Hierarchical Classification for Large-Scale Learning

Round 1

Reviewer 1 Report

Authors propose a framework called Hierarchical PPCA for classification with a large number of classes. The method uses a self-supervised pretrained feature extractor to obtain meaningful features and trains Probabilistic PCA models on the extracted features for each class separately.

(1)   The theoretical background of the proposed method is adequately detailed in the paper.

(2)   The abstract should be improved. Your point is your own work that should be further highlighted.

(3)   The parameters in expressions are given and explained.

(4)   More statistical methods are recommended to analyze the experimental results.

    (5) I would suggest adding some recent literature in the manuscript.  For example, https://doi.org/10.3390/rs15133402; http://dx.doi.org/10.1109/TCSS.2022.3152091; https://doi.org/10.1109/JIOT.2023.3296460 and so on.

      (6) In 5. Experiments, how to determine parameter values for used methods?

Correct typological mistakes and mathematical errors.

Author Response

(1)   The theoretical background of the proposed method is adequately detailed in the paper.

(2)   The abstract should be improved. Your point is your own work that should be further highlighted.

Reply: We rewrote the abstract to better emphasize the advantages of our proposed approach.

(3)   The parameters in expressions are given and explained.

(4)   More statistical methods are recommended to analyze the experimental results.

Reply: We updated our code and rerun all experiments and updated the results, adding standard deviations over 5 runs. The statistical significance of the differences can be now evaluated using two sample t-tests.

(5) I would suggest adding some recent literature in the manuscript.  For example, https://doi.org/10.3390/rs15133402; http://dx.doi.org/10.1109/TCSS.2022.3152091; https://doi.org/10.1109/JIOT.2023.3296460 and so on.

Reply: We added the first two references because they use PCA for different applications. However, the third reference does not use PCA and is not hierarchical, so we could not figure out how it fits into the related work for this paper. We found https://doi.org/10.3390/electronics12153284 instead as a more relevant recent paper.

(6) In 5. Experiments, how to determine parameter values for used methods?

Reply: Section 5.6 presents an analysis of the influence of main parameters: number S of super-classes, number of principal components (q and r), and the parameter T deciding how many top super-classes are used for final classification. Based on those analyses, we chose the parameters that give the best results for all methods.

Comments on the Quality of English Language

Correct typological mistakes and mathematical errors.

Reply: We proofread the paper again and corrected some typos.

Reviewer 2 Report

This paper has interesting experimental result which has been shown on the detail explanation. However there is minor comment to improve the quality of the paper.

1. In the line 4-5, this statement does not represent the goal of this paper. It is too hyperbolic.

2. This paper used the pre-trained model of CNN. Then, please compare the performance with original CNN model with proposed method.

1. Please check the ?? symbol on the pdf output from latex.

Author Response

1. In the line 4-5, this statement does not represent the goal of this paper. It is too hyperbolic.

Reply: We replaced “millions of classes” with “tens of thousands of classes”, which is within the goal of the paper.

2. This paper used the pre-trained model of CNN. Then, please compare the performance with original CNN model with proposed method.

Reply: The original CLIP CNN is not a classification model. It is a feature extractor trained together with a NLP feature extractor to produce similar features for millions of image-text pairs. It can be used for image classification by training a linear projection head (fully connected layer). We trained such a FC layer for each dataset and added it to our evaluation.

Comments on the Quality of English Language

1. Please check the ?? symbol on the pdf output from latex.

Reply: We fixed the ?? symbol, which was a missing reference.

Reviewer 3 Report

I read the article and have a couple of suggestions for authors:

1. I suggest the authors to include in the introduction the aim of the study, the research questions. 

2. The literature review section should be enriched with more studies related to the research topic. \

3. The methodology is not very clearly defined, it is very difficult for the reader to follow what the authors want to explore in the article. 

4. Sections 3 and 4 present the theoretical part of some algorithms, but it would be more interesting if the authors would customize these algorithms on the data they have. I suggest shortening these sections.

5. What does RBF mean?

6. What are the limitations of the study? 

7. I suggest that the conclusions section states whether or not the research questions set out in the introduction have been answered.

Author Response

1. I suggest the authors to include in the introduction the aim of the study, the research questions. 

Reply: We added a paragraph in the introduction summarizing the main research question of the study.

2. The literature review section should be enriched with more studies related to the research topic.

Reply: We added to the literature review section more recent papers related to our work.

3. The methodology is not very clearly defined, it is very difficult for the reader to follow what the authors want to explore in the article. 

Reply: We tried to improve the methodology description by adding a paragraph to the introduction that describes the organization of the paper. We also rewrote parts of the methodology to try to make them easier to understand.

4. Sections 3 and 4 present the theoretical part of some algorithms, but it would be more interesting if the authors would customize these algorithms on the data they have. I suggest shortening these sections.

Reply: We shortened these sections by removing the log determinant from Theorem 1, the definition of the Bhattacharyya distance from Section 4.5 and the definition of the KL divergence from Section 4.6. We could move the proof of Theorem 2 to an appendix, but it seems the journal does not allow appendices, only Supplementary Materials.

5. What does RBF mean?

Reply: We clarified that RBF means radial basis function neurons and added a citation for them.

6. What are the limitations of the study? 

Reply: We added a paragraph in the conclusion summarizing the limitations of the study.

7. I suggest that the conclusions section states whether or not the research questions set out in the introduction have been answered.

Reply: We added a paragraph in the conclusion stating that the research question has been answered positively.

Round 2

Reviewer 1 Report

I have appreciated the deep revision of the contents and the present form of this manuscript. All my previous concerns have been accurately addressed. I think that this paper can be accepted.

Minor editing of English language required

Reviewer 3 Report

From my point of view the paper can be accepted for publication in the present form.