Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

Credit Risk Identification in Supply Chains Using Generative Adversarial Networks

T0 review · 5 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims a GAN-based model, trained with synthetic default data, outperforms SVM, BP network, RNN, and LSTM on supply-chain credit risk identification, reporting 0.96 accuracy, 1.0 recall, 0.97 precision, and 0.97 F1.

desk verdict A plausible data-augmentation idea sunk by a circular test set and internally inconsistent metrics. read the letter →

arxiv 2501.10348 v4 pith:BJRBCKVM submitted 2025-01-17 cs.LG

Zizhou Zhang , Xinshi Li , Yu Cheng , Zhenrui Chen , Qianying Liu This is my paper · ORCID
classification cs.LG
keywords GenerativeAdversarialNetworksSupplyChainRiskCreditIdentificationMachineLearningDataAugmentationWassersteinGANImbalanced
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

Credit risk in supply chains spreads from one firm to its business partners, so identifying it early matters for financial stability. The paper tries to establish that a generative adversarial network can do this better than standard classifiers by learning the distribution of real credit-risk records and manufacturing additional synthetic default scenarios where real defaults are scarce. Using a Wasserstein GAN with multilayer-perceptron generator and discriminator, the authors report accuracy 0.96, recall 1.0, precision 0.97, and F1 0.97 on data from steel manufacturing, pharmaceutical distribution, and e-commerce, ahead of SVM, BP network, RNN, and LSTM. They also report that removing the synthetic data lowers model performance by roughly 5 percent. If the claim holds, GAN-based augmentation is a practical way to train default detectors in data-poor supply-chain settings.

What carries the argument

The load-bearing mechanism is a Wasserstein GAN, a generative adversarial network in which a generator multilayer perceptron maps noise to synthetic credit-risk scenarios and a discriminator attempts to tell synthetic from real records; the two are trained by a minimax objective shown in Equation 1. The paper adds batch normalization, label smoothing, and an Adam optimizer with learning rate 0.0002 and batch size 64 to keep training stable. The synthetic scenarios are used as augmented training data: the authors state that removing them drops performance by approximately 5 percent, which is what makes data generation, rather than any single architectural tweak, the active ingredient in the reported improvement.

What would settle it

Re-train the GAN model exactly as described, then score it on a test set composed only of real default records that were never shown to the generator; if accuracy and recall fall to the level of the LSTM or SVM baselines, the claim of GAN superiority is falsified.

Watch

Extended reading notes

Core claim

The central discovery claimed here is that adding GAN-generated credit-risk samples to the training set materially improves a classifier's ability to flag supply-chain default risk, and that the resulting model captures temporal dependencies in transaction data better than the compared baselines. On a test set that contains both real and synthetic samples, the GAN model reaches accuracy 0.96, recall 1.0, precision 0.97, and F1 0.97, edging out the strongest baseline, LSTM, which reaches 0.92 accuracy and 0.97 recall. The authors interpret this as evidence that generative modeling of the underlying data distribution, not just better discriminative architectures, is what drives the gain.

Load-bearing premise

The reported superiority is measured on a test set that mixes real records with synthetic ones produced by the same generator used in training; the claim depends on those synthetic samples being as hard to classify as real, unseen defaults.

Editorial extensions

If this is right

  • If the reported results hold, firms with sparse default histories can train credit-risk models by generating plausible default scenarios instead of waiting for more real defaults.
  • The approach can be tuned per industry, so steel, pharmaceutical, and e-commerce supply chains can each have a model fitted to their own contagion patterns.
  • The observed drop of about 5 percent after removing synthetic data indicates that augmentation is a necessary part of the model's advantage, not a minor add-on.
  • A reported recall of 1.0 on the test set implies the model flags every default it encounters, making it suitable as an early-warning screening tool if the result generalizes.

Reading between the lines

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

  • A direct extension would evaluate the GAN model on a hold-out set containing only real, never-generated default records; this would separate the model's discriminative skill from the generator's ability to produce easy-to-classify samples.
  • The same generator-plus-classifier recipe could transfer to adjacent imbalanced problems, such as fraudulent invoices or supplier payment delays, where positive cases are rare.
  • Because the underlying data come from commercial market databases, real-world deployment would also need to check whether GAN-generated samples stay representative when macro-financial conditions shift, which the paper does not address.
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

5 major / 5 minor

Summary. The manuscript proposes a Wasserstein GAN-based approach for credit risk identification in supply chains, motivated by data scarcity and class imbalance. It describes a GAN architecture with an MLP generator and discriminator, trained on financial and non-financial indicators from three industries (steel, pharmaceutical distribution, e-commerce), and reports that the GAN model outperforms SVM, BP network, RNN, and LSTM in accuracy, recall, precision, and F1 (Table II). The evaluation is said to use a test set containing both real and synthetic data, and the paper claims that removing GAN-generated data drops performance by about 5%. The central contribution claimed is that GAN-generated synthetic data improves credit risk prediction relative to traditional and sequential deep-learning baselines.

Significance. If the performance claim were supported by a valid evaluation, the paper would be a moderate contribution to supply chain finance and applied machine learning: it applies WGAN-style training (weight clipping, batch normalization, label smoothing) with specified hyperparameters to a practically important problem, and it targets three distinct industries. However, the evidence in the manuscript is not sufficient to establish the central claim. There is no dataset description, no real-only versus synthetic-only evaluation, no error bars or statistical tests, and the reported F1 scores are inconsistent with the precision and recall values. The paper also does not provide code or data for reproducibility. The idea is plausible but the current numerical results cannot be interpreted as evidence of real-world generalization.

major comments (5)
  1. [Section III.C] The evaluation protocol is circular and load-bearing. Section III.C states that the test set 'contains both real and synthetic data,' while the model is trained with GAN-generated data from the same generative process. If the synthetic test samples are drawn from the generator's learned distribution, they can be much easier to classify than real defaults, especially for a model trained to recognize that distribution. The paper provides no split counts, no source for the synthetic test samples, and no performance breakdown on real-only versus synthetic-only subsets. The Table II claims (accuracy 0.96, recall 1.0) therefore do not establish generalization to unseen real credit risk events.
  2. [Section IV.B] The claim that performance drops by approximately 5% when GAN-generated data is removed is unsupported. No table, figure, metric definition, or experimental protocol is given for this ablation, so the reader cannot verify the effect size or even know whether it refers to accuracy, F1, or another metric. This claim is used to justify the value of synthetic data and must be either removed or substantiated with a proper ablation study.
  3. [Table II] The F1 scores in Table II are internally inconsistent with the reported precision and recall. For example, the SVM row gives precision 0.84 and recall 0.88, whose harmonic mean is 0.86, not 0.89; the LSTM row gives 0.93 and 0.97, whose harmonic mean is 0.95, not 0.96; and the GAN row gives 0.97 and 1.00, whose harmonic mean is 0.98, not 0.97. This inconsistency undermines confidence in the numerical results and suggests the metrics were not computed from the same confusion matrix.
  4. [Section III.A and IV.B] The experimental setup is insufficiently described for the results to be reproducible or interpretable. The paper names Wind, Bloomberg, and Reuters as data sources but gives no sample size, time period, industry-level counts, class balance, or preprocessing steps. In addition, the abstract and Section IV.B say the model is compared with logistic regression and decision trees, but Table II reports only SVM, BP network, RNN, LSTM, and GANs. The missing baselines and missing data characterization are not minor omissions; they are necessary to assess the central comparison.
  5. [Section III.B] Equation (1) is garbled: the min and max operators are not attached to G and D, and the expectation subscripts are not typeset correctly. The minimax objective is a standard GAN formulation and can be stated clearly in one line; the current rendering makes it impossible to verify the exact objective used for the WGAN variant described in the text.
minor comments (5)
  1. [Table I] The table header reads 'VRIABLE DEFINITIONS' instead of 'VARIABLE DEFINITIONS', and the 'Net Profit Growth Rate' row appears twice.
  2. [Section II/References] Reference [30] is truncated mid-sentence ('A word is worth a thousand dollars: Adversarial attack on tweets fools...') and appears to merge two separate citations; the reference list also contains several entries unrelated to the topic, such as MRI reconstruction and OpenMP code generation.
  3. [Figures] Figures 1 and 2 are referenced in the text but the manuscript does not contain actual figure content, only placeholders; Figure 4 is described but its axis labels and AUC value are not reported in the text.
  4. [Section IV.B] The sentence 'RNN LSTM also yielded favorable results' is missing a comma between RNN and LSTM.
  5. [Section III.C] The sentence about the test set says it 'contains both real and synthetic data, ensuring that the model is evaluated on its ability to generalize to new, unseen credit risk events,' which is internally contradictory unless the synthetic samples are explicitly shown to be held out from training.

Circularity Check

1 steps flagged · score 6.0 of 10

The GAN's reported advantage is partially circular: the test set includes synthetic data from the same generative process used to augment training, so Table II metrics do not cleanly measure real-world generalization.

  1. fitted input called prediction [Section III.C (Model Evaluation); Section IV.B (Comparison of Models)]
    "These metrics are computed on a test set that contains both real and synthetic data, ensuring that the model is evaluated on its ability to generalize to new, unseen credit risk events. ... When the GAN-generated data is removed, the performance of the model drops by approximately 5%, highlighting the importance of synthetic data in improving predictive accuracy."

    The paper trains the model with GAN-augmented data (removing GAN-generated data costs about 5% performance) and then evaluates all reported metrics on a test set that explicitly contains synthetic data. If those synthetic test samples are produced by the GAN's generator or drawn from its learned distribution, the model is being scored on data whose generative process was an input to training. High scores on such samples can reflect recognition of the generator's output rather than generalization to real credit risk events. Because the central claim that GANs outperform SVM, BP, RNN, and LSTM rests on Table II, and no real-only versus synthetic-only breakdown is reported, the advantage is at least partly forced by the evaluation design rather than demonstrated on independent real data.

full rationale

The only load-bearing circularity is the evaluation protocol. Section III.C states the test set contains both real and synthetic data, while Section IV.B reports that removing GAN-generated data from training lowers performance by about 5%. Taken together, the synthetic test samples appear to come from the same GAN-based generative process used for augmentation, so the Table II metrics mix generalization to real events with classification of generator outputs. This is a reduction by construction of the evaluation, not of the mathematical derivation. There is no significant self-citation chain: reference [14] includes co-authors but is used only as background, not to justify the model choice. No uniqueness theorem is imported. Separately, the F1 values in Table II are internally inconsistent with the reported precision and recall (e.g., SVM F1 should be about 0.86, not 0.89), which is a numerical credibility concern rather than a circularity. Overall, the central superiority claim is partially circular, so the score is 6.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a handful of unstated assumptions: the indicator set captures credit risk, the GAN-generated data is a valid test distribution, and the named data sources are representative. The evaluation design also assumes that mixing synthetic data into the test set is acceptable, which is not demonstrated.

free parameters (3)
  • learning_rate = 0.0002
    Adam optimizer learning rate, stated in Section III.B; hand-chosen, affects training dynamics and final performance.
  • batch_size = 64
    State in Section III.B; standard choice, not justified by experiments.
  • training_epochs = 120
    Figure 3 caption shows 120 epochs; early stopping rule is described qualitatively (stopped once the generator produces sufficiently realistic data), so the effective epoch count is ambiguous.
assumptions (4)
  • domain assumption GAN-generated synthetic data is distributionally similar to real credit risk data and can be used as valid test data.
    Section III.C evaluates on a test set that contains both real and synthetic data; this assumption is load-bearing for the reported scores.
  • domain assumption The variable set in Table I (profitability, liquidity, efficiency, contract status) is sufficient to capture supply chain credit risk across the three industries.
    Section III.A builds the indicator system on this set; no feature importance or external validation is provided.
  • standard math WGAN with weight clipping enforces the Lipschitz constraint sufficiently for stable training.
    Section III.B relies on standard WGAN theory; the paper adds batch normalization and label smoothing without ablation.
  • domain assumption The Wind, Bloomberg, and Reuters data cover representative firms in steel, pharmaceutical distribution, and e-commerce supply chains.
    Section III.A names data sources but does not report sample sizes, time windows, or inclusion criteria.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Credit Risk Identification in Supply Chains Using Generative Adversarial Networks." pith.science (2026). https://pith.science/paper/BJRBCKVM

@misc{pith2026250110348,
  author       = {Pith},
  title        = {Pith review of: Credit Risk Identification in Supply Chains Using Generative Adversarial Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BJRBCKVM}},
  note         = {Machine review of arXiv:2501.10348}
}
read the original abstract

Credit risk management within supply chains has emerged as a critical research area due to its significant implications for operational stability and financial sustainability. The intricate interdependencies among supply chain participants mean that credit risks can propagate across networks, with impacts varying by industry. This study explores the application of Generative Adversarial Networks (GANs) to enhance credit risk identification in supply chains. GANs enable the generation of synthetic credit risk scenarios, addressing challenges related to data scarcity and imbalanced datasets. By leveraging GAN-generated data, the model improves predictive accuracy while effectively capturing dynamic and temporal dependencies in supply chain data. The research focuses on three representative industries-manufacturing (steel), distribution (pharmaceuticals), and services (e-commerce) to assess industry-specific credit risk contagion. Experimental results demonstrate that the GAN-based model outperforms traditional methods, including logistic regression, decision trees, and neural networks, achieving superior accuracy, recall, and F1 scores. The findings underscore the potential of GANs in proactive risk management, offering robust tools for mitigating financial disruptions in supply chains. Future research could expand the model by incorporating external market factors and supplier relationships to further enhance predictive capabilities. Keywords- Generative Adversarial Networks (GANs); Supply Chain Risk; Credit Risk Identification; Machine Learning; Data Augmentation

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Regression and Forecasting of U.S. Stock Returns Based on LSTM

    q-fin.ST 2025-02 reject novelty 2.0 of 10

    The authors fit standard factor models and an LSTM to U.S. sector returns and report that the five-factor model and LSTM each look best in different sectors.

Reference graph

Works this paper leans on

47 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [1]

    Moolchandani, S. (2024). The Integration of Generative AI in Credit Risk Management. Journal Homepage: http://www. ijmra. us, 14(02)

    work page 2024

  2. [2]

    Yusof, S. A. B. M., & Roslan, F. A. B. M. (2023). The Impact of Generative AI in Enhancing Credit Risk Modeling and Decision-Making in Banking Institutions. Emerging Trends in Machine Intelligence and Big Data, 15(10), 40-49

    work page 2023

  3. [3]

    S., Son, B

    Sodhi, M. S., Son, B. G., & Tang, C. S. (2012). Researchers' perspectives on supply chain risk management. Production and operations management, 21(1), 1-13

    work page 2012

  4. [4]

    In-Trajectory Inverse Reinforcement Learning: Learn Incrementally Before an Ongoing Trajectory Terminates

    Liu, S., & Zhu, M. In-Trajectory Inverse Reinforcement Learning: Learn Incrementally Before an Ongoing Trajectory Terminates. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    work page

  5. [5]

    Heckmann, I., Comes, T., & Nic`l, S. (2015). A critical review on supply chain risk–Definition, measure and modeling. Omega, 52, 119-132

    work page 2015

  6. [6]

    Chen, Y., Liu, L., & Fang, L. (2024). An Enhanced Credit Risk Evaluation by Incorporating Related Party Transaction in Blockchain Firms of China. Mathematics, 12(17), 2673

    work page 2024

  7. [7]

    (2024, March)

    Zhao, P., & Wan, Z. (2024, March). Robust nonparametric regression under poisoning attack. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 38, No. 15, pp. 17007-17015)

    work page 2024

  8. [8]

    Brindley, C. (2017). Supply chain risk. Routledge

    work page 2017

Show all 47 references
  1. [9]

    Manuj, I., & Mentzer, J. T. (2008). Global supply chain risk management. Journal of business logistics, 29(1), 133-155

    2008

  2. [10]

    (2024, December)

    Yu, Q., Ke, Z., Xiong, G., Cheng, Y., & Guo, X. (2024, December). Identifying money laundering risks in digital asset transactions based on ai algorithms. In 2024 4th International Conference on Electronic Information Engineering and Computer Communication (EIECC) (pp. 1081-10...

    2024

  3. [11]

    R., Wang, Z

    Ağca, Ş., Birge, J. R., Wang, Z. A., & Wu, J. (2023). The impact of COVID‐19 on supply chain credit risk. Production and Operations Management, 32(12), 4088-4113

    2023

  4. [13]

    Gurtu, A., & Johny, J. (2021). Supply chain risk management: Literature review. Risks, 9(1), 16

    2021

  5. [14]

    Hu, Z., Yu, R., Zhang, Z., Zheng, H., Liu, Q., & Zhou, Y. (2024). Developing Cryptocurrency Trading Strategy Based on Autoencoder-CNN-GANs Algorithms. arXiv preprint arXiv:2412.18202

    2024 arXiv

  6. [15]

    A., Wang, Z., Zhou, Y., & Hassan, S

    Abbasi, W. A., Wang, Z., Zhou, Y., & Hassan, S. (2019). Research on measurement of supply chain finance credit risk based on Internet of Things. International Journal of Distributed Sensor Networks, 15(9), 1550147719874002

    2019

  7. [16]

    S., Mani, V., Benkhati, I., & Touriki, F

    Belhadi, A., Kamble, S. S., Mani, V., Benkhati, I., & Touriki, F. E. (2021). An ensemble machine learning approach for forecasting credit risk of agricultural SMEs’ investments in agriculture 4.0 through supply chain finance. Annals of Operations Research, 1-29

    2021

  8. [17]

    Chen, B. (2025). Leveraging Advanced AI in Activity-Based Costing (ABC) for Enhanced Cost Management

    2025

  9. [18]

    Bi, B., Liu, S., Mei, L., Wang, Y., Ji, P., & Cheng, X. (2024). Decoding by Contrasting Knowledge: Enhancing LLMs' Confidence on Edited Facts. arXiv preprint arXiv:2405.11613

    2024 arXiv

  10. [19]

    Generative Adversarial Nets,

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., et al. (2014). "Generative Adversarial Nets," Advances in Neural Information Processing Systems (NeurIPS), 27, 2672–2680

    2014

  11. [20]

    Moretto, A., Grassi, L., Caniato, F., Giorgino, M., & Ronchi, S. (2019). Supply chain finance: From traditional to supply chain credit rating. Journal of Purchasing and Supply Management, 25(2), 197-217

    2019

  12. [21]

    Li, Z., Qiu, S., & Ke, Z. (2025). Revolutionizing Drug Discovery: Integrating Spatial Transcriptomics with Advanced Computer Vision Techniques. In 1st CVPR Workshop on Computer Vision For Drug Discovery (CVDD): Where are we and What is Beyond?

    2025

  13. [22]

    Zhang, W., Shen, Y., Lu, W., & Zhuang, Y. (2023). Data-copilot: Bridging billions of data and humans with autonomous workflow. arXiv preprint arXiv:2306.07209

    2023 arXiv

  14. [23]

    C., & De, S

    Mahata, G. C., & De, S. K. (2017). Supply chain inventory model for deteriorating items with maximum lifetime and partial trade credit to credit-risk customers. International Journal of Management Science and Engineering Management, 12(1), 21-32

    2017

  15. [24]

    (2020, November)

    Hu, Y., Cao, H., Yang, Z., & Huang, Y. (2020, November). Improving text-image matching with adversarial learning and circle loss for multi-modal steganography. In International Workshop on Digital Watermarking (pp. 41-52). Cham: Springer International Publishing

    2020

  16. [25]

    (2024, November)

    Ke, Z., & Yin, Y. (2024, November). Tail risk alert based on conditional autoregressive var by regression quantiles and machine learning algorithms. In 2024 5th International Conference on Artificial Intelligence and Computer Engineering (ICAICE) (pp. 527-532). IEEE

    2024

  17. [26]

    L., Favara, G., & Abbate, L

    Nigro, G. L., Favara, G., & Abbate, L. (2021). Supply chain finance: The role of credit rating and retailer effort on optimal contracts. International Journal of Production Economics, 240, 108235

    2021

  18. [27]

    Yang, J., Liu, J., Yao, Z., & Ma, C. (2024). Measuring digitalization capabilities using machine learning. Research in International Business and Finance, 70, 102380

    2024

  19. [28]

    H., & Liao, C

    Chen, L., Ding, X., Emani, M., Vanderbruggen, T., Lin, P. H., & Liao, C. (2023, November). Data race detection using large language models. In Proceedings of the SC'23 Workshops of the International Conference on High Performance Computing, Network, Storage, and Analysis (pp. 215-223)

    2023

  20. [29]

    R., & Amiri, B

    Rishehchi Fayyaz, M., Rasouli, M. R., & Amiri, B. (2021). A data-driven and network-aware approach for credit risk prediction in supply chain finance. Industrial Management & Data Systems, 121(4), 785-808

    2021

  21. [30]

    Y., Xiong, J., Liu, S., & Koyejo, S

    Xie, Y., Wang, D., Chen, P. Y., Xiong, J., Liu, S., & Koyejo, S. (2022). A word is worth a thousand dollars: Adversarial attack on tweets fools Zhao, S., Wang, D., & Douady, R. (2024). PolyModel for Hedge Funds' Portfolio Construction Using Machine Learning. arXiv preprint arX...

    2022 arXiv

  22. [31]

    D., & Swart, T

    Mienye, I. D., & Swart, T. G. (2024). A Hybrid Deep Learning Approach with Generative Adversarial Network for Credit Card Fraud Detection. Technologies, 12(10), 186

    2024

  23. [32]

    H., & Zhang, R

    Ke, Z., Zhou, S., Zhou, Y., Chang, C. H., & Zhang, R. (2025). Detection of ai deepfake and fraud in online payments using gan-based models. arXiv preprint arXiv:2501.07033

    2025

  24. [33]

    Lyu, W., Lin, X., Zheng, S., Pang, L., Ling, H., Jha, S., & Chen, C. (2024). Task-agnostic detector for insertion-based backdoor attacks. arXiv preprint arXiv:2403.17155

    2024 arXiv

  25. [34]

    R., & Wu, J

    Agca, S., Babich, V., Birge, J. R., & Wu, J. (2017). Credit risk propagation along supply chains: Evidence from the CDS market. Georgetown McDonough School of Business Research Paper, 3078752, 54

    2017

  26. [35]

    (2024, August)

    Chen, L., Bhattacharjee, A., Ahmed, N., Hasabnis, N., Oren, G., Vo, V., & Jannesari, A. (2024, August). Ompgpt: A generative pre-trained transformer model for openmp. In European Conference on Parallel Processing (pp. 121-134). Cham: Springer Nature Switzerland

    2024

  27. [36]

    & Chen, C

    Lyu, W., Yao, J., Gupta, S., Pang, L., Sun, T., Yi, L., ... & Chen, C. (2024). Backdooring Vision-Language Models with Out-Of-Distribution Data. arXiv preprint arXiv:2410.01264

    2024 arXiv

  28. [37]

    & Tomizuka, M

    Wang, Y., Zhang, Y., Huo, M., Tian, R., Zhang, X., Xie, Y., ... & Tomizuka, M. (2024). Sparse diffusion policy: A sparse, reusable, and flexible policy for robot learning. arXiv preprint arXiv:2407.01531

    2024 arXiv

  29. [38]

    Qiu, S., Wang, Y., Ke, Z., Shen, Q., Li, Z., Zhang, R., & Ouyang, K. (2025). A Generative Adversarial Network-Based Investor Sentiment Indicator: Superior Predictability for the Stock Market. Mathematics, 13(9), 1476. https://doi.org/10.3390/math13091476

    2025 doi

  30. [39]

    Cao, Z., Zhao, S., & Huang, S. (2024). One-X Property Conjecture, Stochastic Orders and Implied Volatility Surface Construction

    2024

  31. [40]

    Lyu, W., Zheng, S., Ma, T., & Chen, C. (2022). A study of the attention abnormality in trojaned berts. arXiv preprint arXiv:2205.08305

    2022 arXiv

  32. [41]

    Bian, W., & Tamilselvam, Y. K. (2024). A review of optimization-based deep learning models for mri reconstruction. AppliedMath, 4(3), 1098-1127

    2024

  33. [42]

    Lai, Y., Yu, Z., Yang, J., Li, B., Kang, X., & Shen, L. (2024). Gm-df: Generalized multi-scenario deepfake detection. arXiv preprint arXiv:2406.20078

    2024 arXiv

  34. [43]

    Li, S., Li, B., Sun, B., & Weng, Y. (2024). Towards visual-prompt temporal answer grounding in instructional video. IEEE Transactions on Pattern Analysis and Machine Intelligence

    2024

  35. [44]

    Li, B., Weng, Y., Xia, F., & Deng, H. (2024). Towards better Chinese-centric neural machine translation for low-resource languages. Computer Speech & Language, 84, 101566

    2024

  36. [45]

    Yu, Q., Yin, Y., Zhou, S., Mu, H., & Hu, Z. (2025). Detecting Financial Fraud in Listed Companies via a CNN-Transformer Framework

    2025

  37. [46]

    Gu, J., Zhang, Y., & Hu, Z. (2025). FAN-TSF: A Frequency Adaptive Normalization Approach for Non-stationary Time SeriesForecasting on Stock Market Data (May 27, 2025)

    2025

  38. [47]

    Li, B., Sun, B., Li, S., Chen, E., Liu, H., Weng, Y., ... & Hu, M. (2024). Distinct but correct: generating diversified and entity-revised medical response. Science China Information Sciences, 67(3), 132106

    2024

  39. [48]

    Lyu, W., Zheng, S., Pang, L., Ling, H., & Chen, C. (2023). Attention-enhancing backdoor attacks against bert-based models. arXiv preprint arXiv:2310.14480

    2023 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.