Clustering Algorithm with a Greedy Agglomerative Heuristic and Special Distance Measures
Round 1
Reviewer 1 Report
The paper is interesting. However, a few aspects should be improved before publication.
Algorithm 3 is not mentioned nor commented in the text. Some comments would be useful also to understand some of the symbology used (e.g. i%(SN+1)=0.
The same for Algorithm 4.
Please clarify the meaning of “parameter” at lines 298-299 and how they contribute to the total amount of 446 devices.
The dissertation about the numerical analyses should be extended with the purpose to explain the authors’ position on the results obtained. In particular, since the performance of the Greedy method seems to be strongly correlated to the characteristics of the problem under study and the way in which the neural network is initialized, the authors should draw some conclusive remarks about the field of application of the approach and its advantages compared to other solutions.
Author Response
Dear Reviewer,
Thank you very much for your interest to our paper and your review.
We tried to follow all your recommendations. Namely:
Algorithm 3 is not mentioned nor commented in the text. Some comments would be useful also to understand some of the symbology used (e.g. i%(SN+1)=0. The same for Algorithm 4.
Answer: We added the information about learning rate calculation and referred to Algorithms 3 and 4 in the text of the article
Please clarify the meaning of “parameter” at lines 298-299 and how they contribute to the total amount of 446 devices.
Answer: We added the following information “Four-batch mixed lot contains 62 parameters (features), three-batch mixed lot and two-batch mixed lot contain 41 parameters. The difficulty of the sample is that the number of parameters in it is large enough relative to the number of sample elements”
The dissertation about the numerical analyses should be extended with the purpose to explain the authors’ position on the results obtained. In particular, since the performance of the Greedy method seems to be strongly correlated to the characteristics of the problem under study and the way in which the neural network is initialized, the authors should draw some conclusive remarks about the field of application of the approach and its advantages compared to other solutions.
Answer: We have added information about clustering results and have discussed advantages and disadvantages of the proposed algorithms
Reviewer 2 Report
The authors discuss the application of Kohonen network in the clustering problem and propose an original modification of the agglomerative algorithm. In addition to introducing the self-organizing principle of Kohonen network, specifying different ways to express the distance of two points in space (Minkowski, Euclidean, Manhattan, Chebyshev, Mahalanobis), the proposed algorithm is tested on a dataset consisting of four different homogeneous batches of electronic radio components with tens of parameters, for different combinations of batches and metrics, resulting in a large (almost uncluttered) set of results in tabular form with calculated values of the parameters minimum, maximum, mean, standard deviation, coefficient of variation and the span factor of the objective function.
In contrast to the many quantitative results, the conclusion "Computational experiments showed that the use of the greedy agglomerative heuristic in online mode SLIGHTLY IMPROVES the accuracy of homogeneous batch separation" is not specific enough and should be more detailed.
Algorithms should at least tentatively mention time complexity, implementation issues (e.g. was MATLAB used for calculations?).
The text needs to be better edited.
Typography - page 3: "i-th variable" - $i$-th variable, similarly parameter p, (pages 3-4) should be italicized in the text; page 4: brackets and numbers (here number 1) should not be italicized; vectors should be bold;
page 5: the quality of figure 1 is low, it is a bitmap scan;
misprint in Algorithm 2: instead of "heighbors" "neighbors" should be; in Fig. 2: "Euclidian" - "Euclidean".
Author Response
Dear Reviewer,
Thank you very much for your interest to our paper and your comments.
We tried to improve our paper in accordance with them. Namely:
In contrast to the many quantitative results, the conclusion "Computational experiments showed that the use of the greedy agglomerative heuristic in online mode SLIGHTLY IMPROVES the accuracy of homogeneous batch separation" is not specific enough and should be more detailed.
Answer: We have completed the study with information about statistical significance of the results and have reformulated the conclusion
Algorithms should at least tentatively mention time complexity, implementation issues (e.g. was MATLAB used for calculations?).
Answer: We have added algorithms implementation details: “Algorithms were implemented in Java. For the computational experiments, we used the following test system: AMD Ryzen 5-1600 6C/12T 3200MHz CPU, 16 CB RAM. Each experiment took an average of 1 minute of computer time”
The text needs to be better edited.
Typography - page 3: "i-th variable" - $i$-th variable, similarly parameter p, (pages 3-4) should be italicized in the text; page 4: brackets and numbers (here number 1) should not be italicized; vectors should be bold;
Answer: fixed
page 5: the quality of figure 1 is low, it is a bitmap scan;
Answer: we agree with this comment and have removed this figure from the paper
misprint in Algorithm 2: instead of "heighbors" "neighbors" should be;
Answer: fixed
in Fig. 2: "Euclidian" - "Euclidean".
Answer: fixed
Reviewer 3 Report
This paper proposed algorithms for products clustering based on Kohonen network and Self-Organizing Kohonen Maps using greedy agglomerative heuristic procedure in online and batch mode
The authors may need to take into consideration the following issues:
- References only have 6 references have 3-years (since 2019) publications. Thus, I suggest the author may add more description about the state-of-the-art techniques.
- The abstract section is a little bit short. I suggest the authors should add some paragraph.
- I would like to see the Wilcoxon rank sum test of proposed algorithms of Tables 1 to 8. In addition, the comparison models should cited in these Table.
- I suggest the algorithm 2, 3 and Table 5 did not across the two pages.
- The authors should provide more SOTA models in this article for Table 6 to 8.
Author Response
Dear Reviewer,
Thank you very much for your interest to our paper and your review.
We tried to take into account all your notes. Namely:
1. References only have 6 references have 3-years (since 2019) publications. Thus, I suggest the author may add more description about the state-of-the-art techniques.
Answer: We supplemented the Introduction section with some amount of recent sources
2. The abstract section is a little bit short. I suggest the authors should add some paragraph.
Answer: We expanded the abstract a little more
3. I would like to see the Wilcoxon rank sum test of proposed algorithms of Tables 1 to 8. In addition, the comparison models should cited in these Table.
Answer: We supplemented Tables with information about statistical significance of the results
4. I suggest the algorithm 2, 3 and Table 5 did not across the two pages.
Answer: fixed
5. The authors should provide more SOTA models in this article for Table 6 to 8.
Answer: We performed additional experiments for k-means model and included the results in Tables 6-8
Round 2
Reviewer 1 Report
All the reviewer's comments have been met by authors in the revised paper.
