Analysis of Corrected Graph Convolutions

Machine learning for node classification on graphs is a prominent area driven by applications such as recommendation systems. State-of-the-art models often use multiple graph convolutions on the data, as empirical evidence suggests they can enhance performance. However, it has been shown empirically and theoretically, that too many graph convolutions can degrade performance significantly, a phenomenon known as oversmoothing. In this paper, we provide a rigorous theoretical analysis, based on the two-class contextual stochastic block model (CSBM), of the performance of vanilla graph convolution from which we remove the principal eigenvector to avoid oversmoothing. We perform a spectral analysis for $k$ rounds of corrected graph convolutions, and we provide results for partial and exact classification. For partial classification, we show that each round of convolution can reduce the misclassification error exponentially up to a saturation level, after which performance does not worsen. We also extend this analysis to the multi-class setting with features distributed according to a Gaussian mixture model. For exact classification, we show that the separability threshold can be improved exponentially up to $O({\log{n}}/{\log\log{n}})$ corrected convolutions.

Paper

References (55)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer kcVu6/10 · confidence 4/52024-06-19

Summary

This paper studies the effects of removing the top eigenvector of the adjacency matrix used for aggregation in graph convolutions. For the contextual stochastic block model (CSBM), this is theoretically shown to be beneficial. The authors provide several theoretical statements describing misclassification ratios and chances and chances of achieving linear separation. Experiments confirm the superiority of removing the smooth eigenvector for the CSBM.

Strengths

* The background, literature review, and preliminaries are nicely presented. * The ideas behind the proofs are interesting. * The conducted experiments confirm the claimed benefits of removing the top eigenvector for CSBM.

Weaknesses

* The structure of the paper is confusing. In Section 4, two Theorems are provided without proofs. Proofs are also not provided in the Appendix. Then, there is a proof sketch for the Theorems in Section 6. * A clear focus would help the accessibility of this work. It would suffice to consider either $\hat{A}$ or $\tilde{A}$ and Theorem 4.1 or Theorem 4.2 while providing more details for the selected case. * Theoretical statements and their implications are hard to understand. Clearly stating all symbols and providing additional details would help. * The implications of this work are not clear to me. We know that $M^kx$ gets dominated by the top eigenvector of M with rate $\lambda_2/\lambda_1$. Setting the largest eigenvalue to zero, results in dominance of the second eigenvector. As pointed out in this work, this eigenvector corresponds to sparse and balanced bipartitions. This analysis seems like a complicated way to come to the same conclusion. * The practical usage of removing the top eigenvector seems very limited, as one dominating signal is exchanged for another. This is confirmed in Figure 3.

Questions

* What are the benefits of this analysis against existing insights that the signal corresponding to the top eigenvector of a matrix gets amplified the most and dominates representations? * What are potential practical implications that can now be developed? * l. 332: Why is the O-notation used for $p=O(log^3 n/n)$?

Rating

6

Confidence

4

Soundness

2

Presentation

1

Contribution

3

Limitations

Limitations are not explicitly presented.

Reviewer kcVu2024-08-11

I thank the reviewers for their detailed feedback. I can now better understand why this work can be interesting from a mathematical and statistical point. I would like to hear more details about the following: * In the referenced work [ref line 443, Keriven: Not Too Little, Not Too Much], the author shows that some rounds of aggregation can be helpful while too many iterations are hurtful for their considered task. Here, it seems like more rounds of aggregation are always beneficial, but fewer rounds can be sufficient. Can this study be extended in the future to include cases where the task does not align so well with the dominating eigenvector? For Figure 3, performance seems to be monotonously decreasing, so (zero or one?) iterations seem to be optimal. What would tasks look like for which some iterations of your corrected convolutions are beneficial while the performance in the limit becomes worse? Two additional comments on terms that I found irritating but not critical: * Convolution: There is no definition of what convolution means. It seems like it's just the aggregation part $\mathbf{A}$ and powers of it. Referring to this as convolution is confusing to me, as no filter is considered. In the referenced work above, the author calls this $k$ rounds of aggregation, which would also make the statements in this work clearer. * Corrected convolution: Was this term used previously in the literature? What makes this aggregation more "correct"? I agree that it seems more correct for the tasks considered in this work. However, I would argue that depending on the graph and task, various aggregation matrices can be more "correct".

Reviewer kcVu2024-08-13

The above comment was apparently not visible to the authors. I apologize and hope my question still comes in time.

Authorsrebuttal2024-08-13

We thank the reviewer for their reply. However, we would like to point out that this post was made very close to the deadline and during the authors' nighttime. A proper reply to the reviewer takes several hours. We provide a reply to the best of our ability given the limited time. We want to re-iterate the fact that the corrected convolution is not ours. We mention this in our first reply above and we point this out in lines 348 to 350. This idea is already implemented in widely cited practical papers [ref line 512 (ICLR 2020)]. We provide a rigorous analysis. > In the referenced work [ref line 443, Keriven: Not Too Little, Not Too Much], ... task. The result that the reviewer claims seems to be Theorem 2. We would like to clarify what is proved in the main theorem of this paper. The authors don't provide a non-asymptotic analysis as we do in our paper. They only provide an existential result. They prove that 1 convolution can be better, under specific assumptions, than 0 convolutions. They also prove that 1 convolution is better than infinite convolutions. This implies that there might exist a $k^* \ge 1$ which is optimal. However, they never rigorously analyze the performance of convolution for $k>1$ and $k<\infty$. In fact, the authors mention in their paper that "In the next section, we also derive an intuitive expression for $R(k)$ (although without rigorous proof), which we observe to match the numerics quite well." We would like to point out that it is technically extremely challenging, and requires a whole different set of methodology, to achieve non-asymptotic analysis as we do in our paper, and we hope the review appreciates it. > Can this study be extended ... in the limit becomes worse? We will reply to this question by focusing on other tasks within node-classification, since node-classification is the focus of our paper. By dominating eigenvector we assume that the reviewer implies the second or higher eigenvectors. The dominant (first) eigenvector does not hold meaningful classification information, therefore removing it helps. We prove this in our Theorems 4.1 and 4.2. Also, Fig 3 shows that removing the top eigenvector allows the model to have better performance as the number of convolutions increases. We interpret the reviewer's comment as a question about why the performance decreases for $k$ larger than 1 or 2. The reason is not because we remove the dominant (first) eigenvector. As we point out this is beneficial, since the first eigenvector holds no information about the classes. The reason is that in the particular real data which we used there are eigenvectors with small eigenvalues which might contain meaningful information. Asymptotically, the corrected convolution, which we repeat is not our idea, see comments above, might miss the information from such eigenvectors. Intuitively, this could be the case in a graph where there are small communities within larger communities in the given graph structure. The top eigenvectors are likely to be correlated with the larger communities. To capture the information from small eigenvalues, a different convolution is needed which filters eigenvectors in an appropriate way. However, we would like to point out again, that the first eigenvector will be filtered out in this case too since it's not meaningful. To analyze such a scenario, we would have to change our random data model. That's because in the current data model the first few top eigenvectors (excluding the first one) are the most meaningful ones, and eigenvectors for small eigenvalues are just noise. Technically, a non-asymptotic spectral analysis could also work in this hypothetical new data model. The corrected convolution in our paper is likely to lose performance as $k$ increases. A different convolution would be needed to better capture the information of eigenvectors of small eigenvalues. This study is indeed interesting, however, we would like to mention that it is far from the goal of our present paper. We would be happy to mention the above in our paper. > Convolution: ... in this work clearer. We will make sure to clarify this in our paper. We come from a graph neural network background where this definition is ubiquitous and we didn't explicitly specify it as a definition. > Corrected convolution: ... can be more "correct". We chose the word corrected since in the literature of statistical community detection (no node features), the word is often used for modified versions of the original SBM, for example: B. Karrer and M. E. J. Newman, Stochastic blockmodels and community structure in networks, Phys. Rev. E 83 (2011). In this case, the fact that one does degree "correction" does not necessarily mean that it is a universal correction for any possible problem. It's a name which makes sense within a specific context. We can offer to change the title to "Analysis of Eigen-Corrected Graph Convolutions" if the reviewer believes that this is more specific.

Reviewer kcVu2024-08-13

I want to thank the reviewers for their quick and detailed answers. I am now convinced of this work's benefits to the graph learning community and am open to accepting this paper. I have changed my score to 6. I want to state my final opinions below, to which the authors do not need to reply: >The dominant (first) eigenvector does not hold meaningful classification information, therefore removing it helps. [...] Fig 3 shows that removing the top eigenvector allows the model to have better performance as the number of convolutions increases. The best-achieved performance of the original aggregation and corrected aggregation are quite similar (both for k=1). Stating that the corrected aggregation retains more task-related information would better describe the behavior. Identifying non-SBM tasks for which "removing the top eigenvector allows the model to have better performance as the number of convolutions increases" is actually true, and the max performance for the corrected aggregation is better compared to the standard aggregation, which is still open and interesting for future work. > We come from a graph neural network background where this definition is ubiquitous and we didn't explicitly specify it as a definition. In my graph neural network background, a graph convolution consists of a feature transformation and an aggregation part. As this work considers the aggregation part, I would personally find it more precise to call it $k$ rounds of aggregation instead of $k$ rounds of convolution. >We can offer to change the title to "Analysis of Eigen-Corrected Graph Convolutions" if the reviewer believes that this is more specific. I don't mind if the authors keep the original title. To me personally, something like "Analysis of CSBM-Corrected Graph Aggregations" would seem clearer.

Authorsrebuttal2024-08-13

We are grateful. Thank you. We will make appropriate modifications to address your concerns in the revised version.

Reviewer uSHp7/10 · confidence 3/52024-07-03

Summary

This paper studies the concept of oversmoothing via a CSBM modeling of a GNN structure, and views the behaviors of vectors after repeated multiplications of a graph. Importantly, they consider a scheme where a dominant eigenvector is "left out", so that it does not dominate the behavior of the evolution so much, and show enhanced performance over real world datasets.

Strengths

This paper explores an important problem from a simple enough framework that it is tractable. There are a ton of statistical results, which all look reasonable and tell a reasonable story about the balance between learning and oversmoothing. The idea of discarding the dominant eigenvector is also interesting, and clearly does show improved results in the numerics.

Weaknesses

It's not clear to me how informative / predictive the various statistical results are to oversmoothing in practice. Can there be some figures that show the theoretical bounds, compared with true performance over some simulated examples? For example, one of the theorems uses Davis/Kahan, which is pretty pessimistic in practice. In general, perhaps the spectral gap alone is not informative enough to showcase the entirety of the oversmoothing behavior, and while that is certainly fine given the complexity of the problem, it still would be useful to see the theorems compared with in practice behavior.

Questions

What is CSBM? I know SBM but what makes it C? Have you considered scenarios where the data has clear block structure, but it does not 100% overlap with the true labels? will it cause undesirable biases? can those be quantified? Actually, one interpretation of this is that the adjacency spectrum somewhat mirrors the Laplacian spectrum, but in reverse order. So, the dominant eigenvector corresponds to the nullspace of L, and the second one is the one associated with the mixing parameter in L. So, in that sense, it makes sense why focusing on that eigenspace gives better performance. I would be interested in hearing the authors' view on this.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

no ethical issues

Reviewer eGdd7/10 · confidence 3/52024-07-11

Summary

This paper studies over-smoothing from $k$ rounds of graph convolutions in the Contextual Stochastic Block Model by considering vanilla graph convolutions and a corrected one where the principal eigenvector is removed. Using spectral analysis, the authors derive the partial and exact recovery for both cases and show that the corrected convolution avoids over-smoothing. They derive the classification error for different densities $p$, $q$ and separability $\gamma$ of the underlying graph model.

Strengths

1. The theoretical analysis is thorough and comprehensive. The studied setting is also of practical significance and the authors provide rigorous theoretical analysis. 2. The presentation is clear. The discussions on the theoretical results and assumptions are very helpful in understanding the paper. 3. Experiments also support the theoretical results.

Weaknesses

There are no critical weaknesses in the work. I have a few comments/questions which I list below. 1. One aspect of the result is the dependence on the density of the graph as well as good SNR ratio. Can the authors comment on the need for denser connectivity in the graph for recovery? 2. The performance of the corrected convolution decreases with $k$ in the case of a multi-class setting. The authors reason it as the projection onto the second eigenvector cannot capture the full class information. However, in the balanced data setting, all the eigenvalues of the corrected convolution are the same, with eigenvectors having information about the classes. So, in expectation, shouldn't the performance be unaffected? 3. The analysis considers homophily structure of the graph $p>q$. Can it be extended to heterophily case as well? 4. In the definition of $\hat{A}$ in equation 1, shouldn't the last $D$ be $D^{-1/2}$? The negative sign is missing.

Questions

Please refer to the Weaknesses.

Rating

7

Confidence

3

Soundness

4

Presentation

3

Contribution

3

Limitations

The analysis is in linear setting, and extending it to non-linear activations may be challenging.

Reviewer xoAh5/10 · confidence 3/52024-07-12

Summary

In this paper, the authors present a comprehensive theoretical analysis using the contextual stochastic block model (CSBM) to evaluate the performance of vanilla graph convolution after removing the principal eigenvector to prevent over-smoothing. They conduct a spectral analysis for k rounds of corrected graph convolutions and provide findings for both partial and exact classification.

Strengths

1. The paper provides both detailed theoretical analysis and experiments on three datasets: CORA, CiteSeer, and Pubmed. 2. The paper provides a novel insight on why corrected convolution can mitigate over-smoothing and improve classification accuracy.

Weaknesses

1. The synthesized data by CSBM might be adequate to illustrate the binary classification case, but the multi-class case could be more complicated. For instance, it is illustrated in [1] that the effects of signed propagation under binary-class case and multi-class case could be quite different. It is mentioned that the authors would like to analyze the multi-class CSBM using more sophisticated architectures and I look forward to the further analysis. 2. Instead of stacking more than 20 layers of GNNs, it is a common practice to limit the number of layers to 3 to 5. According to Figure 3, the accuracy of the three learning methods appears more complex when the number of layers is limited to 5 or fewer. Therefore, the conclusion may not be applicable to real-world data applications. [1] Choi, Yoonhyuk, et al. "Improving Signed Propagation for Graph Neural Networks." arXiv preprint arXiv:2301.08918 (2023).

Questions

1. (Refer to weakness # 2) Would it be possible to analyze GNNs employing various convolution methods with shallower layers?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors have addressed some of the limitations of their work (limited to the binary classification case). Their work does not present any potential negative societal impact.

Authorsrebuttal2024-08-08

Correction of typo in the response

Dear reviewer, we apologize for the typo in the response to Question 2: What makes it C in CSBM? The C (context) here refers to the node features, since the data model consists of node features in addition to the graph. This term has been used in the statistics community to describe such data with a combination of two components: the graph and the node attributes. Please refer to Section 3.1 for a detailed description of the CSBM, and to [Deshpande, Y. and Sen S. and Montanari, A. and Mossel, E. Contextual stochastic block models. Advances in Neural Information Processing Systems, 2018.] where it was introduced.

Reviewer xoAh2024-08-09

Thank you for the response

The authors' response has adequately addressed my main concerns on the multi-class scenario and the analysis on shallower networks. I have raised my rating.

Authorsrebuttal2024-08-10

We thank the reviewer for raising their score. We will add the additional result in the revised version of the paper and address all comments.

Reviewer eGdd2024-08-09

I thank the authors for their clarifications and for providing analysis for multi-class setting. I retain my score as I didn't have any major concerns in my initial review, and I recommend acceptance of the paper.

Authorsrebuttal2024-08-10

We thank the reviewer for recommending acceptance of the paper. We will address all comments in the revised version of the paper.

Reviewer uSHp2024-08-10

Thanks, I am happy with all the responses. The new figures showing the bound vs performance are also very interesting, and a great improvement on the paper. I have no further questions.

Authorsrebuttal2024-08-12

Thank you! We will incorporate the modifications in the revised version of the paper.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC