Beyond Redundancy: Information-aware Unsupervised Multiplex Graph Structure Learning

Summary

- This paper presents an information theory approach to obtain a single graph fused from a multiplex graph, which preserves - sufficient task-relevant information - while removing task-irrelevant noise. - A learnable graph augmentation strategy is also developed. - The learned graph and representation can be applied to different types of tasks. - The effectiveness is supported by extensive experimental results.

Strengths

- This paper is well-motivated. - The authors find that each graph contains much unique task-relevant information, which is ignored by mainstream contrastive learning-based methods. - This paper develops multiple graphs non-redundancy principle, which lays the foundation for multiplex graph data process. - Two random and generative graph augmentation strategies are accordingly built to capture view-unique task information. - The experimental results are promising. - The framework demonstrates a clear advantage over existing methods, including advanced supervised approaches, highlighting its potential for broad application. - This paper provides the code and all experimental settings for reproducing the results.

Weaknesses

- The difference between the existing non-redundancy principle and multiplex graph non-redundancy is unclear. Please clarify it. - The proposed InfoMGF-LA runs out-of-memory on MAG data. The reason should be given. - It is possible that the proposed method cannot handle real-world large-scale graph. It should be addressed in the future and discussed in the conclusion part. - The difference between the proposed method and DGM is unclear.

Questions

I list them in **Weaknesses**.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Some imitations are addressed in $\S5$.

Summary

The paper introduces InfoMGF (Information-aware Unsupervised Multiplex Graph Fusion), a novel framework aimed at addressing the issue of graph structure reliability in Multiplex Graphs. The primary goal is to refine graph structures to eliminate noise and maximize task-relevant information. Theoretical analysis and comprehensive experimental results validate its effectiveness.

Strengths

1. Originality: The paper addresses a critical gap in Unsupervised Multiplex Graph Learning (UMGL) by focusing on the reliability of graph structures, which is often overlooked in existing research. 2. Quality: The proposed InfoMGF framework effectively refines graph structures to eliminate noise and maximizes both view-shared and view-unique task-relevant information. Theoretical analyses provided in the paper validate the effectiveness of InfoMGF in capturing task-relevant information and improving graph fusion. Extensive experiments demonstrate that InfoMGF outperforms various baselines and even sophisticated supervised approaches in different downstream tasks. 3. Clarity: The paper is generally clearly written and well organized.

Weaknesses

1. Scalability: The framework involves several steps. Though the paper provides the complexity analysis in Appendix for each step, it is still unclear what is the overall complexity. 2. Reproducibility: The authors share the code for reproducibility. However, I didn’t see the datasets. 3. Accuracy: The authors should check for the few grammatical and spelling errors that occur in the text.

Questions

As above.

Rating

7

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes.

Summary

The paper introduces InfoMGF, an innovative framework for Unsupervised Multiplex Graph Learning (UMGL) that addresses the often-overlooked issue of graph structure reliability. InfoMGF refines graph structures by removing task-irrelevant noise and maximizing task-relevant information through mutual information maximization. Extensive experiments demonstrate its superior performance over various baselines and even some supervised methods, validating its effectiveness in enhancing node representation learning.

Strengths

- New Problem Formulation: The paper pioneers the investigation of graph structure reliability in multiplex graphs, which is a significant advancement in the field. Multiplex graphs enrich the representation of real-world systems and its analysis is very difficult inherently. - Theoretical Analysis: The several theorems are quite interesting and provide a solid foundation for the proposed method. In particular, Theorem 3 proves the necessity of fusing multiplex graphs. - Extensive Evaluation: The framework is thoroughly tested against various state-of-the-art methods on both node clustering and classification tasks, showcasing its robustness and effectiveness across different tasks. The comparison methods are representative and new.

Weaknesses

- Robustness: Fig.4 shows that the proposed method is very robust to structure noise. However, more analysis is needed. Both InfoMGF and SUBLIME are structure learning methods. Compared to InfoMGF，Why does the performance of SUBLIME degrade rapidly in the case of edge deletions? - Clarity: The paper develops two algorithms in this paper: InfoMGF-RA and InfoMGF-LA. However, it is a little confusion that what is the difference in their objective functions.

Questions

1. There are some small errors in Algorithm 1. In particular, the title of it is InfoMGF-LA, however, line 11 also includes the operation for InfoMGF-RA. 2. The proposed method depends on the assumption of optimal augmentation. How to guarantee that the used feature and structure augmentations are optimal? It is still unclear to me. 3. The authors discuss the robustness against structure noise. How about feature noise? Could you share your intuition on this matter?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The paper discusses the limitations.

Summary

The authors develop a novel approach to improve Unsupervised Multiplex Graph Learning by refining graph structures to eliminate noise and maximize relevant information. The method utilizes mutual information maximization to integrate multiple graph views effectively. Theoretical validation and comprehensive experiments show that the proposed method outperforms existing methods.

Strengths

1. Multiplex graph provides an efficient representation of complex systems. This paper focuses on non-redundancy issue, which is a new perspective and opens up a new avenue for future research. 2. The proposed method adopts an unsupervised and generalized approach. Its performance surpasses several supervised approaches, underscoring its potential for practical applications. 3. The framework’s performance is validated through comprehensive experiments and compared with more than 20 methods. 4. Visualization is also a strong point of this paper. The figures of node correlation, heatmaps of the subgraph, and unique relevant edge ratio are very illustrative.

Weaknesses

1. According to Table 1 and 2, it seems that the proposed method improves more on clustering than classification. 2. Overall, this paper is well-organized. However, the writing could be improved in terms of tone and words. 3. There are too many notations, which are confusing.

Questions

1. Is there any explanation about why the method performs better on clustering than classification? 2. How to solve the above issue? 3. The font of k in the caption of fig.5a is not correct. 4. In the Appendix, the authors proof Proposition 1, however, there is no corresponding one in the main paper.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

NA

Reminder to Check Authors' Reply

Dear Reviewers, Would you mind checking the authors' responses? - AC

Thanks for the raising

We are grateful to Reviewer gjsJ for suggesting this improvement! If you have any additional questions or concerns, please feel free to discuss them with us.