acceptodds
Under review as a conference paper at ICLR 2027

Incomplete Gromov-Wasserstein Alignment for Multi-View Clustering

Abstract

Incomplete multi-view clustering (IMVC) struggles with heterogeneous feature spaces and severe instance missingness. Current methods typically project disparate views into a shared Euclidean space and rely on static graphs to preserve local structures. However, these point-to-point embeddings ignore intrinsic metric discrepancies across modalities and suffer from topological distortion when samples are missing. To address this, we propose Incomplete Gromov-Wasserstein Alignment (IGWA), a novel optimal transport framework for IMVC. IGWA abandons forced Euclidean consensus representations; instead, it directly couples heterogeneous intra-view metric spaces by learning a probabilistic transport plan. Building on established partial Gromov–Wasserstein transport, we incorporate view-specific partial couplings into a learnable global consensus framework, allowing reliable structural mass to be aligned without explicit data imputation. By matching intra-view distance matrices rather than raw features, IGWA preserves geometric invariants and maintains robustness against missing nodes. Furthermore, we establish relabeling invariance and balanced recovery of the partial alignment, together with uniqueness, boundedness, and perturbation stability of the consensus update under fixed transport plans and view weights. Extensive experiments on diverse benchmarks demonstrate that IGWA consistently outperforms State-of-The-Art IMVC methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Incomplete Multi-View Clustering, Heterogeneous Structural Alignment, Topological Consensus Learning.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_gX5rOpenReview ↗
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 21449PCS16Difference
Eq. (4)Eq. (7), the GWε\mathrm{GW}_\varepsilon objectivePartial GW plans, Penalty γ∥C∥F2\gamma \lVert C \rVert_F^2
Eq. (6)Eq. (11), joint objective over CC and (Ts)s(T_s)_sMinimization over α\alpha
Eqs. (9) and (12)Equations 13, and 14per-view marginals q(v)q^{(v)}, consequence of partial GW
Eq. (10)Eqs. (8)–(9),KL projection onto Π≤\Pi_{\le} instead of set of mass-preserving transport plans Π\Pi
BCD (Sec. 3.3)Algorithm 1additional optimization over barycentric coordinates α\alpha

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 CvC_v, CC and TvT_v, corresponding roughly to equation (7), with TT/TvT_v 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 γ∥C∥F2\gamma \lVert C \rVert_{F}^2.

  • Equation (6), that is, J\mathcal{J}, is roughly equation (11). Here, the main difference is that the authors propose optimizing over the barycentric coordinates α\alpha.

  • 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 pp of PCS doesn't factor out, so you'll end up with q(v)q^{(v)}.

  • 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 α\alpha.

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 α\alpha. In my opinion, this makes the methodological contribution rather incremental.

Minor issues

  • CC, 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 C∗C^{*}), 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?

Reviewer_7PmyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes IGWA for incomplete multi-view clustering. Experiments show strong clustering performance against existing baselines.

Critical strengths
  1. The joint formulation provides a coherent way to integrate heterogeneous and incomplete views through partial GW alignment and adaptive view weighting. It also avoids the need to reconstruct missing features.
  2. The paper derives an analytical solution for the consensus update, with guarantees of uniqueness, boundedness, and stability under fixed transport plans and view weights. Relabeling invariance and recovery of balanced GW are also established.
  3. IGWA reports strong clustering performance across six datasets and four missing rates. Ablation studies, sensitivity analyses, and experiments under structural corruption provide a fairly comprehensive evaluation.
Critical weaknesses
  1. IGWA extends partial GW alignment with a regularized consensus, but its main methodological advance over existing barycenter methods is not entirely clear. GWMAC [1] already combines GW alignment, adaptive view weighting, and barycenter learning, while PGW barycenters [2] use a consensus update closely related to Eq. (9).
  2. In standard IMVC, sample correspondences are already known. Table 4 shows improvements from learning (T(v)T^{(v)}), but what additional information does the learned transport capture beyond these known correspondences?
  3. The transported mass (s(v)=Nv/Ns^{(v)}=N_v/N) reflects sample coverage rather than structural reliability. Figure 5 shows little sensitivity to this parameter, making its role in selective structural alignment less clear.
  4. The paper attributes its performance gains to repairing incomplete topology and selecting reliable structural relations. Tables 3 and 4 mainly report clustering improvements, but provide limited insight into which structural relations are recovered or discarded under instance missingness.
  5. Table 2 does not include directly related GW/PGW baselines. The additional comparisons cover only a limited range of datasets and missing rates.
  6. The experiments on large datasets rely on Nyström approximations, but the transport and consensus updates in Eqs. (9) and (11) are presented only in their dense forms. The paper should explain how these updates are implemented within the reported memory budget.
  7. The Introduction shifts abruptly from graph regularization in IMVC to GW, and Section 2.2 could better organize the progression of PGW methods. In Figure 2, the blue dashed lines and the sphere representing global consensus need clearer explanations, and some text is too small. A few citations also contain formatting errors.

References

  • [1] Gong, F., Nie, Y., and Xu, H. Gromov-Wasserstein Multi-modal Alignment and Clustering. CIKM, 2022.
  • [2] Bai, Y., Diaz Martin, R., Kothapalli, A., Du, H., Liu, X., and Kolouri, S. Partial Gromov-Wasserstein Metric. ICLR, 2025.
Questions for authors

See the weakness.

Reviewer_JejKOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The author proposed IGWA to handle incomplete multi-view clustering. IGWA is capable of avoiding the correspondence of the original space and alleviates the accumulation of errors caused by the absence of structure. Numerous experiments have demonstrated the superiority of IGWA.

Critical strengths
  1. IGWA effectively addresses the issue of feature dimension inconsistencies between heterogeneous views. The authors ingeniously solve the heterogeneous view problem using GWOT, transforming dimension matching into structural matching and leveraging structural matching to prevent the propagation of incorrect structures.

  2. IGWA effectively integrates GW into the optimization process rather than optimizing the GW component separately and then proceeding with subsequent optimization. Furthermore, the authors provide a large number of experiments demonstrating the effectiveness of IGWA.

Critical weaknesses
  1. The optimization section could be moved to the appendix, and the main text could include some additional validation experiments—such as performance on comprehensive multi-view datasets and visualizations of the quality of the transfer matrices.

  2. The experimental analysis is insufficient. First, as both are alignment methods, why does IGWA outperform the IMVC method—which uses prototype matching—in terms of performance? Second, at what proportion of missing data did IGWA and the suboptimal method outperform the IMVC method? Third, this point should be addressed in other experiments rather than analyzed in Table 2.

Questions for authors
  1. Have the authors analyzed whether the transition matrix has a solution when the samples are too dense (i.e., when samples within a cluster are tightly packed)?

  2. The authors should provide a list of symbols to improve the readability of the paper.

  3. Is there a risk of weight alpha collapsing? Have the authors considered providing curves showing how the weights change over iterations from different perspectives? Additionally, how do the weights affect clustering performance?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.