NeurIPS 2020

Design Space for Graph Neural Networks

Review 1

Summary and Contributions: This paper systemtically summarizes (1) a general GNN design space, (2) a GNN task space with quantitative similarity metric (3) an efficient and effective design space evaluation. Use this framework, they conduct GNN architecture search over a wide range of tasks, and the model discovered by them achieved sota results on large-scale benchmark.

Strengths: (1) Comprehensive and systemic summarization of existing GNN design space and tasks. (2) Get several interesting key findings that are beneficial to the communities. (3) The code is provided.

Weaknesses: (1) Some essential design choices are missing in the design space selection, such as different aggregation operations (for example, which attention mechanism to use, additive or product), pooling strategies (for example, diff pool and set pool), sampling strategies, etc. Even if the authors don't want to cover any part of these components, it would be better to discuss them and let it as future work. (2) Though the authors claim it achieved best result on the ogbn leaderboard, it would be better to add comparison with some existing standard GNN architectures. More quantitative evaluation on the proposed architecture search framework is needed. ================================================================================= Update: The authors address many of my concerns in the rebuttal, so I decide to raise up my score. But I still have some comments: It's interesting to show the comparison of attention machnism (not sure whether the product-based attention is implemented as that in Transformer, with KQV transformation). Also, I still think it would be better to also add pooling into the design space in the later version. The comparison with default architecture is very interesting, looking forward to seeing that in the later version.

Correctness: Yes, they are correct.

Clarity: Yes, the paper is well motivated and written.

Relation to Prior Work: Yes, the related work is well summarized.

Reproducibility: Yes

Additional Feedback:

Review 2

Summary and Contributions: ================ POST-REBUTTAL The rebuttal has addressed some of my concerns. The absence of theory guiding the search is really an issue not specific to this paper, but with the area. Yes, 1000+ GNN papers but only a handful of them are actually worth reading. The authors should add the “Issue of multiple hypothesis testing” remark in the final version. It will help the literature. I will be raising my score. ================ This work explores a few design options of current graph neural network architectures and reports the options that tend to reliably give better GNN architectures across most tasks.

Strengths: * The community has been proposing a few options in the design space but each paper amounts to a very small contribution. It is good to see a paper that tries to take a more systematic approach, effectively combining small contributions into a larger one. * Looking at the design decisions that give consistent improvements across architectures and tasks is a great idea. * The evaluation methodology is very good. I particularly enjoyed the violin plots and the similarity matrix between tasks. The proposed task similarity metric shows that there are only effectively three tasks in all these datasets. * It also shows that task transfer occurs on tasks that a similar but not on tasks that are different. * Significance and novelty of the contribution: The contribution is not novel. It is relevant since no other work provides a systematic evaluation.

Weaknesses: * Theoretical grounding: There is no theory guiding these design principles. There are a few tasks that test the expressiveness of neural networks. I guess all these GNNs would fail in the expressiveness tasks. * The GNN layer equation, described in the equation at the bottom of page 3, does not use the v’s own embedding. That is a strange choice, since it significantly reduces the expressiveness of the GNN. It would be good to see that as yet another hyperparameter for the model. Moreover, edge features / attention is not described by the framework, even though GAT is mentioned in the beginning of the paper. * There is no mention to the issue of multiple hypothesis testing. Anyone trying to test these many hypotheses should use a family-wise error rate correction technique to control the false discovery rate. Something simple like Bonferroni correction or Holm's step-down procedure should be described.

Correctness: The work is straightforward. There is nothing obviously wrong with the approach.

Clarity: The paper is well written.

Relation to Prior Work: The work only encompasses a small set of graph representation learning methods. Some of the more wild designs like WK-k [1], Tensor Equivariant NNs [2], and Relational Pooling [3] are not contemplated, but should be described as related works with alternative designs to message-passing GNNs. [1] Morris, C., Ritzert, M., Fey, M., Hamilton, W.L., Lenssen, J.E., Rattan, G. and Grohe, M., 2019, July. Weisfeiler and leman go neural: Higher-order graph neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 33, pp. 4602-4609). [2] Maron, H., Ben-Hamu, H., Serviansky, H. and Lipman, Y., 2019. Provably powerful graph networks. In Advances in Neural Information Processing Systems (pp. 2156-2167). [3] Murphy, R.L., Srinivasan, B., Rao, V. and Ribeiro, B., 2019. Relational pooling for graph representations. Proceedings of the International Conference on Machine Learning (pp. 4663-4673).

Reproducibility: Yes

Additional Feedback:

Review 3

Summary and Contributions: The paper summarizes the hyper-parameters currently included in GNN, which is called design space by the author. And proposes: (1) method to find the best hyper-parameters ; (2) metrics to measure different tasks (datasets more precisely) similarity; (3) method to find optimal hyper-parameters according the tasks similarity. All of them have a certain contribution to the research community of GNN.

Strengths: (1) This paper comprehensively summarizes hyper-parameters in the current general GNN , and systematically divides them into three categories. (2) This paper proposes a method to measure tasks (datasets) similarity from an experimental point of view: to judge data-sets' similarity through the performance comparison on several anchor models. Although this method has been applied in other fields (such as NLP) for a long time, as far as I know, this paper is the first to apply this method to GNN. (3) The article does many experiments to show: (a) the overall performance of each hyper-parameter; (b) the task similarity proposed in the paper has certain guiding significance for the selection of hyper-parameters. Based on the above, this paper is a relatively complete work, and it is experiment-driven.

Weaknesses: On the whole, the biggest weakness of this paper relies on experiments too much, and lacks deep insights. The relatively new insight in this paper is: for the first time, the method of using the anchor model to judge the similarity of tasks is introduced to GNN, and the similarity is connected with the selection of hyper-parameters. However, this method of measuring similarity is experimentally driven and not elegant, nor can it accurately give the similarity. In addition, the paper gives conclusions on the effectiveness of many modules (such as dropout, BN-layer) based on experiments in section 7.3. But there is no theoretical explanation for these conclusions. For example, the GIN paper ("HOW POWERFUL ARE GRAPH NEURAL NETWORKS?") argues that injectiveness should be guaranteed as much as possible in the graph classification task, so the aggregation function chooses the sum function (this is consistent with the experimental conclusion of the author of this paper). If the author can explain these conclusions convincingly, this paper will be much more valuable.

Correctness: Yes

Clarity: Yes, the paper is presented well and easy to understand. Some small writing mistake: Line 121, it's 'inter-layer', rather than 'intra-layer'.

Relation to Prior Work: Yes

Reproducibility: Yes

Additional Feedback: --After reading author response---- This paper is experiment-driven, aiming at gaining empirical knowledge about GNN through experiments. Despite the lack of sufficient theoretical support and complete design space including attention mechanisms, the experiments designed are relatively adequate, and some interesting conclusions have been drawn. On the whole, this paper is a relatively complete submission: first, the design space is defined, then experiments are carried out and many interesting conclusions are drawn, and finally, the task similarity is defined to help new tasks quickly find the optimal hyperparameters. Furthermore, no paper has done similar work before, and this paper can promote the development of the GNN field in my opinion. Therefore, I think this paper should be published. As for the drawbacks mentioned above, I think it can be used as future work. About theoretical analysis: This paper draws many interesting conclusions. Some have already given explanations in previous work, more have not yet. The explanations for these conclusions are very valuable. About task similarity: Assuming that the metrics given in this paper are correct, it’s meaningful to introduce statistical information of the graph data (such as the distribution of degrees) to analyze which graph tasks are more similar and give convincing explanations. I have increased my score to 7.

Review 4

Summary and Contributions: The work presents a design space of GNN as while as a GNN task space. The built spaces enable a quick transfer to find the best GNN model for a new coming GNN task. === POST REBUTTAL REVIEW Thanks for the author's effort in the rebuttal. The additional experiment result looks cool. I increase my score to 8.

Strengths: The paper is novel. Instead of optimizing GNN for one specific task. It takes the step back and trying to build a space for the GNN task and looking for a systematic way to discover best GNN model for all tasks. Good perspective. The technique is sound. The way it uses ranking to evaluate the model’s performance and the design of task similarity based on the anchor model’s performance are reasonable and empirically validated. The paper also provides a code platform, UBIGRAPH, which facilitates the GNN model design for the practitioners. It is very significant. Finally, the paper is clearly written.

Weaknesses: === diversity of the space === The GNN design space proposed in the paper is still in the range of the common GNN structure. It limits the system to “invent” GNN which beyond handcrafted GNN designs. Also, some existing GNN structures such as attention[0] are not included in the current design space. Similarly, the diversity issue also exists in the task space. Currently, the task space includes 32 tasks but as the later analysis shows they mainly lie in two groups. For example, one common GNN task, link prediction is not considered in this task space. I totally understand building a system to including everything is hard. I guess the current system already requires a significant amount of effort. Appreciate that. == comparison to the NAS === In the case study, the paper shows the system can be leverage to find a good GNN model a new task. I am curious about the result of applying NAS (such as auto-GNN[1]) on this ogbg-moihiv dataset. I would like to see the difference between the performance as while as the computation power and model search time required. Including such comparison would make us understand the pros and cons better about the system. [0] graph attention networks [1] Auto-GNN: Neural Architecture Search of Graph Neural Networks

Correctness: yes

Clarity: yes

Relation to Prior Work: yes

Reproducibility: Yes

Additional Feedback: