Combining Fuzzy C-Means Clustering with Fuzzy Rough Feature Selection

Round  1

Reviewer 1 Report

The paper introduced a new method of fuzzy rough feature selection by using membership function determination method (based on Fuzzy C-Mean Clustering) and fuzzy equivalence. We highly appreciated the efforts of the authors for the research. The manuscript is well-written and has some noticeable findings.

There are some points that the author can improve as follows:

1.    The authors are combining Introduction and Literature Review into the Introduction part. Normally they should be separated, maybe to different parts or different paragraphs. The Introduction should state the context, the current research, the current research limit, and then state the purpose of the paper. Whereas the Literature Review part should review a little more detail papers in the Literature.

2.    The Introduction has not referred the current advancement of rough set yet. Authors should reference newer and more recent publications on rough set advancement in general and FRFS in particular. I can recommend some as follows:

Ø  https://www.mdpi.com/2073-431X/7/3/44

Ø  https://www.sciencedirect.com/science/article/abs/pii/S0950705118305744

Ø  https://www.fujipress.jp/jaciii/jc/jacii002100071221/

Ø  https://www.sciencedirect.com/science/article/pii/S0166361517302907

Ø  https://www.sciencedirect.com/science/article/abs/pii/S0950705118301631

Ø  https://www.sciencedirect.com/science/article/pii/S016501141400219X

….

3.    Authors should name your method, instead of the “new method”.

4.    In Table 4 and 5, methods with better results should be highlighted so that we can see the superiority of your method.

5.    If possible, we encourage authors to share their work and/or source code (e.g. github) on any public repository so that we can review the results better.

Author Response

Response to Reviewer 1 Comments

Point 1: The authors are combining Introduction and Literature Review into the Introduction part. Normally they should be separated, maybe to different parts or different paragraphs. The Introduction should state the context, the current research, the current research limit, and then state the purpose of the paper. Whereas the Literature Review part should review a little more detail papers in the Literature.

Response 1: Thanks for your suggestion. we think this is a good idea. we ignored the part of Literature Review and combined it to the part of Introduction. To correct this mistake, we separate Introduction and Literature Review to different part and supplement the literatures. we add some relatively new literatures to the part of Literature Review and introduce new progress of rough set and FRFS briefly in order to make the readers know more about this field. Specifically, in the manuscript we reserve a brief introduction of feature selection, rough set and FRFS in section 1, and move the detail of papers to the section 2. Additionally, some recent publications are list in section 2. For more details, you can see clearly with the Track Changes function of Word.

Point 2: The Introduction has not referred the current advancement of rough set yet. Authors should reference newer and more recent publications on rough set advancement in general and FRFS in particular. I can recommend some as follows:

Ø  https://www.mdpi.com/2073-431X/7/3/44

Ø  https://www.sciencedirect.com/science/article/abs/pii/S0950705118305744

Ø  https://www.fujipress.jp/jaciii/jc/jacii002100071221/

Ø  https://www.sciencedirect.com/science/article/pii/S0166361517302907

Ø  https://www.sciencedirect.com/science/article/abs/pii/S0950705118301631

Ø  https://www.sciencedirect.com/science/article/pii/S016501141400219X

….

Response 2: We appreciate for your suggestion. We have read the publications you recommend. All of them are excellent and make contributions in the field of rough set and FRFS. We add most of the above papers to our manuscript and some other outstanding production which are published in these two years. The detail changes are primary in section 2 and reference part.

Point 3: Authors should name your method, instead of the “new method”.

Response 3: We think this is a constructive recommendation. Shame to say that we ignored to given a name to our new method. For careful thinking in consideration of easy to remember and understand, the method named after C-FRFS for short of clustering fuzzy rough feature selection.

Point 4: In Table 4 and 5, methods with better results should be highlighted so that we can see the superiority of your method.

Response 4: Thanks for your suggestion. We have highlighted the better results in Table 4 and 5. Details are shown in manuscript.

Point 5: If possible, we encourage authors to share their work and/or source code (e.g. github) on any public repository so that we can review the results better.

Response 5: We appreciate for you recommendation. We are sorry to say that because the follow-up study of our national key program it is inconvenient to share all the source code on public yet. But with the boost of our program, we will sort out the materials and data, share on public repository.

Author Response File: Author Response.docx

Reviewer 2 Report

Cross validation should be used for the experiments.

A statistical test should be used for the comparison of the examined methods.

The authors should explain why the proposed methodology seems to work well and present the time efficiency of their method.

Author Response

Response to Reviewer 2 Comments

Point 1: Cross validation should be used for the experiments.

Response 1: Thanks for your recommendation. We used 10 folds cross validation in the modelling of each classifier method but forgot to state in the manuscript. We add this important point after the introduction of classifiers in Experiments part of our manuscript, “To obtain more accurate results, we used 10 folds cross validation in the modelling of each classifier method.”

Point 2: A statistical test should be used for the comparison of the examined methods.

Response 2: We appreciate for your suggestion. Refer to our experiments and other publications, we choose Friedman test as statistical test of our manuscript. The result shows that the examined methods are statistical difference. We add average ranks of methods about Friedman test in Table 6. More details are shown in the last paragraph of Experiments.

Point 3: The authors should explain why the proposed methodology seems to work well and present the time efficiency of their method.

Response 3: Thanks for your recommendation. Firstly, the Figure 2 and Figure 3 are about the time efficiency of our method. In Figure 2 we listed the running time of datasets used in the experiments and compared the difference between three methods. In Figure 3, we tested methods efficiency on datasets with objects from 200 to 5000, and concluded there were no obvious difference. Time efficiency of our method is almost same as the existing methods.

Secondly, our proposed method seems to work well because clustering replaced the original membership function. Clustering is a process of constantly adjusting its fuzzy membership, iteration will eventually reach a relatively good membership function. According to the fuzzy equivalence relation, the similarity degree of two elements can be determined. For example, the equation of similarity degree in other method is                                                it looks like complex, but it is the membership function of ,  and the relation .  In some cases, such as the concentrated distribution of data points, the equation may fail to get a good result. We use the clustering to construct membership function so that there is no need to consider the points distribution.

Author Response File: Author Response.docx

Round  2

Reviewer 2 Report

The paper has been improved after the revision