Pith. sign in

REVIEW 4 major objections 4 minor 22 references

Linear Discriminant Analysis in Credit Scoring: A Transparent Hybrid Model Approach

T0 review · 4 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper claims that LDA feature reduction plus an XGBoost-DNN hybrid reaches 99.45% accuracy on Lending Club credit data while keeping explanations intact.

desk verdict LDA feature reduction in credit scoring is a reasonable question, but this paper's SMOTE-before-split leakage and impossible LDA dimensionality make the headline accuracy untrustworthy. read the letter →

arxiv 2412.04183 v1 pith:IABUJ7LD submitted 2024-12-05 cs.LG

Md Shihab Reza , Monirul Islam Mahmud , Ifti Azad Abeer , Nova Ahmed This is my paper · ORCID
classification cs.LG
keywords creditscoringLinearDiscriminantAnalysisLendingClubdatasetXG-DNNhybridXGBoostexplainableAILIMEMorrissensitivity
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 tries to show that Linear Discriminant Analysis can be used as a feature-reduction step in credit scoring without sacrificing accuracy or explainability. On the Lending Club dataset, it reduces 107 features to 21 and then trains an XGBoost-and-deep-network hybrid called XG-DNN, reporting 99.45% accuracy and a 99% F1 score, marginally above the same hybrid without LDA. The paper compares eight models with and without LDA, and applies LIME and Morris Sensitivity Analysis to show that the reduced model still explains its decisions through interpretable loan features. If correct, this matters for resource-constrained lenders who want cheaper, faster scoring models that remain transparent.

What carries the argument

The machinery is LDA as a supervised projector: it finds linear combinations of the 107 original features that best separate loan-status classes and keeps 21 dimensions. Onto that lower-dimensional space the paper fits XG-DNN, a hybrid that feeds XGBoost's refined feature representation into a deep neural network to capture higher-order interactions. Explanations come from two post-hoc tools: LIME, which builds local surrogate explanations for individual predictions, and Morris Sensitivity Analysis, a one-step-at-a-time global sensitivity method that ranks features by elementary effects. Together, the LDA projection is the load-bearing step that lowers model complexity, and the two XAI tools are what the paper uses to claim transparency is preserved.

What would settle it

Re-run the pipeline with SMOTE applied only inside the training split and with out_prncp, total_rec_prncp, and recoveries removed; if accuracy and F1 fall well below 99.45%, the reported result depends on leakage rather than on LDA.

Watch

Extended reading notes

Core claim

The central claim is that LDA-based feature reduction is not a cost paid in performance: the XG-DNN hybrid trained on the 21-dimensional LDA projection reaches the best accuracy (99.45%) and ties the best F1 score (99%) among all tested configurations, while the same model without LDA reaches 99.44% accuracy, and the paper reports a slight sensitivity trade-off (72.79% sensitivity). The paper also claims that the reduced model remains explainable: LIME attributes individual predictions mainly to features like the outstanding principal and principal received to date, and Morris Sensitivity Analysis ranks recoveries, out_prncp, and total_rec_prncp as the most influential. Thus the paper concludes that dimensionality reduction can improve or maintain predictive performance and that post-hoc XAI can recover transparency lost in the projection.

Load-bearing premise

The results assume that SMOTE applied before the 80/20 split does not leak synthetic samples into the test set, and that features such as remaining principal, principal received, and recoveries are available when a loan is being scored.

Editorial extensions

If this is right

  • If the claim holds, lenders can drop from 107 to 21 features before training and still get top accuracy, cutting storage and computation.
  • The XG-DNN hybrid becomes a practical option for resource-constrained settings where training a full-size deep model is too costly.
  • Feature reduction via LDA does not automatically destroy local or global interpretability; LIME and Morris still identify meaningful loan features.
  • The measured gain from LDA is small on accuracy (99.45 versus 99.44 percent), so the main practical benefit is computational rather than predictive.

Reading between the lines

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

  • A natural next experiment is to compare LDA with PCA or variance-based feature selection under the same split; the paper does not isolate whether the gain comes from dimensionality reduction itself or from LDA's class-separating objective.
  • Because LDA components are linear mixtures of the original features, the LIME and Morris explanations may not exactly describe what the reduced model sees unless the projection is mapped back to original feature space.
  • The very high accuracy levels suggest that some influential features, such as remaining principal and recoveries, may be known only after the loan is issued; dropping them would give a sterner test of true scoring performance.
  • If the approach transfers, other high-dimensional tabular credit datasets would be natural test beds, but the paper demonstrates the method only on Lending Club.
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

4 major / 4 minor

Summary. The paper proposes using Linear Discriminant Analysis (LDA) as a dimensionality-reduction step before training several machine learning and deep learning models on the Lending Club loan dataset, with an emphasis on a hybrid XGBoost-DNN model (XG-DNN). The authors report that LDA-based XG-DNN achieves 99.45% accuracy and a 99% F1 score, outperforming all other models, and they use LIME and Morris Sensitivity Analysis to interpret the model decisions. The central empirical claim is that LDA can reduce model complexity without harming performance or explainability.

Significance. If the reported results were valid, the paper would offer a potentially useful demonstration that a simple linear dimensionality-reduction technique can cut computational cost while preserving accuracy and explainability in credit scoring, a domain where both are valued. The work also attempts to combine local (LIME) and global (Morris) explanations, which is a commendable breadth. However, the paper is purely empirical and supplies no code, data splits, or reproducibility artifacts, and the evaluation protocol as described contains several load-bearing flaws. These flaws—SMOTE leakage into the test set, target leakage through post-loan outcome features, and an impossible LDA component count—mean that the headline numbers cannot be interpreted as valid out-of-sample credit-scoring performance. The potential significance is therefore largely unsubstantiated by the manuscript as written.

major comments (4)
  1. [III-A and III-B] The preprocessing section (III-A) applies SMOTE to the full dataset, and only afterward does the modeling section (III-B) describe splitting into 80% training and 20% test data. This order means synthetic instances generated from the entire dataset can appear in the test set. Because SMOTE creates interpolations of real neighbors, the model is effectively evaluated on points close to its training distribution, so the reported test accuracy and F1 are not out-of-sample estimates. The central claim of 99.45% accuracy for LDA-based XG-DNN rests on this contaminated evaluation. The paper should split first, apply SMOTE only to the training portion, and then evaluate on the untouched test set.
  2. [IV-C] The explainability analysis identifies out_prncp (remaining outstanding principal), total_rec_prncp (principal received to date), and recoveries as the dominant features for both LIME and Morris Sensitivity Analysis. These are post-loan outcome variables that are not available at the time a credit application is scored; using them to predict current loan status is a different task from credit scoring and constitutes target leakage. The model may be largely reading off loan repayment status from variables that are themselves consequences of repayment. The paper should restrict features to those available at application time and repeat the performance and XAI analyses, otherwise both the predictive-accuracy claim and the interpretability discussion are misleading.
  3. [III-C] Section III-C states that applying LDA as a feature reduction technique leaves 21 features from 107 features. The target variable loan_status has 10 classes, so LDA can produce at most 9 discriminant dimensions (c-1). A pipeline that outputs 21 LDA components cannot be reproduced as described. This is a load-bearing inconsistency for the central method. The authors should clarify whether they used a different technique (e.g., PCA, or selecting a subset of original features) or correct the description to match the actual component count.
  4. [IV-B] LDA is applied only to XGBoost and XG-DNN after these two models were observed to be the best performers without LDA. This post-hoc selection, made after seeing the non-LDA results, raises a multiple-comparisons concern: the reported improvement of XG-DNN with LDA is selected from a larger implicit set of possible model-LDA combinations. The paper should either pre-register the selection rule or report results for all models both with and without LDA, so the comparison is not conditioned on the outcome it is claiming to demonstrate.
minor comments (4)
  1. [IV-A] The text in Section IV-A says the Deep Neural Network achieved 100% sensitivity and specificity, but Table I lists DNN sensitivity as 99.20 and specificity as 99.10; these numbers should be reconciled.
  2. [IV-B] The phrase 'the LDA-based XG-DNN hybrid model shows the lowest specificity at 99.85' is confusing because 99.85 is a very high value; rephrasing to compare it to the XGBoost specificity of 99.98 would be clearer.
  3. [Literature Review and References] Several in-text citation numbers do not match the reference list: the citation to Jasmina Nalić et al. appears as [21], which is out of range, and the citation to Shrawan Kumar Trivedi appears as [16], which corresponds in the bibliography to Nalić et al.; the reference list should be carefully renumbered and cross-checked.
  4. [Figures] Figure 6 is referenced in Section IV-A ('shown in Figure 6') but the figure is not included in the manuscript text, making the accuracy comparison difficult to follow.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported accuracies are empirical benchmark results, not derivations that reduce to their inputs.

full rationale

The paper's central claim is a comparative benchmark: eight models are trained and evaluated on the Lending Club dataset with and without LDA feature reduction, and the best observed accuracy (99.45% for LDA-based XG-DNN) is reported from test predictions under the Section III-B split. There is no derivation in which an output quantity is defined in terms of the claimed conclusion. LDA is used as a standard supervised projection; its output dimensionality and the subsequent model scores are measured, not forced by construction. The only author-overlapping citation ([19], used in the literature review to note that Lasso feature selection improves predictive systems) is not load-bearing for the 99.45% claim. The SMOTE-before-split protocol and the use of post-loan features such as out_prncp and recoveries are data-leakage and feature-validity concerns, and the statement that LDA leaves 21 components for a 10-class target is arithmetically suspect (LDA can produce at most 9 components), but these are correctness and reproducibility issues rather than circular reasoning. Choosing to apply LDA only to the two best non-LDA models is post-hoc selection, not a reduction of the reported numbers to model inputs. No circular step can be exhibited from the paper's own equations or citations.

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

The central claim rests on questionable free parameters: the LDA component count is mathematically impossible for 10 classes, the SMOTE setup is not specified, and the feature removal threshold is arbitrary. The key assumptions about feature availability and SMOTE-before-split are violated, undermining the reported performance.

free parameters (3)
  • LDA n_components = 21 = 21
    The number of LDA components is chosen to reduce 107 features to 21, but with 10 classes LDA can yield at most 9 components, making this choice inconsistent.
  • Null-value threshold for feature removal = 50%
    Features with over 50% null values are removed, a threshold choice that determines the 107-feature input.
  • SMOTE sampling parameters = not reported
    SMOTE is applied to balance the 10-class target, but the sampling strategy and ratios are not reported.
assumptions (3)
  • domain assumption Features in the dataset are known at loan origination and can be used for prediction.
    The models use features such as out_prncp and recoveries that reflect post-loan outcomes, so this assumption fails.
  • ad hoc to paper SMOTE can be applied before the train-test split without contaminating the test set.
    The paper applies SMOTE to the full dataset (Section III-A) before splitting (Section III-B), which leaks synthetic samples into the test set.
  • standard math LDA can yield up to one component fewer than the number of classes.
    With 10 classes, LDA can produce at most 9 components, contradicting the reported 21 features.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Linear Discriminant Analysis in Credit Scoring: A Transparent Hybrid Model Approach." pith.science (2026). https://pith.science/paper/IABUJ7LD

@misc{pith2026241204183,
  author       = {Pith},
  title        = {Pith review of: Linear Discriminant Analysis in Credit Scoring: A Transparent Hybrid Model Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IABUJ7LD}},
  note         = {Machine review of arXiv:2412.04183}
}
read the original abstract

The development of computing has made credit scoring approaches possible, with various machine learning (ML) and deep learning (DL) techniques becoming more and more valuable. While complex models yield more accurate predictions, their interpretability is often weakened, which is a concern for credit scoring that places importance on decision fairness. As features of the dataset are a crucial factor for the credit scoring system, we implement Linear Discriminant Analysis (LDA) as a feature reduction technique, which reduces the burden of the models complexity. We compared 6 different machine learning models, 1 deep learning model, and a hybrid model with and without using LDA. From the result, we have found our hybrid model, XG-DNN, outperformed other models with the highest accuracy of 99.45% and a 99% F1 score with LDA. Lastly, to interpret model decisions, we have applied 2 different explainable AI techniques named LIME (local) and Morris Sensitivity Analysis (global). Through this research, we showed how feature reduction techniques can be used without affecting the performance and explainability of the model, which can be very useful in resource-constrained settings to optimize the computational workload.

Figures

Figures reproduced from arXiv: 2412.04183 by the authors.

Figure 2
Figure 2. Distribution plot of the target variable "loan_status" after applying SMOTE. Then, we analyzed the bar plot and histogram of categorical and numerical features to observe data distribution [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 5
Figure 5. Architecture of hybrid XG-DNN model. For result evaluation, we have used six state-of-the-art evaluation techniques for credit scoring: accuracy, sensitivity (true positive rate), specificity (true negative rate), G-mean, H-measure, and F1 score. C. Linear Discriminant Analysis (LDA) We used LDA as a classification model and also as a feature reduction technique. While preserving the feature information as much as p… view at source ↗
Figure 3
Figure 3. Bar Plot of categorical features - (a) purpose (b) emp length. (a) (b) [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Histogram of numerical features (a) loan_amnt, (b) installment [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 13 canonical work pages

  1. [1]

    Rana Alasbahi and XiaoLin Zheng. 2022. An Online Transfer Learning Framework with Extreme Learning Machine for Automated Credit Scoring. IEEE Access (2022), 1 –1. DOI:https://doi.org/10.1109/access.2022.3171569

    arXiv 2022

  2. [2]

    Lara Marie Demajo, Vince Vella, and Alexiei Dingli. 2020. Explainable AI for Interpretable Credit Scoring. Computer Science & Information Technology (CS & IT) (November 2020), 185 –203. DOI:https://doi.org/10.5121/csit.2020.101516

    arXiv 2020

  3. [3]

    Multi dimensional Hidden markov model for credit scoring systems in Peer-To-Peer (P2P) lending

    Monir El Annas, Mohamed Ouzineb, and Badreddine Benyacoub. "Multi dimensional Hidden markov model for credit scoring systems in Peer-To-Peer (P2P) lending." Big Data and Networks Technologies

    work page

  4. [4]

    Springer International Publishing, 2020

    2020

  5. [5]

    Monir El Annas, Badreddine Benyacoub, and Mohamed Ouzineb

    work page

  6. [6]

    Branka Hadji Misheva, Joerg Osterrieder, Ali Hirsa, Onkar Kulkarni, and Stephen Fung Lin. 2021. Explainable AI in Credit Risk Management. arxiv.org (March 2021). DOI:https://doi.org/10.48550/arXiv.2103.00949

    arXiv doi:10.48550/arxiv.2103.00949 2021

  7. [7]

    Štefan Lyócsa, Petra Vašaničová, Branka Hadji Misheva, and Marko Dávid Vateha. 2022. Default or profit scoring credit systems? Evidence from European and US peer -to-peer lending markets. Financial Innovation 8, 1 (April 2022). DOI:https://doi.org/10.1186/s40854-022- 00338-5

    doi:10.1186/s40854-022- 2022

  8. [8]

    Dong-Her Shih, Ting-Wei Wu, Po-Yuan Shih, Nai-An Lu, and Ming- Hung Shih. 2022. A Framework of Global Credit -Scoring Modeling Using Outlier Detection and Machine Learning in a P2P Lending Platform. Mathematics 10, 13 (June 2022), 2282 –2282. DOI:https://doi.org/10.3390/math10132282

    work page doi:10.3390/math10132282 2022

Show all 22 references
  1. [9]

    Vincenzo Moscato, Antonio Picariello, and Giancarlo Sperlí. 2021. A benchmark of machine learning approaches for credit score prediction. Expert Systems with Applications 165, (March 2021), 113986. DOI:https://doi.org/10.1016/j.eswa.2020.113986

    2021

  2. [10]

    Petr Teply and Michal Polena. 2019. Best classification algorithms in peer-to-peer lending. The North American Journal of Economics and Finance (January 2019). DOI:https://doi.org/10.1016/j.najef.2019.01.001

    2019 doi

  3. [11]

    Jen-Ying Shih and Ze -Han Chin. 2023. A Fairness Approach to Mitigating Racial Bias of Credit Scoring Models by Decision Tree and the Reweighing Fairness Algorithm. (April 2023). DOI:https://doi.org/10.1109/iceib57887.2023.10170339

    2023

  4. [12]

    Zaimei Zhang, Kun Niu, and Yan Liu. 2020. A Deep Learning Based Online Credit Scoring Model for P2P Lending. IEEE Access 8, (2020), 177307–177317. DOI:https://doi.org/10.1109/access.2020.3027337

    2020

  5. [13]

    Swati Tyagi. 2022. Analyzing Machine Learning Models for Credit Scoring with Explainable AI and Optimizing Investment Decisions. arXiv.org. DOI:https://doi.org/10.48550/arXiv.2209.09362

    2022 doi arXiv

  6. [14]

    15, 2020

    Lending Club 2007 -2020Q3, “Kaggle,” Dec. 15, 2020. [Online]. Available: https://www.kaggle.com/datasets/ethon0426/lending -club- 20072020q1

    2007

  7. [15]

    Tanantong, T., and Loetwiphut, P. (2023). Association rule mining framework for financial credit risk analysis in peer -to-peer lending platforms. Science, Engineering and Health Studies, 17, 23020006

    2023

  8. [16]

    Jasmina Nalić, Goran Martinović, and Drago Žagar. 2020. New hybrid data mining model for credit scoring based on feature selection algorithm and ensemble classifiers. Advanced Engineering Informatics 101130. https://doi.org/10.1016/j.aei.2020.101130

    2020

  9. [17]

    Stefan Lessmann, Bart Baesens, Hsin‐Vonn Seow, and Lyn C. Thomas

  10. [18]

    Improving the Morris method for sensitivity analysis by scaling the elementary effects,

    G. Sin and K. V. Gernaey, “Improving the Morris method for sensitivity analysis by scaling the elementary effects,” in Computer - aided chemical engineering/Computer aided chemical engineering, 2009, pp. 925–930. doi: 10.1016/s1570-7946(09)70154-3

    2009 doi

  11. [19]

    Optimizing StrokeDetection: An Analysis of Different Feature Selection Approaches,

    M. I. Mahmud, M. S. Reza, and S. S. Khan, “Optimizing StrokeDetection: An Analysis of Different Feature Selection Approaches,”in Companion of the 2024 on ACM International Joint Conference onPervasive and Ubiquitous Computing, 2024, pp. 142 – 146

    2024

  12. [20]

    Shrawan Kumar Trivedi. 2020. A study on credit scoring modeling with different feature selection and machine learning approaches. Technology in Society 101413. https://doi.org/10.1016/j.techsoc.2020.101413

    2020

  13. [2015]

    European Journal of Operational Research 124–136

    Benchmarking state -of-the-art classification algorithms for credit scoring: An update of research. European Journal of Operational Research 124–136. https://doi.org/10.1016/j.ejor.2015.05.030

    2015 doi

  14. [2022]

    Computational Statistics (May 2022)

    Semi-supervised adapted HMMs for P2P credit scoring systems with reject inference. Computational Statistics (May 2022). DOI:https://doi.org/10.1007/s00180-022-01220-9

    2022 doi

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.