acceptodds
Under review as a conference paper at ICLR 2027

The Equivalence of Multi-view Clustering via Latent Co-association Matrix

10 views

Abstract

Studying the relationship between multi-view clustering and ensemble clustering is of great practical and theoretical significance in achieving the equivalence of multi-view clustering. The problem of ensemble clustering is usually tackled by generating multiple base clusterings and then combining them to achieve the clustering result. Most existing ensemble clustering methods either linearly combine connective matrices to construct the co-association matrix or refine the co-association matrix with the built coherent-link matrix. However, directly casting ensemble clustering as a multi-view clustering problem becomes challenging. In this paper, we show that the equivalence of multi-view clustering and ensemble clustering can be established based on the formulation of objective function. To be specific, we have multiple co-association matrices by defining -sensitive feature descriptor and introduce a latent co-association matrix shared by different co-association matrices. The -sensitive feature descriptor can refine the original co-association matrix in a general manner and alleviate the impacts brought by poor base clusterings. We then define the -sensitive margin and prove that -sensitive feature descriptor generates the largest -sensitive margin, which helps lead to a discriminative co-association representation. These multiple co-association representations are collected into a tensor and the low-rank constraint is used to better explore the global relationship among entries in the tensor. Then the relationship between multi-view clustering and ensemble clustering is built in this way. Extensive experiments on several data sets validate the above analysis in terms of different metrics.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-view clustering, effectiveness

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_K5AJOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper shows that the equivalence of multi-view clustering and ensemble clustering can be established based on the formulation of objective function. To be specific, the authors have multiple co-association matrices by defining $ \delta $-sensitive feature descriptor and introduce a latent co-association matrix shared by different co-association matrices. The $ \delta $-sensitive feature descriptor can refine the original co-association matrix in a general manner and alleviate the impacts brought by poor base clustering. The authors then define the $ \delta $-sensitive margin and prove that $ \delta $-sensitive feature descriptor generates the largest $ \delta $-sensitive margin, which helps lead to a discriminative co-association representation. Then the relationship between multi-view clustering and ensemble clustering is built in this manner.

Critical strengths
  1. The authors have multiple co-association matrices by defining $ \delta $-sensitive feature descriptor and introduce a latent co-association matrix shared by different co-association matrices.
  2. The $ \delta $-sensitive feature descriptor can refine the original co-association matrix in a general manner and alleviate the impacts brought by poor base clustering.
Critical weaknesses
  1. The authors can add representative multi-view clustering methods to the comparison methods in the experiment.
  2. How to justify that “the relationship between multi-view clustering and ensemble clustering has never been studied”? It seems that a work [1] addresses both multi-view clustering and ensemble clustering. [1] From ensemble clustering to multi-view clustering
Questions for authors
  1. The authors should add representative multi-view clustering methods to the comparison methods in the experiment.
  2. The authors simply described the experimental results without in-depth analysis in section 5.2, which will weaken the convincing of the results.
  3. The results of some methods in Tables 1 and 2 are not statistically significant, such as GTLEC, CELTA, and the proposed method.
3 comments

Q1: The authors can add representative multi-view clustering methods to the comparison methods in the experiment.

A1: Thanks for the comment! We have added a compared method in the experiment, i.e., FastMICE [1]. The means are 0.900, 0.937, 0.890, 0.900, 0.992, 0.982, 0.980, 0.990, 0.979, 0.980, 0.973, 0.980, 0.986, 0.978, 0.975, 0.980, 0.495, 0.705, 0.420, 0.458, 0.995, 0.997, 0.994, 0.994 based on ACC, NMI, ARI and F1-score and 0.890, 0.830, 0.859, 0.992, 0.985, 0.987, 0.982, 0.980, 0.979, 0.985, 0.975, 0.970, 0.810, 0.757, 0.518, 0.994, 0.990, 0.995 based on other three metrics on all datasets.

[1] Fast multi-view clustering via ensembles: Towards scalability, superiority, and simplicity

Q2: How to justify that “the relationship between multi-view clustering and ensemble clustering has never been studied”? It seems that a work [1] addresses both multi-view clustering and ensemble clustering.

[1] From ensemble clustering to multi-view clustering

A2: Good question! As reviewer pointed, it is needed to justify that the relationship between multi-view clustering and ensemble clustering. The work in [1] proposes a multi-view ensemble clustering framework to solve multi-view clustering problem in an ensemble clustering way. It mainly focuses on the combination of multi-view clustering and ensemble clustering, which differs from ours since we independently study the relationship between multi-view clustering and ensemble clustering. We will add these comparisons for the camera-ready version.

Q3: The authors simply described the experimental results without in-depth analysis in section 5.2, which will weaken the convincing of the results.

A3: Thanks for the comment! As reviewer mentioned, we describe the experimental results without in-depth analysis and it can include more in-depth analysis for the revision, i.e., the performance improvements can be explained by considering consistency and complementarity issues among co-association matrices in ensemble clustering. We will add the above analysis for the camera-ready version.

Q4: The results of some methods in Tables 1 and 2 are not statistically significant, such as GTLEC, CELTA, and the proposed method.

A4: Good question! As reviewer suggested, we have slightly improved the final results by adjusting hyperparameters and the mean results are 0.925, 0.950, 0.900, 0.914, 0.998, 0.990, 0.990, 0.995, 0.989, 0.992, 0.985, 0.988, 0.995, 0.985, 0.990, 0.992, 0.517, 0.725, 0.437, 0.480, 0.998, 0.999, 0.996, 0.995 under ACC, NMI, ARI and F1-score and 0.920, 0.852, 0.885, 0.995, 0.990, 0.995, 0.991, 0.990, 0.985, 0.992, 0.985, 0.979, 0.825, 0.789, 0.600, 0.999, 0.996, 0.999 under other three metrics on all datasets .

The manuscript has been substantially improved after revision and, in my view, now meets the standard of ICLR. However, to ensure transparency and reproducibility, the authors must add the public code link as the final sentence of the Abstract. In addition, the authors must ensure that the released code is complete and runnable, including all necessary scripts/instructions to reproduce the reported experiments and figures.

Thanks very much for your comment and we will add the public code link as the final sentence of the Abstract for the camera-ready version that the released code is complete and runnable, including all necessary scripts/instructions to reproduce the reported experiments and figures.

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

This paper presents a novel integrated clustering method. The expected potential co-correlation matrix is achieved by stacking the co-correlation representations of different views to form a three-dimensional tensor and applying a tensor low-rank norm to this constructed tensor in order to study the relationship between different entries in the tensor. Experiments demonstrate that the proposed algorithm outperforms other SOTA methods in terms of clustering performance on some benchmark datasets.

Critical strengths
  1. This paper proves the equivalence of multi-view clustering and ensemble clustering.
  2. This paper provides rigorous theoretical analysis and detailed computational flow.
Critical weaknesses
  1. How were the parameters selected for the compared methods? The paper mentions that "the parameters of the compared methods are set the same as the original papers," which may not be optimally tuned for the datasets used in your experiments as they are for your method. And still the performance of the proposed algorithm remains comparable to some compared methods?
  2. How robust is this method to extreme outliers and noise, considering it imposes the squared Frobenius norm on the error tensor?
Questions for authors
  1. What about the scalability of this method? To address it with the inexact augmented Lagrangian multiplier, V auxiliary variables need to be introduced to solve for the co-association tensor, and it requires elementwise analysis in each update iteration to obtain results.
  2. Could you compare the performance of the proposed method with the recent study about multi-view clustering via ensemble in [1]? [1] Huang, Dong, Chang-Dong Wang, and Jian-Huang Lai. "Fast multi-view clustering via ensembles: Towards scalability, superiority, and simplicity." IEEE Transactions on Knowledge and Data Engineering (2023).
  3. What about the scalability of this method? To address it with the inexact augmented Lagrangian multiplier, V auxiliary variables need to be introduced to solve for the co-association tensor, and it requires elementwise analysis in each update iteration to obtain results.
3 comments

Q1: How were the parameters selected for the compared methods? The paper mentions that "the parameters of the compared methods are set the same as the original papers," which may not be optimally tuned for the datasets used in your experiments as they are for your method. And still the performance of the proposed algorithm remains comparable to some compared methods?

A1: Good question! It is expected to explain why we use the same parameters of the compared methods as the original papers in selecting the parameters selected for the compared methods as reviewer mentioned. The reason why we use the same parameters as in the original papers is that these parameters achieve better performance in our parameter selection and these selected parameters are actually consistent with the ones given in the original papers. Then the performance of the proposed algorithm can remain comparable to some compared methods.

Q2: How robust is this method to extreme outliers and noise, considering it imposes the squared Frobenius norm on the error tensor?

A2: Thanks for the comment! As reviewer mentioned, we should clearly show how robust is our method to extreme outliers and noise considering it imposes squared Frobenius norm on the error tensor. In the experiment, we have added the robust analysis based on the Gaussian noise with 90 percentages on MNIST and the corresponding results are 0.900, 0.885, 0.897, 0.872, 0.916, 0.920 and 0.918 on this dataset based on all metrics. Then the robustness of our method to extreme outliers and noise can be validated.

Q3: What about the scalability of this method? To address it with the inexact augmented Lagrangian multiplier, V auxiliary variables need to be introduced to solve for the co-association tensor, and it requires elementwise analysis in each update iteration to obtain results.

A3: Good question! As reviewer pointed, we should clearly present the scalability of the proposed method. To address it with the inexact augmented Lagrangian multiplier, V auxiliary variables need to be introduced to solve for the co-association tensor, and it requires elementwise analysis in each update iteration for results. The overall computational complexity for our method is O(n^2log(n)) with n being the dataset size as dataset scales when the auxiliary variables are introduced. Considering the obtained performance, the scalability of our method is relatively acceptable.

Q4: Could you compare the performance of the proposed method with the recent study about multi-view clustering via ensemble in [1]? [1] Huang, Dong, Chang-Dong Wang, and Jian-Huang Lai. "Fast multi-view clustering via ensembles: Towards scalability, superiority, and simplicity." IEEE Transactions on Knowledge and Data Engineering (2023).

A4: Thanks for the comment! We have added a compared method in the experiment as reviewer suggested, i.e., FastMICE [1]. The means are 0.900, 0.937, 0.890, 0.900, 0.992, 0.982, 0.980, 0.990, 0.979, 0.980, 0.973, 0.980, 0.986, 0.978, 0.975, 0.980, 0.495, 0.705, 0.420, 0.458, 0.995, 0.997, 0.994, 0.994 based on ACC, NMI, ARI and F1-score and 0.890, 0.830, 0.859, 0.992, 0.985, 0.987, 0.982, 0.980, 0.979, 0.985, 0.975, 0.970, 0.810, 0.757, 0.518, 0.994, 0.990, 0.995 based on other three metrics on all datasets.

Reviewer_rtJVOpenReview ↗

87977

The authors have properly responded to my concerns and I would keep my original rating score.

Thanks very much for your insightful comments, which helps improving our work greatly.

Reviewer_U65uOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This work defines the $ \delta $-sensitive margin and proves that $ \delta $-sensitive feature descriptor generates the largest $ \delta $-sensitive margin, which helps lead to a discriminative co-association representation. These multiple co-association representations are collected into a tensor and the low-rank constraint is used to better explore the global relationship among entries in the tensor.

Critical strengths
  1. The work in this paper is able to study the relationship between multi-view clustering and ensemble clustering.
  2. The authors extend the traditional construction of the co-association matrix in ensemble clustering to multi-view cases by extracting multiple co-association representations as different views based on $ \delta $-sensitive feature descriptor.
Critical weaknesses
  1. More analysis for the visualization part can be given in the experiment, i.e., why the proposed method can achieve more satisfied performance.
  2. The rationale for the specific aggregation used to construct the ensemble co-association matrix, and the fragmentation step can be better clearly motivated and the current presentation is not very clear.
Questions for authors
  1. More analysis for the visualization part can be given in the experiment, i.e., why the proposed method can achieve more satisfied performance.
  2. The rationale for the specific aggregation used to construct the ensemble co-association matrix, and the fragmentation step can be better clearly motivated and the current presentation is not very clear.
  3. The symbol ‘,’ in Eq. (12) and Eq. (24) are expected to be consistent with the symbols in other formulations and the authors are expected to check the whole paper to avoid the similar issues.
  4. The authors are expected to show how the tensor Z\mathcal{Z} is actually computed and can highlight the most significant results for the tables in experiment.
1 comment

Q1: More analysis for the visualization part can be given in the experiment, i.e., why the proposed method can achieve more satisfied performance.

A1: Good question! We should give more analysis for the visualization part in the experiment as reviewer pointed. The reason why the proposed method achieves more satisfied performance is that a discriminative co-association representation is achieved by defining the δ\delta-sensitive margin and proving that δ\delta-sensitive feature descriptor generates the largest δ\delta-sensitive margin. We will add these details for the camera-ready version.

Q2: The rationale behind the specific aggregation used to construct the ensemble co-association matrix, and the fragmentation step can be clearly motivated and the current presentation is not very clear.

A2: Thanks! The rationale behind the specific aggregation used to construct the ensemble co-association matrix is supported by Eq. (5) raised by Fred & Jain in Related Work for the submitted manuscript. The fragmentation step introduced in Equation (9) is motivated by aspects as follows. $ A_{v}(i,j) $ is influenced by the threshold δv\delta_{v} and Eq. (9) is called as $ \delta $-sensitive feature descriptor. A large $ \delta_{v} $ produces a desired co-association representation, resulting in the satisfied clustering result.

Q3: The symbol ‘,’ in Eq. (12) and Eq. (24) should be consistent with the ones in other formulations and the authors are expected to check the whole paper to avoid the similar issues.

A3: Thanks for the comment! As reviewer pointed, the symbols ‘,’ in Eq. (12) and Eq. (24) are not consistent with the ones in the other formulations for the paper. We will correct this issue and check the whole paper to avoid the similar issues for the camera-ready version.

Q4: The authors are expected to show how the tensor Z\mathcal{Z} is actually computed.

A4: Good question! In the submitted manuscript, we show how the tensor Z\mathcal{Z} is computed from the last line for the left side of page 4, which is presented as “Based on $\mathcal{A} $, we attempt to remove the existing incorrect connections depicted by $ \mathcal{E} $, resulting in a refined co-association tensor $ \mathcal{Z} $”.

Q5: The authors can highlight the most significant results for the tables in experiment.

A5: Thanks for the comment! As reviewer pointed, the most significant clustering results on tables are not highlighted and we can highlight the most significant ones on these tables. We will highlight the most significant clustering results on tables to make the performance gains more obvious for the camera-ready version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.