Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
This paper describes a family of optimal transport metrics that are efficiently computable. In particular, these are 1-Wasserstein distances in which the ground metric is tree-structured. The paper gives a closed form for the distance whose computation scales as the number of edges in the tree. Several facts about this family of OT metrics are presented, including negative definiteness. Empirical results investigate a "tree-slicing" scheme, in which multiple trees are constructed and the resulting distances are averaged, showing favorable tradeoffs between computational cost and performance in several statistical tasks. Overall I like this paper and believe it should be accepted. Strengths: 1. The paper is written fairly clearly and is easy enough to understand. 2. It is a nice idea, exploiting tree structure in the ground metric, and yields an efficient algorithm. 3. The empirical results in Section 6 demonstrate the utility of tree-slicing. Weaknesses: 1. It appears in Sections 6.1 and 6.2 that the tree-sliced Wasserstein distance outperforms the original optimal transport distance, which is surprising. Could you explain why this occurs? 2. The proof in the main text of Proposition 1 looks more like a proof sketch, particularly as the existence of a function f having the property you claim isn't immediately obvious. Could you include (in the supplement, at least) the full proof? --- UPDATE: I have read and I appreciate the authors' response. I will not be changing my score.
The paper introduces the three-slice Wasserstein metric as a generalization of the well known sliced-Wasserstein metric. The new metric can be computed very efficiently and it is more flexible than the sliced metric. The underlying idea is simple and appealing and it can potentially have a sizable impact on the generative modeling and variational inference. The paper also introduces a new positive-definite kernel. In general, I think that the contributions of the paper are relevant and original. The writing is clear and the basic ideas are properly presented to the reader. However, I have some concerns concerning the theoretical results and the experimental analysis. Major comments: - Most of the theoretical results (proposition 1 and 2) are not original as they follow directly from previous work on the tree OT metrics. In my opinion, the bound on the Wasserstein distance given at line 185 is an interesting result as it partially justifies the minimization of the sliced-tree metric as a proxy for the minimization of the intractable Wasserstein metric. However, this result should be complemented with further analysis concerning the tightness of the bound and the limiting behavior for the number of trees tending to infinity under several possible tree distributions. Specifically, the properties of this limit under the hypercube sampling model deserves further investigation. In absence of theoretical result, it would be useful to run a numeric analysis comparing the limit of the hypercube tree-sliced metric to the Wasserstein metric and the sliced Wasserstein metric. - The experimental section is fairly well done. However, the first experiment on topological data analysis is a bit to esoteric and it is not easy to interpret for the many readers not familiar with the field. All experiments concern with the tree-sliced kernel but most of the paper is concerned with the metric itself. It would have been very useful to have an experiment where the metric is used in a more conventional OT problem such as color transfer or generative modeling. Minor comments: - The results of the topological analysis are scattered across too many figures. I would suggest to organize them on a single multi paneled figures. - The word embedding experiment should probably be presented before the TDA experiment as it is easier to understand for a wider audience.
The paper suggests to replace the Wasserstein distance in applications with a "tree-sliced" variant, which embeds the points into a tree metric and computes Wasserstein on the tree, which takes a closed-form solution on tree metrics and can be easily computed. In particular, the Wasserstein distance on the tree is an l_1-distance between the pointsets, and then the paper suggests to use the associated Laplacian kernel, i.e., exp(-l_1-distance). The application to Wasserstein distance is natural and also generalizes the "sliced approximation" of Wasserstein which uses 1-dimensional embeddings of the pointsets (i.e., path metrics instead of tree metrics). It should be noted however that in a broader context, embedding the input pointset in a tree metric (or average of trees) and solving the problem on the tree (where many problems become dramatically more tractable) is a standard approach that and has been applied to countless metric problems, with great success. This is not to say that instantiating it for Wasserstein distance is less significant, but to put originality in context. What is most troubling is that the paper seems to be completely unaware any literature of embedding points into a distribution over tree metrics, and claims some standard and well-known techniques and novelties. This has been a vast and prominent field of research for over two decades with countless papers; some classic references are [1,2,3], but there are many more. In particular, "algorithm 1" is an extremely standard and well-known construction called a quadtree. It is also well-known that shifting the initial hypercube at random (equivalent to your random expanding of it) preserves the distances in expectation with bounded distortion, i.e., that averaging over several such trees is a provably good probabilistic embedding (eg. [4, section 2.3]). I don't see why such a classic notion needs to be laid out in full in the main text and presented as a contribution; I can only assume this is a case of unaware "rediscovery". Indeed, what the paper calls "partition-based methods" form the bulk of literature for bounded-dimensional pointsets, and "clustering-based methods" form the bulk for general finite metric spaces  and in particular high-dimensional ones . This part of the paper appears to warrant substantial revision. Moving on to a different topic, I found some parts of the paper overly obscure and unclear. In particular: 1. Definite-negativity is mentioned and highlighted so many times, that the notion should probably be defined in the paper and not just by reference, and also perhaps explain why is it important to you. Is this to ensure that the kernel is positive-definite? 2. I am unclear on where slicing comes into play. The TW kernel is defined for one tree, and afterwards you define STW citing empirical motivation. So, which kernel did you actually use in the experiments -- exp(-TW) or exp(-STW)? If it's the former, then why define STW, and what are the empirical considerations mentioned in line 107? If it's the latter, then why is it apparently not mentioned (and the kernel is called k_TW and not k_STW)? Is there a reason to use one and not the other? (I'm noting that STW is also negative-definite, simply because the average of l_1-metrics is an l_1-metric -- right?) 3. The description of the clustering-based tree construction (line 141) is too obscure. How do you set the number of clusters in each level? What is the connection to fast Gauss transform (why is it more related than any other clustering method)? Why do you use further-pair clustering (a.k.a k-center) instead of, say, the more standard k-means, or any other method? Does it have downstream motivation and does it effect the results? Conclusion: The upside is that idea at the base of this paper, while simple, is nice and potentially useful, and the experiments show advantage. I do like the suggested approach and find it interesting. On the downside, the paper has a significant issue with wheel-reinventing and apparent unfamiliarity with relevant literature, and some overly obscure parts in the presentation. ==Update== I have read the authors' response. It was somewhat difficult to understand and I am unsure of the extent to which it acknowledges the issue of duplication of prior work. I emphasize that there are prior works on embedding specifically OT into tree metric: For example Indyk-Thaper , ref  in your submission, with the same tree-of-hypercubes construction, l_1-embedding of the resulting tree metric, and application to nearest neighbor search. See in particular their "our techniques" section and more references therein (eg. Charikar'02). I would strongly advise to revise the paper so as to correctly position it in the existing literature with due attribution, and highlight its actual contributions which are--as I see it--exploring the empirical application of these techniques to real-world classification tasks (whereas the aforementioned prior work was more theoretically-minded), and in particular using the exponential kernel associated with the l_1-embedding of the tree metric. References:  Y. Bartal, Probabilistic approximation of metric spaces and its algorithmic applications, FOCS 1996.  M. Charikar, C. Chekuri, A. Goel, S. Guha, S. Plotkin, Approximating a finite metric by a small number of tree metrics, FOCS 1998.  J. Fakcharoenphol, S. Rao, K. Talwar, A tight bound on approximating arbitrary metrics by tree metrics, JCSS 2004.  P. Indyk, Algorithmic applications of low-distortion geometric embeddings, FOCS 2001.  P. Indyk, N. Thaper, Fast image retrieval via embeddings, SCTV 2001.