- Rating
- 2
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
This paper proposes an algorithm for incomplete multi-view clustering, a problem where objects are represented through multiple views (multiple modalities, different embeddings, etc), each with its own intrinsic geometry. The authors use the Gromov-Wasserstein formulation of Optimal Transport for solving this problem, most importantly through the Gromov-Wasserstein distance. Empirics of the paper are strong, with important gaps with respect 2nd best across various datasets. Nonetheless, as covered below, there are non-trivial intersections with prior art that are not-so-well discussed in the paper.
Critical strengths
The main strength of its paper are its strong empirical results, and comprehensive benchmarking. As a second strength, I found the paper generally well written and easy to read.
Critical weaknesses
My main problem with this paper is how it discusses its relationship with prior state-of-the-art. In particular, I think that the intersections with (Peyré, Cuturi and Solomon, 2016), henceforth denoted PCS16, are non-trivial and should have been more well discussed.
For convenience, I refer the authors to the following equivalences between their formalism and that paper. Note that these equivalences cover most of their Section 3 (3.1. Partial GW Coupling, 3.2. Unified Topology Integration and 3.3. Optimization Strategy).
| Submission 21449 | PCS16 | Difference |
|---|---|---|
| Eq. (4) | Eq. (7), the objective | Partial GW plans, Penalty |
| Eq. (6) | Eq. (11), joint objective over and | Minimization over |
| Eqs. (9) and (12) | Equations 13, and 14 | per-view marginals , consequence of partial GW |
| Eq. (10) | Eqs. (8)–(9), | KL projection onto instead of set of mass-preserving transport plans |
| BCD (Sec. 3.3) | Algorithm 1 | additional optimization over barycentric coordinates |
Analysis. In general, there's a substantive intersection with prior work that is not well discussed in the paper. For instance, while the authors do cite PCS, the only place that this is done is in Section 2, Related Work, around lines 153/154. Based on the table above,
-
The loss in (4) is the GW loss with variables , and , corresponding roughly to equation (7), with / understood as an additional variable, and the marginals fiexed. I concede on 2 points: (i) the authors use partial GW, and (ii) they add a sparsity penalty through .
-
Equation (6), that is, , is roughly equation (11). Here, the main difference is that the authors propose optimizing over the barycentric coordinates .
-
Equations (9), and (12), which are repeated, BTW, correspond to equations (13) and (14, the L$_{2}$ especialization of 13) in PCS. The main difference here is that, due the plans being partial, the marginal of PCS doesn't factor out, so you'll end up with .
-
Equation (10) corresponds roughly to equations (8) and (9), with the difference, again, coming from the fact that the authors are solving a partial GW problem.
-
The BCD procedure as well is the same, except for the optimization over .
So, as you see, in general the main differences with respect to PCS are (i) the fact that the authors do partial GW, and (ii) the optimization over barycentric weights . In my opinion, this makes the methodological contribution rather incremental.
Minor issues
- , the consensus kernel, is used with an asterisk across the text. This gives the false impression of an optimal variable everywhere (e.g., minimizations over ), and thus make some mins/argmins look awkward.
References
Peyré, Gabriel, Marco Cuturi, and Justin Solomon. "Gromov-wasserstein averaging of kernel and distance matrices." International conference on machine learning. PMLR, 2016.
Questions for authors
Q1. Is it possible that the optimization of barycentric coordinates can ignore outlier views? Or, conversely, can it ignore views whose geometry is informative but too dissimilar?
Q2. Did the authors consider manifold-aware metrics in place of the Euclidean distance in equation 3?