Exact Representation of Sparse Networks with Symmetric Nonnegative Embeddings

Many models for undirected graphs are based on factorizing the graph's adjacency matrix; these models find a vector representation of each node such that the predicted probability of a link between two nodes increases with the similarity (dot product) of their associated vectors. Recent work has shown that these models are unable to capture key structures in real-world graphs, particularly heterophilous structures, wherein links occur between dissimilar nodes. In contrast, a factorization with two vectors per node, based on logistic principal components analysis (LPCA), has been proven not only to represent such structures, but also to provide exact low-rank factorization of any graph with bounded max degree. However, this bound has limited applicability to real-world networks, which often have power law degree distributions with high max degree. Further, the LPCA model lacks interpretability since its asymmetric factorization does not reflect the undirectedness of the graph. We address these issues in two ways. First, we prove a new bound for the LPCA model in terms of arboricity rather than max degree; this greatly increases the bound's applicability to many sparse real-world networks. Second, we propose an alternative graph model whose factorization is symmetric and nonnegative, which allows for link predictions to be interpreted in terms of node clusters. We show that the bounds for exact representation in the LPCA model extend to our new model. On the empirical side, our model is optimized effectively on real-world graphs with gradient descent on a cross-entropy loss. We demonstrate its effectiveness on a variety of foundational tasks, such as community detection and link prediction.

Paper

Similar papers

Peer review

Reviewer zLuj7/10 · confidence 3/52023-07-01

Summary

This paper extends for by Chanpuriya et al (2020) for networks with homophilous and heterophilous edges. They show that their model is able to yield interesting results and they give theoretical underpinnings as well.

Strengths

The model is straightforward but seems to work well. Due do its relative simplicity and its relationship to previous work it is possible to give theoretical guarantees.

Weaknesses

The paper seems to ignore relationships to multilayer and multiplex networks which are also able to capture different types of relationships through different layers. Mason Porter, Peter Mucha and others have written many papers on such networks. There is also a literature on spectral methods for signed networks which can also capture heterophily; Mihai Cucuringu comes to mind. These areas should be mentioned in the related works section. Some of the presentation is not clear: In (4) it seems that A is a graph but the right hand side is a probability matrix. How is a 0-1 matrix reconstructed this way? The overall method is competitive but does not outperform any of the other methods. The conclusion does not give any indication of how the performance could be improved. Is the issue related to independence assumptions? A discussion would be helpful.

Questions

In the motivating example there are 10 cities but the model uses 12 communities. How is the number of communities chosen? Equation 4: what do the wriggly lines mean? Theorem 5.1: can it not happen that X Y^T has entries taking the value 0, in which case H would give the value 1/2 and not 0 or 1? Proof of Theorem 5.2: In my understanding a forest is a collection of trees. What is a partition into forests? It would be good to give a worked example in the appendix. How are the children found? Are the trees supposed to be rooted, choosing a root at random, and then orienting the tree from there? The proof seems to be constructive but it is not clear what the resulting matrices X and Y would be; the final step in the proof seems to be missing.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed 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

3 good

Limitations

There is no mention of societal impact. For example when assessing credit risk by assigning customers to communities of different risk, would extra caution be advised?

Reviewer SSbD5/10 · confidence 3/52023-07-04

Summary

This paper proposes and studies a community-based factorization model for exact representation of sparse networks. The authors extend a prior result on exact factorization which is based on logistic principal component analysis (LPCA), and show that an LPCA factorization can be converted to the proposed community-based graph factorization. The new factorization model explicitly captures both homophily and heterophily among nodes in the graph, and hence it is more interpretable. The authors carry out experiments using a different training algorithm to fit the model, and illustrate the advantage of the proposed decomposition method over existing ones for community detection and link prediction tasks.

Strengths

- The paper provides an intuitive community-based graph factorization model that is able to explicitly capture both homophily and heterophily among nodes in the graph. The expressiveness of the new factorization is backed by a reduction from LPCA which is shown to be able to exactly represent any sparse graph with bounded arboricity. Although it may not be something of extremely high theoretical significance, I liked this aspect of technical development. It is an interesting result to have.

Weaknesses

- There is a clear gap between the exact factorization studied up to Section 5 and the actual training algorithm evaluated in Section 6. Although the authors have mentioned the algorithmic gap explicitly at the end of Section 5. In order to well approximate the heavyside function which is used to achieve exact representation by LPCA, the input of the logit function should have entries that are arbitrarily large in magnitude. However, the training algorithm uses regularization to prevent this. The regularization does exactly the opposite of what is needed for exact representation. Therefore, there is a significant gap between theoretical development and empirical verifications. - Since the paper studies exact representation, there should be at least an experiment (e.g. the synthetic one) which shows that the reconstruction error goes to 0. - The overall writing and clarity can be improved. There are a few typos. For example, Line 32, principal components analysis -> principal component analysis. I also could not find where in the paper the term arboricity is clearly defined.

Questions

- For the synthetic experiment shown in Figure 2, what happens if you increase the embedding length further? Will SVD beat your method afterwards? Can you make your method has close to 0 error (since it is supposed to be exact)? - What happens if you don't regularize your training algorithm for the experiments in Section 6, and how would the methods compare in that case?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

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

2 fair

Contribution

2 fair

Limitations

I cannot find where the authors discuss the limitations or potential negative societal impact of their work, but I think it is fine.

Area Chair px9n2023-08-17

Please provide additional feedback

Hi, You seem to have the lowest score for this paper. Could you please acknowledge that you have read the rebuttal and let us know if you still have concerns or not. If not, I would encourage you to raise your score.

Reviewer VwbN4/10 · confidence 4/52023-07-05

Summary

This paper proposes to use the logistic PCA model to represent. Paper's main contribution is theoretical – improving the embedding dimensionality bounds from maximum degree to the arboricity of a graph. Overall, I believe the contribution is significant, but currently the paper does not highlight why too much – see more in the weaknesses section. The evaluation is severely limited, which might be forgivable for a more theory-oriented paper.

Strengths

* Theoretical guarantees are interesting to (potentially) broader research community. * The methods offers grounded interpretability of node embeddings in terms of the community structure. * The paper is written clearly and should be easy to understand to the general audience.

Weaknesses

* While the theoretical results are interesting, they are not well connected to the literature on either: ** Spanning trees and effective resistances and graph sparsification would be exciting to sparsification folks. ** Curvature (through bounds on Ollivier-Ricci curvature from effective resistance) would yield interesting connections to GNNs * The proposed algorithm is not really practical computationally speaking and does not scale to large graphs. * Experiments are not really indicative of real-world performance of the algorithm and rather focus on provide some free-form support for the theoretical claims of the paper.

Questions

* Would node classification experiments on the datasets like POS/PPI/Blog (these are already present in the experimental section) strengthen the paper? Suggested comparison is some early neural embedding such as DeepWalk. This would open up interest from the subfield of more heuristic node embedding methods. * Can you detail the comparison with Chanpuriya et al., especially in terms of the proof?

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

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

2 fair

Limitations

Yes; although I would stress limited practicality of the proposed algorithm due to its time complexity.

Reviewer gKGt6/10 · confidence 4/52023-07-06

Summary

This paper concerns several main results/themes: - the authors show a theoretical result on a prior model known as LPCA. Their result indicates that exact factorizations for graphs under the LPCA model is possible under a bounded arboricity assumption, which is more generally applicable than the prior degree-based assumption. - the authors propose a graph generative model that uses two non-negative vectors (one for homophily, one for heterophily) per node as well as a non-linear linking function to embed a given graph in an interpretable manner. This models generalizes the prior LPCA model, and has nice properties such as having factorizations that are non-linear, captures heterophily in addition to homophily, and is non-negative. - they show that the theoretical results that are available for LPCA (including the one that they derived) is applicable to their more general model as well. - they show promising experimental results on datasets.

Strengths

- Originality: The two original contributions are mainly as follows. 1. the theoretical result on LPCA representations which uses a new arboricity criteria rather than a max degree criteria. 2. The proposed generative model (which is more general than LPCA). To the best of my knowledge, these two contributions have not been priorly proposed. I believe that the level of originality of this article, while not groundbreaking, is sufficient for publication at NeurIPS. - Quality: The quality of the experiments and the theoretical results are sound and technically solid, to the best of my knowledge. The results of the experiments supports the authors' claims that their proposed approach capture heterophily in an interpretable manner and that in specific cases it outperforms competitors. - Clarity: the paper is presented in a clear and easy to read manner, which is much appreciated. The authors motivated their results well. - Significance: Graph embeddings are certainly a relevant area of research nowadays. The authors' contributions have moderate impact, given their expressive, interpretable and more generally applicable model. It is significant enough for publication at NeurIPS.

Weaknesses

- The authors covered virtually all the bases. Overall, I do not see any obvious shortcomings/omissions in the paper. I do have some comments/questions below: - From a theoretical and originality perspective, the advances that they make over the prior result (the arboricity vs degree assumption) appears rather marginal (similar to Chanpuriya 2020). Yet a new result is a new result. I do not find this to disqualify this article from being accepted. - I find a small disconnect between the theoretical results and the proposed methods. The theory says that exact representations exist. The methods provide one way to find a representation. But there seems to be little linking the two: there seems to be no guarantee that what the algorithm actually finds in practice is the exact representation. It is possible that I might be missing something here, and any clarifications from the authors would be appreciated.

Questions

- See above section

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

4 excellent

Contribution

3 good

Limitations

The main limitations of these classes of methods is scalability and computational complexity. The authors briefly address this in the paper in A.3 for example by mentioning a stochastic training approach. I think the author's discussion of the limitations is sufficient but could be improved, for example, by discussing computational complexity (big Oh run time) in more explicit manners in both the deterministic and the stochastic case.

Reviewer zLuj2023-08-11

The main result from Chanpuriya et al (2020)

Thank you for your reply. Regarding (4), it would be better to rephrase as it sounds as if (4) was a way to construct \tilde{A}, which is not the intention I understand. Instead (1) is the definition of \tilde{A} when B and C are given. Looking more into Theorem 5.1, this is not actually the main result from Chanpuriya et al (2020); they prove it for the function \sigma(s) = max( 0, min (1,x)) instead of the Heaviside function H(x). So why does Theorem 5.1 hold? Thank you for expanding on your notion of a forest. What is the difference between a forest and a tree in your notion? Usually forests are collection of trees, and hence a partition into forests is ambiguous. Your example in the pdf seems to indicate that you equate forest and tree.

Authorsrebuttal2023-08-11

Response to follow-up questions

Thank you for following up. >Regarding (4), it would be better to rephrase as it sounds as if (4) was a way to construct \tilde{A}, which is not the intention I understand. Instead (1) is the definition of \tilde{A} when B and C are given. Thank you for the suggestion. We agree that it would improve clarity to modify (4) by replacing $\mathbf{A} \approx \sigma( \mathbf{X} \mathbf{Y}^\top )$ with $\mathbf{A} \approx \tilde{\mathbf{A}} = \sigma( \mathbf{X} \mathbf{Y}^\top )$. We will make this change and modify the surrounding text appropriately. >Looking more into Theorem 5.1, this is not actually the main result from Chanpuriya et al (2020); they prove it for the function \sigma(s) = max( 0, min (1,x)) instead of the Heaviside function H(x). So why does Theorem 5.1 hold? This is a great, precise question. The proof of the theorem in Chanpuriya et al. (2020) is actually in terms of the Heaviside nonlinearity $H(z)$ (they call it ($s(z)$), and they just note how it implies the theorem in terms of the clipping nonlinearity $\sigma(z)$ (see the second paragraph of the proof). We lift the theorem from their work without the extra step of this implication because $H(z)$ is more suitable for this paper; we will note this clearly in the revision. (The preceding answers the question of why Theorem 5.1 holds, but if you are curious, the implication is as follows: suppose you were given an adjacency matrix $\mathbf{A}$ and found $\mathbf{X},\mathbf{Y}$ such that $H( \mathbf{X} \mathbf{Y}^\top ) = \mathbf{A}$. Then you can scale up the factorization by replacing $\mathbf{X}$ with $\mathbf{X}' = c \cdot \mathbf{X}$ for a large enough positive constant $c$ such that $\sigma( \mathbf{X}' \mathbf{Y}^\top ) = H(\mathbf{X} \mathbf{Y}^\top ) = \mathbf{A}$. This is because, for each entry $z$ in $\mathbf{X} \mathbf{Y}^\top$, either 1) $z$ is negative, in which case $c \cdot z$ is still negative and hence $\sigma(c \cdot z)=H(z)=0$; or 2) $z$ is positive, in which case, for a large enough $c$, you have $c \cdot z \geq 1$ and hence $\sigma(c \cdot z)=H(z)=1$.) >Thank you for expanding on your notion of a forest. What is the difference between a forest and a tree in your notion? Usually forests are collection of trees, and hence a partition into forests is ambiguous. Your example in the pdf seems to indicate that you equate forest and tree. We use the standard notion of a forest, which is a union of *disjoint* trees. The example in the PDF shows a partition of the edges of a graph into two forests - call them $F_ \text{blue}$ and $F_ {\text{orange}}$. In this case, since we wanted a small example, $F_ \text{blue}$ and $F_ \text{orange}$ both happen to be trees. This does not mean that the union of $F_ \text{blue}$ and $F_ \text{orange}$ itself is a forest, because the union is exactly the original graph, which has a cycle and hence cannot be a forest. (Note that if the definition of forest was *any* union of trees, rather than a union of *disjoint* trees, then the union would be a forest. This is disallowed because the two trees are touching and not disjoint.) Perhaps this thought will help clarify: Suppose we took the example in the PDF and added two extra nodes which have an edge between them but are disconnected from the rest of the graph. These two extra nodes will definitely demand an extra tree to cover (which will just comprise the single edge between them), but they will not demand an extra forest, because this extra tree will be disjoint from both $F_ \text{blue}$ and $F_ {\text{orange}}$ and hence can be added to either forest.

Reviewer zLuj2023-08-12

Thank you for your explanation. Trusting that you will amend the paper as you suggested, I am happy to change my score from 6 to 7.

Authorsrebuttal2023-08-15

We are glad the explanation was helpful. Thank you for following up and reconsidering your score.

Reviewer gKGt2023-08-11

reply to authors

I thank the authors for reading and address my comments. I am satisfied with the authors' responses.

Reviewer VwbN2023-08-14

Thanks for the rebuttal. I hope you find connections interesting/useful. In the current form without additional experimental evidence I do not think I should raise the score.

Authorsrebuttal2023-08-15

We thank you for reviewing, responding, and raising interesting connections. Related to your comment, we briefly summarize our experiments for the reviewers and chairs: - We evaluate NMF models on their ability to represent a synthetic network, as well as several real-world networks. In the rebuttal PDF, we also confirm our model's ability to do exact embedding. - We then evaluate on two other *unsupervised learning* tasks (more precisely, tasks without info beyond network structure): 1) community detection and 2) link imputation. - In the appendix, to investigate a more realistic application, we evaluate scalable ($O(n)$) versions of the NMF methods for tabular data completion (via link imputation) on industry datasets. These experiments span a range in terms of corroborating our theoretical results directly to supporting the graph modeling concepts we introduce. While we believe we have added sufficient experimental support, in order to maintain clarity and readability, we are careful to keep our experimental section closely related to the theoretical and conceptual contributions of the work.

Area Chair px9n2023-08-17

Comment by the AC

Dear Authors, Thank you for your detailed answers and the effort that you put in your rebuttal. Could the reviewers who haven't responded to the authors please do this as soon as possible.

Reviewer SSbD2023-08-17

Thank you for the responses and additional experiments. Most of my questions have been addressed, I slightly raised my score. But I still don't like fact that the actual exact factorization may have very limited practical use, as the authors pointed out in the above. It seems that what works in practice is the training algorithm in Section 6. The theory-to-practice gap is still there.

Authorsrebuttal2023-08-17

We thank you for reviewing, following up, and reconsidering your score. We are glad we have addressed the point from your review about our training algorithm's capability for exact factorization. Regarding the point from your comment on the usefulness of exact factorization itself, we briefly expand on this: One of the core ideas of the paper is the benefit of adding to nonnegative embedding models a second, heterophilous embedding in addition to the usual homophilous embedding. This allows for the representation of heterophilous structures, which models like BigClam and SymNMF (which lack a heterophilous embedding) will struggle with at practical embedding lengths, as is clearly illustrated in Figures 1-3 for the synthetic plot and also evidenced for real-world networks in Figure 4. The point of the exact embedding result in this context (though it may also be of independent theoretical interest) is to show that our model, and more broadly, nonlinear factorizations with *both* kinds of embeddings, are enough to represent *any* kind of network structure. In other words, representing homophilous and heterophilous structure is enough, and there is no need for, e.g., a third embedding. While the exact embedding results (both theoretical and empirical) are in the regime of no/low regularization and tight fitting, they are reflected in the practically regularized regime which tries to capture signal and ignore noise: as we show throughout the experiments, there is benefit to also capturing heterophilous "signal."

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC