Emerging Trends in Deep Learning for Credit Scoring: A Review

Round 1

Reviewer 1 Report

I enjoyed reading the manuscript, which is well written and quite comprehensive. However, I do have a concern of having another systematic review on the subject on deep learning and credit scoring despite several similar studies recently. 

I agree with the author about the depth of analysis of this study. However, to ameliorate this concern, I strong recommend that the author create a Sub-Section in Section 2 with a caption such as "Previous Reviews." Here, the author can demonstrate how this review study is different from the previous reviews undertaken on the subject.

Abstract: Comment about the uniqueness of this review.

This review differs from most systematic reviews of the literature in that it aims to provide deep insights by focusing on emerging trends in, and the potential of, advanced deep learning techniques, such as machine learning being partially replaced by deep learning (DL) because it can achieve higher accuracy for credit scoring.

Author Response

Electronics-192674R1

Emerging trends in deep learning for credit scoring: A review

Author’s comments to Review#1 comments:

I enjoyed reading the manuscript, which is well written and quite comprehensive. However, I do have a concern of having another systematic review on the subject on deep learning and credit scoring despite several similar studies recently. 

Reviewer comment #1-1

I agree with the author about the depth of analysis of this study. However, to ameliorate this concern, I strong recommend that the author create a Sub-Section in Section 2.1 with a caption such as "Previous Reviews." Here, the author can demonstrate how this review study is different from the previous reviews undertaken on the subject.

Author’s comment #1-1

Thank you very much for your professional comments. Based on your comments, I made a new sub-section 2.1 “Previous Review”. Here, I briefly described how this review study is different from the previous reviews from the conceptual viewpoints in section 2.1.

Reviewer comment #1-2

Abstract: Comment about the uniqueness of this review. This review differs from most systematic reviews of the literature in that it aims to provide deep insights by focusing on emerging trends in, and the potential of, advanced deep learning techniques, such as machine learning being partially replaced by deep learning (DL) because it can achieve higher accuracy for credit scoring.

Author’s comment #1-2

Thank you for your professional comments and constructive proposal. The reviewer’s revision is invaluable. I changed as suggested.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper conducts a survey using deep learning techniques for credit scoring. The motivation of this paper is good and relevant to the scope of "Electronics". However, there still has some flaws in this paper.

1.        The author mentioned "This review differs from most systematic reviews of the literature in that it aims to provide deep insights by focusing on emerging trends in, and the potential of, advanced deep learning techniques," in the abstract. However, most of the content in this article is a brief introduction of individual models and experimental simulation comparisons, without the “deep insights” you mentioned. Therefore, it is recommended that the author add more explanations about the core concepts of the models and the reasons why the resulting models are good or bad.

2.        The comments to Section 6, are quite naive. Authors seem just to describe them and not explain the reason for the provided results. More explanations should be given here.

3. To give readers a more comprehensive understanding of the future development of the proposed topic. Can the author provide a tree diagram in Section 7 to organize the future technologies you mentioned?

4.        In Section 2.8, is there a typo in the sentence "2D CNNs are used more frequently than 2D CNNs”?

5.        Is Section 3.3 incomplete, because you have mentioned that "we investigated these performances separately" at the end of this section, but there are no other descriptions for the Performance in the follow-up

6.        Please redraw Figures 7 and 8. It is too vague to understand.

7.        In Figures 11 and 15, the legend is missing.

8.        The font size of all tables in this paper is too large.

9.        On page 27 of the paper, there is a paragraph of text that does not conform to the paper specification.

Author Response

Electronics-192674R1

Emerging trends in deep learning for credit scoring: A review

Author’s comments to Review#2 comments:

Reviewer comments #2-1

The author mentioned "This review differs from most systematic reviews of the literature in that it aims to provide deep insights by focusing on emerging trends in, and the potential of, advanced deep learning techniques," in the abstract. However, most of the content in this article is a brief introduction of individual models and experimental simulation comparisons, without the “deep insights” you mentioned. Therefore, it is recommended that the author add more explanations about the core concepts of the models and the reasons why the resulting models are good or bad.

Author’s comments #2-1

Thank you very much for your professional comments. Based on your comments, I added and enriched more explanations about the core concepts of the models and the reasons why the resulting models are suitable, promising, not adequate etc. in sections 2, 6.4, 6.5., 6.6 and 7.

Reviewer comments #2-2

The comments to Section 6, are quite naive. Authors seem just to describe them and not explain the reason for the provided results. More explanations should be given here.

Author’s comments #2-2

Thank you very much for your professional comments. Based on your comments, I added and enriched explanations in section 6.4, 6.5 and 6.6.

 In 6.4, I added “Generally, CNN can achieve reasonable performance with default hyperparameter settings; however, extensive hyperparameter tuning is typically required to achieve the best performance. Thus, for highly imbalanced datasets with many nominal attributes, the predictive model generated by the CNN may have formidable challenges [96]”.

 In 6.5, Rudin also showed that simple models such as linear regression and rule-based learners, achieved comparable performance to the complicated models such as deep learning models, ensemble models, and random forests. Moreover, there exists no noticeable difference in their performance [97]. I also added two recent literatures that criticizes the interpretability of XGBoost [104]. I introduced a literature [105] proposed a heterogeneous deep forest model that combines deep learning architecture and tree-based ensemble classifiers is proposed as the modeling approach. This means that the authors want to know about deep learning meets decision trees (includes complex trees)? This challenge is really fundamental to use deep learning instead of complex trees such as XGBoost and random forest.

In 6.6, I also added the evidence of realization of hardware processors synthesized on Field-Programmable Gate Array (FPGAs) and Convolution Neural Networks (CNN) [109]. This is not proof of concept. This paper was published by Xilinx which is a famous FPGA hardware company.

Reviewer comments #2-3

To give readers a more comprehensive understanding of the future development of the proposed topic. Can the author provide a tree diagram in Section 7 to organize the future technologies you mentioned?

Author’s comments #2-3

Thank you very much for your faithful comments. Based on your comments, I provided a tree diagram (Fig.17) to organize the future technologies of Deep Learning-based credit scoring in section 7.

Reviewer comments #2-4

In Section 2.8, is there a typo in the sentence "2D CNNs are used more frequently than 2D CNNs”?

Author’s comments #2-4

Thank you very much for your careful comments. You are right. There is a typo. I apologize. That is, "2D CNNs are used more frequently than 1D CNNs”

Reviewer comments #2-5

Is Section 3.3 incomplete, because you have mentioned that "we investigated these performances separately" at the end of this section, but there are no other descriptions for the Performance in the follow-up.

Author’s comments #2-5.

Thank you very much for your careful comments. I apologize for careless mistake for placement. Section 3.3 should be section 3.4 vice versa.

Reviewer comments #2-6.

Please redraw Figures 7 and 8. It is too vague to understand.

Author’s comments #2-6.

Thank you for your constructive comments. I apologize. Based on your comments, I redrawn Figure 7 and 8 to make the figures more understandable.

Reviewer comments #2-7

In Figures 11 and 15, the legend is missing.

Author’s comments #2-7

Thank you for your constructive comments. Based on your comments, I revised them appropriately sized legends for two figures.

Reviewer comments #2-8

The font size of all tables in this paper is too large.

Author’s comments #2-8

Thank you for your constructive comments. Based on your comments, I revised font size to see all tables properly smaller.

Reviewer comments #2-9

On page 27 of the paper, there is a paragraph of text that does not conform to the paper specification.

Author’s comments #2-9

Thank you for your constructive comments. Based on your comments, I revised the last paragraph to meet the paper specifications.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

In this revision, the author provided responses to my comments and revised the paper accordingly. So, I recommend an acceptance of the paper.