A Hybrid Two-Phase Recommendation for Group-Buying E-commerce Applications

Round 1

Reviewer 1 Report

Details of the well-known K-means algorithm (p.8, 10, 11) are not necessary in m opinion.

Eq17 is not necessary.

I am not convinced by the interest of figures 5-7. It could be useful for readers to comment more deeply these figures.

Other experiments with other datasets would help readers to be convinced about the method proposed. Especially if the dataset considered is not publicly available.

Moreover, authors have to give statistics about the results presented, especially abut significance.

Authors insist on "scalability" and "data sparsity" but no specific experiment are given about these elements.

Author Response

Point 1: Details of the well-known K-means algorithm (p.8, 10, 11) are not necessary in m opinion.

Response 1: Thanks for your suggestion. In this paper, the K-Means clustering algorithm is applied for item clustering and user clustering, and their processes are listed in detail.

 

Point 2: Eq17 is not necessary.

Response 2: Thanks for your suggestion. We deleted Eq17.

 

Point 3: I am not convinced by the interest of figures 5-7. It could be useful for readers to comment more deeply these figures.

‍Response 3: Thanks for your suggestion, we made some revision marked in red. Figures 5-7 verified the effectiveness of HTPR in term of item clustering and user clustering.‍

Fig. 5 shows more nearest neighbor items can be found out with less searched items through item clustering, the greater the number of clusters is, the higher the efficiency of neighbors searching is.

In Fig. 6 the curves of MAE values in different cases as a whole show down first and then up. They decrease obviously when the number of clusters increases from 10 to 40 or 50, and afterwards go up slowly and tend to stabilize. Moreover, it is noticed that the curve whose nearest neighbors is 30 is lower than other curves, MAE achieves a lowest value when the number of clusters is 40. Therefore, for item clustering, the optimal values of parameters, i.e., the number of clusters and nearest neighbors could be 40 and 30 respectively.

Fig. 7 shows the diagram of search efficiency with different clusters varied from 30 to 50 with a step of 10. Y-coordinate denotes the ratio of searched nearest neighbors of users, and X-coordinate denotes the percentage of users that have been searched.

 

Point 4: Other experiments with other datasets would help readers to be convinced about the method proposed. Especially if the dataset considered is not publicly available.

‍Response 4: Thanks for your suggestion. In the future work, we will focus on a comprehensive comparison and application of the HTPR method and analyze the strengths and weaknesses by the other datasets and the real-time Group-buying e-commerce application.‍

 

Point 5: Moreover, authors have to give statistics about the results presented, especially abut significance.

Response 5: Thank you for your suggestion. The purpose of experiments is to verify effectiveness of HTPR in term of item clustering and user clustering, and to determine the number of nearest neighbors and the number of clusters, and so the statistics about the results is not described.

 

Point 6: Authors insist on "scalability" and "data sparsity" but no specific experiment are given about these elements.

Response 6: Thank you for your suggestion. The dataset comes from an offline dataset used by an operation team of a group-buying website in China. It is taken from a half year (from June to December 2014) and includes ratings of 3,043 users on 1,628 items, its time span and scope are scale and data content is sparse.

Reviewer 2 Report

The paper presents a recommendation algorithm for e-commerce applications, the HTPR method. The main novelty of the proposal is that it is a hybrid recommendation method considering both item clustering and user clustering.

In general, the paper is well written, well presented, and easy to follow. However, there are some aspects of the format that can be improved:

-   Some section or table headers have been left orphan at the end of a page. See line 128 (page 3), line 255 (page 7), and line 432 (page 12).

-   The format of the equations is not correct according to the template in https://www.mdpi.com/journal/applsci/instructions

-   The quality of Figures 5 to 13 can be improved. Use a vector format to export the figures, or at least a high resolution, lossless compression format such as PNG. The figures seem to be in JPEG format and at a not very high resolution.

-   I don’t know if the format of the subsubsections is correct according to the journal guidelines (big bullet and bold font, e.g. line 457, page 13, or line 523, page 16), but it sounds strange to me.

Regarding the content of the paper, I found it technically sound and interesting. However, there are a three points I would like the authors to clarify.

-   First, if the dataset you mention in section 5.1 is publicly available or not. If it is not, the authors should define a mechanism by which other researchers interested in your work could access the dataset, in order to reproduce your results. For example, by contacting one of the authors and signing some kind of confidentiality agreement or something like that.  The authors should explain this mechanism in the paper.

-   Second, for the evaluation HTPR is compared with other three recommendation methods: basic CF, user clustering based CF (UCCF) and item clustering based CF (ICCF). The election of these mechanisms is not justified. Moreover, in future work you state that a comprehensive comparison of the HTPR method against state-of-the-art methods is needed. What methods do the authors refer to and why has not the comparison been made with them instead of those presented in the paper? Please, clarify.

-   Finally, can you give more detail on in which platform you have done the implementation, and on CPU time and memory consumed by the different methods to do the process?

Author Response

Point 1: The paper presents a recommendation algorithm for e-commerce applications, the HTPR method. The main novelty of the proposal is that it is a hybrid recommendation method considering both item clustering and user clustering.

In general, the paper is well written, well presented, and easy to follow. However, there are some aspects of the format that can be improved:

Response 1: We appreciate for the reviewers’ work and hope that the revisions meet with his/her approval.

 

Point 2: -   Some section or table headers have been left orphan at the end of a page. See line 128 (page 3), line 255 (page 7), and line 432 (page 12).

Response 2: Thank you for your suggestion. We have re-edited the whole contents.

 

Point 3-   The format of the equations is not correct according to the template in https://www.mdpi.com/journal/applsci/instructions

Response 3: We edit all the equations using the Equation editor version 3.0 from Microsoft.

 

Point 4:-   The quality of Figures 5 to 13 can be improved. Use a vector format to export the figures, or at least a high resolution, lossless compression format such as PNG. The figures seem to be in JPEG format and at a not very high resolution.

Response 4: Thank you for your suggestion. The figures are from the export of matlab and are saved as the JPEG format.

 

Point 5:- -   I don’t know if the format of the subsubsections is correct according to the journal guidelines (big bullet and bold font, e.g. line 457, page 13, or line 523, page 16), but it sounds strange to me.

Response 5: Thank you for your suggestion. We have re-formated the article using the applsci-template.

 

Point 6:Regarding the content of the paper, I found it technically sound and interesting. However, there are a three points I would like the authors to clarify.

Response 6: Thank you for your suggestion. We hope that our response meet with your approval.

 

Point 7:-   First, if the dataset you mention in section 5.1 is publicly available or not. If it is not, the authors should define a mechanism by which other researchers interested in your work could access the dataset, in order to reproduce your results. For example, by contacting one of the authors and signing some kind of confidentiality agreement or something like that.  The authors should explain this mechanism in the paper.

Response 7: The dataset in section 5.1 is not publicly available and comes from an offline dataset used by an operation team of a group-buying website in China. It is taken from a half year (from June to December 2014) and includes ratings of 3,043 users on 1,628 items. The dataset can be gotten by contacting the corresponding author.

 

Point 8:-   Second, for the evaluation HTPR is compared with other three recommendation methods: basic CF, user clustering based CF (UCCF) and item clustering based CF (ICCF). The election of these mechanisms is not justified. Moreover, in future work you state that a comprehensive comparison of the HTPR method against state-of-the-art methods is needed. What methods do the authors refer to and why has not the comparison been made with them instead of those presented in the paper? Please, clarify.

‍Response 8: In our experiments, we compared HTPR with other three recommendation methods: basic CF, user clustering based CF (UCCF) and item clustering based CF (ICCF) because the last three approaches had been widely applied in the e-commerce website, especially in the group-buying website from which we acquired the testing dataset. In the future work, we will focus on the application of the HTPR method and analyze the strengths and weaknesses by the real-time Group-buying e-commerce application.‍

 

Point 9:-   Finally, can you give more detail on in which platform you have done the implementation, and on CPU time and memory consumed by the different methods to do the process?

Response 9: All the experiments are carried out with the spark framework and run on a workstation equipped with an Intel 12-core 3.5GHz CPU, a GTX1080TI GPU, and 16GB memory.

Reviewer 3 Report

Thank you for providing the chance to review this interesting paper.
The paper presents a hybrid two phase recommendation for group-buying e-commerce application by combining collaborative filtering with cluster technique.
The introduction section provides enough information to motivate the study. However, the presented gap resulted dated. Still, authors based the research gap only three papers. Please strength and comment this part. (line 46-51)
The literature review part shows the recommendation techniques. The part contains some English mistakes and a proof reading is required.
Problem analysis and proposed HTPR method are clear and have a solid base.
Experimental results show initially the dataset used for the experiments and the metrics for evaluate the quality of recommendation. Then, Accuracy evaluation and effectiveness evaluation are farther discussed by means of graph. This part required a proof reading since the narrative flow is not easy to follow.
Conclusion part is weak. There are not implications for practitioners and very few implications for researchers. Please add those parts.

Author Response

Point 1: Thank you for providing the chance to review this interesting paper.

The paper presents a hybrid two phase recommendation for group-buying e-commerce application by combining collaborative filtering with cluster technique.

Response 1: We appreciate for the reviewers’ work and hope that the revisions meet with his/her approval.

Point 2: The introduction section provides enough information to motivate the study. However, the presented gap resulted dated. Still, authors based the research gap only three papers. Please strength and comment this part. (line 46-51)

Response 2: Thanks for the valuable suggestion. We have adopted this suggestion, and added the reference [43] and some comments in this part (line 51-56).

 

Point 3: The literature review part shows the recommendation techniques. The part contains some English mistakes and a proof reading is required.

Response 3: Thanks for your suggestion. We read the manuscript many times and tried our best to polish the language, including grammar, expression, formula, table, and structure. We hope that the revisions meet with your approval.

 

Point 4: Problem analysis and proposed HTPR method are clear and have a solid base.

Experimental results show initially the dataset used for the experiments and the metrics for evaluate the quality of recommendation. Then, Accuracy evaluation and effectiveness evaluation are farther discussed by means of graph. This part required a proof reading since the narrative flow is not easy to follow.

Response 4: Thanks for your suggestion. We read the experimental results part and tried our best to make the proof reading, and hope that the revisions meet with your approval.

 

Point 5: Conclusion part is weak. There are not implications for practitioners and very few implications for researchers. Please add those parts.

Response 5: Thanks for your suggestion. We re-write this part and hope that the revisions meet with your approval.

Round 2

Reviewer 3 Report

Authors improved the narrative flows.

The article is clearer.

The introduction presents the study in a better way. The gap is now update.

Furthermore, the added discussion in section 5 makes the section more interesting.

The conclusion now contains implication.

I only suggest  a last proof reading prior to the acceptance.

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

The paper presents an interesting approach (HTPR method) on group buying services. However, I would like to suggest the following issues to improve. 

1) The mathematical modeling of this paper is very poor. 

In particular, the modeling does not seem to be original. 

2) The presentation of this paper is very poor. Authors need to significantly improve it. 

3) The main problem of group buying is NOT clearly explained. Please look at the abstract more carefully. 

4) Most importantly, I am NOT sure whether this paper is in the scope of this journal "Sensors".

Author Response

Point 1: The mathematical modelling of this paper is very poor.  In particular, the modelling does not seem to be original.

Response 1: Our proposed model is different from some existing hybrid models. Although there are some hybrid recommendation algorithms based on clustering and collaborative filtering, they seldom consider both the characteristics of items and the user behaviors on items. Our two-stage hybrid recommendation algorithm has two main innovations: (1) our method separates item clustering and user clustering, and the latter uses the information of the former, which not only simplifies the recommendation process and solves the data sparsity problem, but also facilitates parallelization to improve the efficiency of the algorithm; (2) our method provides higher prediction accuracy than three recent recommendation methods: traditional CF, user-clustering based CF, and item-clustering based CF.

 

Point 2: The presentation of this paper is very poor. Authors need to significantly improve it. 

Response 2: Thanks for your suggestion. We read the manuscript many times and tried our best to polish the language, including grammar, expression, formula, table, and structure, and the modified places were marked in red in the revised manuscript. We hope that the revisions meet with your approval.

 

Point 3: The main problem of group buying is NOT clearly explained. Please look at the abstract more carefully.

Response 3: Thank you for your suggestion. We have rewritten the abstract and made it more clear. “This leads to the poor recommendation effect of traditional collaborative filtering (CF) methods in group-buying applications” have been added, and “It has become increasingly difficult to recommend services or items efficiently and effectively” have been deleted.  

 

Point 4: Most importantly, I am NOT sure whether this paper is in the scope of this journal "Sensors".

Response 4: Thank you for pointing this out. Many papers related to Recommendation Systems and Collaborative Filtering have been published in Sensors recently. Here are some examples:

1)       Exploring IoT Location Information to Perform Point of Interest Recommendation Engine: Traveling to a New Geographical Region. https://doi.org/10.3390/s19050992

2)       An Exception Handling Approach for Privacy-Preserving Service Recommendation Failure in a Cloud Environment. https://doi.org/10.3390/s18072037

3)       A Smart Recommender Based on Hybrid Learning Methods for Personal Well-Being Services. https://doi.org/10.3390/s19020431

4)       A Personalized QoS Prediction Approach for CPS Service Recommendation Based on Reputation and Location-Aware Collaborative Filtering. https://doi.org/10.3390/s18051556

5)       An Efficient Recommendation Filter Model on Smart Home Big Data Analytics for Enhanced Living Environments. https://doi.org/10.3390/s16101706

6)       HyRA: A Hybrid Recommendation Algorithm Focused on Smart POI. Ceutí as a Study Scenario. https://doi.org/10.3390/s18030890 

Reviewer 2 Report

This paper proposes a hybrid recommendation system for e-commerce applications.

1) This paper is written well overall. However, the authors should improve the presentation of Section 4 significantly.

2) There are many broken symbols in Lines, e.g. 250, 343, 346, 347, 348, and so on.

3) Notations are very confusing.

- Line 250: Attributes of item u should be denoted as s_{u, k} instead of s_{mu,k}.

- Line 271: What is 'e'? The symbol 'e' defined in Line 267 is valid only until Line 269.

- Line 277: Does the notation 'I' represent the intersection of sets? If so, please state it explicitly.

- Line 293: What is K_l NN(e)? Does it indicate KNN(e) in Eq. (7)?

- Line 296: What is 'f'?

- Lien 320: What is 'E' in the denominator?

- Line 339: What is \rho and 'cursive d', respectively?

There still remain many to be modified.

4) The authors has defined the similarity of two items in Eq. (4). However, the metric does not seem to be used in Algorithm 1 (in Line 274) when comparing the distance of two items. What is the reason?

Author Response

Please check the point-by-point response to the reviewer’s comments  in the below Word file because there exist equations in the reponse file.

Author Response File: Author Response.docx

Reviewer 3 Report

this paper proposed two Two-Phase Recommendation system for an E-commerce application. it is somehow interesting but need to improve related citation and some major change.

I suggested the following comments to the authors;

1: in the title, you mention "hybrid" authors need to explain, why and how is your method hybrid?

2: abstract need to improve based on method and result.

3: manuscript lack some related references, authors need to improve based on a concept, you can use the following papers.

Intelligent E-commerce logistics platform using hybrid agent based approach

Web-Based Recommendation System for Smart Tourism: Multiagent Technology

4: results are not clear, I suggested provided tables for each result, authors can be used first recommended paper to improve this manuscript.   

Good Luck 

Author Response

Point 1: In the title, you mention "hybrid" authors need to explain, why and how is your method hybrid?

Response 1: Hybrid means the combination of two different methods, collaborative filtering and clustering, in this paper.

Why we combine two different approaches? Because Hybrid can make use of the advantages of both methods. Clustering can classify items according to their features, and its advantage is that it does not need manual annotation. Collaborative filtering can recommend the interest products to other users who have the same shopping experience. The advantage of this method is that it is fast and can discover potential user interests.

How we combine two different approaches? Firstly, item features and user behaviors are applied to cluster items, in which we use the K-Means algorithm. Secondly, the user-item rating matrix is supplemented where items without rating is rated by the rating of their rated nearest neighbors. Thirdly, the preference of a user for an item category, the concern degree of the user for the item category, and the tendency of the user for the item category can be calculated successively, and then the integrated similarity can be computed. Then, we cluster users according to the user-item category tendency matrix using K-Means. Lastly, a personalized service recommendation list is generated by a collaborative filtering method (here we use KNN algorithm) and the most appropriate items are recommended to the target user.

 

Point 2: Abstract need to improve based on method and result.

Response 2: Thank you for your suggestion. We have rewritten the abstract and made it more clear.

 

Point 3: Manuscript lack some related references, authors need to improve based on a concept, you can use the following papers.

Intelligent E-commerce logistics platform using hybrid agent based approach

Web-Based Recommendation System for Smart Tourism: Multiagent Technology

Response 3: Thank you for your suggestion. We have added these two references to the Literature Review section ([5], [25]).

 

Point 4: Results are not clear, I suggested provided tables for each result, authors can be used first recommended paper to improve this manuscript.  

Response 4: Thank you for your suggestion. The results will be clearer if we provided tables for each result, but the depiction on the 9 comparison results will be repetitive with the figures.