acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Multi-View Clustering as Symmetric Incremental Reasoning

10 views

Abstract

Deep multi-view clustering seeks shared structure from heterogeneous observations without supervision; however, existing methods often rely on shallow view-level coupling, limiting progressive cross-view reasoning. We reformulate multi-view fusion as view-wise incremental reasoning and propose VIR-SCI, which infers complementary information for each view via anchor-excluded symmetric cross-view interactions. To improve discriminability and robustness, we further introduce feature-wise incremental reasoning to adaptively modulate cross-view information. Mutual-information-constrained semantic alignment and contrastive learning jointly preserve global–local semantic coherence throughout reasoning. Extensive experiments show that VIR-SCI consistently outperforms state-of-the-art methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Deep Multi-View Clustering, Incremental Reasoning, Cross-View Interaction, Feature-Wise Fusion, Semantic Alignment

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_RETXOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes VIR-SCI, which re-models the information fusion in multi-view clustering as an incremental reasoning process. This method exploits complementary information through symmetric cross-view interactions and achieves selective fusion by using feature-level adaptive modulation. Additionally, it combines mutual information constraints with contrastive learning to maintain semantic consistency. Experiments on six benchmark datasets demonstrate that this method has excellent clustering performance, and ablation experiments also validate the effectiveness of the main modules.

Critical strengths
  1. The method design is innovative. Compared with the traditional one-time multi-view fusion, the paper proposes a symmetrical cross-view incremental interaction mechanism, combined with feature-level modulation, to more meticulously model the complementary relationship between different views.
  2. The framework design is relatively complete. Mutual information constraints, cross-view interaction, and feature-level fusion cooperate with each other, resulting in a method with good logical coherence.
  3. The experimental results are persuasive. The method achieves the best results in terms of ACC, NMI, and PUR metrics on six benchmark datasets, and provides sufficient ablation experiments, parameter sensitivity analysis, and visualization results.
Critical weaknesses
  1. The paper can further analyze the computational complexity and running time of the model to more fully illustrate the scalability of the symmetric cross-view interaction mechanism when the number of views increases.
  2. It is suggested to further clarify the essential difference between incremental reasoning and conventional attention-weighted fusion to highlight the uniqueness of the method.
  3. The current experiments are mainly based on standard multi-view clustering datasets. The generalization ability in higher dimensions or more complex noise environments still deserves further research.
Questions for authors
  1. As the number of views increases, how does the computational cost of symmetric cross-view interaction change?
  2. Compared to directly using cross-view attention fusion, does the proposed incremental reasoning mechanism have independent and stable performance advantages?
1 comment
AuthorsOpenReview ↗

Reply to Reviewer RETX

Thanks for your careful review. We are glad to address your questions one by one.

W1: We have provided a detailed analysis of the computational complexity and runtime of VIR-SCI. Please refer to YX5g W2. We have also conducted additional experiments to evaluate the scalability of symmetric cross-view interaction as the number of views increases. For details, please refer to Q1.

W2: Conventional attention-weighted fusion directly calculates view importance weights and aggregates different views into a fused representation, focusing on weighted information aggregation. In contrast, VIR-SCI first excludes the current view and takes the remaining views as anchors in turn to infer complementary information through anchor–candidate interactions (Eqs. (8)–(9)). It then combines the current view with the incremental representations to generate feature-wise importance weights and selectively injects the complementary information into the current view (Eqs. (10)–(13)). Therefore, the key difference is that VIR-SCI focuses on cross-view complementary information inference and selective feature-wise injection, rather than simply improving the calculation of attention weights.

W3: We have conducted additional noise robustness experiments by adding 10%–50% Gaussian noise to one input view of each of the three datasets: MNIST-USPS, Fashion, and Cifar100, and re-evaluated the clustering performance of VIR-SCI. The results show that VIR-SCI maintains relatively stable clustering performance even under high noise levels. For example, on Fashion, the ACC decreases by only 1.47% when the noise level reaches 50%. These results demonstrate the robustness of VIR-SCI against noise interference.

NoiseMNIST-USPSFashionCifar100
0.199.36 / 98.26 / 99.3699.11 / 97.74 / 99.1199.98 / 99.97 / 99.98
0.298.82 / 96.84 / 98.8298.65 / 96.84 / 98.6599.98 / 99.96 / 99.98
0.398.20 / 95.49 / 98.2098.18 / 95.89 / 98.1899.97 / 99.96 / 99.97
0.497.90 / 94.87 / 97.9098.15 / 95.87 / 98.1599.97 / 99.96 / 99.97
0.597.44 / 93.83 / 97.4498.05 / 95.72 / 98.0599.98 / 99.97 / 99.98
No Noise99.78 / 99.32 / 99.7899.52 / 98.65 / 99.5299.98 / 99.97 / 99.98

Values are reported as ACC / NMI / PUR.


Q1: We have conducted additional experiments on Caltech-5V to evaluate runtime and memory consumption with different numbers of views. As the number of views increases from 2 to 5, memory consumption rises from 133.36 MB to 374.11 MB, while runtime shows an overall modest increase from 194.33s to 246.01s. These results indicate that although symmetric cross-view interaction introduces additional computational costs, VIR-SCI maintains manageable runtime and memory consumption as the number of views increases, demonstrating its scalability in few-view scenarios.

Caltech-5VTime (s)Memory (MB)
View = 2194.33133.36
View = 3189.79208.92
View = 4210.04275.81
View = 5246.01374.11

Q2: Conventional cross-view attention fusion directly aggregates view information based on their correlation weights, making it difficult to explicitly identify complementary information beneficial to the current view. In contrast, VIR-SCI first infers cross-view complementary information through symmetric interactions that exclude the current view, and then selectively injects it through feature-wise modulation to reduce redundancy. Meanwhile, normalized additive fusion and semantic consistency constraints help maintain representation stability. The existing ablation study (M20) also demonstrates the contribution of incremental reasoning over simple one-shot fusion. Therefore, VIR-SCI improves not only information weighting but also complementary information inference and selective fusion.

Reviewer_YX5gOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes VIR-SCI, a symmetric cross-view incremental reasoning framework for deep multi-view clustering. The method uses a mutual information constraint to align semantics across views. Different views take turns serving as anchors to extract complementary information through cross-view interaction. Feature-level modulation and contrastive learning further filter information and improve representation consistency.

Critical strengths
  1. The paper studies multi-view information fusion from the perspective of incremental reasoning and designs a symmetric cross-view interaction mechanism. Compared with approaches that directly combine representations from different views, this approach seems novel to me.

  2. The experiments are relatively comprehensive. The method is compared with several existing methods on six datasets, and fairly detailed ablation studies are provided. The overall results are good and provide some support for the effectiveness of the method.

Critical weaknesses
  1. Although the authors compare VIR-SCI with several deep multi-view clustering algorithms, comparisons with the latest methods from 2026 are still limited. The current comparisons are not enough to fully show the advantages of this method over the latest research.

  2. VIR-SCI introduces symmetric cross-view interaction and feature-level modulation mechanisms. These may increase the computational cost compared with simple fusion methods. However, the paper does not provide a specific analysis of time complexity.

  3. The paper uses contrastive learning to improve semantic consistency between different representations, but it does not discuss the potential problem of false negatives. In unsupervised clustering tasks, different samples may belong to the same class. This may have some impact on contrastive learning.

Questions for authors
  1. Could the authors include some of the latest multi-view clustering methods proposed in 2026 for comparison to more fully verify the competitiveness of the proposed method?

  2. Does VIR-SCI require much more running time than simple fusion methods, such as one-shot fusion?

  3. The paper uses contrastive learning to improve semantic consistency across different views, but it does not further discuss the selection of negative samples. Are there any ways to improve negative sampling?

  4. Figure 2(d) shows the feature-level cross-view modulation process. However, it is still not clear how the incremental representations generated by multiple anchors are used to compute the feature weights. Could the authors further explain the specific computation process of this module with reference to Equations (10)-(12)?

2 comments
AuthorsOpenReview ↗

Reply to Reviewer YX5g

Thanks for your careful review. We are glad to address your questions one by one.

W1: We have added comparisons with two recent methods, MoMVC [1] and DMVC_CSG [2]. MoMVC focuses on noise filtering, while DMVC_CSG uses cluster-level semantic guidance. In contrast, VIR-SCI infers and selectively fuses cross-view complementary information through symmetric interactions and feature-wise modulation. VIR-SCI outperforms both methods on all metrics across six datasets.

MethodsMNISTNGsCaltech5VFashionCifar100Reuters
ACCNMIPURACCNMIPURACCNMIPURACCNMIPURACCNMIPURACCNMIPUR
MoMVC98.8996.5998.8975.4069.9175.6081.5069.7981.5080.2583.5282.9698.6499.3398.6450.7025.6751.21
DMVC_CSG96.6895.1096.9886.2079.9787.2079.2976.1082.4392.8594.0393.4689.7897.8091.7144.6520.6451.24
VIR-SCI99.7899.3299.7896.4088.8496.4092.8686.2192.8699.5298.6599.5299.9899.9799.9867.7848.3973.76

[1] Trusted Mixture-of-Mambas Contrastive Network for Multi-View Clustering, IEEE Transactions on Multimedia, 2026.

[2] Deep Multi-View Clustering via Cluster-Semantic Guidance, IEEE Transactions on Image Processing, 2026.

W2: We have compared VIR-SCI's runtime with simple fusion methods. Please refer to Q2. The computational complexity of VIR-SCI is analyzed as follows. Let nn, mm, BB, and dd denote the number of samples, views, batch size, and semantic feature dimension, respectively. Each fully connected layer in the autoencoders and semantic mappings requires O(nDinDout)O(nD_{\mathrm{in}}D_{\mathrm{out}}) operations. The mutual information constraint and contrastive learning require O(nBm(m−1)d)O(nBm(m-1)d) and O(nBmd)O(nBmd), respectively. The symmetric cross-view interaction costs O(nm(m−1)d2+nm(m−1)(m−2)d)O(nm(m-1)d^2+nm(m-1)(m-2)d), while feature-wise incremental reasoning costs O(nm2dh+nmh2)O(nm^2dh+nmh^2), where hh is the maximum hidden-layer width. The final K-means clustering requires O(nIKd)O(nIKd), where II and KK denote the number of iterations and clusters. Overall, the complexity grows linearly with nn and at most cubically with mm. Since common multi-view clustering tasks typically involve few views (e.g., m≤6m\leq6), the computational overhead remains manageable.

W3: We acknowledge that false negatives may affect contrastive learning in unsupervised clustering. VIR-SCI uses standard instance-level contrastive learning to maintain semantic consistency across views and before and after incremental reasoning, while its main contribution is complementary information fusion through symmetric cross-view interactions and feature-wise modulation. Although better negative sample selection may further improve clustering performance, it is not the main focus of this work. In future work, we will explore ways to reduce false negatives and improve representation learning.


Q1: We have added the corresponding experiments. Please refer to W1.

Q2: We have compared the runtime of VIR-SCI with one-shot fusion. VIR-SCI incurs only slight runtime increases on some datasets and is even faster on Cifar100. Although the increases are more noticeable on Caltech-5V and Reuters, the overall overhead remains reasonable while providing more effective cross-view complementary information fusion.

MethodsMNIST-USPSNGsCaltech-5VFashionCifar100Reuters
One-shot Fusion138.8247.2293.24470.531533.06579.16
VIR-SCI141.6548.16246.01492.581472.48735.95

All runtime values are measured in seconds (s).

Q3: VIR-SCI currently adopts standard cross-sample negative sampling. Possible improvements include using high-confidence pseudo-labels to identify and remove potential false negatives, or dynamically adjusting negative sample weights based on semantic similarity to avoid separating samples from the same class. In future work, we will explore integrating these strategies into VIR-SCI to improve contrastive learning.

Q4: For each current view Rv\mathbf{R}^{v}, we map the concatenated incremental representations from the other m−1m-1 anchors into FM1\mathbf{FM}_{1} (Eq. (10)). By combining FM1\mathbf{FM}_{1} with Rv\mathbf{R}^{v}, we generate the feature-wise importance map Fv\mathbf{F}^{v} (Eq. (11)), which weights each anchor's incremental representation (Eq. (12)). The selected complementary information is then injected into Rv\mathbf{R}^{v} (Eq. (13)).

Thanks very much for the detailed response that solved my most concerns, I will keep my positive rating.

Reviewer_XZvQOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper combines complementary information by selecting anchor views in turn and using feature-level weights, while maintaining semantic consistency through a mutual information constraint and contrastive learning. Comparisons with twelve methods on six public datasets and ablation studies verify the effectiveness of the method.

Critical strengths
  1. The feature-level modulation mechanism selectively combines cross-view features based on information from the current view. This design is reasonable and helps reduce interference from redundant information.
  2. The paper has a relatively clear structure. The role and implementation of each module are described, making the method section relatively easy to understand.
  3. The authors conduct comprehensive ablation studies on different components of the proposed method.
Critical weaknesses
  1. The authors mention that multi-view data often contain noise and redundant information. However, the paper lacks experiments on clustering performance under different noise conditions. It is therefore unclear whether VIR-SCI can maintain stable performance in the presence of noise.
  2. The paper generates fused view representations through cross-view incremental reasoning, but it uses the global representation H for the final clustering. The authors do not explain the reason for this choice clearly enough.
  3. The paper does not clearly state the specific parameter settings of the twelve comparison methods or the sources of their experimental results. This may affect the fairness of the comparisons and the reproducibility of the results.
Questions for authors
  1. Have the authors considered the impact of noise on the clustering performance of VIR-SCI?
  2. In Section 3.3, the authors exclude the current view from symmetric cross-view interaction. This can reduce the direct influence of the current view, but how can the authors ensure that the resulting complementary information is relevant to the current view?
  3. The paper uses both a mutual information constraint and contrastive learning to maintain semantic consistency. Could the authors further explain the role of each constraint and the differences between them?
2 comments
AuthorsOpenReview ↗

Reply to Reviewer XZvQ

Thanks for your careful review. We are glad to address your questions one by one.

W1: We have conducted additional noise robustness experiments by adding 10%–50% Gaussian noise to one input view of each of the three datasets: MNIST-USPS, Fashion, and Cifar100, and re-evaluated the clustering performance of VIR-SCI. The results show that VIR-SCI maintains relatively stable clustering performance even under high noise levels. For example, on Fashion, the ACC decreases by only 1.47% when the noise level reaches 50%. These results demonstrate the robustness of VIR-SCI against noise interference.

NoiseMNIST-USPSFashionCifar100
0.199.36 / 98.26 / 99.3699.11 / 97.74 / 99.1199.98 / 99.97 / 99.98
0.298.82 / 96.84 / 98.8298.65 / 96.84 / 98.6599.98 / 99.96 / 99.98
0.398.20 / 95.49 / 98.2098.18 / 95.89 / 98.1899.97 / 99.96 / 99.97
0.497.90 / 94.87 / 97.9098.15 / 95.87 / 98.1599.97 / 99.96 / 99.97
0.597.44 / 93.83 / 97.4498.05 / 95.72 / 98.0599.98 / 99.97 / 99.98
No Noise99.78 / 99.32 / 99.7899.52 / 98.65 / 99.5299.98 / 99.97 / 99.98

Values are reported as ACC / NMI / PUR.

W2: We choose the global semantic representation H\mathbf{H} for final clustering because it is jointly learned from the latent representations of all views and can capture the global semantics shared across views. In contrast, R~v\tilde{\mathbf{R}}^{v} is a view specific representation obtained through incremental reasoning and information injection, which still retains some view-specific characteristics. Therefore, we use contrastive learning to align each R~v\tilde{\mathbf{R}}^{v} with H\mathbf{H}, maintaining global semantic consistency while preserving complementary information. Finally, clustering on H\mathbf{H} directly utilizes the shared semantic information across views and avoids the need for additional fusion of multiple view-specific representations.

W3: For those comparison methods, we mainly use the publicly available code provided by the original authors and the parameter settings recommended in their papers to avoid bias caused by manual parameter adjustments. We also evaluate all methods on the same datasets using the same evaluation metrics (ACC, NMI, and PUR). These settings follow the original experimental configurations as closely as possible to ensure the fairness and reproducibility of the comparisons.


Q1: We have conducted additional clustering robustness experiments under different levels of Gaussian noise to evaluate the impact of noise on VIR-SCI's performance. For detailed experimental results and analysis, please refer to W1.

Q2: Excluding the current view only occurs during the inference of cross-view complementary information, which does not mean that its relevance is ignored. First, the mutual information constraint maintains semantic consistency across different view representations. Then, Eq. (11) combines the current view Rv\mathbf{R}^v and the cross-view incremental representation FM1\mathbf{FM}_{1} to generate the feature-wise importance map Fv\mathbf{F}^{v}, which adaptively selects complementary information relevant to the current view. The selected information is then injected into the current view through Eq. (13). Therefore, the relevance of complementary information is enhanced through both semantic alignment and feature-wise modulation guided by the current view.

Q3: The two constraints mainly differ in the stages where they are applied and the representations they align. The mutual information constraint (Eqs. (4)–(5)) is applied before incremental reasoning. It maximizes the mutual information between different view representations Rv\mathbf{R}^v to establish cross-view semantic consistency. In contrast, the contrastive learning constraint (Eq. (14)) is applied after incremental reasoning. It aligns the fused representations R~v\tilde{\mathbf{R}}^{v} with the global representation H\mathbf{H} to reduce semantic drift during information fusion. Therefore, the two constraints focus on cross-view semantic alignment before reasoning and global–local semantic consistency after reasoning, respectively, playing complementary roles.

Thanks for the response. I have read the comments from the other reviewers and the newly added experimental results. The authors have addressed all of my concerns, and I will maintain my positive score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.