All reviews of published articles are made public. This includes manuscript files, peer review comments, author rebuttals and revised materials. Note: This was optional for articles submitted before 13 February 2023.
Peer reviewers are encouraged (but not required) to provide their names to the authors when submitting their peer review. If they agree to provide their name, then their personal profile page will reflect a public acknowledgment that they performed a review (even if the article is rejected). If the article is accepted, then reviewers who provided their name will be associated with the article itself.
The authors have addressed all concerns.
no comment
no comment
no comment
The authors have thoroughly addressed the comments raised in the first-round review. Good work!
None
None
None
Authors have modified all required content. No more comments for that.
Please respond to all the reviewers' comments.
**PeerJ Staff Note:** Please ensure that all review and editorial comments are addressed in a response letter and that any edits or clarifications mentioned in the letter are also inserted into the revised manuscript where appropriate.
**Language Note:** The review process has identified that the English language must be improved. PeerJ can provide language editing services - please contact us at [email protected] for pricing (be sure to provide your manuscript number and title). Alternatively, you should make your own arrangements to improve the language quality and provide details in your response letter. – PeerJ Staff
The manuscript would benefit from more thorough proofreading, as there are many instances of unnatural or awkward English. For example, expressions involving "default" and "non-default" sometimes result in grammatically correct but semantically unclear sentences. In addition, several sentences are unnecessarily long, and some formulas are overly complex. These issues collectively give the impression that the paper may have been written by a non-native speaker.
The writing is generally accessible, which is a strength, but in some cases, the level of detail may be excessive. For instance, since the deep learning model used is a standard fully connected neural network, the inclusion of a detailed explanation of the chain rule and diagrams such as Figure 2 may not be essential.
There are also several descriptions that lack proper citations. A thorough check of references is recommended throughout the paper to ensure academic rigor.
The study proposes a hybrid modeling approach that combines logistic regression and deep learning to enhance interpretability in credit scoring. The methodology is clearly positioned within the scope of interpretable machine learning and addresses the challenge of balancing predictive accuracy with model transparency.
The authors use logistic regression as a base model due to its interpretability and propose a strategy to distinguish between noise and boundary samples among misclassified cases. Noise samples are removed, while boundary samples are further analyzed using a separate deep learning model. To compensate for the lack of interpretability in the deep learning model, SHAP values are used to provide an explanation.
Finally, the use of agglomerative clustering to calculate the distance between a new sample and cluster centers provides a basis for deciding whether logistic regression or deep learning should be applied. The number of cluster centers is defined using √(N(Dₗ)), where N(∙) is said to be a rounding function. However, the notation suggests that the rounding should be applied after taking the square root rather than inside it.
Although SMOTE-ENN is briefly mentioned, it is neither explained in the methods section nor included in the experimental results. This suggests that the component may currently be missing from the implementation and should be clarified.
The experiments conducted on three different datasets appropriately demonstrate the interpretability of the proposed methodology.
The proposed framework is conceptually sound and well motivated. The use of separate models for boundary and non-boundary samples is reasonable, and the incorporation of SHAP values adds credibility to the interpretability claims.
However, the listing of variables in the Results – Interpretability of the Model section may not provide significant insight. It is suggested that this part be revised to highlight more meaningful interpretations.
The results are presented in alignment with the proposed methodology, and the selective use of deep learning for boundary samples adds a novel dimension to the overall modeling approach.
no comment
no comment
no comment
1. The paper astutely identifies the critical paradox in credit risk assessment models—the trade-off between the accuracy of black-box models and their interpretability. This research motivation aligns well with the practical demand for model transparency under the backdrop of increasingly stringent financial regulations.
2. In the experimental results section, it would be beneficial to include parameter sensitivity analysis (e.g., evaluating model performance under varying hyperparameter settings) and a comparative analysis with state-of-the-art deep learning models proposed in existing literature.
3. The current analysis utilizes three datasets. It is worth considering whether this sample size is sufficient to ensure the statistical robustness of the hypothesis tests conducted (e.g., Friedman tests). Generally, a minimum of four independent datasets is recommended and enhance the generalizability of conclusions.
4. While SHAP values effectively quantify feature contributions, the result analysis would benefit from a deeper exploration of feature interactions (e.g., using SHAP interaction values or partial dependence plots) and domain-specific interpretations (e.g., linking high-impact features to real-world credit risk factors such as debt-to-income ratios or payment history).
5. The research objectives could be articulated with greater specificity (e.g., explicitly stating the threshold for "interpretability" and "accuracy" in model design). Furthermore, parameter sensitivity analysis should be systematically integrated into the results to demonstrate the model's resilience across configurations. Finally, discussing the limitations (e.g., sensitivity to data distribution shifts, computational costs of ARPD) and future directions (e.g., extending to dynamic credit environments, integrating macroeconomic indicators) would significantly strengthen the paper.
All text and materials provided via this peer-review history page are made available under a Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.