Guided Semi-Supervised Non-Negative Matrix Factorization

Round 1

Reviewer 1 Report

This paper proposes a guided semi-supervised non-negative matrix factorization method that performs classification and topic modeling by incorporating supervision from both pre-assigned document class labels and user-designed seed words. This paper is well motivated and structured. The formulation is clear and reasonable. But, I still have the following concerns. 

1, The proposed approach is only evaluated on one dataset, more datasets are suggested to demonstrate the performance of the proposed approach. 

2, There are two hyper-parameters lambda and mu in the proposed approach, their influence on the proposed model should be investigated.

3, Some recently published related works should be cited and discussed. For example:
[1] Positive and Negative Label-Driven Nonnegative Matrix Factorization, IEEE TCSVT 2021
[2] Semisupervised Adaptive Symmetric Nonnegative Matrix Factorization, IEEE Transactions on Cybernetics, 2021
[3] Semi-Supervised Non-Negative Matrix Factorization with Dissimilarity and Similarity Regularization, IEEE TNNLS 2020
[4] Simultaneous Dimensionality Reduction and Classification via Dual Embedding Regularized Nonnegative Matrix Factorization, IEEE TIP, 2018

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors present a new semi-supervised strategy for NMF that incorporates information from possible seed configurations of topics in the context of topic modeling. The paper is clear and the work is well-motivated and explained. The results are interesting, but I believe there are few points to be addressed before publication:

1) Related work: The topic supervised NMF method from "https://arxiv.org/pdf/1706.05084.pdf" seems closely related to this work and should be discussed in the context of the Semi-supervised and Guided NMF frameworks presented.

2) Applications: A more in-depth comparison of methods are needed. As it stands, only two forms of NMF are compared on the data; however, there should be a comparison (at least across summary metrics like coverage and average F1 score) for standard NMF, the topic-supervised NMF mentioned in (1), and potentially (though perhaps not necessary) against semi-supervised Latent Dirichlet Allocation methods. 

3) Applications: I get the point of Figure 3, but it does not clearly present the differences between the two methods and the advantage of GSSNMF. I would consider visualizing a heat map of differences between the real and recovered labels and give a summary metric associated with each. Also, if additional methods are included in the analysis, I would add them to this figure as well.

4) Theory: I think this is straightforward and perhaps obvious from the derivation of the algorithm but it would be worthwhile to mention the monotonicity of the GSSNMF algorithm. 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Kindly see the attached file for comments.

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

My previous comments have all been well solved. I suggest to accept this paper. 

Author Response

Dear Reviewer,

We appreciate your time and previous pieces of advice in improving our paper.

Thank you so much.

Best,
Pengyu

Reviewer 3 Report

All comments have been resolve. Add the description of Algorithm in the final version as it is necessary to understand its working.

Author Response

Please see the attachment

Author Response File: Author Response.pdf