Hypernetwork Link Prediction Method Based on Fusion of Topology and Attribute Features

Round 1

Reviewer 1 Report

Summary

The authors propose an approach for link prediction in hypernetworks with attributes. Both structural information and attribute information is taken into account and combined in a model using attention. The approach is evaluated on common benchmarking datasets and an ablation study is performed.

Pros

- Model outperforms baselines

- Interesting model using both structural information and attribute information.

- Nice ablation study investigating the effects of different parts of the models

Cons

- The related work and baselines appear to be outdated. The best baseline model HyperGCN is from 2018, NHP from 2019. However, there are newer works, for example:

  - Zhang, R., Y. Zou, and J. Ma. "Hyper-SAGNN: a self-attention based graph neural network for hypergraphs." International Conference on Learning Representations (ICLR). 2020.

  - Bai, Song, Feihu Zhang, and Philip HS Torr. "Hypergraph convolution and hypergraph attention." Pattern Recognition 110 (2021): 107637.

- How was the hyperparameter optimization done? It seems that there was no separate validation set which might lead to overfitting on the test set.

Details

- Clearly define hypernetworks in the introduction. Doing this in the related work section is a bit late: "Hypernetworks are the generalization of networks in which hyperedges can represent the relationship between multiple nodes."

- What exactly is the difference between a hyperlink and a hyperedge. This should be stated more clearly.

- The matrix D_v is not defined.

- Table 3: reaches into the right margin of the page. Reduce the space between columns or choose shorter column headings?

- The conclusion only summarizes the article but does not discuss the impact of this work.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Please see the attachment.

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

comments:

* presentation needs to be improved. The figures are in low resolution, some text is not readable.

* Table 3, right part is cut out of the page. 

* Figure 2, why split into two figures?

* reproducibility is low: the authors did not provide how to reproduce the result 

* what is the threshold for recall? it is important but not reported

* many minor things (definition 4 repeated, equation 14 |e|, etc.)

* Equation 4, what is || "splicing operation"? Please use a more common term or explain it in detail.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Weaknesses- 

 

Major weaknesses : 

** Include a figure to represent the architecture of the hypergraph neural network developed.

  

A graphical representation of how hyperedge-level attention mechanisms are performed with an example of  joint embedding using one of the real world 6 datasets is required for better understanding of the mechanism.

 

Line 18: “Complex networks provide a unique research idea for depicting various relationships 18 in the real world.” Instead of this definition, give a more meaningful introduction to what a complex network is.

 

Line 23: Give proper references to “coauthorship network” and “metabolic network”.

 

Line 24: “Obviously, in this case,” should be as “Obviously, in these cases,”, which refers to the practical problems included in the previous sentence.

 

Line 24: what does a “single network” refer to? Is it a simple network which is the counterpart of a complex network?

 

Line 30 :

Paraphrase this sentence to convey the meaning well.

“In recent years, because of the efficiency and sufficiency of its modeling, hypernetworks have received extensive attention from scholars.”

Ex: “Scholars have paid extensive attention to hypernetworks in recent years owing to the efficiency and sufficiency  of its modeling.”

 

Line 55: What is “paper coauthorship network”? Is it a paper ? Give proper citations.

 

Line 58: Grammatical error-binary relationship should be as binary relationships.

 

Line 61: Give proper citations to metabolic networks.

 

Line 139: what is ​​aggregatiotab3n?

 

Line 294, Line 298, Line 302, Line 307 - Give proper citations for these datasets.

 

Line 357 - Is Recall@k in line 357 and R@k in equation 18 the same or different?

 

Line 345, “Because the above methods does not ” better replace with “Since the above methods do not”

 

Line 545 - Ambuiguity in reference. What does “ref8” means?

 

Minor weaknesses

 

Line 354- Recall, R to be simple letter

 

Line 355 - Should be as “Here, we rank the”, missing comma

 

Line 411 - “But compared with…” comma should be after “But,”

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 3

Reviewer 2 Report

Authors have addressed the previous issues in the review processes.