NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:8678
Title:Multi-Criteria Dimensionality Reduction with Applications to Fairness

Reviewer 1

I found this paper difficult to read. Some of the reasons were structural - I found the main and readily usable contribution of this paper to be the actual algorithm( the multi-criteria dimensionality reduction algorithm). The problem is that this information is written in paragraph form (between line 259 to 269) instead of pseudo-code so that it is readily visible. I also did not like the naming and organisation of the sections. The authors used the chapter name 'results' to mean mathematical foundational results it might be more helpful to rename this chapter into something like 'theoretical background'. It might also be more useful to have this chapter after the related research chapter so that it becomes even more evident why that background theory is necessary and where is plugs into the current state of the art solutions. I am not familiar enough with the literature in this area to comment on originality of this work.

Reviewer 2

The paper makes interesting connection between fairness and multi-parameter optimization, bringing to bear powerful techniques (such as SDP relaxation) from the latter. It supports its theoretical findings with experimental validation whose results suggest that in practice the number of extra dimensions is even smaller than the analytical guarantees. The paper is notable for diversity of its tools (SDP relaxation, geometry of quadratic functions, reduction of MAX-CUT to Fair PCA, experimental results).

Reviewer 3

Originality: The authors’ solve a problem left open in a previous paper and made strictly improve on previous work for approximation algorithms. They do so by giving new insights into structural properties of extreme points of semi-definite programs and more general convex programs. As far as I understand, the algorithms presented in the paper are not substantially new, but the analysis of these algorithms is novel. Quality: The paper seems complete and the proofs appear correct. The paper tackles interesting problems and reaches satisfying conclusions. They essentially close the book on the k=2 case, make significant improvements for k>2 and leave open some questions for structured data. Clarity: The paper is well-written. The introduction is easy to follow. It does a good job introducing and motivating the problem. The contributions of the work are made obvious with clear theorem statements. I appreciated that the authors’ focus first on the simpler case of Fair-PCA. The use of this as a building block to understand the more general case is really helpful for the reader. The progression from Thm 1.1 to Thm 1.2 to Thm 1.3 made each result easier to understand. Significance: The problem they study, dimensionality reduction, is an important and common technique for data analysis. The authors’ provide convincing motivation for the multi-criteria variant of the problem. They focus on the fairness as a main motivation. The authors’ resolve an open problem. Samadi et al. introduced the problem of Fair-PCA and gave an approximation algorithm and left the computational complexity of finding the exact optimal solution as an open problem. The authors’ improve upon that approximation for all k (number of constraints) and solve the open problem for k=2. The previous interest in this problem suggest that it was be interesting to the community. The authors’ also make several contributions of independent interest. They give new insights into structural properties of extreme points of semi-definite programs and more general convex programs. This is not my area of expertise but the authors’ say that similar results have been used to study fast algorithms, community detection and phase synchronization so this does seem to indicate that these results may have applications elsewhere. They do seem like general enough results to be useful to other researchers. There are some avenues for future work. In their experiments indicate that the rank bounds presented in the paper (while better than those in previous papers) may be loose on real datasets. They prove in the appendix that it is not the case in general for k>2 that the SDP solution obtains the exact rank. However, this does seem to be the case in their experiments. This leaves open the question what what properties of the data would ensure this? Minor comments: - I thought the progression to Thm 1.4 was a bit harder to understand. It would be nice to have a formal statement of what you mean by an approximation algorithm for Fair-PCA. It’s sort of clear from the proceeding prose that it has to satisfy the rank condition but it would be nice to have this stated in the thm. - Similarly, you define the marginal loss function in words but it would be nice to have it defined in an equation. It would also be nice to have a reminder of this definition and some intuition for it because Figure 1. I found this figure a little confusing. - Integrality gap is a common term in the optimality community but may not be as common elsewhere. It would be nice to have a brief definition of it. - It would be nice to have pointers to where particularly proofs occur in the appendix. - Typo: line 183. Singular-> singular value - In line 186 it should be noted that Figure 2 is in the appendix.