Adaptive Two-Index Fusion Attribute-Weighted Naive Bayes

Round 1

Reviewer 1 Report

A Naive Bayes based on a two-index attribute is proposed. In addition, a β factor is proposed to weigh the influence of each index on the classification. A thorough analysis regarding different indexes is carried out illustrating the advantages of a variable β. 

 

I have the following suggestions to improve this article:

> The English need to improve to remove typos and grammatical errors in the entire manuscript. For example, there are missing spaces between the text and citations (almost every time an [ ] is present). On line 307, it should be "Similarly" instead of "Similar". 

> On page 10, line 210, which state-of-the-art filters do the authors mean?

> Section 4.4: what does the parameter p, α, n, R+ and R- mean? the authors need to mention where the parameters' values come from. If they are using other studies as the reference for the chosen values, that should be mentioned and provide the original work that obtained such values. 

> Page 14, line 271, why 434?

> Table 8 and Table 12 are not called/invoked anywhere in the text nor discussed.

> Page 16, line 317, it is not clear what the authors meant by "is higher than the number of instances is greater than or equal to 500 (56.25%)," It Can probably be fixed just by adjusting the English.

> Line 327, "By contrast, ATFNB does not perform well on datasets with large instances and attributes." Do the authors know why? I mean, what is the key bottleneck this proposal faces with big datasets?

Also, is there any idea of the dataset size limit that this proposed approach can be applicable? I suggest the authors add in the text possible applications that rely on small datasets to validate the need for such modifications. This could be added to the Introduction to better explore this proposal's relevant background and needs.

> Please, add the hyperlink to Section calls. For example, in lines 331 and 341, it is not possible to click on the number of the section and immediately jump to it.

> Table 16: Please add the magnitude unity to the column head as well and not only to the table subtitle. For example: Stage 1 (s) | Stage 2 (s) | ...

> Table 4 discusses the processing time of each algorithm. I suggest the authors provide information regarding the algorithms' implementation. Were both developed with the same language and on the same platform? This could be addressed in Subsection 5.4 to clarify that a fair comparison was performed.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

In contrast to the frequently used approaches in literature, e.g. gradient descent or evolutionary algorithms, authors choose a conventional approach. With this approach, features in the data are extracted via filters.

The unique approach of this paper is that two values are used to characterize a feature instead of one. Thus, the paper provides a new approach to scientific literature.

I am currently of the opinion that this approach cannot achieve a general improvement in performance for theoretical reasons and if it does (according to the evaluation section in the paper), then only for reasons of randomness or non-comparable configurations.

If a feature is assigned a value, it is hard-wired and absolutely determinable.

If - in an extreme case - an infinite number of values is assigned to a feature, then the following happens:

Some of the values are dependent among themselves and some of the values are independent among themselves. The dependent values are trivial, they can be combined to a single hyper-value. But the independent ones represent different features of the data. From the combination of independent features, a new and relevant feature could arise in rare cases. In the vast majority of cases, it results in data garbage.

In this paper, two values for one feature are used. This only works, if the two values are dependent. If not, it does not work. Further, the approach used in the paper uses an arbitrarily selected value for the feature. For a feature that is combined by two independent features, that means, one is randomly set as the dominant. This must always be worse in terms of performance over time than if both properties are included in the optimization process.

I attribute the good performance of the approach compared to other approaches achieved in the evaluation to the fact that for the optimization processes in machine learning, hyperparameters have to be determined, which depend on the type and number of properties. I suspect optimally chosen hyperparameters for this approach, and that these hyperparameters are less optimally for the comparison approaches.

Although the paper is very well written, abstract and introduction, background, method and evaluation is done well; I actually have to reject the paper, but give the authors a chance to respond with a resubmission.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

All my comments are addressed in the revised version.