On the Black-Box Challenge for Fraud Detection Using Machine Learning (I): Linear Models and Informative Feature Selection

Round 1

Reviewer 1 Report

The article is devoted to the development of a methodology to overcome the difficulty in applying machine learning to the problem of detecting credit fraud using modern algorithms capable of quantifying information about variables and their relationships. The relevance of the task is dictated by the fact that artificial intelligence is rapidly shaping the global financial market and its services, including when detecting credit fraud. The need for strict adherence to non-discriminatory legislation and data protection rules restricts almost all possible applications of artificial intelligence to simple and easily traceable neural networks, thus preventing the use of more advanced and modern methods. The authors propose a new methodology to address the difficulty in applying machine learning to the loan fraud detection problem using state-of-the-art algorithms capable of quantifying information about variables and their relationships. The approach proposes a new concept of interpretability to deal with this multifaceted situation. To do this, we first adapt and apply the feature selection method, the identifier of an informative variable, which can distinguish between informative, redundant, and noisy variables. Second, a set of innovative recurrent filters is applied to minimize training data bias. The output is classified using machine learning techniques such as gradient boosting, support vector machines, linear discriminant analysis, and linear regression. These models are applied both to a synthetic database for better descriptive modeling and fine-tuning and to a real database. The results of the study confirm that the proposal provides valuable interpretability by defining weights of informative features that link input variables to final goals, indicating the weights of informative features. 76% accuracy in CFD, representing an improvement of over 4% over existing studies. Using the presented methodology, one can not only reduce the dimension but also increase the accuracy, as well as trace the relationship between input and output features.

Despite the satisfactory quality of the article, some shortcomings need to be corrected.

1. The aim of the article should be defined.
2. The state-of-the-art methods should be separated from the ones proposed by the authors.
3. Figures 3 and 6 are not informative. It is recommended to separate each of them into four ones, increase the font and describe figures more in text.
4. The methodology proposed by the authors should be described in more detail.
5. The scientific novelty of the paper should be highlighted.

In summarizing my comments I recommend that the manuscript be accepted after major revision, including improving the article design and focusing on the methodology description. 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Dear authors,

Please, see the attached file. 

Cheers

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper is well-written and well-structured and clearly demonstrates its main contributions. The discussion is sound and the results are consistent with the claims made by the authors.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Thanks for the authors for considering my comments and recommendations. I believe, now the article can be accepted.