Simple and Asymmetric Graph Contrastive Learning without Augmentations

Graph Contrastive Learning (GCL) has shown superior performance in representation learning in graph-structured data. Despite their success, most existing GCL methods rely on prefabricated graph augmentation and homophily assumptions. Thus, they fail to generalize well to heterophilic graphs where connected nodes may have different class labels and dissimilar features. In this paper, we study the problem of conducting contrastive learning on homophilic and heterophilic graphs. We find that we can achieve promising performance simply by considering an asymmetric view of the neighboring nodes. The resulting simple algorithm, Asymmetric Contrastive Learning for Graphs (GraphACL), is easy to implement and does not rely on graph augmentations and homophily assumptions. We provide theoretical and empirical evidence that GraphACL can capture one-hop local neighborhood information and two-hop monophily similarity, which are both important for modeling heterophilic graphs. Experimental results show that the simple GraphACL significantly outperforms state-of-the-art graph contrastive learning and self-supervised learning methods on homophilic and heterophilic graphs. The code of GraphACL is available at https://github.com/tengxiao1/GraphACL.

Paper

References (63)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer gZUY6/10 · confidence 4/52023-06-15

Summary

This paper proposes an asymmetric contrastive learning framework for the homophilic and heterophilic graphs, which does not rely on graph augmentations and homophily assumptions. The theoretical analysis and empirical results further support the effectiveness of the proposed method.

Strengths

1. This paper is easy to understand and the motivation is clear and reasonable. 2. The proposed method is simple yet effective. The authors further demonstrate the effectiveness of the proposed method from the information theory and downstream tasks perspectives. Such results give a theoretical understanding of this work. 3. Extensive experimental results including comparison experiments, ablation study, visualization, and case study empirically verify the effectiveness of the proposed method. Moreover, the proposed method shows superior performance on both the homophilic and heterophilic graph datasets.

Weaknesses

1. This paper is somewhat over-claim and some important references are missed. In lines 45-47, the authors claim that this work makes the first attempt to design contrastive learning objectives by neither relying on explicit nor implicit homophily assumptions. Actually, there are several works already done. For example, [1] Wang H, Zhang J, Zhu Q, et al. Augmentation-free graph contrastive learning. arXiv preprint arXiv:2204.04874, 2022. 2. The proposed objective function seems like a variant of the GCL with representation smoothing diagram and the GCL with augmented views diagram. Actually, the difference between them and the proposed method is the definition of the positive pairs. 3. In this work, the authors claim that the proposed method does not rely on the homophily assumption and conduct a theoretical analysis to further support the proposed method. Are these theorems hold without any assumptions?

Questions

See above

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

None

Reviewer cxpZ6/10 · confidence 4/52023-07-02

Summary

This work first points out that existing GCL can fail to generalize to heterophilic graphs, then develops a new framework called GraphACL based on an encoder capturing one-hop neighbourhood context and two-hop monophyly. Experiments validate the effectiveness of the proposed method.

Strengths

The paper is well organized. Each part has a clear goal and supports the authors' claim well. The proposed method, Graph Asymmetric Contrastive Learning, is not only explained in detail but also justified by solid theoretical analysis. The authors adequately answered how and why it works. In the empirical study, the choice of baselines has good coverage. The selection of datasets respects the theme of each subsection, and the experiment setup is reasonable. The experiments are well explained, making it easy to reproduce them.

Weaknesses

There are no significant weaknesses in this paper. The application scenario of the proposed method may be a little bit limited, but this is completely acceptable.

Questions

Does GADC have the potential to be applied to graph tasks other than node classification shown in the paper?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors adequately addressed their work's limitations and potential societal impact.

Reviewer ycWA6/10 · confidence 3/52023-07-06

Summary

This paper presents GraphACL, which aims to tackle limitations of other graph contrastive learning works which have implicit or explicit homophily assumptions, and suffer in learning effective representations for heterophilic tasks. The approach is designed to leverage the principle of monophily, and the authors evaluate their work on several homophilic and heterophilic datasets, achieving at-par performance on the former and improved performance on the latter, on downstream node classification tasks.

Strengths

- The approach presented here is intuitive once the monophilic principle is understood, and the implementation seems straightforward. - Results suggest the approach is strongly effective (e.g. Table 1), and hte improvement on heterophilic datasets without any material performance loss / delta on homophilic datasets is compelling. - The paper is well-written, and the figures help clarify the principles motivating the design.

Weaknesses

- Lines 69-80 could greatly benefit from a toy example with nodes illustrated; it is confusing to understand the principles of monophily and the design of GraphACL without such an example early in the paper. - "we are faced with the challenge of simultaneously capturing the one-hop neighborhood context and monophily in the contrastive objective" (line 74) -- it's not clear why this is a particular challenge; is there something about capturing these two together that is difficult to reconcile in an objective? If so, I didn't understand this from the text. - Lines 40 and 97 rely quite a bit on the (in paper) reference [22] to reference issues with contrasting view-based GCL approaches on heterophilic graphs; it would be helpful to include some more formal claims or concrete examples in this work to make it more self-contained, since the work strongly seems to draw motivation from [22]'s findings. Without this, it is difficult to fully appreciate the supposed homophilic biases in those GCL approaches. - The section in Related Work from line 137 onwards seems to categorize BGRL as a contrastive method. BGRL is not conventionally contrastive due to its lack of utilizing negative examples. There are other such non-contrastive methods (several mentioned in [1]) which are used in graph SSL which should probably be mentioned in the related work in this paper. In fact, I believe the authors' intent to pursue a graph SSL approach which does not use augmentations but instead uses induced asymmetry is actually the objective of many "non-contrastive" methods. I would encourage they look into the literature and evaluate whether they indeed think their method is contrastive or not, given it's lack of need of negative samples. - The words "context" and "preference" are used quite often throughout the paper (I think sometimes to mean the same thing). It would be helpful to formalize these words and just use the same word frequently, as it can be confusing to disambiguate which representation is which. - Can the authors clarify if/why the loss in Eqn 4 is required? Methods like SimSiam, BYOL, BGRL etc. which use this kind of predictor structure and stop-gradient show they avoid collapse through the asymmetric predictor and the EMA procedure on target weight update. Does GraphACL not naturally avoid collapse for the same reasons? Does it require this uniformity regularization? - Figure 3 could go earlier to help introduce monophily and the intuition behind what existing GCL methods prioritize and what this method prioritizes. - There are a few very strong self-supervised GNN approaches missing from baselines, e.g. [2] or the approach mentioned in [1]. - There are opportunities to better evaluate the quality of self-supervised representations using link-level tasks (e.g. link prediction), or several of the tasks mentioned in [2], which may offer improved understanding beyond node classification and node clustering (referenced in Sec 6.1) - Evaluation on some larger datasets with greater data diversity and lower homogeneity would have been more compelling, e.g. ogbn-arxiv or ogbn-products. Flickr might also be a useful moderate-size heterophilic dataset to try out. - The node clustering results should be included in the main paper in my opinion -- the evaluation on just 1 downstream SSL task is fairly "light" and could use greater support to evidence that GraphACL is a generally strong SSL method, rather than "SSL for node classification" method. This is one of the biggest gripes in this paper -- the results demonstrating that GraphACL is a very effective SSL method are sparse, with only 1 demonstrated task in the main (evaluated) paper in terms of node classification. Other graph SSL work typically considers several downstream tasks and evaluates across them. [1] Link Prediction with Non-Contrastive Learning (Shiao et al, ICLR'23) [2] Multi-task Self-supervised Graph Neural Networks Enable Stronger Task Generalization (Ju et al, ICRL'23)

Questions

Please see "Weaknesses" for general comments / concerns which would be helpful during rebuttal.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

These are addressed in Appendix E.

Reviewer 1K5Q6/10 · confidence 4/52023-07-06

Summary

This paper propose a simple and effective contrastive learning framework named GraphACL for both homophilic and heterophilic graphs. In particular, GraphACL can capture both one-hop local neighborhood context and two-hop monophily similarties in one single objective. The authors theoretically analyze the learning process of the GraphACL and show that GraphACL can explicitly maximize the mutual information between representations and one-hop neighborhood patterns. The experiments and theoretical analysis show the effectiveness of the proposed approach.

Strengths

-This paper is well written and well structured. The authors first analyze the common limitations of existing GCL frameworks, and then propose some ideas and validate them on many datasets, and finally solve these limitations through the proposed methods and theorems, making this paper well understandable. -The experiments and theories are sufficient to demonstrate the effectiveness. -This paper is an early investigation of GCL on heterophilic graphs, which provides a new direction for the subsequent development of GCL.

Weaknesses

-It is common to know about the graph contrastive learning with graph augmentations mentioned in paragraph 3 of the introduction. For the first contrastive scheme mentioned in paragraph 2, the authors should give more details. -The authors should make a more specific presentation and explanation of the high-frequency and low-frequency information mentioned in the introduction and give a specific instance that high-frequency information is more beneficial for heterophilic graphs. -Some nodes mentioned in the description of the monophily property of the graphs in lines 190 to 192 can be corresponded to the subgraph d in Figure 3, thus making it less difficult for the reader to read. -The experimental setup is inconsistent. In the 296 line the authors introduce experiments on five random seeds, while in section 6.2 it is changed to ten runs.

Questions

1.The 184 lines of "node identity and node preference in real-world graphs" is not very understandable. Is this a metaphor of your own? 2.In a general GCL methods (such as GraphCL), the representations are usually mapped into the contrastive space via a projection head before the contrastive loss is calculated. Is this operation similar to the one you mentioned in this paper. Some typos: -In Eq.9, $\mathcal{L}_{A C L}(q)$ -> $ \mathcal{L}_{A}(q) $

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

NA

Reviewer fMCY6/10 · confidence 4/52023-07-08

Summary

The paper presents GraphACL, a contrastive learning method for graph representation learning. It aims to address the limitations of current methods that only consider the homophily property of graphs or rely heavily on graph augmentation methods. The authors propose an asymmetric predictor approach where the one-hop local neighborhood context and the two-hop monophily similarity are captured.

Strengths

- The premise of the need for a new contrastive learning method that does not overly rely on graph augmentation methods or the homophily property of graphs is reasonably argued and valid. Evidence from other studies or more data to support this argument could further its strength.

Weaknesses

- The explanation of how GraphACL works and how it captures both the one-hop local neighborhood context and the two-hop monophily similarity might be difficult to comprehend for readers lacking technical background. The use of concrete examples or analogies could strengthen this argument. - The argument strength could be improved through more detailed explanation of the theoretical analysis of GraphACL, giving more justifications or linking it to established theoretical or empirical studies in the field of contrastive learning. - The paper's argument that GraphACL is a superior tool when compared to current state-of-the-art graph contrastive learning methods is backed by empirical evidence on both homophilic and heterophilic graphs. However, some rencent proposed method are missing (i.e., [1] and [2]) [1] Spectral Augmentation for Self-Supervised Learning on Graphs [2] Spectral Feature Augmentation for Graph Contrastive Learning and Beyond

Questions

see weakness

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Overall, the paper presents an interesting and valid argument. However, to enhance its strength, the author could provide more empirical data, concrete examples, and clearer explanation of the concepts and methodology.

Reviewer fMCY2023-08-13

I am satisfied with author's rebuttal. It would be good to see the new result are included in the revision. I rasie my score from 5 to 6

Authorsrebuttal2023-08-13

Thank you for your comments and updating the score!

Thank you very much for carefully reading our response and increasing your score! We are glad our response has addressed your comments. We genuinely appreciate your support and will include the new results in the main paper. Best wishes, Authors

Reviewer ycWA2023-08-14

Thank you

Thanks for your responses and comprehensive rebuttal. In particular, I appreciate the experiments and references on low-frequency claims about existing GCL methods, which helps strengthen this work's motivation considerably (I'd make sure you add this in the new version, as it helps this work a lot). R6: Thanks for the clarification. I missed this experiment and indeed it helps understand that the loss helps but is not required to avoid collapse. You may want to reference [1] from my response -- this work also introduces some auxiliary augmentations (instead of a uniformity loss term), but for the same effect of helping methods which have such asymmetric predictor structure avoid collapse. It may help the discussion around why the uniformity loss works in the paper and how it helps the reference GraphACL method achieve better performance. R9: Thanks for adding this. It strengthens the work and helps support the argument of generality of representations. In light of the above updates, I will raise my score to 6.

Authorsrebuttal2023-08-14

Thanks for your updates and updating the score!

Thank you very much for reviewing our paper and reading our rebuttal. We sincerely appreciate your recognition of our clarifications and the increase in your score! In the new version, we will include the experiments and references related to the low-frequency claims about existing GCL methods, as well as cite the work [1] from your response to further illustrate the uniformity loss and its relation to our work. Best wishes, Authors

Reviewer 1K5Q2023-08-15

Thanks for the authors' responses. I am satisfied with the responses that address my concerns. I raise my score to 6.

Authorsrebuttal2023-08-15

Thank you for your comments!

Thank you so much for your efforts! We sincerely appreciate the reviewer for checking our responses and the increase in your score! Best wishes, Authors

Reviewer gZUY2023-08-15

Response to the rebuttal

Thanks for your efforts, I will maintain my positive score.

Reviewer cxpZ2023-08-15

Thanks for the clarification, and I appreciate the authors' effort in this. I will maintain my positive score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC