Quasi-Monte Carlo Graph Random Features

We present a novel mechanism to improve the accuracy of the recently-introduced class of graph random features (GRFs). Our method induces negative correlations between the lengths of the algorithm's random walks by imposing antithetic termination: a procedure to sample more diverse random walks which may be of independent interest. It has a trivial drop-in implementation. We derive strong theoretical guarantees on the properties of these quasi-Monte Carlo GRFs (q-GRFs), proving that they yield lower-variance estimators of the 2-regularised Laplacian kernel under mild conditions. Remarkably, our results hold for any graph topology. We demonstrate empirical accuracy improvements on a variety of tasks including a new practical application: time-efficient approximation of the graph diffusion process. To our knowledge, q-GRFs constitute the first rigorously studied quasi-Monte Carlo scheme for kernels defined on combinatorial objects, inviting new research on correlations between graph random walks.

Paper

References (58)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer qCjY7/10 · confidence 3/52023-07-06

Summary

Kernel methods are important in many ML applications, but they suffer from scalability issues. In the Euclidean setting, random features methods (Rahimi & Recht NeurIPS 2017) can be used to "sketch" the kernel matrices and speed up computations. In the graph setting, however, no such random feature seems to have been available until a 2023 preprint by Choromanski. The present paper present an improvement over Choromanski's method. To explain it, note that the basic idea in the 2023 preprint is to consider random features for the regularized graph Laplacian, and to obtain an approximation of an operator of the form $(I-U)^{-d}$ via importance-weighted random walks. These random walks have geometrically distributed lengths. The main theoretical contribution of the present paper can now be explained: a simple modification of Choromanski's method, where the the geometrics of pairs of random walks become correlated, has the same zero bias and a smaller variance than the above method. This requires an elementary, but long proof with a long case analysis. Experimentally, the gains observed were: * visible, but not too big for approximation of the kernel; * impressive for graph diffusions (which is natural because this requires exponentiating the Laplacian); * fairly significant for clustering and node attribute prediction methods.

Strengths

The paper is clear and seemingly original. In general it is fairly well-written. Moreover, the "anthitetic" coupling is a simple little idea that could be applicable more broadly.

Weaknesses

The present paper has been submitted before Choromanski's preprint appeared. As a result, applications of graph kernel random sketches are quite incipient still. It is hard to say whether the present paper will have a significant direct impact. _[Edit on 08/11: this was partly addressed in the rebuttal.]_ The proof of variance reduction is very long and difficult to check. (Though I believe it.) (An "actionable version" of this point is that it would pay off to make this paper a bit less dependent on Chorumovski's preprint by eg. adding more experiments or theory.) _Less important comments_ The notation $t_{1,2}$ is a bit weird as it represents a pair of RVs. Moreover, if I understand correctly, these r.v.s are resampled at each step of the walk. In terms of exposition, it seems that all the authors need to do is to correlate the coin flips for the two walks at each step. Could you write their joint distribution purely in terms of the probabilities of 00, 01, 10 and 11?

Questions

Can you point to more specific and realistic applications where this methodology would be useful? (A somewhat speculative discussion would be nice.) -- Edit (08/11) -- addressed in rebuttal.

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

These are properly discussed.

Reviewer mBGP7/10 · confidence 4/52023-07-09

Summary

The author(s) extend the work of Choromanski [2023] and introduce quasi monte carlo graph random features. To estimate the 2-regularized Laplacian kernel, they now use random walk features for each node which have coupled probabilities of terminating after a certain number of steps. This increases diversity in the features and interestingly leads to provable reduction of variance of the global estimates. An empirical evaluation supports the theoretical findings by improved practical variance, as well as in a clustering and supervised node learning scenario.

Strengths

The paper is well-written and mostly easy to follow. While proofs are moved to the appendix, a sketch of proof of the main result is included in the paper. I like that. The introduced way of increasing diversity in random walks is (up to my knowledge) novel and very easy to implement. While it is evaluated in the context of estimating the 2-regularized laplacian of a graph, I agree with the authors that this method is of independent interest. An interesting question is if this kind of sampling improves results of other random walk based approaches, as well, or if the parctical impact is restricted to just a few scenarios (GRF being one of them). As a result, I see potential impact for further research beyond this immediate application.

Weaknesses

The introduction is structurally very similar to that of Choromanski [2023]. I'm assuming now that I know one of the authors of this submission. Otherwise, I would recommend the authors to revise the introduction to decrease similarity.

Questions

The last sentence of the proof sketch seems to encompass quite some work. I suggest it to be extended to mention at least what kind of matrix is supposed to be negative definite. Suggestions l. 107: 'the the' l. 164: 'the event that event that' l. 207 and l. 222: 'path' should probably be 'walk' References: I suggest the author(s) revise the references for consistent formatting. E.g. Arxiv vs Corr.Abs., as another example, Choromanski [2023] does not list any venue, at all.

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

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 provide a dedicated section discussing potential (far away) societal impacts, as well as open questions with the current approach. The discussion is adequate.

Reviewer SAEb5/10 · confidence 3/52023-07-27

Summary

This work proposes an efficient random-walk sampling approach that accelerates the estimation of the feature mapping for 2-regularized Laplacian kernels. The random walk sampling is based on antithetic termination, which is a variance reduction technique and sounds novel when being applied to the estimation of graph random features. The technique sounds solid. The authors also conducted experiments from different angles to verify the effectiveness of the proposed approach in practice.

Strengths

1. Using random walk with antithetic termination to estimate graph random features is novel, to the best knowledge of me. 2. Although I did not check the appendix, the technique and the benefits it may bring look reasonable to me. 3. The authors did a great job to explain the key insights behind. The paper is very well written. 4. The experiments are extensive. The benefits of the proposed approach is justified from various angles.

Weaknesses

1. The biggest weakness of the work is the limited scope of this work. It is unclear why the 2-regularized Laplacian kernel is important and why we need to compute it. It seems that 1-regularized Lap kernel can be efficient computed by Pagerank algorithms, why we may need 2-regularized Lap kernel. 2. It is unclear how much the estimation of the 2-regularized Laplacian kernel relies on the random walk with antithetic termination. Can this random walk approach be only applied to 2-regularized Laplacian kernels or a broader range of graph kernels? This question is kind of related to whether the paper has over-claimed the contributions. Neither the title nor the introduction reflect that the improvement is only for 2-regularized Laplacian kernels, and they claim a broader range of contributions. I expect the authors to make the actual contributions more clear in the revised version.

Questions

1. Why is the 2-regularized Laplacian kernel important? Adding experiments to compare with other graph kernels (in the aspects of both inference accuracy and computation) may improve the draft. 2. Is the random walk approach only applied to the estimation of 2-regularized Laplacian kernel?

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

3 good

Contribution

3 good

Limitations

N/A

Reviewer wGeX9/10 · confidence 4/52023-08-05

Summary

The paper introduces a method for graph random features (GRFs). Graph random features is the graph equivalent (i.e. kernels defined on graphs) of kernel matrix approximation using Random Features (Rahimi & Recht). Recently Chromanski et al. had proposed a method for GRF which uses a series of random walks. In the Euclidean setting, the quasi-Monte Carlo variants of the Random Features methods have been developed since they provide better convergence properties. So the goal of this paper is to devise an algorithm for the quasi-Monte Carlo GRFs setting. The idea is to correlates the length of the random walks by imposing 'antithetic termination' which is effectively a procedure to obtain a more diverse ensemble of graph random walks. The paper provides several theoretic results related to the error guarantees as well as a brief empirical study.

Strengths

+ Original idea with strong practical impact and wide ranging downstream applications as time-efficient approximation of the graph diffusion process is used in many scientific fields. + The paper is of very high quality: Well written and clear. + Good literature review and problem justification, as well as empirical analysis.

Weaknesses

- Some typos/grammatical mistakes (esp. definite/indefinite articles), please proof read before publishing - Some questions below as well.

Questions

* It seems that all the experiments are on synthetic datasets. Can the authors confirm that or clarify why they didn't test this on at least on real dataset? * Sec 4.3 cites the Dhillon paper regarding the kernel matrix they used but that paper has experiments on both handwritten image data and a gene dataset. It is not clear what dataset this experiment is based on. Please clarify this point. * Additionally, I would suggest adding a sentence in the text describing the nature of the datasets so that the reader doesn't have to go back to the original papers to understand the eval/test data. Similarly for Sec 4.4 what are the matrices used by Dawson-Haggerty, 2023? Please add a 1 sentence clarification.

Rating

9: Very Strong Accept: Technically flawless paper with groundbreaking impact on at least one area of AI/ML and excellent impact on multiple areas of AI/ML, with flawless evaluation, resources, and reproducibility, and no unaddressed 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

4 excellent

Limitations

The authors address the limitations and potential downstream impact in a dedicated section at the end.

Reviewer wGeX2023-08-17

thanks for your response

Thanks for your detailed response. It addresses my questions and given the questions from the other reviewers and the subsequent responses and discussions I am leaving my scores unchanged.

Authorsrebuttal2023-08-17

Thanks for your reply

We thank the reviewer for their reply, and again for their detailed reading of the text and strong review.

Reviewer SAEb2023-08-11

Thanks for the response

I appreciate the authors' detailed response. I agree that antithetic termination is fundamentally a procedure to sample more diverse discrete random walks, which is a great contribution. The empirical observation on PageRank approximation is also interesting. I still have questions about how wide the range of graph kernels could be where the proposed approach may be applied. 1. "As we show explicitly in Sec. 8.1,..." Sorry, I did not see Sec. 8.1. Do you mean Sec. 2.1? 2. "construct scalable estimates to regularized Laplacians for arbitrary integer $d$, ..." I do not think the paper given its current content shows a general applicability of the proposed approach. The analysis is mainly on 2-regularized Laplacians if I did not miss anything. Also, I know that heat kernel has a wide range of applications, but I did not see how the proposed approach can estimate the heat kernel. The discussion on heat kernel in Sec 2.1 is just to motivate d-regularized Laplacians where 2-regularized Laplacian is actually what to approximate in this work. Please correct me if I am wrong.

Authorsrebuttal2023-08-12

Application to other graph kernels

We thank the reviewer for their reply. Estimates of the $2$-regularised Laplacian kernel $\mathbf{K}_L^{(2)}$ can be used to construct estimates to the $d$-regularised Laplacian $\mathbf{K}_L^{(d)}$ and heat kernels $\mathbf{K}_D$ (note the slight change in notation compared to the manuscript because of markdown compiler problems). For the $d$-regularised Laplacian kernel, the section of interest is 7.1 from the supplementary material: `on the approximation of the $d$-regularised Laplacian using GRFs' (with apologies to mistakenly referring to Sec. 8.1 after using a newer version of the draft). We summarise it briefly for the reviewer's convenience. As they correctly point out, we dedicate time in the manuscript to approximating efficiently $\mathbf{K}_L^{(2)} \coloneqq (\mathbf{I} - \mathbf{U})^{-2}$. By also multiplying by $\mathbf{I} - \mathbf{U}$, it is trivial to obtain $\mathbf{K}_L^{(1)} \coloneqq (\mathbf{I} - \mathbf{U})^{-1}$. We can then take multiplicative combinations of (independent) estimators for an unbiased approximation to $\mathbf{K}_L^{(d)} \coloneqq (\mathbf{I} - \mathbf{U})^{-d}, d \in \mathbb{N}$. Since antithetic termination suppresses the variance on $\mathbf{K}_L^{(2)}$ and $\mathbf{I} - \mathbf{U}$ is known exactly, it follows that such estimates of ${\mathbf{K}_L^{(d)}}$ will enjoy lower variance. We will make this clearer in the manuscript. The argument for the heat kernel is very similar. Indeed, we note that `simulation of diffusion on a graph' (introduced in Sec. 2.1 and tested in Sec. 4.2) is exactly equivalent to approximating the action of the heat kernel on a vector corresponding to the initial state. To briefly summarise, $\mathbf{K}_D= \text{exp}(- \widetilde{\mathbf{L}}t) \simeq \left [(\mathbf{I}_N + \frac{t}{N_t} \widetilde{\mathbf{L}})^{-2}\right]^\frac{N_t}{2}=(\mathbf{K}_L^{(2)})^\frac{N_t}{2}$ with equality becoming exact in the limit of a large number of discretisation time steps $N_t$. This is shown in Eq. 21 of the main text. In Sec. 4.2 we probe the quality of this approximation by comparing the error on $\mathbf{u}_t \coloneqq {\mathbf{K}}_D \mathbf{u}_0$ when ${\mathbf{K}}_D$ is estimated with GRFs and q-GRFs, and find it to be substantially improved in the latter case (see Table 1). It is also clear from the expression above that $\mathbf{K}_D$ is a special case of $\mathbf{K}_L^{(d)}$ in the large $d = N_t$ limit with $\mathbf{U}$ scaled appropriately, so theoretical guarantees follow. We hope this clarifies the concrete applicability of the approach to other kernels, and how q-GRFs are theoretically and empirically found to provide better estimates. We will try to make this more obvious in the manuscript. We are also pleased to report that the new PageRank application admits straightforward theoretical analysis, with antithetic termination enjoying strong performance guarantees. The new section will now include a full proof rather than just being presented as an empirical observation.

Authorsrebuttal2023-08-18

Any further questions?

Could the reviewer please kindly confirm whether the response above answers their remaining questions, or whether anything remains unclear?

Reviewer SAEb2023-08-18

Thanks for the response

> From 2-regularized Lap to d-regularized Lap Thanks for the explanation! I still have the following questions. I think one key usage of random feature based estimation of the kernel is that we do not need to estimate the entire kernel matrix $K_L^{(2)}$. For each entry (two nodes) of interest, we can estimate separately. However, based on the authors' response, it seems that we have to estimate the entire $K_L^{(2)}$ and get $K_L^{(d)}$ via multiplication. I am not quite sure that such multiplication may have good variance. I saw the authors do some experiments shown in Table 2. But it is a comparison between GRF and q-GRF. I buy the results because q-GRF is more accurate than GRF. I am curious how q-GRF is compared with other ways to estimate $K_L^{(d)}$. Regarding the multiplication idea, I am curious why not estimate $K_L^{(1)}$ (effective resistance) and do multiplication to get $K_L^{(2)}$ and $K_L^{(d)}$ in general. Why do we need $K_L^{(2)}$?

Authorsrebuttal2023-08-18

Thanks for the further questions

We thank the reviewer for their response and further interesting questions. We are pleased that they are convinced that q-GRFs outperform the previously introduced class of GRFs, which is the central claim of the paper. The question of how to generalise GRFs (and their q-GRF variants) to other graph kernels is not our chief focus, but since it is an interesting problem we are happy to discuss it in more detail below. Scalable estimation of graph kernels with random walks is a new field and we will be grateful if the reviewer could point us to any other specific methods for computing $\mathbf{K}_L^{(d)}$ (including $d=1$). Should they exist and use terminating random walks, we strongly suspect that antithetic termination will still prove an effective QMC mechanism. We would be happy to run further experiments and attempt theoretical analysis in the future, but we were unable to find other such methods in the literature. The reviewer is correct that, though unbiased, multiplying independent estimators of $\mathbf{K}_L^{(2)}$ to obtain estimators of $\mathbf{K}_L^{(d)}$ is likely not the most efficient approach. It is straightforward to calculate the variance: for example, $\textrm{Var} (\widehat{\mathbf{K}}_L^{(4)}) = \textrm{Var} (\widehat{\mathbf{K}}_L^{(2)})^2 + 2\textrm{Var} (\widehat{\mathbf{K}}_L^{(2)}) {\mathbf{K}_L^{(2)}}^2$. As the reviewer says, this will be suppressed with q-GRFs because we have proved that in this case $\textrm{Var} (\widehat{\mathbf{K}}_L^{(2)})$ is smaller. Since we derived closed forms for $\textrm{Var} (\widehat{\mathbf{K}}_L^{(2)})$ with both mechanisms it is trivial to explicitly obtain $\textrm{Var} (\widehat{\mathbf{K}}_L^{(d)})$, but even just inspecting the small errors in Table 2 which uses as many as $500$ such multiplications we can see that the approximation quality is still good, even with just $10$ walkers per node. That said, we agree that a method to *directly* estimate $\mathbf{K}_L^{(d)}$ without recourse to multiplying several estimators would be preferable. Adjusting the GRF mechanism to do this is nontrivial and a problem we intend to dedicate substantial time to. We have made some progress, at a high-level by introducing an *extra function of walker path length that modulates its load*, separate from the existing functions $\widetilde{\omega}(\omega)$ and $p(\omega)$. We intend to carry out a detailed theoretical and empirical exploration of these methods in the future; our strong intuition is that antithetic termination will once again prove very effective at reducing kernel estimator variance. We thank the reviewer for prompting this interesting discussion, but trust that they appreciate that the central contribution of the present manuscript is a novel QMC method for random walkers including for Choromanski's GRF mechanism, *not* a new GRF mechanism for direct estimation of different kernels. This is a substantial, open research problem and deserves to be properly addressed in a separate paper. To update the reviewer, we have also now proved that inducing antithetic termination suppresses the error on estimates of the PageRank vector for arbitrary graphs and arbitrary termination probabilites $p$. This surprisingly strong result is in the manuscript.

Authorsrebuttal2023-08-20

End of discussion period

As the discussion period draws to a close, we respectfully ask that, if satisfied with the responses above, the reviewer kindly considers raising their score. We again thank them for their interesting questions, to which we hope we have provided exhaustive answers.

Reviewer qCjY2023-08-11

Thank you

Thank you for your response. I have raised my score slightly. I will look carefully at the discussion with other reviewers to see if any further changes to my score make sense.

Authorsrebuttal2023-08-14

Thank you and further update

We thank the reviewer for their prompt response and for raising their score. For the reviewer's interest and as further evidence of the applicability of antithetic termination beyond Choromanski's GRF setting, **we have now finished a proof that antithetic termination suppresses the variance on estimates of the PageRank vector for any graph**. This explains the excellent empirical results sent earlier in the rebuttal. We have added these new theoretical insights to the manuscript and are excited about further future applications in algorithms using terminating random walkers.

Reviewer mBGP2023-08-18

Thank you for your reply. My questions have been addressed. I like the authors extension to the computation of page rank scores. I believe this indeed shows that antithetic termination is a technique that can be relevant in many scenarios. I hence have increased my confidence and stand by my suggestion to accept this paper.

Authorsrebuttal2023-08-18

Thanks

We thank the reviewer for their response and for increasing their confidence score.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC