, H. R. (2005). Opposition-based learning: A new scheme for machine intelligence. In International Conference on Computational Intelligence for Modelling, Control and Automation and International Conference on Intelligent Agents, Web Technologies and Inter net Commerce (CIMCA -IAWTIC'05) (Vol. 1, pp. 695 -701). IEEE

  5. [9]

    R., & Salama, M

    Rahnamayan, S., Tizhoosh, H. R., & Salama, M. M. A. (2008). Opposition-based differential evolution. IEEE Transactions on Evolutionary Computation, 12(1), 64-79

  6. [10]

    Z., Rahnamayan, S., & Deb, K

    Mahdavi, S. Z., Rahnamayan, S., & Deb, K. (2018). Opposition based learning: A literature review. Swarm and Evolutionary Computation , 39, 1 -23. (Note: A comprehensive review showing the breadth of OBL applications)

  7. [11]

    Alpaydin, E. (2020). Introduction to Machine Learning. MIT Press

  8. [12]

    Murphy, K. P. (2012). Machine Learning: A Probabilistic Perspective. MIT Press

Show all 50 references
  1. [13]

    Cover, T., & Hart, P. (1967). Nearest neighbor pattern classification. IEEE transactions on information theory, 13(1), 21-27

  2. [14]

    Hastie, T., Tibshirani, R., & Friedman, J. (2009). The Elements of Statistical Learning: Data Mining, Inference, and Prediction. Springer

  3. [15]

    O., Hart, P

    Duda, R. O., Hart, P. E., & Stork, D. G. (2001). Pattern Classification. Wiley

  4. [16]

    Dudani, S. A. (1976). The distance-weighted k-nearest-neighbor rule. IEEE Transactions on Systems, Man, and Cybernetics, SMC-6(4), 325-327

  5. [17]

    Cortes, C., & Vapnik, V. (1995). Support-vector networks. Machine learning, 20(3), 273-297

  6. [18]

    Vapnik, V. (2000). The Nature of Statistical Learning Theory. Springer

  7. [19]

    W., Lemeshow, S., & Sturdivant, R

    Hosmer Jr, D. W., Lemeshow, S., & Sturdivant, R. X. (2013). Applied Logistic Regression. Wiley

  8. [20]

    Ng, A. Y. (2011). Lecture Notes on Logistic Regression . Stanford University. (Often cited from machine learning courses)

  9. [21]

    Friedman, J. H. (2001). Greedy function approximation: a gradient boosting machine. Annals of statistics , 1189-1202

  10. [22]

    Friedman, J. H. (2002). Stochastic gradient boosting. Computational Statistics & Data Analysis, 38(4), 367- 378

  11. [23]

    Chen, T., & Guestrin, C. (2016). XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 785-794)

  12. [24]

    Shorten, C., & Khoshgoftaar, T. M. (2019). A survey on image data augmentation for deep learning. Journal of Big Data, 6(1), 60

  13. [25]

    D., Zoph, B., Vasudevan, V., & Le, Q

    Cubuk, E. D., Zoph, B., Vasudevan, V., & Le, Q. V. (2019). Autoaugment: Learning augmentation strategies from data. In Proceedings of the the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 113-123)

  14. [26]

    & Huang, K

    Wen, S., Zhang, X., Gao, X., Gu, L., Sun, J., Ma, X., ... & Huang, K. (2020). Data augmentation for tabular data. arXiv preprint arXiv:2007.03780

  15. [27]

    T., & Li, B

    Wang, Y., Ding, Y., Jiang, J., Kwok, J. T., & Li, B. (2017). Understanding and improving deep learning for biomedical image analysis. IEEE Transactions on Biomedical Engineering, 65(4), 901-909

  16. [28]

    Morel, P., Adam, S., & Melançon, G. (2021). A survey on data augmentation for tabular data. arXiv preprint arXiv:2104.10657

  17. [29]

    Wang, J., & Perez, L. (2017). The effectiveness of data augmentation in image classification using deep learning. arXiv preprint arXiv:1712.046 augmentation

  18. [30]

    Shrivastava, A., Pfister, T., Oncel Hacioglu, O., Tulyakov, S., & Leibig, C. (2017). Data augmentation by pairing samples for images. arXiv preprint arXiv:1710.07198

  19. [31]

    Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R. (2014). Dropout: a simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research, 15(1), 1929-1958

  20. [32]

    V., Bowyer, K

    Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002). SMOTE: synthetic minority oversampling technique. Journal of artificial intelligence research, 16, 321-357

  21. [33]

    He, H., & Garcia, E. A. (2009). Learning from imbalanced data. IEEE Transactions on knowledge and data engineering, 21(9), 1263-1284

  22. [34]

    A., & Li, S

    He, H., Bai, Y., Garcia, E. A., & Li, S. (2008). ADASYN: Adaptive synthetic sampling approach for imbalanced learning. In 2008 IEEE international joint conference on neural networks (IEEE World Congress on Computational Intelligence) (pp. 1322-1328). IEEE

  23. [35]

    Y., & Mao, B

    Han, H., Wang, W. Y., & Mao, B. H. (2005). Borderline-SMOTE: a new oversampling method in imbalanced data sets learning. In International conference on intelligent computing (pp. 878-887). Springer

  24. [36]

    Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer

  25. [37]

    Efron, B., & Tibshirani, R. J. (1993). An Introduction to the Bootstrap. Chapman and Hall/CRC

  26. [38]

    N., & Ganguli, S

    Zhang, H., Cisse, M., Dauphin, Y. N., & Ganguli, S. (2017). mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412

  27. [39]

    Gui, J., Sun, Z., Wen, Y., Tao, D., & Ye, J. (2021). A review on generative adversarial networks: Algorithms, theory, and applications. IEEE Transactions on Knowledge and Data Engineering, 33(7), 3576-3601

  28. [40]

    & Bengio, Y

    Goodfellow, I., Pouget -Abadie, J., Mirza, M., Xu, B., Warde -Farley, D., Ozair, S., ... & Bengio, Y. (2014). Generative adversarial nets. Advances in neural information processing systems, 27

  29. [41]

    P., & Welling, M

    Kingma, D. P., & Welling, M. (2013). Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114

  30. [42]

    Xu, L., Skoularidou, M., Anthony, J., Sun, Y., & van der Schaar, M. (2019). Modeling tabular data using conditional gan. Advances in Neural Information Processing Systems, 32

  31. [43]

    Park, T., Cornelius, S., Phillips, J., & Lee, J. (2018). Data augmentation using generative adversarial networks (DAGAN) for improved training. arXiv preprint arXiv:1805.08201. (Note: While DAGAN is image-focused, the concept of GANs for augmentation applies and TabGAN built u...

  32. [44]

    Xu, L., Skoularidou, M., Anthony, J., Sun, Y., & van der Schaar, M. (2019). Modeling tabular data using conditional gan. Advances in Neural Information Processing Systems , 32. (Duplicate, already [42]. Let's find another relevant survey or specific tabular GAN pap er if possi...

  33. [45]

    TGAN is another op tion

    (Revisiting Tabular GANs) Let's use a different, well -known tabular GAN. TGAN is another op tion. Or simply cite a survey. [26] and [28] are good surveys. Let's cite [26] and [28] for the concept of using generative models like GANs for tabular data

  34. [46]

    (Revised citation plan): Use [26] and [28] for the general idea of generative models for tabular data, and keep

  35. [47]

    for CTGAN as a specific example

  36. [48]

    Wang, S., Cao, J., Zhang, X., Wang, X., & Zheng, Y. (2020). Graph data augmentation for graph neural networks. arXiv preprint arXiv:2005.10213

  37. [49]

    Verma, V., Raghunathan, A., Steinhardt, J., Liang, T., & Ma, T. (2019). Manifold mixup: Better representations by interpolating hidden states. In International Conference on Machine Learning (pp. 6432-6441). PMLR

  38. [50]

    Differential evolution - based neural network training incorporating a centroid-based strategy and dynamic opposition-based learning

    MOUSAVIRAD, Seyed Jalaleddin, OLIVA, Diego, HINOJOSA, Salvador, et al. Differential evolution - based neural network training incorporating a centroid-based strategy and dynamic opposition-based learning. In : 2021 IEEE congress on evolutionary computation (CEC). IEEE, 2021. p...

  39. [51]

    Kalra, S., Sriram, A., Rahnamayan, S., & Tizhoosh, H. R. (2016, December). Learning opposites using neural networks. In 2016 23rd International Conference on Pattern Recognition (ICPR) (pp. 1213 -1218). IEEE

  40. [52]

    Hyper -parameter tuning for support vector machine using an improved cat swarm optimization algorithm,

    A. W. Hadi and I. I. P. Singh, "Hyper -parameter tuning for support vector machine using an improved cat swarm optimization algorithm," Journal of Natural Sciences and Practical Medicine , vol. 6, no. 1, 2023, doi: 10.46481/jnsps.2023.1007

  41. [53]

    UCI machine learning repository https://archive.ics.uci.edu/datasets

  42. [54]

    Datasets : https://jundongl.github.io/scikit-feature/datasets.html

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.