Gradient-Optimized Fuzzy Classifier: A Benchmark Study Against State-of-the-Art Models

Kelly Cohen; Magnus Sieverding; Nathan Steffen

REVIEW 4 major objections 5 minor 17 references

Gradient-Optimized Fuzzy Classifier: A Benchmark Study Against State-of-the-Art Models

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a gradient-trained fuzzy classifier can beat or match state-of-the-art models on structured classification benchmarks while training in seconds.

desk verdict The paper's central claim of GF superiority is not supported: the comparison uses UCI chart bests with unknown protocols against the authors' own 5-fold CV, and the GF method is never described in enough detail to reproduce. read the letter →

arxiv 2504.16263 v1 pith:LQ6A2KKR submitted 2025-04-22 cs.LG cs.AI

Magnus Sieverding , Nathan Steffen , Kelly Cohen This is my paper · ORCID

classification cs.LGcs.AI

keywords fuzzyinferencesystemgradientdescentoptimizationclassificationbenchmarkinterpretableAITakagi-Sugeno-Kangcross-validationexplainablemachinelearning

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 claims that a fuzzy inference system trained with gradient descent can match or beat established classifiers such as Random Forest, XGBoost, logistic regression, support vector machines, and neural networks on structured classification tasks, while training in seconds rather than minutes or hours. The claim is tested on five public datasets with different feature types and class counts: credit risk, breast cancer, car evaluation, heart disease, and wine. The reported results have the gradient-optimized fuzzy classifier reaching the top mean accuracy on the breast-cancer and heart-disease datasets, tying the best model with 100 percent accuracy on the wine dataset, coming 0.075 percent below the best on credit data, and falling behind XGBoost only on the all-categorical car dataset. Across all five datasets, its accuracy varies little between cross-validation folds, which the paper reads as evidence that the model handles noisy and small data without the brittleness often attributed to deep networks. If the claim holds, interpretable fuzzy classifiers become a practical default for applications where transparency and training speed matter.

What carries the argument

The Gradient-Optimized Fuzzy Inference System (GF): a fuzzy classifier whose membership functions and rule base are tuned by gradient descent, specifically the ADAM optimizer, on a cross-entropy classification loss. This is what carries the argument, because standard fuzzy systems use derivative-free optimizers that converge slowly and scale poorly. Gradient descent gives the GF fast training (single-digit seconds in most cases) and stable behavior across folds; the rule base and membership functions give it interpretability. The paper also treats the number of membership functions and rules as adaptable per dataset (for example, 6 membership functions and 85 rules on German credit, 27 and 128 on car evaluation), which lets the same mechanism adapt to tasks of different complexity.

What would settle it

Run GF and each comparison model on the same five datasets with identical preprocessing, the same five-fold split, and similar hyperparameter budgets; if GF's mean accuracy no longer beats or ties the best model on the breast-cancer, heart-disease, wine, and credit datasets, the paper's central competitive claim fails.

Watch

Extended reading notes

Core claim

The central claim is that replacing the derivative-free optimization normally used to train a fuzzy inference system with gradient descent removes the main practical drawback of fuzzy classifiers and makes them competitive with modern machine learning models. The paper's GF classifier tunes fuzzy membership functions and rule parameters by minimizing cross-entropy loss with the ADAM optimizer, and it is evaluated with five-fold cross-validation on five datasets. On the German credit dataset it reaches a maximum accuracy of 83.125 percent, within 0.075 percent of the top Random Forest score; on breast cancer it reaches a mean accuracy of 98.901 percent, the highest among all compared models, and 100 percent maximum; on heart disease it tops every reported measure; on wine it achieves 100 percent accuracy on every fold; and on car evaluation its 96.029 percent maximum trails XGBoost's 99.769 percent, with the paper attributing the gap to the dataset's all-categorical, few-category inputs. Training times average between 3.6 and 16.7 seconds across the five datasets. The paper concludes that gradient-optimized fuzzy models are viable alternatives to more complex models in real-world classification, particularly where interpretability and speed are priorities.

Load-bearing premise

The load-bearing premise is that the other models' published best-accuracy figures, produced under unknown preprocessing, splitting, and hyperparameter settings, are directly comparable to the GF model's own five-fold cross-validated results.

Editorial extensions

If this is right

Structured-data applications that need explainable decisions, such as credit scoring or medical screening, can use a classifier that trains in seconds and still matches top accuracies.
The interpretability-versus-accuracy trade-off often assumed for fuzzy systems is, under this evidence, not a strict trade-off: the GF reaches top-tier accuracy on four of the five datasets.
On datasets with few categorical inputs and several classes, a gradient-optimized fuzzy classifier may need feature extraction or tuned membership functions to stay competitive, as the car-evaluation result shows.
The consistently low fold-to-fold variance reported for GF suggests it should be evaluated on small, noisy datasets where neural networks tend to overfit.
Because GF converged before the 250-epoch cap in every dataset and trained in under 17 seconds, early stopping is unnecessary and further training-time gains are available on newer hardware.

Reading between the lines

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

The paper leaves implicit that the GF's speed advantage is most visible at the small-to-mid dataset scale; a natural extension would benchmark it on larger, higher-dimensional tasks, where gradient-based fuzzy systems may need more rules or features.
A direct test of the categorical-input hypothesis would run GF on the car-evaluation data after applying feature extraction or an embedding encoding while keeping the same cross-validation folds; if the accuracy gap to XGBoost closes, the bottleneck is fuzzification of already-categorical inputs.
A matched-protocol head-to-head, in which every model is trained on the exact same folds and preprocessing pipeline, would be a stronger comparison than the published best-accuracy figures used in the paper.
One practical consequence not drawn by the authors is that the same gradient-trained fuzzy architecture could be extended to regression or multi-label tasks, since the cross-entropy objective is only one of several differentiable losses applicable to the fuzzy rule parameters.

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.

Desk Editor's Note

read the letter

Two things to know up front. First, the headline claim—that the gradient-optimized fuzzy classifier (GF) is competitive or superior to random forests and XGBoost—rests on an uncontrolled comparison. GF numbers come from the authors' 5-fold CV with one preprocessing pipeline; baseline numbers are taken from UCI performance charts, which may reflect different splits, preprocessing, hyperparameter tuning, or test-set reporting. The paper even admits these protocols are "not always uniform or fully documented," then assumes they represent each model's optimal usage. Every "outperformed" statement in Table 3 inherits that assumption, so the central conclusion is not established. No baseline is re-run, no significance test is done, and no training times are reported for the baselines, which undercuts the "exceptionally low training times" claim as well.

Second, the GF method itself is under-specified. The paper gives no membership function type, no rule-generation procedure, no fuzzification/defuzzification details, no loss equations, and no gradient-flow explanation. Per-dataset rule counts (85, 202, 128, 300, 300) appear without justification, raising selection-bias concerns. No code is released. This is not reproducible in any practical sense. The missing ANFIS citation is also a real gap, since gradient-trained fuzzy inference systems are an established idea in the neuro-fuzzy literature; the paper's framing as a novel contribution is misleading.

To give credit where it is due: the paper reports min/mean/max validation accuracy rather than a single number, which is more transparent than many benchmark studies. It also explicitly acknowledges the limitations of the UCI benchmark data, and the dataset choices are standard and varied. Those are good habits, but they do not fix the load-bearing comparison flaw.

This is not a minor issue. The paper's own evidence cannot support its abstract. As is, I would desk reject it. But the path to a publishable paper is clear: rerun the baselines under identical folds and preprocessing, report the GF architecture with full equations, and supply code or at least pseudocode. With that, a controlled benchmark of a properly described GF could be a useful empirical note for the interpretable-ML community. Without it, there is nothing for a referee to verify.

Referee Report

4 major / 5 minor

Summary. The manuscript introduces a Gradient-Optimized Fuzzy Inference System (GF) classifier and reports a benchmarking study against Random Forest, XGBoost, Logistic Regression, SVM, and Neural Networks on five UCI datasets (German Credit, Breast Cancer Wisconsin, Car Evaluation, Heart Disease, Wine). The GF is trained with the ADAM optimizer and evaluated with 5-fold cross-validation; the authors report minimum, mean, and maximum validation accuracy per dataset, together with GF training times. Baseline numbers are taken from UCI performance charts rather than being re-run. The central claim is that the GF achieves competitive or superior classification accuracy with high consistency and very low training times, and is therefore a viable interpretable alternative to state-of-the-art models.

Significance. If the comparison were controlled, the paper would provide a useful empirical data point for gradient-trained fuzzy systems as interpretable classifiers on small tabular benchmarks, and the reporting of per-dataset dataset characteristics and GF configurations is a helpful starting point. The authors are also honest about the heterogeneity of UCI benchmark protocols. However, the central comparative claim is not currently established: no baseline model is re-run under identical conditions, no uncertainty or significance measures are reported, and the 'low training time' claim lacks any baseline timings. The paper is an empirical benchmark whose main conclusion rests on uncontrolled external numbers, so the significance of the reported results is substantially lower than the abstract suggests. With a controlled re-benchmark, the contribution could be salvageable, but the current manuscript does not support its headline claims.

major comments (4)

[Section 3, Table 3] The benchmark comparison is uncontrolled. GF results come from the authors' 5-fold cross-validation with a single preprocessing pipeline (integer encoding plus min-max normalization, Section 3), while baseline results are transcribed from UCI performance charts whose preprocessing, train/test splits, and hyperparameter optimization are unknown and, in the authors' own words, 'assumed to represent each model's optimal usage.' Comparing the GF fold distribution to a single published accuracy value is not a valid statistical test. For example, the Heart Disease claim of outperforming XGBoost and Logistic Regression rests on a 0.109 percentage point difference between the GF's best fold and an external chart value (89.583 vs. 89.474). Every comparative statement in the abstract and conclusion depends on this comparison. The authors should re-run all baseline models under the same 5-fold CV and preprocessing protocol, report mean and standard deviation, and use a paired significance test where applicable.
[Sections 2 and 3.1–3.5] The 'exceptionally low training times' claim is not supported by comparative evidence. The paper reports only GF training times (e.g., 8.985 s on Statlog, 6.005 s on Breast Cancer, 16.691 s on Car Evaluation) and provides no training-time measurements for Random Forest, XGBoost, Logistic Regression, SVM, or Neural Networks under the same hardware and stopping conditions. Without baseline timings, the efficiency advantage asserted in the abstract and conclusion is not established.
[Sections 3.1–3.5] The number of membership functions and rules is reported per dataset (6/85, 13/202, 27/128, 13/300, 13/300), but no procedure is described for selecting these values, and the ADAM hyperparameters (learning rate, initialization, regularization, batch handling) are not given. Without a stated model-selection protocol, the per-dataset choices could reflect test-set tuning, which would compromise the generalization claims. The authors should report all hyperparameters and describe a validation-based selection rule.
[Section 3.1, Section 3.3] Statements such as 'outperformed the mean accuracy of all other models by 2.625%' compare the GF 5-fold mean against a point value from an external chart with no measure of uncertainty. Additionally, the paper uses 'precision' to mean a narrow range of fold accuracies (e.g., Section 3.3), which is not the standard classification-precision metric; classification precision as a per-class metric is never actually reported, despite being mentioned in the abstract.

minor comments (5)

[Abstract] The phrase 'significantly improving training efficiency' should be 'significantly improve training efficiency' for grammatical parallelism with the surrounding clauses.
[Section 1] The sentence ending 'between a head full of hair and being bald. something which other models...' has a capitalization and punctuation error; 'something' should begin a new sentence and be capitalized.
[General] No code, data, or hyperparameter table is provided, which makes the GF implementation difficult to reproduce. A link to a repository or an appendix with the full configuration would be needed for a journal submission.
[Table 3] The baseline values from UCI charts lack citations to the original benchmark papers and are presented without any indication of their source protocol. The table would be clearer if GF rows were visually separated and if a note indicated which entries come from the authors' 5-fold CV versus external charts.
[Section 3.4] The Heart Disease benchmark values for XGBoost and Logistic Regression are identical to three decimal places (89.474); the paper should explain whether this is a data-entry artifact or an actual coincidence, since treating them as independent evidence is questionable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical benchmark against external UCI chart values; the protocol mismatch is a correctness threat, not a circular derivation.

full rationale

The paper's claimed contribution is an empirical comparison, not a derivation. The GF model's accuracy and training-time numbers come from the authors' own 5-fold cross-validation runs (Section 2: 'a 5 fold K-Fold was used for each of the datasets'), while the comparison targets are externally sourced UCI performance charts (Section 3: 'Benchmark results for each dataset were sourced from the performance charts provided by the UCI Machine Learning Repository'). There is no equation in the paper that defines a benchmark value in terms of the GF's fitted parameters, no fitted parameter that is renamed as a prediction, and no load-bearing self-citation: the reference list contains no works by Sieverding, Steffen, or Cohen, so the claimed result is not forced by a self-citation chain. The closest issue is the passage in Section 3 stating that UCI benchmarks 'are not always uniform or fully documented, they are assumed to represent each model's optimal usage.' That is a genuine validity threat to the superiority claim, because GF's 5-fold CV results are compared to best-reported external numbers under unknown preprocessing and evaluation protocols; it is also possible, though not stated in the paper, that the per-dataset numbers of membership functions and fuzzy rules were selected using the same data used for evaluation. However, the manuscript does not say those hyperparameters were tuned on the test folds, and there is no exhibited reduction of any claimed result to a fitted input. Under the hard rule that circularity must be shown by quote and specific reduction, these methodological weaknesses are correctness risks, not circularity. The paper is self-contained against external benchmarks, so the appropriate finding is no significant circularity, score 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new theoretical entities. Its free parameters are the per-dataset architectural choices (membership functions, rules) that are not justified or disclosed as tuned, and the unspecified optimizer settings. The key domain assumption is that UCI benchmark charts are a valid baseline, which is explicitly stated and is load-bearing for the central claim.

free parameters (5)

Number of membership functions per dataset = 6, 13, 27, 13, 13
Reported per dataset in Sections 3.1-3.5. The paper does not explain how these counts were chosen; if tuned on the evaluation data, the accuracy is partly a result of model selection.
Number of fuzzy rules per dataset = 85, 202, 128, 300, 300
Reported per dataset. Same concern as membership functions.
ADAM optimizer hyperparameters
The paper only states ADAM was used (Section 2). Learning rate, beta values, and epsilon are not given, so default or hand-chosen values are implied.
Maximum training epochs = 250
Section 2 states training ran up to 250 epochs; although convergence was earlier, this maximum is a free parameter.
Features removed from Heart Disease = 'ca' and 'thal'
Section 3.4 states these attributes were removed due to missing values; this is a modeling choice affecting results.

assumptions (4)

domain assumption UCI repository benchmark values represent optimal usage of each baseline model
Stated in Section 3: 'they are assumed to represent each model's optimal usage.' This is load-bearing for the comparison.
domain assumption Min-Max normalization and integer encoding of categorical variables are sufficient preprocessing
Stated in Section 3. This may not match the preprocessing used for the UCI benchmarks.
domain assumption Gradient descent is applicable to the fuzzy inference system
The paper states this but does not provide the differentiability or architecture details needed to verify it.
domain assumption 5-fold cross-validation yields stable estimates on these datasets
Used throughout; on small datasets like Wine (178 samples) 5-fold estimates have high variance.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Gradient-Optimized Fuzzy Classifier: A Benchmark Study Against State-of-the-Art Models." pith.science (2026). https://pith.science/paper/LQ6A2KKR

@misc{pith2026250416263,
  author       = {Pith},
  title        = {Pith review of: Gradient-Optimized Fuzzy Classifier: A Benchmark Study Against State-of-the-Art Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LQ6A2KKR}},
  note         = {Machine review of arXiv:2504.16263}
}

read the original abstract

This paper presents a performance benchmarking study of a Gradient-Optimized Fuzzy Inference System (GF) classifier against several state-of-the-art machine learning models, including Random Forest, XGBoost, Logistic Regression, Support Vector Machines, and Neural Networks. The evaluation was conducted across five datasets from the UCI Machine Learning Repository, each chosen for their diversity in input types, class distributions, and classification complexity. Unlike traditional Fuzzy Inference Systems that rely on derivative-free optimization methods, the GF leverages gradient descent to significantly improving training efficiency and predictive performance. Results demonstrate that the GF model achieved competitive, and in several cases superior, classification accuracy while maintaining high precision and exceptionally low training times. In particular, the GF exhibited strong consistency across folds and datasets, underscoring its robustness in handling noisy data and variable feature sets. These findings support the potential of gradient optimized fuzzy systems as interpretable, efficient, and adaptable alternatives to more complex deep learning models in supervised learning tasks.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 10 canonical work pages

[1]

URL: https://nvlpubs.nist.gov/nistpubs/ai/nist.ai (2023)

NIST: Artiﬁcial intelligence risk management framework (AI RMF 1.0). URL: https://nvlpubs.nist.gov/nistpubs/ai/nist.ai (2023)

work page 2023
[2]

L.: Brittle AI, causal confusion, and bad mental models: challenges and successes in the XAI program

Druce, J., Niehaus, J., Moody, V., Jensen, D., Littman, M. L.: Brittle AI, causal confusion, and bad mental models: challenges and successes in the XAI program. arXiv preprint arXiv:2106.05506 (2021)

arXiv 2021
[3]

Estimating the Brittleness of AI: Safety Integrity Levels and the Need for Testing Out-Of-Distribution Performance

Lohn, A.J.: Estimating the brittleness of AI: Safety inte grity levels and the need for testing out-of-distribution performance. arXiv preprint arXiv:2009.00802 (2020)

work page Pith review arXiv 2020
[4]

In : 2007 International Symposium on Computational Intelligence and Intelligent I nformatics, pp

Habbi, A.H., Zelmat, M.: Fuzzy logic based gradient desce nt method with appli- cation to a PI-type fuzzy controller tuning: New results. In : 2007 International Symposium on Computational Intelligence and Intelligent I nformatics, pp. 93–97. IEEE (2007)

work page 2007
[5]

Integrated Computer-Aided Engineering 32(1), 25–38 (2025)

Succetti, F., Rosato, A., Panella, M.: Multi-label class iﬁcation with imbalanced classes by fuzzy deep neural networks. Integrated Computer-Aided Engineering 32(1), 25–38 (2025)

work page 2025
[6]

A Preprint 1 (2020) 12 M

Masoumi, M., Hossani, S., Dehghani, F., Masoumi, A.: The c hallenges and advan- tages of fuzzy systems applications. A Preprint 1 (2020) 12 M. Sieverding

work page 2020
[7]

Pekaslan, D., Chen, C., Wagner, C., Garibaldi, J.M.: Perf ormance and Inter- pretability in Fuzzy Logic Systems–can we have both? In: Information Processing and Management of Uncertainty in Knowledge-Based Systems: 18th International Conference, IPMU 2020, Lisbon, Portugal, June 15–19, 2020, Proceedings, Part I, pp. 571–584. Springer (2020)

work page 2020
[8]

In: 2023 14th International Conference on Intelligent Systems: Theories and Applications (SITA) , pp

Letrache, K., Ramdani, M.: Explainable Artiﬁcial Intell igence: A Review and Case Study on Model-Agnostic Methods. In: 2023 14th International Conference on Intelligent Systems: Theories and Applications (SITA) , pp. 1–8. IEEE (2023)

work page 2023

Show all 17 references

[9]

arXiv preprint arXiv:1609.04747 (2016)

Ruder, S.: An overview of gradient descent optimization a lgorithms. arXiv preprint arXiv:1609.04747 (2016)

2016 arXiv
[10]

Journal of Global Optimization 56(3), 1247–1293 (2013)

Rios, L.M., Sahinidis, N.V.: Derivative-free optimiza tion: a review of algorithms and comparison of software implementations. Journal of Global Optimization 56(3), 1247–1293 (2013)

2013
[11]

Information and Control 8(3), 338–353 (1965)

Zadeh, L.A.: Fuzzy sets. Information and Control 8(3), 338–353 (1965)

1965
[12]

Information Fusion 101, 101977 (2024)

Zhang, Y., Wang, G., Zhou, T., Huang, X., Lam, S., Sheng, J ., Choi, K.S., Cai, J., Ding, W.: Takagi-Sugeno-Kang fuzzy system fusion: A sur vey at hierarchical, wide and stacked levels. Information Fusion 101, 101977 (2024)

2024
[13]

UCI Machine L earning Repository (1994)

Hofmann, H.: Statlog (German Credit Data). UCI Machine L earning Repository (1994). https://doi.org/10.24432/C5NC77

1994 doi
[14]

UCI Machine Learning Reposito ry (1993)

Wolberg, W., Mangasarian, O., Street, N., Street, W.: Br east Can- cer Wisconsin (Diagnostic). UCI Machine Learning Reposito ry (1993). https://doi.org/10.24432/C5DW2B

1993 doi
[15]

UCI Machine Learning Repos itory (1988)

Bohanec, M.: Car Evaluation. UCI Machine Learning Repos itory (1988). https://doi.org/10.24432/C5JP48

1988 doi
[16]

UCI Machine Learning Repository (1989)

Janosi, A., Steinbrunn, W., Pﬁsterer, M., Detrano, R.: H eart Disease. UCI Machine Learning Repository (1989). https://doi.org/10.24432/C 52P4X

1989 doi
[17]

UCI Machine Learning Re pository (1992)

Aeberhard, S., Forina, M.: Wine. UCI Machine Learning Re pository (1992). https://doi.org/10.24432/C5PC7J

1992 doi

Pith tools

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

[1] [1]

URL: https://nvlpubs.nist.gov/nistpubs/ai/nist.ai (2023)

NIST: Artiﬁcial intelligence risk management framework (AI RMF 1.0). URL: https://nvlpubs.nist.gov/nistpubs/ai/nist.ai (2023)

work page 2023

[2] [2]

L.: Brittle AI, causal confusion, and bad mental models: challenges and successes in the XAI program

Druce, J., Niehaus, J., Moody, V., Jensen, D., Littman, M. L.: Brittle AI, causal confusion, and bad mental models: challenges and successes in the XAI program. arXiv preprint arXiv:2106.05506 (2021)

arXiv 2021

[3] [3]

Estimating the Brittleness of AI: Safety Integrity Levels and the Need for Testing Out-Of-Distribution Performance

Lohn, A.J.: Estimating the brittleness of AI: Safety inte grity levels and the need for testing out-of-distribution performance. arXiv preprint arXiv:2009.00802 (2020)

work page Pith review arXiv 2020

[4] [4]

In : 2007 International Symposium on Computational Intelligence and Intelligent I nformatics, pp

Habbi, A.H., Zelmat, M.: Fuzzy logic based gradient desce nt method with appli- cation to a PI-type fuzzy controller tuning: New results. In : 2007 International Symposium on Computational Intelligence and Intelligent I nformatics, pp. 93–97. IEEE (2007)

work page 2007

[5] [5]

Integrated Computer-Aided Engineering 32(1), 25–38 (2025)

Succetti, F., Rosato, A., Panella, M.: Multi-label class iﬁcation with imbalanced classes by fuzzy deep neural networks. Integrated Computer-Aided Engineering 32(1), 25–38 (2025)

work page 2025

[6] [6]

A Preprint 1 (2020) 12 M

Masoumi, M., Hossani, S., Dehghani, F., Masoumi, A.: The c hallenges and advan- tages of fuzzy systems applications. A Preprint 1 (2020) 12 M. Sieverding

work page 2020

[7] [7]

Pekaslan, D., Chen, C., Wagner, C., Garibaldi, J.M.: Perf ormance and Inter- pretability in Fuzzy Logic Systems–can we have both? In: Information Processing and Management of Uncertainty in Knowledge-Based Systems: 18th International Conference, IPMU 2020, Lisbon, Portugal, June 15–19, 2020, Proceedings, Part I, pp. 571–584. Springer (2020)

work page 2020

[8] [8]

In: 2023 14th International Conference on Intelligent Systems: Theories and Applications (SITA) , pp

Letrache, K., Ramdani, M.: Explainable Artiﬁcial Intell igence: A Review and Case Study on Model-Agnostic Methods. In: 2023 14th International Conference on Intelligent Systems: Theories and Applications (SITA) , pp. 1–8. IEEE (2023)

work page 2023