Enhanced Multi-Level Recommender System Using Turnover-Based Weighting for Predicting Regional Preferences
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe paper proposes a collaborative recommendation model based on weights derived from the turnover of participating branches, with application in distributed multi-sector recommendation systems. The idea is to use the weight of each data source as a criterion to generate more relevant recommendations, particularly when regional variations exist. The authors apply the approach to the MovieLens Dataset and compare their model with a traditional approach.
The presented proposal is relevant and aligns with the scope of the Computers journal, addressing aspects of big data, recommendation systems, and distributed modeling. However, the article has limitations that need to be addressed before it can be accepted.
Recommendations for acceptance
A direct comparison with existing approaches in the literature directly best demonstrates the originality of the proposal.
The writing requires significant revision. There are grammatical errors, incorrect use of technical terms, and confusing sentence structure in several passages. Below are some examples:
Informal use of "Though" at the beginning of the scientific sentence.
“data warehouse based” should be “data warehouse-based” (with a hyphen).
In: "...computed using equation 3 as the item(104,101) is 0.453.", the term "item(104,101)" is unclear — does this refer to the similarity between items 104 and 101? The writing is confusing.
In: “as the item is 0.453” it seems to say that the item has a value of 0.453, when in fact it should be the similarity between two items.
There is an excess of jargon and confusing sentence structure; a general revision is needed.
The “Research Contribution” section presents good intentions but lacks clarity in explaining how each contribution is implemented in practice. It is recommended to better divide the text between motivation, hypothesis, and technical execution.
The criterion for defining “branches” in MovieLens (user clusters?) is not justified.
Improve Figure 2 to enhance readability. My recommendation is not to use colors; instead, follow the approach taken in Figure 1.
Tables 3 and 4 should be reduced so that the page number does not overlap the data.
The discussion is very focused on describing the results, without delving into why the model performs better or under what conditions it is justified, which is necessary for a publication like this.
Author Response
Dear Reviewer,
We changed the paper in line with your comments. We appreciate the time invested and your comments, both of which helped improve our paper.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsSummary:
The manuscript proposes a turn-over-based weighting model that augments classical item-based collaborative filtering with a branch-level weight proportional to each branch’s sales revenue, and an “effective voting threshold’’ that prunes rarely-rated items. Three algorithms are presented for computing branch weights, voting rates, and multi-level predicted ratings (global / sub-global / local)
. The approach is evaluated on the MovieLens-100K benchmark after artificially splitting it into three “branches’’ with synthetic revenue figures . Reported results show higher weighted predicted ratings than a vanilla item-based CF baseline.
Strengths:
1. Pseudocode for all three core procedures is provided, so practitioners could re-implement the method.
2. Well-structured writing.
Weaknesses:
1. The authors do not test their method on an actual multi-branch retail dataset. Instead, they slice the public MovieLens-100K movie ratings into three artificial “branches” and then make up turnover figures of $100 k, $50 k, and $20 k for those partitions. Because the underlying items, inventories, and user populations are identical, this setup cannot reproduce real-world problems such as regional catalogue differences or branch-specific popularity skews, so the reported gains may not hold in practice.
2. The paper never reports standard accuracy measures like RMSE, MAE, precision@K, or recall@K. Instead, it shows only raw “weighted predicted rating” values and bar charts, which do not reveal whether the recommendations customers receive are actually better. This problem is amplified by the fact that the only baseline is a very simple item-based collaborative-filtering model; well-known stronger methods such as matrix factorisation, neural collaborative filtering, or graph-based recommenders are left out, so the true performance gap remains unknown.
3. The two key hyper-parameters, i.e., the linear revenue weight and the minimum voting threshold, are introduced heuristically. The authors multiply rating similarities by revenue in a straight line without testing whether a different curve (e.g., logarithmic) would work better, and they pick the cut-off “Min.Ds = 0.20” purely by expert opinion without any tuning study. Because only one train-test split is run, there are no confidence intervals or significance tests to show that the observed improvements are statistically reliable.
Minor issues:
1. Tables 2–4 that illustrate user ratings for the three sites still contain “?” placeholders, making the raw data unreadable and impossible to check.
2. Table 5 and Table 7 headings overflow into the running text, and some numeric columns mis-align under their labels, especially where multi-line item lists are wrapped .
Author Response
Dear Reviewer,
We changed the paper in line with your comments. We appreciate the time invested and your comments, both of which helped improve our paper.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsLack of Clarification in Figure 2 and figure refinement needed:
- Figure 2: The schematic representation of the proposed weighting model requires improved clarity, particularly in the labeling. We suggest that to enhance the textual components of the figure (e.g., analytical model text is not clear, please enhance the figure) and provide a more thorough, detailed explanation of the data flow and functional responsibilities at each layer (e.g., big data layer, analytical layer, pattern detection layer/discovery). Please add key concepts (data flow and explanation of Figure 2 with step by step).
- No Discussion on limitations and future work suggestions needed
Suggest the potential limitations related to data distribution and turnover as weighing factors. In addition, please provide a brief suggestion on future research directions, such as (ex: dynamic weighing schemes, Integration technique with advanced DL models, adaptive weighing scheme, etc.)
- Expansion of literature review
Please add a brief description related to multi-level/ distributed recommendations-based system, such GNN (Graph neural network) and others, federated learning, and more light on weighing models (turnover-based), and current existing software frameworks.
Author Response
Dear reviewer,
We changed the paper in line with your comments. We appreciate the time invested and your comments, both of which helped improve our paper.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe revised version of the manuscript shows substantial improvements compared to the original submission. The authors have addressed several of the initial concerns effectively, both in terms of content depth and technical validation.
Comments on the Quality of English Language
The manuscript still requires language and style revision to meet the standards of a scientific publication in English. The following issues should be addressed:
- Grammatical inconsistencies and awkward phrasing
Several sentences remain structurally weak or overly convoluted. Examples include: - “The rating of the active user in each branch, who has purchased and rated the maximum number of items…”
Suggested: “…the user who has rated the highest number of items in the branch…” - “has been designated as a higher turnover branch…”
Consider simplifying: “…was designated as the high-turnover branch…” - Redundant or repetitive expressions
Phrases such as “proposed weighting model,” “turn-over based rating,” and “collaborative filtering technique” are used repetitively throughout the text, often multiple times in the same paragraph. Please revise to avoid redundancy and improve readability. - Overuse of generic, overly formal expressions
For instance, expressions like “in the realm of,” “it is imperative that,” or “to ascertain the interestingness” can be simplified or rephrased in more natural academic English. They give the text a formulaic and artificial tone. - Clarification of some examples
While the use of the MovieLens dataset is acceptable for benchmarking, the analogy to multi-branch business data remains conceptually weak. It would be helpful to briefly acknowledge this limitation and justify its use more explicitly in the discussion.
