Paper ID: | 1742 |
---|---|

Title: | Scalable Gromov-Wasserstein Learning for Graph Partitioning and Matching |

In this paper, the authors propose a framework for scalable graph partitioning and matching. They use the pseudometric Gromov-Wasserstein in a scalable setup by utilizing a recursive mechanism. They provide detailed analysis of the complexity of the method. They justify the design of this method and they cover the related work in depth. In the detailed experimental setup, they mention all the comparison methods that are being used, the real and synthetic data, as well as the metrics that those methods will be evaluated. In their experiments for the graph partitioning, they focus both on analyzing the complexity/runtime for their method and the comparison methods, and they provide detailed results on a variant of mutual information metric. For the graph matching task, again the authors provide an analysis for node correctness metric and runtime for their method and the comparison methods. Overall, the paper is about a problem interesting for the NeurIPS community and the authors propose a novel setting to use the Gromov-Wasserstein distance in the graph matching and graph partitioning tasks in a scalable way. The paper is well-written and each claim is well supported by the authors comments', proofs or references. The reviewer has only concerns regarding the experiments and the metrics used in the analysis (see below for metric recommendations).

=============== Post-response Update: I thank the authors for their response. As I pointed out in my original review, I think this is an interesting (if somewhat limited in novelty) work, therefore, I maintain my score, and recommend acceptance on the understanding that: 1) The additional results and modifications mentioned in the rebuttal are included in the final version (in particular, details about the node measure). 2) The redundancy pointed out by R3 is discussed in the final version. 3) The "initialization" strategies section is updated and clarified to reflect the additional information provided in the rebuttal =============== Summary: This paper investigates the use of the Gromov-Wasserstein (GW) distance for large-scale graph partitioning and matching. The GW distance produces as a byproduct of its computation a transportation coupling, which can be used to infer node-to-node correspondences between graphs. In addition, the optimal transport framework has the appealing property of generalizing to multi-distribution comparisons thorough barycenters, which is exploited in this work to yield joint multi-graph approaches to partitioning and matching. Instead of a the more traditional projected gradient descent approach, the authors rely on a regularized proximal gradient method to compute the GW distance and barycenters. In order to scale up to large graphs, they propose a recursive divide-and-conquer approach. Various experiments on benchmark graph/network partitioning and matching tasks are performed, showing that the proposed method compares favorably (both in terms of accuracy and runtime) to various popular baselines. Strengths: - Strong theoretical foundations (the Gromov-Wasserstein distance) to a task often approach with heuristic methods - Superbly written paper: clear and concise argumentation, easy to follow, and a pleasure to read - The thorough experimental results, which show that the proposed approach is effective and efficient in practice - Rigorous and comprehensive review of computational complexities of the proposed + alternative methods Weaknesses: - Limited novelty of methods / theory Major Comments/Questions: 1. Novelty/Contributions. While GW has been used for graph matching repeatedly in previous work (albeit for small tasks - see below), I am not aware of other work that uses it for graph partitioning, so I would consider this an important contribution of this paper. It should be noted that most of the individual components used in this work are not novel (the GW itself, its application to graph matching, the proximal gradient method). However, I see consider its main contribution combining those components in a coherent and practical way, and producing as a consequence a promising and well-founded approach to two important tasks. 2. The Scalable method. While recursive partitioning methods are a common fixture of discrete optimization (and thus its use here provides limited novelty), it is satisfactorily applied in this case, and it seems to work well. My main concern/question about this component its is robustness/reliability. Recursive partitioning methods are prone to be unforgiving: I a wrong decision is made in the early stages, it can have disastrous consequences downstream as there is no possibility to revise poor early decisions. This seems to be the case here, so I would be interested to see if and whether the authors observed those catastrophic early mistakes in their experiments, and whether a best-of-k version of their method (e.g., like beam search for sequence decoding) would be able to lessen these. 3. Adjacency matrices. The paper continuously refers to the C matrices as adjacency matrices, yet they are defined over the reals (ie., not binary, as adjacency matrices usually are). I assume they are using soft edges or distances for the entries of these matrices, and that's why they are continuous, but the authors should clarify this. Also on this note, I would have liked to see some intuitive explanation of Eq (4), e.g., I understand it as a soft-averaged (barycenter-weighted) extrapolated similarity matrix. I would suggest the authors to discuss its interpretation, even if briefly. 4. The paper mentions node measures being estimated from node degrees (e.g. L58). How exactly is this done? 5. Initialization. I might be missing something, but I just don't see how the approaches described in the "Optimal Transports" and "Barycenter Graphs" paragraphs are initialization schemes. In particular, the former looks like a regularization scheme applied to every iteration. Could the authors please provide more details about these? 6. Complexity Analysis. I appreciate the thorough yet concise complexity analysis of related methods. This is sadly less and less common in ML papers, so it's encouraging to see it here. One comment: I would suggest reminding the reader what d is in Table 1. 7. Results. I have various minor comments about the experimental results: * The adjusted mutual information should probably be spelled out explicitly (perhaps in the Appendix). I suspect many readers are not familiar with it or don't remember its exact form (I didn't). * Fig 1(e) would be more meaningful with two-sided error bars * What is q%|V| in L273? Is this a product? * Why are there no results for MultiAlign for > 3 graphs? Was it because of timeout? Please mention this in the paper. * NC@1 and NC@all could be better explained - it took me a while to understand what was meant by these Minor Comments/Typos: - L83. "discrepancy" repetition - L107. Two arbitrary nodes - L138. Convergence is not properly linear, but nearly-linear - a detail, yes, but an important one.

While I think that the idea definitely worth it, I have some doubts about the fact that the paper is ready for publication. Indeed, it raises some questions that should be treated. Here is a list below. 1) The node distribution is set as the normalized node degree (as in [48]). This choice should be discussed in more details as it is not obvious and may be redundant with matrix C. Why a node with more connections should have a greater weight than others? Why the adjacency matrix not enough for enforcing nodes with similar node degree to be matched? Anyway, this choice deserves a detailed discussion in the paper. 2) The method is applicable to non-attributed graphs (this should be mentioned in the paper). Nevertheless, in section 3.1, authors provide a extra term in the GW formulation, C_node, that involves the differences between the 2 node distributions. The formulation then seems to come down to a GW term + a W term as in the fused Gromov-Wassertein method in [43]. Is this correct? If not, the differences should be highlighted. In addition, would it be possible to consider an other C_node matrix that would involve node labels? 3) Authors consider an entropic version of the GW distance, allowing a faster resolution of the problem. Nevertheless, at least for the graph partitioning problem, considering a non-regularized problem seems to be more intuitive: a node is matched to only one (in most cases) node of the disconnected graph. This is illustrated in figure 1b), where we would have preferred to see the first 3 nodes matching to one node and the 3 others to an other one, instead of having 2 nodes with splitted mass. Authors should justify this choice in more details. 4) the multi-graph partitioning scheme is unclear to me. Does it consists in i) first estimating the barycenter of all the graphs ii) then partitioning the barycenter? In figure 1d), it is not clear what does the transport matrices represent (transport to the barycenter or to the disconnected nodes?). The motivation behind the multi-graph partitioning should also be better explained. 5) A scalable algorithm is given in section 3.2. A discussion about how the results are close of the original solution, and in which cases it can be/should not be used is needed. Indeed, in the experiments, it leads to degraded performances, and this behavior should be better understood. 6) Algorithm in section 3.1 seems to be a direct extension of the algorithm provided in [48]. Originality of the algorithm should be better highlighted. Minor comments: - page 2 "we propose a GW learning framework to unify these two problems": the method proposed to solve these problems is the same but the two problems are definitely different. - regarding the density \mu: what happens if the graph contains isolated nodes? Are they discarded? - page 3 "the maximum in each row indicates the cluster of a node": what happens if some quantities are equal, as it seems to be in fig. 1b)? - page 3: the derivation of the node distribution \mu_dc is probably the most important quantity to be set and its computation details should not appear only in the appendix - for the graph partitioning problem, how do you choose the K value? - in several parts of the paper, assumption that the observed graphs have comparable size is made. Is this a reasonable assumption? **** UPDATE AFTER REBUTTAL**** Thanks for your feedback that I read carefully. It adresses some of my concerns (points 3-4-5-6). I believe that the choice of the node distribution, the cost matrix and C_node should be discussed in more details (all of these seem somehow redundant and some insight about how to set "good" definitions should be added). As such, I modified my score and set it to 6.