Layer-Neighbor Sampling -- Defusing Neighborhood Explosion in GNNs

Graph Neural Networks (GNNs) have received significant attention recently, but training them at a large scale remains a challenge. Mini-batch training coupled with sampling is used to alleviate this challenge. However, existing approaches either suffer from the neighborhood explosion phenomenon or have poor performance. To address these issues, we propose a new sampling algorithm called LAyer-neighBOR sampling (LABOR). It is designed to be a direct replacement for Neighbor Sampling (NS) with the same fanout hyperparameter while sampling up to 7 times fewer vertices, without sacrificing quality. By design, the variance of the estimator of each vertex matches NS from the point of view of a single vertex. Moreover, under the same vertex sampling budget constraints, LABOR converges faster than existing layer sampling approaches and can use up to 112 times larger batch sizes compared to NS.

Paper

References (36)

Scroll for more · 24 remaining

Similar papers

Peer review

Reviewer rBrD6/10 · confidence 3/52023-07-02

Summary

The paper proposes the LABOR sampling algorithm to address the NEP problem in GNN. LABOR combines layer sampling (LADIES), neighbor sampling (NS), and Poisson sampling. In contrast to prior works, LABOR calculates the sampling probabilities for importance sampling “iteratively” for each layer. Moreover, the Poisson sampling method (PLADIES) enhances the existing LADIES method by using “sampling without replacement” instead of “sampling with replacement.” Also, the authors derive the hidden layer variance of PLADIES and find that, under certain conditions, the variance can converge to 0, showing better convergence properties. Experiments show that LABOR has better sampling efficiency, using only 1/7 vertex to achieve the same F1-score as NS and using 112 times larger batch sizes during training.

Strengths

• LABOR has a better estimator of sampling with lower variances than prior work LADIES, which might not have a vertex with its neighbor among all sampled vertices. • LABOR has better vertex/edge efficiency than NS. As a result, LABOR can reach the same F1-score by using fewer sampled vertices under the same hyperparameter settings.

Weaknesses

The experimental evaluation is not sufficiently detailed. Some technical details are not clear.

Questions

Figure 1 shows that the vertices efficiency of LADIES/PLADIES might be higher than LABOR in Reddit dataset. Why?

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

3 good

Limitations

While the paper addresses an important topic, it has several issues that need to be addressed before publication: • LABOR outperforms NS in a graph with a high average degree (same as the assumption in the NEP problem), such as the Reddit dataset (average degree 493) shown in Table 2. However, when the average degree of the dataset is low (e.g., flickr: 10.09), the performance gap between LABOR and NS will be smaller. • LABOR is more often compared with NS in the paper, but LABOR is highly based on LADIES/layer sampling method. For example, the vertices efficiency only compared to NS in section 4.2, but Figure 1 shows that the vertices efficiency of LADIES/PLADIES might be higher than LABOR in Reddit dataset. • Neighbor Sampling (NS) is an important comparison subject in this paper, but it is not detailed in both Introduction and Background. The background simply states, “Neighbor Sampling: The neighbor sampling approach was proposed by Hamilton et al. [2017]” with a citation to GraphSage. Yet, section 3 describes that “Neighbor Sampling currently seems to be the most popular sampling approach.” The latter statement does not seem to refer to GraphSage. • The paper needs to report the improvement in the total training time explicitly, as the abstract only mentions “7 times fewer vertices” and “use up to 112 times larger batch sizes compared to NS”. • Table 2 does not give detailed hyperparameter settings. For example, how were the hyperparameters obtained? what’s the range for hyperparameter sweeping? LADIES paper shows a 5 times total time difference by only adjusting hyperparameters – LADIES (64) and LADIES (512) on the Cora dataset. Moreover, details need to be provided in line241-242 and Figure2’s caption – “The hyperparameters of LADIES and PLADIES were picked to roughly match LABOR-” • “same sampling budget” needs to be defined more clearly in Table 2.

Reviewer QdGo4/10 · confidence 2/52023-07-07

Summary

This paper proposes a new sampling algorithm called LAyer-neighBOR sampling (LABOR). The authors show the variance of the estimator is controlled. Experiments show this method is much more efficient in sampling and faster in convergence.

Strengths

1. The paper is well-written and clear. 2. Some analysis is provided. The experiments on some large graph datasets verify the advantages of the proposed method.

Weaknesses

1. Only variance reduction analysis is provided. Theoretical convergence analysis or generalization analysis is lacking. 2. This paper mainly compares with LADIES, which is a work in 2019. I don't think it is still a state-of-the-art work about graph sampling that is appropriate for comparison. Some more recent works should be covered in experiments for comparison.

Questions

1. The proposed method allows a larger batch size. This can make convergence faster. However, a larger batch size may not lead to a better generalization. This makes the statement of a faster convergence weaker.

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

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

There is no potential negative societal impact of this work.

Reviewer qFRg6/10 · confidence 5/52023-07-10

Summary

This paper improves the previous layer-wise importance sampling algorithm in GNN training by: 1) introduce poison sampling for debiasing; 2) it modifies the sampling importance by incorporating the probability of node-wise sampling, to make sure all nodes have similar neighbor nodes. It shows that the variance of each sampled nodes is similar to that of node-wise sampling, while enjoying some benefits of efficiency of layer-wise sampling. It shows that with similar budget, the proposed method converges faster and more memory efficient.

Strengths

1. It systemetically studeis the limitation of prior node-wise and layer-wise sampling methods, and introduce two remedy: 1) use poison sampling distribution for debiasing, and 2) add back node-wise sampling probability to reduce per-node variance. 2. Results show improvement against LADIES baseline.

Weaknesses

1. The only two baselines this paper use is neighborhood sampling and LADIES. It's better to include more sampling baselines including FastGCN, GraphSAINT, etc. 2. The paper cites "Calibrate and Debias Layer-wise Sampling for Graph Convolutional Networks", which provides several improvement to LADIES. It seems that the problem this paper tries to solve has many similar part of this paper. It's very important to discuss the difference and add the method (flat & debias) of this paper into evaluation. 3. Regarding training speed & time cost, it's better to include a comparison on total time required for the model to converge. Currently the paper only includes iteration/s, but not how many iterations are needed. It's better to show the total time. 4. It's highly recommended to add a psudo-code to show how to implement the proposed LABOR, to make readers understand the method easily.

Questions

See the concerns in the weakness part. Also, it seems that the pdf format of the submitted paper does not support selecting & searching text, which makes it hard to find some part in the paper. Better to correct this part. Even after reading the paper, I still think the title of this paper is a bit misleading and unclear. I recommend the authors to reconsider this title and highlight your contributions in the title.

Rating

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

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

No limitation provided in the main paper.

Reviewer 99pb5/10 · confidence 4/52023-07-10

Summary

This paper proposes a novel neighbor sampling (NS) method, called layer-neighbor sampling (LABOR), for training GNNs on large-scale graphs. Unlike the representative NS method, which samples neighbors for each node from a uniform distribution, LABOR samples neighbors for each node using Poisson sampling from a modified distribution. This modified distribution is optimized to match the variance of LABOR with that of NS for each single node. As a result, LABOR can reduce the number of sampled nodes by a large margin compared to NS. Experiments demonstrate the effectiveness of LABOR.

Strengths

Originality: The proposed method is novel in applying Poisson sampling and its effort to optimize a modified distribution. Quality: This paper falls within the borderline category, and introducing some in-depth analyses could improve its quality. Clarity: While some unexplained notations and problematic statements exist, this paper is generally easy to follow. Significance: Neighborhood explosion is an important problem in GNNs. The proposed method serves as a complement to existing neighbor sampling methods. However, it appears that the computation and memory complexity of the proposed method still exhibits exponential growth with respect to the depth of GNNs.

Weaknesses

1. Including complexity analysis can help readers to understand how much the proposed method improves over baselines. 2. Wall-clock time is a crucial concern in the efficient training of GNNs. Since the proposed method introduces additional time overhead during neighbor sampling, it may not be advantageous in wall-clock time, despite its superiority in batch size. 3. Baselines for comparison are somewhat old and not state-of-the-art. Introducing some state-of-the-art baselines for comparison may be more convincing.

Questions

1. Equation (8) appears problematic since the variance is independent of the number of sampled nodes, i.e., $\tilde{d}_s$ or n. If I am wrong about this, please correct me. Considering Equation (8) is fundamental in their follow-up analysis, it is important to explain this. 2. In the first row of Figure 2, the proposed method behaves strangely in the last figure (flickr). Is there any explanation for this phenomenon?

Rating

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

Reviewer gLfM6/10 · confidence 4/52023-07-26

Summary

This paper studies the problem of improving the efficiency of GNN training. It first points out that the node-based sampling methods that sample independently ad recursively for each node suffer the neighbor explosion problem (NEP), while the layer-based sampling methods that sample for the whole layer collectively cannot guarantee the approximation of each vertex would be good enough. To address these issues, this paper proposes the LABOR sampler which combines the advantages of node and layer-based sampling using Poisson Sampling. The main idea is to conduct neighbor sampling by letting individual vertices make correlated decisions while sampling their neighbors, and introducing a hyperparameter on each node to afftect the expected number of the sampled neighbors. The sampling algorithm would optimize over the sampling probabilities and the newly introduced hyperparameter, s.t. can sample the minimum number of vertices in an unbiased way. Experiments show that LABOR can reduce the sampled nodes and edges, and thus enjoys better efficiency.

Strengths

1. The idea of letting the algorithm decide the minimum neighbors to be sampled makes sense and works well. 2. According to the experimental results, the proposed method works well in reducing the sampled nodes and edges while maintaining the same-level performance as baselines. 3. Code is provided in the supplementary.

Weaknesses

1. It would be more clear if the authors can put the main algorithm in the main paper instead of in the supplementary. 2. The authors only report the it/s (which can also indicates the run-time per iteration), it would be better if the authors can also report the total run-time or the number of training epochs needed to get converged.

Questions

Q1: For Fig. 1, what is the ``cumulative # sampled nodes/edges``? Is it linear to the number of training iterations? Q2: How would different fanouts affect the performance?

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

Please refer to "Weakness".

Reviewer qFRg2023-08-12

Response

1. I highly suggest authors discuss the difference between "Calibrate and Debias Layer-wise Sampling for Graph Convolutional Networks" more clearly, and add them into your paper (in the revised version). As NOT all readers are experts in this field and are familiar with these literatures. 2. Again, I think convergence time is very important (even more important than #steps) and shall be put in the main table, with all different configs. Different baselines have different time per step, so comparing #steps is not enough. 3. Current Alg 1 is still super unclear. E.g., how do you solve eq 14? what is each notation means? what is the output and how it's used for sampling? The current form is super hard to understand. The reason I think "Defusing Neighborhood Explosion" is not a good title is that I think most existing layer-wise sampling does not have neighbor explosion issue? As long as they control #nodes to be sampled per layer. Could the authors highlight your difference to these methods in the title? (Also "existing approaches either suffer from the neighborhood explosion phenomenon or have poor performance" seem to be a very strong criticism to other layer-dependent methods, but at least in table 2, I didn't see significant improvement to LADIES, and you didn't compare with "Calibrate and Debias Layer-wise Sampling for Graph Convolutional Networks")

Authorsrebuttal2023-08-12

Followup

1. We will revise our paper to include details of the difference between our proposed PLADIES algorithm and [1]. Let us talk about some of those differences in the following 2 paragraphs: Authors of [1] noticed the same bias issue as us, and propose a way to "de-bias" sampling without replacement by computing actual sampling probabilities in quadratic time (Mentioned in [1] Section 5.5). However, PLADIES simply takes linear time because we use Poisson sampling, and with Poisson sampling the end probability of being sampled is simply the probability of the Bernoulli trial of success, simply $\pi_t$. When one uses weighted sampling without replacement as in [1], the sampling probability is actually not the weights provided initially and [1] proposes a way to compute the actual probabilities and a way to de-bias them in quadratic time. There is no convergence difference when it comes to PLADIES and the de-bias procedure proposed in [1], only the computational complexity of sampling is different. They also propose a way to flatten the probabilities and these flattened probabilities can be fed to PLADIES as well. So the flattening approach [1] proposes is orthogonal to our work. To give an example of why weighted sampling without replacement is biased, consider the probabilities [0.7, 0.2, 0.1]. Now, if we want to sample 2 items without replacement, then notice that the probability of the first item, 0.7, is more than 2 times the rest of the items (0.3). Thus, the first item needs to be sampled always. But even if we always sample the first item, the second item is going to be either 0.2 or 0.1, the probability of which one it is going to be is [2/3, 1/3]. Thus, in the end, the probability of sampling individual items is [1, 2/3, 1/3]. However, these resulting probabilities are not linearly proportional to the probabilities we started with. [1] fixes this in quadratic time. We propose to use Poisson Sampling instead. With Poisson sampling, we can have a constant $c$ to adjust how many items we want to sample. We want 2 items in expectation, so $0.7c + 0.2c + 0.1c = 2$. But the solution is $c=2$, and the probability of sampling the first item becomes 1.4, which is not possible. So, we need to solve for $c$ in the following equation: $\min(0.7c, 1) + \min(0.2c, 1) + \min(0.1c, 1) = 2$. The solution is $c=10/3$, and this can be found in linear time. The resulting probabilities become: $\min(0.7c, 1) = 1$, $\min(0.2c, 1) = 2/3$, $\min(0.1c, 1) = 1/3$. Because we use these adjusted probabilities during GNN aggregation, our method is unbiased, and faster than [1] due to computational complexity difference. 2. Note that GNN training runtime depends on how many vertices or edges are processed. Thus, Figure 1 portrays time to convergence in a hardware-agnostic way by scaling the x-axis by the number of processed vertices or edges. All sampling algorithms compared have linear time complexity in our paper. We will add a version of Figure 1 where the x-axis is the wall time. 3. We will improve the clarity of Algorithm 1 by adding more details about variables and how they are computed in the camera-ready version. Section 3.2.2 and Equations 15, 16, and 17 describe one way of solving Equation 14 in linear time. The output is the sampled set of edges $E'$ along with any computed edge weights $A'$ derived from the importance sampling probabilities to be used as edge weights during aggregation (Equation 4a or 4b) to make the estimation unbiased. Table 4 in [1] compares two different configurations of LADIES, "LADIES" uses a fixed number of sampled vertices per layer (512, 512, 512) and "LADIES (2)" when sampling 2 times more than the previous layer (512, 1024, 2048). When we look at their results, it is clear to see that LADIES (2) outperforms LADIES on all datasets. This means that LADIES doesn't work well with fixed numbers of vertices per layer and it doesn't solve the neighborhood explosion issue by doing that because it performs poorly as we mentioned in our paper. What we do in our experiments instead is to follow LABOR-* , when choosing the LADIES variants' number of sampled vertices per layer, as can be seen in Table 2 of our paper. As a result, it is clear to see that LADIES (PLADIES especially because it is unbiased) performs reliably well, matching LABOR and NS variants in accuracy on all datasets. Thus, our work suggests how one should pick the hyperparameters of LADIES variants in a suitable manner, and that is by looking at LABOR-* rather than simply doubling the number of sampled vertices per layer. We think this is important because the graph topology can be very different for various applications. Before our work, how to choose the hyperparameters of LADIES in a manner that is robust to changing the underlying graph did not exist. [1] Yifan Chen, Tianning Xu, Dilek Hakkani-Tur, Di Jin, Yun Yang, Ruoqing Zhu, "Calibrate and Debias Layer-wise Sampling for Graph Convolutional Networks", TMLR 2023

Reviewer qFRg2023-08-12

Response

1. The mentioning of results of Table 4 in (1) is good to support your claim, and it shall be added into your paper with clear discussion (as it's related to your title and main argument). Also I'm not pretty convinced a linear increase of sampling node shall be claimed as "explosion", as this is what we mainly call for exponential increasing. 2. LABOR definitely has many good properties and design, which I'd encourage authors use plain language to summarize and highlight in the intro. 3. (and again, I'd suggest authors not using such subjective words as "poorly" in a research paper, even looking at Table4 in (1), the avg. performance gap between LADIES(1) and (2) seems not that significant. You could simply report the relative performance improvement of your approach compared with prior baselines, as well as NS) 4. In general response, you mention for some other sampling methods like VR-GCN and shadow, NS could be replaced with LABOR. This is an interesting claim, as previously people thought those methods are coupled with NS, and not for Layer-Sampling. Could the authors elaborate such possibility? (and again, would be helpful to be put in the paper) It's very important for the GNN community to accept that Layer-sampling is useful and flexible.

Authorsrebuttal2023-08-12

Followup 2

We appreciate the follow-up response! Please do note that for "LADIES (2)", [512, 1024, 2048] is not a linear increase but an exponential one, at layer $i$, they sample $512 * 2^{i}$ vertices. And it is not clear why one should stop with the exponent $2$, one can go ahead and use $3$ or $4$, etc. as long as there is an accuracy improvement. Similarly, one can choose smaller/larger fanout values for LABOR and NS algorithms if one wanted to change their iteration runtime, although the end accuracy will be slightly affected. > 3. (and again, I'd suggest authors not using such subjective words as "poorly" in a research paper, even looking at Table4 in (1), the avg. performance gap between LADIES(1) and (2) seems not that significant. You could simply report the relative performance improvement of your approach compared with prior baselines, as well as NS) To clarify what we had meant by the accuracy difference between "LADIES" and "LADIES (2)", Table 4 in the Calibrate and Debias paper has the following values: | | reddit | ogbn-arxiv | ogbn-mag | ogbn-proteins| ogbn-products| |--------------|----------------|----------------|----------------|------------------|-------------------| |LADIES | 73.86±0.17| 60.95±0.31| 24.79±0.48| 68.28±0.05| 52.97±1.11| |LADIES (2)|88.34±0.11| 64.01±0.39| 28.59±0.39| 68.17±0.10| 65.24±0.40| The accuracy difference between these two differs by up to 15\%. So, we wanted to say that the difference is in fact significant. And it is not clear that LADIES (3) or LADIES (4) will not be even better. We will revisit our introduction and highlight the good properties of our proposed sampler LABOR. We will also be careful with our wording per your suggestions when talking about related work. We would be more than happy to elaborate. Let us start with VR-GCN, and then talk about SHADOW and GNS. VR-GCN uses NS to sample a fixed number of neighbors (usually 2), and for the neighbors that weren't sampled, it uses their historical embeddings at each layer instead of recomputing them. If you look at our paper's supplementary material, you will see that Appendix A.3 details how to modify LABOR so that it samples a fixed number of neighbors in linear time, just like NS. If we replace NS in VR-GCN with LABOR-0 along with the method in A.3, then VR-GCN will run faster because there will be fewer vertices sampled while still sampling 2 neighbors per vertex. SHADOW [2] paper in its Appendix C describes how it extracts the subgraphs it uses during training. One option is to use an L-hop extract. It says that one can use NS to sample the $L$-hop neighborhood of a given minibatch of vertices, which then is converted to a subgraph to be used in training. Table 1 shows some experimental results with $L=2$, (shadow-SAGE 2-hop). If they had used LABOR-0 instead, their code would have run faster with no noticeable convergence difference as we demonstrate in our paper. For GNS, simply drop-in replacing LABOR-0 in place of NS will give us Global Layer-Neighbor (LABOR) Sampling. And we would expect similar runtime improvements there as well. GNS keeps a cache of vertices that they try to sample a fixed number of vertices from, and they replenish the cache periodically. If they can't find enough vertices in the cache, they sample from the rest of the graph. As you can see, there is nothing in this algorithm that requires you to use NS, and LABOR-0 is designed as a drop-in replacement for NS. Since LABOR-0 along with the method described in Appendix Section A.3 is almost indistinguishable from NS (But it is still technically layer sampling and takes advantage of neighborhood overlaps and samples fewer vertices), one can directly replace any occurrence of NS with LABOR-0. It is hard to even write a test to distinguish the outputs of the two algorithms, statistical correlation testing is required. However, extending other works and using LABOR-0 instead of NS is hard to fit into a single paper and we leave it as future work. We will include some of this discussion in our camera-ready version. [2] Zeng et al. Decoupling the Depth and Scope of Graph Neural Networks, arXiv:2201.07858

Reviewer qFRg2023-08-18

Response

Thanks the authors' responses. Adding these discussion and descriptions for previous baselines are important to be updated. I'd like to increase the score to "weak accept"

Reviewer QdGo2023-08-15

Thank you for your feedback. It is a pity that some theoretical analysis is not offered. However, the empirical results look good. I don't think the baseline issue (my Weanknesses 2) is solved or clearly clarified. I prefer to keep my current score, but I decrease my confidence.

Reviewer gLfM2023-08-18

Thanks to the authors for your response. My questions are addressed. I would like to keep my original score 6.

Reviewer rBrD2023-08-20

Thank you for the clarifications, especially regarding the experiments.

Reviewer 99pb2023-08-20

Thanks for the response

The authors have addressed most of my concerns. Hence, I raise my score by one level. But some concerns still remain unresolved. (1) The baselines used for comparison are somewhat outdated and not representative of the state-of-the-art methods. While the authors mentioned that LABOR is orthogonal to most state-of-the-art samplers, all these samplers aim to improve the efficiency of training GNNs and can be directly compared against each other. If the authors can address this issue, I believe this paper will be more persuasive. (2) It is better to summarize the complexity of different methods in an independent table without considering specific datasets.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC