Music Recommendation Based on “User-Points-Music” Cascade Model and Time Attenuation Analysis

Round 1

Reviewer 1 Report

1. Abstract needs to be re-written and please also address following ones:

2. Please give fig 1 in more resolution. Please change fig 2 with more understandable one.
3. Please give a framework/flowchart of the proposed system.
4. Please give values of all performance metrics in a table instead of figures.

5. Please cite following papers:

-Playlist Generation via Vector Representation of Songs, 2016

-Content-driven music recommendation: Evolution, state of the art, and challenges, 2021.

-An exploratory teaching program in big data analysis for undergraduate students, 2020

-Music Recommendation Systems: Techniques, Use Cases, and Challenges, 2022

-Explainability in Music Recommender Systems, 2022

Author Response

Here is our response, and have also provided the word file in the attachment.

Dear reviewer:

Thank you for your comments concerning our manuscript entitled “Music Recommendation Based on “User – Points - Music” Cas-cade Model and Time Attenuation Analysis” (ID: electronics-1894876). Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in red in the paper. The corrections in the paper are as flowing:

Concern # 1: Abstract needs to be re-written and please also address following ones:

• State the key results/findings of proposed work.
• State the percentage improvement by proposed work, compared with existing works.

Author response: This concern has been adopted.

Author action: We have re-written the abstract.

We state the key results of proposed work as follows:

“And it came to the conclusion that the multi-interest point attenuation model can more accurately simulate the actual music consumption behavior of users and recommend music better”

We state the percentage improvement by proposed work as follows:

“a series of experimental results show that the effect of the proposed model MCTA has improved by 7 percentage points in terms of accuracy compared with existing works”

Concern # 2: Please give fig 1 in more resolution. Please change fig 2 with more understandable one.

Author response: This concern has been adopted.

Author action: We have changed the fig 1 to a more resolution one, and changed the fig 2 to a more understandable one.

Concern # 3: Please give a framework/flowchart of the proposed system.

Author response: This concern has been adopted.

Author action: Based on the original fig 1, we have added more information to help the reader to understand the framework about the proposed system.

Concern # 4: Please give values of all performance metrics in a table instead of figures.

Author response: This concern has been adopted.

Author action: We have expressed the following experiment result with table instead of figure:

Table 1. F1-value of different vector sizes; Table 2. F1-value of different models to extract the word vectors; Table 3. F1-value of different similarity threads; Table 4. Half-lives of different points of interest; Table 5. F1-value of different similarity calculation methods.

As for the following experimental results, we believe that the use of figures may better highlight the advantages of the model proposed in this paper.

  Figure 3. Model precision result; Figure 4. Model recall result; Figure 5. Model F₁ result; Figure 6. Model coverage value result.

Concern # 5: Please cite following papers:

-Playlist Generation via Vector Representation of Songs, 2016

-Content-driven music recommendation: Evolution, state of the art, and challenges, 2021.

-An exploratory teaching program in big data analysis for undergraduate students, 2020.

-Music Recommendation Systems: Techniques, Use Cases, and Challenges, 2022.

-Explainability in Music Recommender Systems, 2022

Author response: This concern has been adopted.

Author action: We have cited these papers as follow:

[20] Kse, B. , S. Eken , and A. Sayar . "Playlist Generation via Vector Representation of Songs." INNS Conference on Big Data 2016.

[8] Deldjoo Y , Schedl M , Knees P . Content-driven Music Recommendation: Evolution, State of the Art, and Challenges, 2021.

[2] Eken, S. An exploratory teaching program in big data analysis for undergraduate students. J Ambient Intell Human Comput 11, pp. 4285–4304, 2020.

[27] Schedl, M., Knees, P., McFee, B., Bogdanov, D. (2022). Music Recommendation Systems: Techniques, Use Cases, and Challenges. In: Ricci, F., Rokach, L., Shapira, B. (eds) Recommender Systems Handbook. Springer, New York, NY.

[5] Afchar D , Melchiorre A B , Schedl M , et al. Explainability in Music Recommender Systems,2022

We have tried our best to improve the manuscript and made some changes in the manuscript. 

We appreciate for your warm work earnestly, and hope that the correction will meet with approval.

Once again, thank you very much for your comments and suggestions.

Author Response File: Author Response.pdf

Reviewer 2 Report

The work is focused on the music recommmendation domain, managing new information sources for music delivery to the end users. Even though it has some merits, we think that it should be improved in some directions.

-We think that the contributions of the research, pointed out at the introduction section, should be revised and focused on the RS domain.

-The research gap should be declared in the related work section, more precisely.

-It is expected that the main novelties should be highlighted across the proposal presentation in Section 3. The authors suggest that the combination of user-point of interest-music, is among the most relevant contribution of the proposal. However, this should be justified in a better way.

-The concept of point of interest, should be explicitly described here.

-The proposal is a bit difficult to follow across the different stages. An overall figure, which specifies the input and output of each stage, would help in this direction. In this way, it is not clear the role of the similarities calculation in this scenario.

-The selection of the baselines to compare with the current proposal is not justified well. Authors should check the chapter focused on music recommender systems in the new RS Handbook 2022, for a better screenshot of the most recent works, and for possibly including new baselines.

Author Response

Here is out response, and the word file is in the attachment.

Dear reviewer:

Thank you for your comments concerning our manuscript entitled “Music Recommendation Based on “User – Points - Music” Cas-cade Model and Time Attenuation Analysis” (ID: electronics-1894876). Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in red in the paper. The corrections in the paper are as flowing:

Concern # 1: We think that the contributions of the research, pointed out at the introduction section, should be revised and focused on the RS domain.

Author response: This concern has been adopted.

Author action: We have re-written the part of contribution as follows:

-       We remodel the traditional "User - Music" recommendation framework into "User - Points of Interest - Music". Compared with single music recommendation, music rec-ommendation using interest points composed of multiple music can achieve a more stable recommendation effect.

-       We develop a music clustering model to extract the interest points for music recommendation system, ignoring the length of music list consumed is short or not, and no need to set the number of clusters in advance. It still works well for the niche music.

-       We propose a music interest attenuation algorithm considering the uneven distribution of music recommendation system. This slows down the decay rate of interest points user prefers and speeds up the decay rate of interest points that users do not like very much in the same time window, mitigating the Matthew effect of system.

Concern # 2: The research gap should be declared in the related work section, more precisely.

Author response: This concern has been adopted.

Author action: We have declared the research gap in the related work section as follows:

The above research recommends other music of this type based on the user's single music. However, in many cases, users consume music because of some special needs, all the music they have consumed should be considered to determine whether the user likes this type of music. We cluster users' historical music to obtain their favorite types, and further recommend other music under this type for users. In addition, we design a decay model to simulate the changing behavior of users to different types of music. The following section explains the proposed approach in detail.

Concern # 3: It is expected that the main novelties should be highlighted across the proposal presentation in Section 3. The authors suggest that the combination of user-point of interest-music, is among the most relevant contribution of the proposal. However, this should be justified in a better way.

Author response: This concern has been adopted.

Author action: We have highlighted the main novelties in Section 3 as follows:

• At the beginning of chapter 3.1 “Music Recommendation Structure”:

Instead of single music, music recommendation using interest points composed of multiple music can achieve a more stable recommendation effect

• At the beginning of chapter 3.2 “Music Recommendation Structure”:

For clustering the historical consumption music of each user to obtain the user's interest points, the following issues need to be considered:(1) Users' consumed music lists are of different length (2) Short historical music list for inactive users (3) The number of interest points of each user is different (4) The clustering speed should be fast enough to meet the online update requirements of the recommendation system. As far as we know, there is no suitable clustering algorithm to solve the above-mentioned problems, and a music clustering model is proposed here.

• At the beginning of chapter 2.1. “User-Music” Implicit Rating Matrix Construction:

In order to get the preference of each user for each point of interest, it is necessary to calculate the user's score for all music under this point of interest, but the user will not take the initiative to score the consumed music, and a scoring function is designed according to the user's music consumption behavior

• At the end of chapter 2.3. Clustering for User Multi-interest:

The cluster model need not to set the number of clusters in advance, and it still works well for the inactive user and niche music.

• At the end of chapter 3. Time Decay Modeling:

The attenuation algorithm slows down the decay rate of interest points user prefers and speeds up the decay rate of interest points that users do not like very much in the same time window. In addition, the niche music that are often consumed recently by people are recommended to more users, mitigating the Matthew effect of system.

Concern # 4: The concept of point of interest, should be explicitly described here.

Author response: This concern has been adopted.

Author action: Point of interest, also called point of information, it includes all the information for the recommended target, such as name, address, coordinate, category and so on.For the music point of interest in this article, it stands for a kind of music, represented by an n-dimensional vector, and n is the feature length of music. And the vector is composed of multiple music with similar characteristics.

Concern # 5: The proposal is a bit difficult to follow across the different stages. An overall figure, which specifies the input and output of each stage, would help in this direction. In this way, it is not clear the role of the similarities calculation in this scenario.

Author response: This concern has been adopted.

Author action: Based on the original fig 1, we have added more information to help the reader to understand the framework about the proposed system.

In addition, we have added the following contents to help readers understand the role of some chapters in this article:

• We have added following content at the chapter of 3.2.1 “User-Music Implict Rating Matrix Construction:

In order to get the preference of each user for each point of interest, it is necessary to calculate the user's score for all music under this point of interest, but the user will not take the initiative to score the consumed music, and a scoring function is designed according to the user's music consumption behavior.

• We have added following content at the chapter of 3.2.4. “Mixed Similarity Evaluation Index”:

This paper proposes a clustering model to obtain users' interest points, and the similarity calculation method between music plays a very important role in clustering effect. In addition, the existing similarity calculation indicators can’t measure the similarity between users' music consumption behaviors well.

Concern # 6: The selection of the baselines to compare with the current proposal is not justified well. Authors should check the chapter focused on music recommender systems in the new RS Handbook 2022, for a better screenshot of the most recent works, and for possibly including new baselines.

 Author response: This concern has been adopted.

Author action: We have carefully checked the chapter focused on music recommender systems in the new RS Handbook 2022, and add following contents to the article:

• We have cited the article “Music Recommendation Systems: Techniques, Use Cases, and Challenges” as follows ([27] in our article ):

Schedl, M., Knees, P., McFee, B., Bogdanov, D. (2022). Music Recommendation Systems: Techniques, Use Cases, and Challenges. In: Ricci, F., Rokach, L., Shapira, B. (eds) Recommender Systems Handbook. Springer, New York, NY.

• We have included a new baseline named “EUE” ([29] in our article):
• At the chapter of 4. “Experimental Framework and Setup”, We have added content as follow:

The approach in [29] includes two steps: pre-processing and prediction. In the preprocessing stage, all the music clips that users listen to in a certain time window and the microblog published by users are extracted, and then analyze the emotional characteristics of the microblog published, and then form a historical association between users, music clips and emotions. In the prediction phase, they obtain the current emotions of users, and then recommend music projects suitable for their emotional background.

• At the chapter of 4.4. “Results”, We have added content as follow:

By introducing information from other platforms, the EUE model can help alleviate the cold start problem of the music recommendation system and improve the coverage of the system, so the recommendation effect is better than the MTS model and CF model.

• We have improved the screenshot of the most recent works:

Recently, research on recommender systems emerged that aims at enhancing the traditional data-driven techniques with psychological constructs [27]. The users’ personality traits were considered in [28], the authors adapt the level of diversity in the recommendation list according to the personality traits of the user by reranking the results of a CF system. The research in [29] exploited the user’s affective state in music recommendation system. Using natural language processing techniques on a corpus of microblogs, they reached the emotional state and music listening information of the target user, and integrate this contextual information into other models such as CF and random rank.

We have tried our best to improve the manuscript and made some changes in the manuscript.  These changes will not influence the content and framework of the paper.

We appreciate for your warm work earnestly, and hope that the correction will meet with approval.

Once again, thank you very much for your comments and suggestions.

Author Response File: Author Response.pdf

Reviewer 3 Report

The manuscript “Music Recommendation Based on “User – Points - Music” Cascade Model and Time Attenuation Analysis” proposes A recommendation model MITA based on User-Point-Music structure is that clusters 16  user's historical behaviour, different interest points, obtained to further recommend 17 high-quality music under the interest points.

·       There are certain typos, grammatical mistakes in the manuscript. The authors need to proofread the paper for any possible mistake.

·       The abstract needs to be more informative. As in the current abstract the size of music database, contribution, etc., are unclear. Also, it is not clear if the music is language agnostic.

·       It is not clear what does “MITA” stands for?

·       Figure 1. Is poor quality. Kindly, provide high resolution images as per the journal’s guidelines.

·       The authors have mentioned in the section 4.2 “W2V model is used to extract music features”. It would have broaden the scope of manuscript if the authors have provided the comparison with other feature extraction models (Fasttext, Glove, etc.)

·        The results need better representation from contribution point of view.     

Conclusion : The manuscript is interesting, but efforts are needed to improve the manuscript holistically to be a potential publication. 

Author Response

Here is the response, and a word file is in the attachment.

Dear reviewer:

Thank you for your comments concerning our manuscript entitled “Music Recommendation Based on “User – Points - Music” Cas-cade Model and Time Attenuation Analysis” (ID: electronics-1894876). Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in red in the paper. The corrections in the paper are as flowing:

Concern # 1: There are certain typos, grammatical mistakes in the manuscript. The authors need to proofread the paper for any possible mistake.

Author response: This concern has been adopted.

Author action: We are very sorry for our negligence of typos, grammatical mistakes in the manuscript, we have checked and revised the article word by word, and the errors after correction are as follows:

(1) “PrefixSpan”  ->  “Prefix-Projected Pattern Growth”at the chapter of “Related Work”

(2) “cbow”  ->  “CBOW” at the chapter of 3.2.2. The Structure of Word2Vector.

(3) “number” -> “size” at the chapter of 4.1 Metrics.

(4) “LastFM” -> “Last.fm” at the chapter of 1 Introduction.

(5) “comparision” -> “comparison” at the chapter of 4.3.Settings

(6) “We design an interest attenuation algorithm considering the uneven distribution of user interest points is”->”We propose a music interest attenuation algorithm considering the uneven distribution of music recommendation system”

Concern # 2: The abstract needs to be more informative. As in the current abstract the size of music database, contribution, etc., are unclear. Also, it is not clear if the music is language agnostic.

Author response: This concern has been adopted.

Author action: We have re-written the abstract. In addition, We have highlighted the data is “after desensitization and encoding”, so the user will be clear that the music is language agnostic.

Concern # 3: It is not clear what does “MITA” stands for?

Author response: This concern has been adopted.

Author action: It’s a mistake, and we have modified it in the aticle.The model name ‘MITA’shoud be ‘MCTA, and ‘MCTA’is the abbreviation of our article title:  

Music Recommendation Based on “User – Points - Music” Cas-cade Model and Time Attenuation Analysis

Concern # 4: Figure 1. Is poor quality. Kindly, provide high resolution images as per the journal’s guidelines.

Author response: This concern has been adopted.

Author action: We have changed the fig 1 to a more resolution one.

Concern # 5: The authors have mentioned in the section 4.2 “W2V model is used to extract music features”. It would have broaden the scope of manuscript if the authors have provided the comparison with other feature extraction models (Fasttext, Glove, etc.)

Author response: This concern has been adopted.

Author action: We have provided the comparison with other feature extraction models of Fasttext, Glove and W2V at the chapter of 4.3. “Settings” as follows:

Besides W2V model, Fasttext and Glove are mostly used to extract the word vec-tors. To verify the effectiveness of W2V model in the system, a comparison experiment is carried out to make recommendation for users in the same dataset, using these three models separately. The experiment result is as follows:

Table 2. F₁-value of different models to extract the word vectors.

Through comparison, it’s found that the recommendation performance based on W2V model is obviously better than Fasttext and Glove. Therefore, the system adapts the W2V model to get the implicit feature of music.

Concern # 6: The results need better representation from contribution point of view.

Author response: This concern has been adopted.

Author action: We have explained the experiment results from  contribution point of view at the chapter of 4.4. “Result”:

MCTA remodel the traditional "User - Music" recommendation model into "User - Points of Interest - Music". Compared with the recommendation of a single music, the recommendation based on the points of interest composed of multiple music will be more accurate. In addition, users' interests will change over time. The interest point time decay model proposed in this paper can more accurately predict users' current preferences.

The clustering model proposed in this paper can obtain users' interest points composed of niche music, and according to this interest point, combined with the cascade model of " User - Points of Interest - Music ", it can recommend other niche music of this type for users, improving coverage of the recommendation system.

We have tried our best to improve the manuscript and made some changes in the manuscript.  These changes will not influence the content and framework of the paper.

We appreciate for your warm work earnestly, and hope that the correction will meet with approval.

Once again, thank you very much for your comments and suggestions.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors have covered all my previous comments. We suggest paper acceptance.