NeurIPS 2020

Fast Epigraphical Projection-based Incremental Algorithms for Wasserstein Distributionally Robust Support Vector Machine

Review 1

Summary and Contributions: This paper studies the Wasserstein distributionally robust support vector machine problems and proposes two efficient methods to solve them. Convergence rates are established by the Holderian growth condition. The updates in each iteration of these algorithms can be computed efficiently, which is the focus of this paper.

Strengths: I did not know the Wasserstein distributionally robust optimization before. So I am not sure whether the problem studied in this paper is significant. 1. Two algorithms are proposed for the DRO problem, one is based on the projected subgradient algorithm and the other is based on the proximal point algorithm. The update in each iteration is implemented efficiently 2. Convergence rates are given by establishing the Holderian growth condition.

Weaknesses: 1. In Table 1, the convergence when c=0 is faster than the one when c>0. However, when c>0, the objective has an additional quadratic term than the case of c=0. From my experimence, adding a quadratic term always makes the algorithms faster. So Table 1 looks strange to me. It might better to give more intuitions and explanations. The sharpness condition may give a faster convergence than the QG condition. However, can the authors prove the convergence beyond the QG condition when c>0, for example, give a possible faster rate by exploiting the specification of the problem? 2. In Section 5.1, the authors claim that IPPA is slower than M-ISG. However, from my experience, the proximal point algorithm is always faster than the subgradient method in general. It might better to explain more. 3. Please confirm the Holderian growth condition is global or local.

Correctness: Correct

Clarity: Yes

Relation to Prior Work: Yes

Reproducibility: Yes

Additional Feedback: After rebuttal: Thanks for the response. It has addressed my questions.

Review 2

Summary and Contributions: The authors propose first order algorithms to solve Wasserstein distributionally robust support vector machine problems. The proposed approach heavily exploits the hidden structure of the resulting reformulation. In particular, when the transportation cost is the Euclidean norm, the authors show that they can solve the sub-problems in projected subgradient and proximal point algorithms analytically. Moreover, when the transportation cost is the \ell_1 or \ell_infty norm, they reformulate the sub-problems as a one dimensional optimization problems, which can be efficiently solved with a modified version of the secant algorithm. ================================= [Update after reading rebuttal]: My evaluation remains the same.

Strengths: I found the technical analysis of the paper very interesting. Numerical results also show that the proposed algorithms consistently outperform general convex optimization solvers by a great margin.

Weaknesses: The only weakness of the paper, in my opinion, is that the algorithms are mainly designed to handle the support vector machine problem. The extension of the proposed algorithm to more general piece-wise linear cost functions is very interesting, but I can understand that the authors have to restrict the problem setting due to lack of space or difficulty in subsequent analyses.

Correctness: I have checked appendix in the paper, and to the best of my knowledge, all proofs are correct. The empirical results are also consistent with theory.

Clarity: The paper is very well-written and well-organized. All steps are clearly explained and the motivation is crystal clear.

Relation to Prior Work: The focus of the current work is support vector machine, and its connection to the previous results are clearly made in the paper.

Reproducibility: Yes

Additional Feedback:

Review 3

Summary and Contributions: The authors have proposed two new algorithms to solve the timely and important distributionally robust SVM problem. These new algorithms are instances of incremental projected subgradient descent and incremental proximal point algorithm. The main novelty of ISG and IPPA proposed in this work is that they developed efficient algorithms (in linear time) for the subproblems in these solving strategies, which are interesting and potentially beneficial for other problems with similar structures. Besides, the authors carefully analyze the iteration complexity of their new algorithms under the so-called BLR condition. They show if the BLR condition is satisfied, the exponent in the Holderian growth condition could be explicitly determined. Thus, the proposed algorithms enjoy sublinear or linear fast rates in these settings. The empirical speedups appear to be substantial. Overall, this paper contributes a well-founded and empirically successful algorithms for accelerating solution of the distributionally robust SVM problem. This could be impactful for both practitioners and researchers, as it extends DRO methodology to large datasets in real machine learning tasks.

Strengths: 1. Fast solution of DRSVM problem is potentially useful and (to my knowledge) previously unaddressed. This work proposed two novel incremental algorithms (incremental projected subgradient descent and incremental proximal point algorithm) to tackle the DRSVM problem. 2. The efficient solution proposed to address the single proximal point update is interesting and highly non-trivial. To the best of my knowledge, this class of problem has not been well addressed before. 3. The convergence rates of the new algorithms are analyzed under the BLR condition for the set of optimal solutions. The theorem shows the new algorithms enjoy sublinear/linear rates, which is unknown before. 4. They conducted extensive experiments that show the newly proposed strategies significantly outperform existing approaches. 5. The authors do a quite good job of the literature review and provide detailed comparisons between theirs and existing techniques. The paper is fairly easy to follow, with the exception of the citation mismatch between main paper and appendix.

Weaknesses: 1. I think the authors might want to provide more details on the Bound Linear Regularity condition in Def.2 as BLR is the central assumption in the main Thm 4.3. I am especially interested in how strong is this assumption? That is to say, can we have some simple examples that satisfy BLR? For example, if g is strongly convex, is arg_x g(Ax) BLR? 2. In Figure 4, dataset a3a, it seems that the Hybrid strategy does not run as fast as GS-ADMM and YALMIP. Is there some reason for that? I am interested in that as in all the other setting the Hybrid run really fast. 3. It might be better if the authors could make their fast solvers available to the practitioners and researchers in the community. Minor points: 1. Figure 1: larger font and legend makes the figure clear 2. In Appendix, there might be some mismatching of the number of references, e.g., [15]->[16], [25]->[26].

Correctness: Seems correct to me.

Clarity: The paper is well written and in good style. Some typos should be fixed.

Relation to Prior Work: Good. The authors do a quite good job of the literature review and provide detailed comparisons between theirs and existing techniques.

Reproducibility: Yes

Additional Feedback: == After Response == Thank you to the authors for the response and I am satisfied with that.

Review 4

Summary and Contributions: This paper presents two efficient optimization algorithms for Wasserstein distributionally robust support vector machine. The authors propose to jointly optimize the learning parameter w and its l_q-norm upper bound \lambda, where l_1, l_2, and l_infty norm-induced transport costs of w are considered. The two optimization methods are based on epigraphical projection-based incremental algorithms: one is incremental projected subgradient descent (ISG) method, and the other is incremental proximal point algorithm (IPPA). Theoretical analysis is shown for ISG and IPPA on linear/sublinear convergence rates. ================ [After reading the rebuttal] I have read the authors' feedback and other reviews. I agree with other reviewers on the theoretical contributions of this paper, and the rebuttal addressed my previous concerns. I updated my score from 5 to 6 accordingly.

Strengths: The proposed method is well motivated and theoretically grounded. The authors provide both theoretical proof and empirical results to support the convergence behavior of both ISG and IPPA. The paper is well written. The summary on the properties of DRSVM problem under different conditions is clear, and the proposed incremental algorithms for improving optimization efficiency is valid.

Weaknesses: My major concern is on the empirical results. Maybe I missed it, but it seems not quite clear to me why the comparing methods are selected under different conditions. For example, why GS-ADMM results are not shown for l_1 and l_2-norm optimization, i.e., Table 2 and 3? It would be helpful to show the objective function curves of the comparing method in Fig. 1. How the parameters c, κ, and \epsilon determined in the experiments?

Correctness: The claims and method look correct to me. More implementation details can be provided to make the empirical results more convincing.

Clarity: The paper is well written.

Relation to Prior Work: The difference from previous contributions is discussed.

Reproducibility: No

Additional Feedback: