Transfer Learning with Social Media Content in the Ride-Hailing Domain by Using a Hybrid Machine Learning Architecture

Round 1

Reviewer 1 Report

The paper presents a study of transfer learning-based social media content analysis in the ride-hailing field using hybrid machine learning. The paper has many issues as follows.

• The structure of the paper is not well organized and rambling. The paper should straight to the point and avoids repetitions. I suggest removing unnecessary parts, for example:
  • Section 2.2, 3.5.1, and 4.2 are the same.
  • Use more dense sections, avoid using as many sections/subsections as possible.
  • LDA is a well-known topic modeling approach, thus it is not necessary to describe it in detail. The readers can refer to the original paper.
  • Standard procedures such as data cleaning, etc as shown in 3.5 should be removed.
• The hybrid machine learning used in the paper is still questionable.  The authors just employ the existing NLP technique without any improvement. In what aspect LDA is considered as a hybrid approach?
• Figure 2, 3, 7 are not visible. All captions of all figures should be self-explained, not just title-like ones.
• I am interested to see the applicability of this work in the real world. The authors can elaborate on the social impact of this research. I have some knowledge of computational social science and I do not see any meaningful impact of this paper.

Author Response

Please see attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The manuscript addresses a novel topic, namely hybrid machine learning-based systems for text classification, presenting the authors' effective contribution to the development of such system that was trained with Reddit ride-hailing posts, then tested on data from Twitter and Google Play, and then being able to perform transfer learning.

The paper has a solid, well-structured construction, starting with the Introduction section, then the Background section summarizing algorithms and processes used, continuing with the third section System Architecture, and then The Ride-Hailing important domain for companies with mobility-based services, the fifth section based on Transfer Learning and the final Conclusions.

I would have liked to read more about the directions of continuation / improvement of the present research, including in the Conclusions part, where this is very little discussed.

It would be interesting to approach data from other social media sites (for example, my students use Facebook for ride-hailing more than other sites).

In a search on the Turnitin portal, the paper has a general similarity of 9%, which is a good thing, but there are certain paragraphs in section 2.Background that appear colored in the text and corresponding nonsense (especially those that refer to mathematical formulas / equations used). Thus are rows 71-77, 95-105, 152-164, 178 ~ 200 (see attached pdf file). I would recommend the authors to put in quotation marks the quoted text / explanations or directly the bibliographic source next to the equations, etc. (since I reckon somebody else created/invented them)

A final reading of the manuscript by a native English speaker is my last recommendation.

Otherwise, I did not notice any other problems, the manuscript looks original, the authors' own work, I recommend its publication with the aforementioned recommendations.

Comments for author File: Comments.pdf

Author Response

Please see attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

In "Transfer Learning with Social Media Content in the Ride-Hailing Domain by using a Hybrid Machine Learning Architecture" authors utilized a hybrid algorithm to classify into domains the comments of users of ride hailing services. The model was trained considering Reddit posts/comments and then utilized to classify text originating from different platforms (twitter and google play). This transfer learning algorithm allows to training a model in domain where it is possible to acquire a large datasets to be used on other domains where data is scarce. 

The authors considered different models and parameters to evaluate the performances considering both English and Spanish languages.

The large number of topics (11) make rather difficult the classification in some case especially because the intrinsic differences in language and occurrence of topics in the different domains. In spite of this authors obtained a quite satisfactory results which can be a promising starting point for future works.

The main concerning of reviewer ( it is more a curiosity) concerns the assignment of each post and comment to a single topic. Do the authors think that this assumption can be a limit to the approach ( sometime multiple topics or mixed topic can be present)? In case do the authors think that the algorithm can assign a single "sample" to more than one class, maybe setting a threshold on the probability of belongings to the single topics? A comment of authors would be appreciated.

Secondly it would be interesting a comment about the misclassified samples. In particular in which topics the misclassified samples of topics are assigned by the models.

In the reviewer opinion, after this minor points/comments the manuscript will be worth to be published.

 

 

Author Response

Please see attachment.

Author Response File: Author Response.pdf