acceptodds
Withdrawn from ICLR 2027

Beyond SVD: Differentiable Low-Rank Matrix Factorization for Multiview Subspace Clustering

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-view subspace clustering (MvSC) has attracted considerable attention due to its ability to uncover underlying subspace structures from heterogeneous multiview data. However, existing methods require time-consuming SVD during each iteration, resulting in quadratic or even cubic computational complexity with respect to the number of samples and thereby limiting their scalability to large-scale datasets. To overcome these limitations, we propose a novel differentiable low-rank matrix factorization method for MvSC, which unifies multi-factor decomposition, low-rank representation learning, and structural regularization within a coherent framework. The proposed method not only characterizes the intrinsic low-rank structure through multiple factor matrices under cross-view consistency and symmetry constraints, but also transforms the factor-wise iterative optimization into differentiable learning steps, avoiding repeated SVD and enabling end-to-end learning through back propagation. The resulting block-wise optimization algorithm is further supported by a theoretical convergence analysis. Extensive experiments on multiple benchmark datasets demonstrate that our method achieves superior clustering performance while substantially reducing computational costs.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multiview Subspace Clustering, Low-Rank Representation Learning, Differentiable Optimization, SVD-Free 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_WC6TOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a differentiable low-rank matrix factorization framework for multi-view subspace clustering. The method aims to reduce the high computational cost caused by repeated SVD operations in conventional low-rank optimization. It factorizes each self-representation matrix into multiple low-rank factors and incorporates cross-view consistency and symmetry constraints to capture shared and complementary structures across views. The factor-wise iterative updates are further transformed into differentiable learning steps, enabling the model parameters to be optimized through backpropagation while avoiding repeated SVD-based updates. The paper also provides a convergence analysis for the resulting block-wise optimization scheme. Experiments on multiple benchmark datasets show that the proposed method achieves competitive clustering performance with substantially reduced computational cost.

Critical strengths

The paper presents a novel and practically meaningful attempt to address the heavy computational cost caused by repeated SVD operations in conventional low-rank multiview subspace clustering methods. In particular, replacing the SVD-based factor update with a differentiable and learnable low-rank mapping is an interesting idea that connects model-based low-rank optimization with trainable architectures.

Another strength is that the proposed framework preserves explicit low-rank and multiview structural priors while improving computational efficiency. The experimental results on multiple benchmark datasets show competitive clustering performance together with substantial reductions in runtime, demonstrating the practical potential of the proposed approach.

Critical weaknesses

The main weakness lies in the theoretical justification of the proposed differentiable low-rank update. The original Schatten-pp proximal subproblem is replaced by an equality-constrained formulation, but the relationship between the two is not rigorously established. Moreover, the Lagrange multiplier is further replaced by a freely trainable weight matrix, which weakens the connection to the original KKT conditions and makes it unclear whether the learned update still approximates the intended proximal operator.

Another concern is the computational claim. Although the method is motivated by avoiding repeated SVDs, the proposed update still involves fractional matrix powers and pseudoinverses, which may themselves require eigendecomposition or SVD in practice. Without a clearer implementation-level complexity analysis, it is difficult to assess whether the proposed operator truly provides the claimed computational advantage.

Questions for authors
  1. The original Schatten-pp proximal update in Eq.(7) is replaced by the equality-constrained formulation in Eq.(8). Could the authors clarify the theoretical relationship between these two problems? In particular, is Eq.(8) an equivalent reformulation, an approximation, or only a surrogate used to derive the differentiable operator?
  2. The Lagrange multiplier Λi(v)\Lambda_i^{(v)} derived from the constrained problem is directly replaced by a trainable matrix Wi(v)W_i^{(v)}. Since Λi(v)\Lambda_i^{(v)} is originally constrained by the corresponding optimality conditions, could the authors clarify what theoretical connection remains between the learned Wi(v)W_i^{(v)} and the original Lagrange multiplier, and whether the resulting update still approximates the original Schatten-pp proximal operator?
  3. The proposed method claims to avoid repeated SVD computations. However, Eq.(13) still involves the fractional matrix power (Q⊤Q)(p−2)/2(Q^\top Q)^{(p-2)/2} and a pseudoinverse. Could the authors clarify how these operations are implemented in practice and whether they require eigendecomposition or SVD? A clearer complexity comparison with the original SVD-based proximal update would be helpful.
Reviewer_nbKTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a novel differentiable low-rank matrix factorization method for MvSC, which unifies multi-factor decomposition, low-rank representation learning, and structural regularization within a coherent framework. The proposed method not only characterizes the intrinsic low-rank structure through multiple factor matrices under cross-view consistency and symmetry constraints, but also transforms the factor-wise iterative optimization into differentiable learning steps, avoiding repeated SVD and enabling end-to-end learning through back propagation.

Critical strengths

Extensive empirical evaluations are conducted on eight public multi‑view benchmark datasets against thirteen representative baselines covering matrix‑based, tensor‑based, anchor‑based, and deep‑unfolding methods. Evaluations include clustering metrics, runtime, affinity‑matrix visualization, t‑SNE embedding plots, parameter sensitivity, and convergence analysis.

Critical weaknesses

1.The research motivation of this work is not clear enough and lacks in-depth analysis. It is suggested to further clarify the research motivation. 2.The related work section should better clarify the distinctions between the proposed method and existing approaches. 3.Anchor‑based methods also avoid SVD on large matrices. What are the fundamental technical differences between the proposed method and anchor‑based MvSC approaches?

Questions for authors

1.The deep unfolding method LargeMvC‑Net is included as one baseline. Please compare your technical differences with LargeMvC‑Net. What are the respective pros and cons? Why is your design more suitable for the subspace‑clustering setting? 2.The main ablation is DLRMF‑1, which removes the trainable network module. However, many intermediate ablations are missing. For example: (1) remove symmetry loss; (2) remove cross‑view consistency loss. Why are these ablations not included? 3.All experiments are conducted on complete multi‑view datasets. Real‑world multi‑view data often contains missing samples or corrupted views. What is the expected failure mode of DLRMF under missing‑view/noisy view conditions?

Reviewer_55WROpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

In this work, the author propose a novel differentiable low-rank matrix factorization method for multi-view subspace clustering, which unifies multi-factor decomposition, low-rank representation learning, and structural regularization within a coherent framework. The proposed idea aims to avoide repeated SVD and enable end-to-end learning through back propagation.

Critical strengths

The proposed method is easy to read. Meanwhile, the motivation is meaningful.

Critical weaknesses

However, there are significant concerns about this paper.

W1. The novelty of this work needs to be deliberated, where the main contributions of this work stay in the optimization part. Such contributions may be not fit for this top-A conference.

W2. The preliminaries about Chen et al. (2023) is missing, whose idea seems similar with this work.

W3. The used datasets are smaller, which can not well demonstrate the scalability of the proposed method. The performance and time cost should be evaluated with the increasing larger data size.

Questions for authors

See weakness.

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

Related papers

Discussion (0)

Sign in to comment.