Heterogeneous Graph Purification Network: Purifying Noisy Heterogeneity without Metapaths

Round 1

Reviewer 1 Report

Referee Report

 

General comment:

This paper proposes a novel Heterogeneous Graph Purification Network (HGPN)  framework for heterogeneous graph representation learning which aims to solve such dilemma by effectively purifying the noisy heterogeneity. Specifically, instead of relying on artificial metapaths, HPGN models heterogeneity by subgraph decomposition and adopts inter-subgraph and intra-subgraph aggregation methods. HPGN can learn to purify noisy edges based on semantic information with parallel heterogeneous structure purification mechanism. Besides, the authors design a neighborhood related dynamic residual update method, a type specific normalization module and cluster-aware loss to help all types of node get high quality representations and maintain feature distribution while preventing feature over-mixing problem. Extensive experiments are conducted on four common heterogeneous graph datasets and results show that the proposed approach and outperforms all existing methods and achieves state-of-the-art performances consistently among all the datasets.

 

Major comment:

Although the problem studied is somewhat novel and interesting, I do not believe the paper is ready to be published yet. Contribution of this study is not convincing and overall structure of this manuscript is not good enough. I have detected major logical and technical errors in the following concerns:

 

(1)   Please check and correct the error of the author's names in use of capital and small letter.

 

(2)   Though the incremental contribution of the paper as compared to the related literature is mentioned, the introduction and literature reviews referred do not provide a detailed analysis to the research status and do not emphasize the background and motivation.

 

(3)   Please check the use of abbreviations such as “Heterogeneous Graph Purification Network (HGPN)”. Usually they occur only in the first mentioned case but should not be repeated once again and again.

 

(4)   In Section 4，the authors should pay more attention to the comparisons for the superiority/novelty of your proposed model. It is not clearly stated which works have addressed this problem, which of them is the closest to your paper and what is the gap between the works and the current.

 

(5)   In Section 5，it is suggested that the authors should inserted some pictures to show your experimental results so that the results are more easily understood and visualized.

 

(6)   The conclusion and outlook part should be appropriately extended.

 

(7)   Please ensure that your reference are consistent in a unified style.

(8)   The paper should be proofread, since there are some typos and errors in the text. For instsnces, “In Proceedings of the Proceedings of the 20th ACM SIGKDD…” should be cut out the superfluous words or phrases in reference [15] on Page 17.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The abstract needs to be revised considering the context, challenge, aim, method, results and implications.

Research questions should be discussed.

Motivation behind the work should be discussed.

Comparison with the previous works might be included.

More details and highlights might be reviewed.

Research gaps and potential future directions should be included.

Review the conclusion considering main point and claim points of this review.

Author Response

Please see the attachment.

Author Response File: Author Response.docx