Pith. sign in

REVIEW 3 major objections 6 minor 43 references

Kolmogorov-Arnold Networks-based GRU and LSTM for Loan Default Early Prediction

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Adding a Kolmogorov-Arnold Network layer to GRU and LSTM predicts loan defaults 3–8 months ahead at over 92%/88% accuracy on out-of-time data, beating all baselines.

desk verdict A careful but flawed empirical study: the KAN-hybrid results look real at each fixed setting, but the early-prediction experiment confounds gap length with input history, so the central robustness claim doesn't hold as stated. read the letter →

arxiv 2507.13685 v1 pith:27NLZBUN submitted 2025-07-18 cs.LG

Yue Yang , Zihan Su , Ying Zhang , Chang Chuan Goh , Yuxiang Lin , Anthony Graham Bellotti , Boon Giin Lee This is my paper
classification cs.LG
keywords loandefaultpredictionKolmogorov-ArnoldNetworksGRULSTMearlyout-of-timedatatimeseriesanomalydetectioncreditrisk
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 aims to show that grafting a Kolmogorov-Arnold Network (KAN) layer onto a GRU or LSTM backbone produces loan-default models that can predict defaults three to eight months before they happen, using training and test data from different years. The proposed GRU-KAN and LSTM-KAN models are compared against GRU, LSTM, LSTM-Attention, and LSTM-Transformer baselines on the Freddie Mac single-family loan-level dataset. The paper reports that GRU-KAN reaches over 92% accuracy with a three-month blank interval, stays above 88% accuracy at an eight-month interval, and beats all baselines on accuracy, recall, and F1 for every interval from 3 to 8 months. If true, this would give banks roughly two extra months of actionable warning compared with the best baselines at the same accuracy, and would make the predictions usable in near-real-time settings where models are trained and tested on different calendar years.

What carries the argument

The load-bearing component is the Kolmogorov-Arnold Network (KAN) layer, defined in the paper as a network of learnable univariate spline functions on each edge, replacing the fixed node activations and linear weights of a multilayer perceptron; the KAN output is a sum of composed univariate functions $\sum_{q=1}^{2n+1}\Phi_q\left(\sum_{p=1}^{n}\phi_{p,q}(x_p)\right)$. The full model is a masking layer, two stacked GRU or LSTM layers (128 units then 64 units with batch normalization between), the KAN layer with ten functions and one output dimension, a Dense(64, ReLU) layer, Dropout(0.3), and a sigmoid output. The recurrent layers provide a 64-dimensional temporal summary of the masked monthly loan features, and the claim is that the KAN layer's learned edge-wise nonlinearities are what let the model keep separating defaulters from non-defaulters when a blank interval separates the input window from the observation period.

What would settle it

Run the blank-interval experiment again with the input feature window held fixed at 18 months for all gaps (3 through 8), letting the total window-plus-gap grow to 26 months instead of 21; if GRU-KAN's accuracy and F1 still decline at the same rate and still exceed the baselines, the early-prediction claim is confirmed, whereas if the decline mostly disappears or the gap between GRU-KAN and LSTM shrinks, the original result was driven by the shrinking input window rather than the gap itself.

Watch

Extended reading notes

Core claim

The central claim is that inserting a KAN layer after the recurrent feature extractor changes how the model uses degraded input. Where GRU and LSTM rely on fixed activation functions and linear output weights, the KAN layer learns a separate univariate spline function for each edge, so the mapping from the 64-dimensional temporal summary to default risk is itself learned and nonlinear. Under a blank-interval design—where the input window ends i months before the 3-month observation window, with i ranging from 3 to 8 and the total window-plus-gap held at 21 months—GRU-KAN records mean accuracy 0.9268 at i=3 and 0.8860 at i=8, with F1 above 0.7 and recall above 0.6 even at the longest gap. The authors interpret this as evidence that the KAN layer preserves the signal that the recurrent layers extract, so the gap between observed history and the default event is less damaging than for the baselines; GRU-KAN at a 5-month gap matches the accuracy, recall, and F1 that the best baselines reach at a 3-month gap. The same models also generalize across cohorts trained on 2018 and tested on 2019–2022, where GRU-KAN has the highest accuracy, F1, and recall in nearly every cohort pair, which the authors attribute to greater robustness to concept drift.

Load-bearing premise

The early-prediction experiments keep the total of input window plus blank interval at 21 months, so as the gap grows from 3 to 8 months the model sees 5 fewer months of borrower history; the paper reads the resulting performance drop as the cost of predicting further ahead, and this confound is the premise the 'robustness to early prediction' claim rests on.

Editorial extensions

If this is right

  • GRU-KAN delivering F1 above 0.7 and recall above 0.6 with an 8-month gap means a lender could issue default warnings roughly two months earlier than with the best baseline at the same accuracy, without waiting for delinquency to begin.
  • Because the models are trained on 2019 data and tested on 2020–2022 cohorts, the result supports using the KAN hybrids in rolling out-of-time deployments where the model is retrained on past years and applied to the current year.
  • At a 5-month blank interval, GRU-KAN and LSTM-KAN reach the same accuracy, recall, and F1 that the non-KAN baselines reach at a 3-month interval, so the KAN layer effectively buys two months of extra lead time for the same predictive quality.
  • LSTM-Attention and LSTM-Transformer lose recall more sharply than the recurrent baselines as the gap grows, so the paper's comparison casts doubt on the value of attention/transformer layers for this particular early-prediction setup.
  • The KAN hybrids' performance is close to each other across window lengths and sample sizes, so the cheaper GRU backbone can be used instead of LSTM without sacrificing much accuracy.

Reading between the lines

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

  • A direct test of the mechanism, not run in the paper, would be to replace the KAN layer with a same-size MLP; if the MLP matches the baselines too, the gain is extra capacity rather than KAN's spline parameterization. This is an editorial inference.
  • The absolute accuracy numbers come from an undersampled balanced test set; my inference is that a lender's real, imbalanced portfolio would show lower accuracy at the 0.5 threshold, so the durable result is the relative ordering, not the raw percentages.
  • The blank-interval protocol generalizes naturally to other 'predict from stale features' settings such as fraud detection or churn; my inference is that KAN's edge-wise splines may help wherever the informative features sit months before the event.
  • Because KAN's univariate functions are inspectable, my inference is that a follow-up could recover which borrower features drive late default signals, something the paper leaves undone.
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

3 major / 6 minor

Summary. The paper proposes two hybrid recurrent architectures, GRU-KAN and LSTM-KAN, that replace the final feature transformation of GRU/LSTM with a Kolmogorov-Arnold Network layer, and evaluates them on the Freddie Mac Single-Family Loan-Level dataset for loan-default prediction. The experimental program is organized around three axes: feature-window length, blank-interval length for early prediction, and training-sample size, with an additional cross-cohort generalization analysis. The authors report that the proposed models outperform LSTM, GRU, LSTM-Attention, and LSTM-Transformer baselines, with headline accuracies above 92% three months in advance and above 88% eight months in advance. The central empirical claim is plausible in its broad outline, but the early-prediction experiment does not cleanly isolate the blank-interval effect, and the headline accuracy is computed on an artificially balanced test set rather than on the real borrower distribution.

Significance. If the claimed early-prediction gains were cleanly established, the paper would be a useful applied contribution to credit-risk time series modeling: it uses a public benchmark dataset, an explicit out-of-time train/test split, a reasonable default definition, and a broad set of recurrent/attention baselines, and it reports means and standard deviations over 20 trials. These design choices are strengths that make the comparisons more informative than a single-run evaluation. However, the central novelty claim, that the blank interval enables robust early prediction, is contaminated by a simultaneous change in input-window length, and the operational accuracy claim is based on a balanced test set that does not reflect real class prevalence. Both issues are fixable within the scope of the manuscript, provided the authors add controlled experiments and prevalence-aware reporting.

major comments (3)
  1. [Section 4.2 (Figure 4, Table A2)] The early-prediction experiment as designed cannot separate the effect of the blank interval from the effect of the input-window length. Section 4.2 fixes x + interval = 21 months, so when the interval grows from 3 to 8 months, the input window shrinks from 18 to 13 months. Since Section 4.1 (Table A1) shows that all models, including the KAN variants, are sensitive to feature-window length, the monotone decline in Table A2 could be caused by reduced input history rather than by the prediction gap. The claim that GRU-KAN and LSTM-KAN display greater robustness and adaptability over different interval settings is therefore not cleanly identified, and the headline 'over 88% eight months in advance' is measured with a 13-month window. Please run a controlled comparison with fixed input window (e.g., x=18 months) and varying interval, or otherwise decouple the two factors.
  2. [Section 3.2.2 and Table 2] The headline accuracy values are computed on a class-balanced test set, not on the real test population. Section 3.2.2 states that random undersampling is applied to the test dataset so that default and non-default instances are equal, while Table 2 gives the actual 2020 test default rate as 0.811%. Accuracy on a 50/50 test set is not the operational accuracy a bank would observe on the borrower population, and it can be inflated by models that over-predict the positive class; precision and F1 on the balanced set do not translate to population prevalence either. Please report metrics on the original test distribution in addition to the balanced-set metrics, or use a cost-sensitive evaluation that accounts for the true class ratio.
  3. [Section 4.2 and Table A2] Claims of consistent superiority and robustness are supported only by point estimates without uncertainty quantification. In Table A2, the GRU-KAN and LSTM-KAN advantages over LSTM at interval 8 are small relative to the reported standard deviations (e.g., F1 for GRU-KAN is 0.7372±0.0211 versus 0.7069±0.0193 for LSTM), and several baseline rows have extremely large standard deviations (e.g., GRU F1 0.0982±0.2285). No paired significance tests are reported. Please add appropriate tests across the 20 trials (e.g., Wilcoxon signed-rank or paired bootstrap), or otherwise temper statements that the proposed models consistently outperformed all baselines.
minor comments (6)
  1. [Sections 3.2.2 and 4.2] Please clarify whether the blank-interval experiments used the undersampled balanced set or an unmodified complete set: Section 3.2.2 says random undersampling is applied to both training and test sets, while Section 4.2 says both training and testing phases utilize the complete dataset.
  2. [Section 3.1, Equation (11) and following text] The sentence after Equation (11) contains a typo ('weather' should be 'whether') and an ungrammatical fragment ('zt here determine weather updates'); please revise.
  3. [Reference [29]] The KAN reference is cited as 'Anonymous... Under review'; this paper has since been published, so the placeholder citation should be replaced with a complete published reference.
  4. [Section 3.1, Algorithms 1-2] Reproducibility would be improved by specifying the KAN hyperparameters beyond the number of functions, such as grid size, spline order, initialization, and any regularization; alternatively, provide a code repository.
  5. [Section 2.2.1] The text says the study compared 'five machine learning classifiers' but then lists seven (Gaussian Naive Bayes, AdaBoost, Gradient Boosting, K-Nearest Neighbors, Decision Trees, Random Forest, and Logistic Regression); please correct the count.
  6. [Acknowledgment] The acknowledgment begins with 'his work' and should read 'This work'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports direct empirical comparisons on a public dataset, and its self-citations are peripheral.

full rationale

The paper's central claims are empirical: GRU-KAN and LSTM-KAN are evaluated against external baselines on the Freddie Mac loan-level dataset, with accuracy, precision, recall, F1, and AUC computed from held-out out-of-time test data. The KAN, GRU, and LSTM equations are standard definitions and are not constructed from the reported outcome. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via citation. The self-citations, namely Wang et al. [2] and Yang et al. [28], are used only in the literature review to motivate OOT evaluation and to note that model performance is dataset-specific; they are not load-bearing premises of the reported results. The Section 4.2 early-prediction protocol fixes the combined input-window-plus-blank-interval length at 21 months, so longer blank intervals correspond to shorter input windows. This is an experimental confound that weakens the causal attribution of performance decline to the blank interval alone, but it is not circularity: the reported accuracy values are direct measurements, not quantities defined in terms of the claim. Accordingly, no circular step is exhibited, and the appropriate circularity score is 0.

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

The central claim rests on modeling choices rather than physical theory. The most important free parameter is the feature window length, selected on the test set; the undersampling ratio defines the metric scale; and the default definition and cohort sampling are domain assumptions. No invented entities are proposed.

free parameters (3)
  • Feature window length = 15 months
    Selected as the 'optimal length' from Section 4.1 based on accuracy on the 2020 out-of-time test set, then reused in Sections 4.2, 4.3, and 4.4. This is test-set-guided model selection, so the final numbers are optimistically biased.
  • Test-set undersampling ratio = 1:1
    Random undersampling creates equal default and non-default instances in the test set (Section 3.2.2). This determines the scale of the headline accuracy (over 92 percent) which would not apply to the real population with a roughly 0.8 percent default rate.
  • KAN layer capacity = output dim=1, num functions=10
    The KAN head configuration is fixed without ablation (Algorithms 1 and 2). Performance may depend on this choice, which is not varied or justified.
assumptions (3)
  • domain assumption Default is defined as CLDS (current loan delinquency status) >= 3, following the Basel II standard.
    Used in Section 3.2.1 to construct the binary target from Freddie Mac data. This is an industry convention, not derived in the paper.
  • domain assumption Accuracy on a class-balanced test set is an acceptable measure of practical prediction performance.
    Section 3.2.2 undersamples the test set to 50/50. The paper reports accuracy, precision, recall, F1, and AUC on this artificial distribution, which is appropriate for comparing classifiers but not for stating real-world error rates.
  • domain assumption The first 1.5 million repayment records from each cohort are representative of that cohort.
    Section 4.4 uses the first 1.5 million records per cohort for cross-year generalization without random sampling or inspection of selection bias.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Kolmogorov-Arnold Networks-based GRU and LSTM for Loan Default Early Prediction." pith.science (2026). https://pith.science/paper/27NLZBUN

@misc{pith2026250713685,
  author       = {Pith},
  title        = {Pith review of: Kolmogorov-Arnold Networks-based GRU and LSTM for Loan Default Early Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/27NLZBUN}},
  note         = {Machine review of arXiv:2507.13685}
}
read the original abstract

This study addresses a critical challenge in time series anomaly detection: enhancing the predictive capability of loan default models more than three months in advance to enable early identification of default events, helping financial institutions implement preventive measures before risk events materialize. Existing methods have significant drawbacks, such as their lack of accuracy in early predictions and their dependence on training and testing within the same year and specific time frames. These issues limit their practical use, particularly with out-of-time data. To address these, the study introduces two innovative architectures, GRU-KAN and LSTM-KAN, which merge Kolmogorov-Arnold Networks (KAN) with Gated Recurrent Units (GRU) and Long Short-Term Memory (LSTM) networks. The proposed models were evaluated against the baseline models (LSTM, GRU, LSTM-Attention, and LSTM-Transformer) in terms of accuracy, precision, recall, F1 and AUC in different lengths of feature window, sample sizes, and early prediction intervals. The results demonstrate that the proposed model achieves a prediction accuracy of over 92% three months in advance and over 88% eight months in advance, significantly outperforming existing baselines.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 39 canonical work pages

  1. [1]

    Attention-based Dynamic Multilayer Graph Neural Networks for Loan Default Prediction

    Zandi S, Korangi K, ’Oskarsd’ottir M, Mues C, Bravo C. Attention-based Dynamic Multilayer Graph Neural Networks for Loan Default Prediction. ArXiv. 2024;abs/2402.00299

    work page Pith review arXiv 2024

  2. [2]

    Discrete-Time Survival Models with Neural Networks for Age–Period–Cohort Analysis of Credit Risk

    Wang H, Bellotti A, Qu R, Bai R. Discrete-Time Survival Models with Neural Networks for Age–Period–Cohort Analysis of Credit Risk. Risks. 2024;12(2):31

    work page 2024

  3. [3]

    Stabilizing machine learning models with Age-Period- Cohort inputs for scoring and stress testing

    Breeden JL, Leonova YA. Stabilizing machine learning models with Age-Period- Cohort inputs for scoring and stress testing. In: Frontiers in Applied Mathematics and Statistics; 2023. Available from: https://api.semanticscholar.org/CorpusID: 259099888

    work page 2023

  4. [4]

    Explainable prediction of loan default based on machine learning models

    Zhu X, Chu Q, Song X, Hu P, Peng L. Explainable prediction of loan default based on machine learning models. Data Science and Management. 2023;6(3):123–133. https://doi.org/https://doi.org/10.1016/j.dsm.2023.04.003

    work page doi:10.1016/j.dsm.2023.04.003 2023

  5. [5]

    A Hierarchical User Intention-Habit Extract Network for Credit Loan Overdue Risk Detection

    Guo H, Ren X, Wang R, Cai Z, Shuang K, Sun Y. A Hierarchical User Intention- Habit Extract Network for Credit Loan Overdue Risk Detection. ArXiv. 2020;abs/2008.07796

    work page Pith review arXiv 2020

  6. [6]

    Loan Default Prediction with Machine Learning Techniques

    Lai L. Loan Default Prediction with Machine Learning Techniques. In: 2020 International Conference on Computer Communication and Network Security (CCNS); 2020. p. 5–9

    work page 2020

  7. [7]

    Machine learning approach for credit score analysis : a case study of predicting mortgage loan defaults; 2019

    Elmasry M. Machine learning approach for credit score analysis : a case study of predicting mortgage loan defaults; 2019. Available from: https://api. semanticscholar.org/CorpusID:182301171

    work page 2019

  8. [8]

    Loan Default Prediction Model Improve- ment through Comprehensive Preprocessing and Features Selection

    Al-qerem A, Al-Naymat G, Alhasan M. Loan Default Prediction Model Improve- ment through Comprehensive Preprocessing and Features Selection. In: 2019 International Arab Conference on Information Technology (ACIT); 2019. p. 235–240

    work page 2019

Show all 43 references
  1. [9]

    An LSTM and GRU based trading strategy adapted to the Moroccan market

    Touzani Y, Douzi K. An LSTM and GRU based trading strategy adapted to the Moroccan market. Journal of Big Data. 2021;8

    2021

  2. [10]

    Forecasting peer-to-peer platform default rate with LSTM neu- ral network

    Liang L, Cai X. Forecasting peer-to-peer platform default rate with LSTM neu- ral network. Electronic Commerce Research and Applications. 2020;43:100997. https://doi.org/https://doi.org/10.1016/j.elerap.2020.100997

    2020

  3. [11]

    Https://github.com/matmcreative/Lending- Club-Loan-Analysis/

    LendingClub.: Lending-Club Dataset. Https://github.com/matmcreative/Lending- Club-Loan-Analysis/

  4. [12]

    Machine Learning and Deep Learning for Loan Prediction in Banking: Exploring Ensemble Methods and Data Balancing

    Hussein Sayed E, Alabrah A, Hussein Rahouma K, Zohaib M, Badry RM. Machine Learning and Deep Learning for Loan Prediction in Banking: Exploring Ensemble Methods and Data Balancing. IEEE Access. 2024;12:193997–194019. 27 https://doi.org/10.1109/ACCESS.2024.3509774

    2024

  5. [13]

    Deep Learning-Based Credit Score Prediction: Hybrid LSTM-GRU Model

    ASL GS, Shamsi K, Thulasiram RK, Akcora C, Leung C. Deep Learning-Based Credit Score Prediction: Hybrid LSTM-GRU Model. In: 2023 IEEE Symposium Series on Computational Intelligence (SSCI); 2023. p. 395–400

    2023

  6. [14]

    GRU-Based Interpretable Multivariate Time Series Anomaly Detection in Industrial Control System

    Tang C, Xu L, Yang B, Tang Y, Zhao D. GRU-Based Interpretable Multivariate Time Series Anomaly Detection in Industrial Control System. Computers and Security. 2023;127:103094. https://doi.org/https://doi.org/10.1016/j.cose.2023. 103094

    2023 doi

  7. [15]

    Network Log Anomaly Detection Based on GRU and SVDD

    Liu S, Chen X, Peng X, Xiao R. Network Log Anomaly Detection Based on GRU and SVDD. In: 2019 IEEE Intl Conf on Parallel and Distributed Process- ing with Applications, Big Data and Cloud Computing, Sustainable Computing and Communications, Social Computing and Networking (ISPA...

    2019

  8. [16]

    Multidimensional Time Series Anomaly Detection: A GRU-based Gaussian Mixture Variational Autoencoder Approach

    Guo Y, Liao W, Wang Q, Yu L, Ji T, Li P. Multidimensional Time Series Anomaly Detection: A GRU-based Gaussian Mixture Variational Autoencoder Approach. In: Zhu J, Takeuchi I, editors. Proceedings of The 10th Asian Confer- ence on Machine Learning. vol. 95 of Proceedings of Mac...

    2018

  9. [17]

    Credit Card Fraud Detection Model Based on LSTM Recurrent Neural Networks

    Benchaji I, Douzi S, Ouahidi BE. Credit Card Fraud Detection Model Based on LSTM Recurrent Neural Networks. Journal of Advances in Information Technology. 2021;12:113–118

    2021

  10. [18]

    Rural micro-credit model design and credit risk assessment via improved LSTM algorithm

    Gao X, Yang X, Zhao Y. Rural micro-credit model design and credit risk assessment via improved LSTM algorithm. PeerJ Computer Science. 2023;9

    2023

  11. [19]

    Internet Financial Credit Risk Assessment with Sliding Window and Attention Mechanism LSTM Model

    Menggang L, Zhang Z, Ming L, Jia X, Liu R, Zhou X, et al. Internet Financial Credit Risk Assessment with Sliding Window and Attention Mechanism LSTM Model. Tehnicki vjesnik - Technical Gazette. 2023

    2023

  12. [20]

    AT-LSTM: An Attention- based LSTM Model for Financial Time Series Prediction

    Zhang X, Liang X, Zhiyuli A, Zhang S, Xu R, Wu B. AT-LSTM: An Attention- based LSTM Model for Financial Time Series Prediction. IOP Conference Series: Materials Science and Engineering. 2019;569

    2019

  13. [21]

    Enhanced credit card fraud detection based on attention mechanism and LSTM deep model

    Benchaji I, Douzi S, Ouahidi BE, Jaafari J. Enhanced credit card fraud detection based on attention mechanism and LSTM deep model. Journal of Big Data. 2021;8

    2021

  14. [22]

    Real-time anomaly detection based on long short-Term memory and Gaussian Mixture Model

    Ding N, Ma H, Gao H, Ma Y, Tan G. Real-time anomaly detection based on long short-Term memory and Gaussian Mixture Model. Comput Electr Eng. 2019;79

    2019

  15. [23]

    Https://www.kaggle.com/datasets/boltzmannbrain/nab

    NABDataset.: NAB Dataset. Https://www.kaggle.com/datasets/boltzmannbrain/nab. 28

  16. [24]

    Are Transformers Effective for Time Series Forecasting? In: AAAI Conference on Artificial Intelligence; 2022

    Zeng A, Chen MH, Zhang L, Xu Q. Are Transformers Effective for Time Series Forecasting? In: AAAI Conference on Artificial Intelligence; 2022. Available from: https://api.semanticscholar.org/CorpusID:249097444

    2022

  17. [25]

    Time-series anomaly detection with stacked Trans- former representations and 1D convolutional network

    Kim J, Kang HG, Kang P. Time-series anomaly detection with stacked Trans- former representations and 1D convolutional network. Eng Appl Artif Intell. 2023;120:105964

    2023

  18. [26]

    Time series prediction model using LSTM- Transformer neural network for mine water inflow

    Shi J, Wang S, Qu P, Shao J. Time series prediction model using LSTM- Transformer neural network for mine water inflow. Scientific Reports. 2024;14

    2024

  19. [27]

    Application of BiLSTM-Transformer in Portfolio Optimiza- tion

    Sun C, Li H, Fu S. Application of BiLSTM-Transformer in Portfolio Optimiza- tion. Advances in Economics, Management and Political Sciences. 2023

    2023

  20. [28]

    A comprehensive study on the interplay between dataset characteristics and oversampling methods

    Yang Y, Fang T, Hu J, Goh CC, Zhang H, Cai Y, et al. A comprehensive study on the interplay between dataset characteristics and oversampling methods. Journal of the Operational Research Society. 2025;0(0):1–22. https://doi.org/10.1080/ 01605682.2025.2450060. https://doi.org/10...

    2025

  21. [29]

    KAN: Kolmogorov–Arnold Networks

    Anonymous. KAN: Kolmogorov–Arnold Networks. In: Submitted to The Thirteenth International Conference on Learning Representations; 2024. Under review. Available from: https://openreview.net/forum?id=Ozo7qJ5vZi

    2024

  22. [30]

    Available from: https: //arxiv.org/abs/2411.00278

    Zhou Q, Pei C, Sun F, Han J, Gao Z, Pei D, et al.: KAN-AD: Time Series Anomaly Detection with Kolmogorov-Arnold Networks. Available from: https: //arxiv.org/abs/2411.00278

    arXiv

  23. [31]

    Kolmogorov-Arnold Networks for Time Series: Bridging Predictive Power and Interpretability

    Xu K, Chen L, Wang S. Kolmogorov-Arnold Networks for Time Series: Bridging Predictive Power and Interpretability. ArXiv. 2024;abs/2406.02496

    2024 arXiv

  24. [32]

    Available from: https://arxiv.org/abs/2405

    Vaca-Rubio CJ, Blanco L, Pereira R, Caus M.: Kolmogorov-Arnold Networks (KANs) for Time Series Analysis. Available from: https://arxiv.org/abs/2405. 08790

  25. [33]

    C-KAN: A New Approach for Integrating Convolutional Layers with Kolmogorov–Arnold Networks for Time-Series Forecasting

    Livieris IE. C-KAN: A New Approach for Integrating Convolutional Layers with Kolmogorov–Arnold Networks for Time-Series Forecasting. Mathematics. 2024;12(19). https://doi.org/10.3390/math12193022

    2024 doi

  26. [34]

    Kan: Kolmogorov-arnold networks

    Liu Z, Wang Y, Vaidya S, Ruehle F, Halverson J, Soljaˇ ci´ c M, et al. Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:240419756. 2024

    2024

  27. [35]

    Https://freddiemac.embs.com

    FreddieMac.: Freddie Mac Dataset. Https://freddiemac.embs.com

  28. [36]

    Available from: http://www.bis

    on Banking Supervision (BCBS) BC.: Basel II: International Convergence of Capital Measurement and Capital Standards. Available from: http://www.bis. org/publ/bcbsca.htm. 29

  29. [37]

    Feature engineering for machine learning and data analytics

    Dong G, Liu H. Feature engineering for machine learning and data analytics. CRC press; 2018

    2018

  30. [38]

    Detection of fraudulent credit card trans- actions: A comparative analysis of data sampling and classification techniques

    Mahesh KP, Afrouz SA, Areeckal AS. Detection of fraudulent credit card trans- actions: A comparative analysis of data sampling and classification techniques. Journal of Physics: Conference Series. 2022;2161

    2022

  31. [39]

    ADASYN: Adaptive synthetic sampling approach for imbalanced learning

    He H, Bai Y, Garcia EA, Li S. ADASYN: Adaptive synthetic sampling approach for imbalanced learning. In: 2008 IEEE International Joint Conference on Neural Networks (IEEE World Congress on Computational Intelligence); 2008. p. 1322– 1328

    2008

  32. [40]

    A New Oversampling Method Based on the Classification Contribution Degree

    Jiang Z, Pan T, Zhang C, Yang J. A New Oversampling Method Based on the Classification Contribution Degree. Symmetry. 2021;13(2). https://doi.org/10. 3390/sym13020194

    2021

  33. [41]

    Anomaly Detection in Medical Time Series with Generative Adversarial Networks: A Selective Review

    Ceki´ c M. Anomaly Detection in Medical Time Series with Generative Adversarial Networks: A Selective Review. null. 2023;https://doi.org/10.5772/intechopen. 112582

    2023 doi

  34. [42]

    A Review of Anomaly Detection Techniques and Applications in Financial Fraud

    Hilal W, Hilal W, Gadsden SA, Gadsden SA, Yawney J, Yawney J. A Review of Anomaly Detection Techniques and Applications in Financial Fraud. Expert Systems With Applications. 2021;https://doi.org/10.1016/j.eswa.2021.116429

    2021

  35. [43]

    OneNet: Enhancing Time Series Forecasting Models under Concept Drift by Online Ensembling

    zhang y, Wen Q, wang x, Chen W, Sun L, Zhang Z, et al. OneNet: Enhancing Time Series Forecasting Models under Concept Drift by Online Ensembling. In: Oh A, Naumann T, Globerson A, Saenko K, Hardt M, Levine S, editors. Advances in Neural Information Processing Systems. vol. 36....

    2023

Pith tools

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