acceptodds
Under review as a conference paper at ICLR 2027

Learning Intra-View Complementarity for High-Dimensional Multi-View Clustering

Abstract

Multi-view clustering mainly exploits complementarity across views, while often overlooking structural diversity within each individual view. This limitation becomes more pronounced in high-dimensional settings, where heterogeneous structural cues can coexist within a single view and may not be adequately captured by a single representation. We propose LIVC, a framework for Learning Intra-View Complementarity in high-dimensional multi-view clustering. LIVC first constructs raw, global-principal, and local-structural representations to expose complementary structures within each view. It then performs granularity-aligned cross-view interaction with learnable gating for selective information transfer. Finally, residual low-rank bilinear fusion captures second-order dependencies with controlled model complexity. A joint self-supervised objective preserves view-specific information while coordinating representations across granularities and views. Experiments on multiple high-dimensional multi-view datasets demonstrate the effectiveness of LIVC for clustering.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-view clustering, High-dimensional data, Intra-view complementarity, Low-rank representation.

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_saW3OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this manuscript, the authors argue that existing multi-view clustering methods primarily focus on cross-view relationships while overlooking structural diversity within individual views. To address this issue, they propose LIVC, which aims to exploit intra-view complementarity in high-dimensional multi-view clustering. LIVC constructs raw, main, and graph branches for each view and coordinates their representations through contrastive learning. It then performs granularity-aligned cross-view enhancement using attention and gating mechanisms, followed by low-rank bilinear fusion. The resulting representations are clustered using K-means. Experiments on five multi-omics benchmarks demonstrate strong clustering performance.

Critical strengths

1.LIVC presents a well-designed framework. Although it builds on conventional techniques such as PCA and KNN, the integration strategy is thoughtful.

2.LIVC achieves strong clustering performance on the evaluated multi-omics benchmarks, demonstrating its effectiveness in high-dimensional, small-sample settings.

Critical weaknesses

1.The motivation of this manuscript is that “relying on a single representation may therefore fail to capture these structures (heterogeneous and complementary structural patterns) simultaneously”. However, it seems not so straightforward to me. In existing MVC research, one important motivation for modeling inter-view complementarity is that excessive emphasis on consistency may suppress view-specific information. However, the manuscript does not establish an equally clear motivation for intra-view complementarity. I think more clarifications are necessary to establish the significance of this concept.

2.It is also unclear how interactive view expansion enhances complementarity within views. In IVE, each view is expanded to three branches: raw branch, main branch and graph branch. However, there is limited evidence on whether the three branches are complementary. Since both the main and graph branches are derived from the original view, which is also retained as a separate branch, the proposed expansion seems closer to feature engineering or representation enhancement.

Questions for authors

1.Why is it necessary to explicitly construct multiple intra-view representations rather than learning these structural properties directly from the original data? What limitations of single-representation learning does LIVC specifically address with intra-view complementarity?

2.What evidence supports the claim that the three branches capture complementary, rather than largely redundant, information?

3.Eq. (5) and Eq. (6) primarily encourage consistency across branches and views, while Eq. (10) aggregates the transformed representations through attention. How does this process specifically identify or exploit complementary information, rather than simply combining multiple representations?

4.LIVC is only evaluated on multi-omics benchmarks which may be highly similar in structures. Since LIVC is proposed as a general method for high-dimensional multi-view clustering, more heterogeneous benchmarks should be included.

3 comments
AuthorsOpenReview ↗

Reply to Reviewer saW3 (Part 1 of 2)

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

W1:

We thank the reviewer for highlighting the need to clarify our motivation. Inter-view complementarity primarily concerns information differences across views, whereas intra-view complementarity concerns heterogeneous structural patterns within the same view. In high-dimensional data, global principal variations, local neighborhood relationships, and potentially discriminative fine-grained information may coexist, but different representations emphasize these structures differently. PCA reconstruction highlights dominant global structure while potentially suppressing low-variance discriminative information; KNN smoothing enhances local neighborhood consistency but may blur cluster boundaries. Meanwhile, the Raw branch preserves original information, although its learned embedding may not adequately express all clustering-relevant relationships. Therefore, intra-view complementarity addresses the limitations of relying on a single structural perspective, rather than introducing new information absent from the original data.

Main and Graph identify distinct same-class neighborhood relationships absent from the learned Raw representation, with only 3.57%–4.50% overlap among their newly identified edges. Please refer to Dpyw W1 for the detailed neighborhood complementarity analysis. These observations support our motivation that a single learned representation may not adequately express different clustering-relevant structural relationships within the same view.

W2:

We thank the reviewer for this important observation. We acknowledge that PCA reconstruction and KNN smoothing are deterministic transformations of the original view and do not introduce new information. However, the purpose of IVE is not to generate additional raw information, but to expose different clustering-relevant structural relationships through complementary inductive biases. Specifically, Raw preserves original information, Main emphasizes global principal structure, and Graph highlights local neighborhood relationships. These representations are encoded through a shared view-specific encoder and coordinated by intra-view contrastive learning, enabling subsequent modules to exploit their distinct structural characteristics.

To evaluate whether these structural transformations provide useful complementary information, we conduct additional experiments. The frozen-representation prediction experiments show that combining different branches can provide additional discriminative value. For example, adding Graph to Raw+Main on STAD reduces held-out prediction log-loss by 0.0184 ± 0.0033 nats, with positive gains across all five seeds. Please refer to Dpyw W1 for details. Furthermore, the branch-combination ablation shows that the full three-branch configuration improves CESC ACC from 51.35% (Raw only) to 93.92%, outperforming both two-branch variants. Please refer to Dpyw Q1 for details.

Although these transformations constitute structural feature engineering, their incremental predictive contributions and clustering improvements support the practical value of jointly exploiting different structural representations within LIVC.

Q1:

The key limitation of single-representation learning is not insufficient expressive capacity, but the difficulty of simultaneously capturing heterogeneous structural relationships without explicit structural constraints. In high-dimensional, small-sample settings, a single encoder trained directly on raw features may favor dominant patterns while insufficiently representing local neighborhood relationships or subtle discriminative information. Consequently, some clustering-relevant structures present in the original data may remain inadequately expressed in the learned representation.

LIVC addresses this limitation by explicitly introducing complementary structural inductive biases rather than relying entirely on implicit structure discovery. Specifically, Raw preserves original information. The Main branch uses PCA reconstruction to emphasize global principal structure, while the Graph branch uses KNN smoothing to highlight local neighborhood relationships. These branches provide distinct structural perspectives before encoding, while the shared view-specific encoder enables coordinated representation learning and subsequent integration.

Thus, explicit multi-branch construction does not introduce new information or imply that a single encoder is inherently incapable of learning these structures. Instead, it reduces reliance on implicit structural discovery and mitigates the risk of overlooking clustering-relevant relationships. The neighborhood complementarity analysis further supports this motivation. Please refer to Dpyw W1 for details.

AuthorsOpenReview ↗

Reply to Reviewer saW3 (Part 2 of 2)

Q2:

We conduct two analyses to examine whether the three branches capture complementary rather than largely redundant information.

First, the neighborhood analysis examines same-class relationships identified by the three learned branch representations. Main and Graph identify distinct same-class KNN edges absent from the learned Raw representation, with their overlap accounting for only 3.57%–4.50% of the newly identified same-class edges across five datasets. Moreover, the fused representation consistently retains a higher proportion of newly identified same-class edges than different-class edges, suggesting that these structural differences are relevant to clustering rather than merely arbitrary variations.

Second, the frozen-representation prediction experiments evaluate the incremental discriminative value of combining branches. For example, on STAD, adding Graph to Raw+Main reduces held-out prediction log-loss by 0.0184 ± 0.0033 nats, with positive gains across all five seeds. Although the gains vary across datasets, these results suggest that the learned branches are not merely redundant copies but expose distinct clustering-relevant relationships that can provide additional predictive value.

Please refer to Dpyw W1 for the detailed experimental results.

Q3:

The contrastive objectives in Eqs. (5) and (6) encourage sample-level semantic consistency rather than forcing different branches to become identical. Since Raw, Main, and Graph are constructed with distinct structural inductive biases, they can retain different neighborhood relationships while learning semantically aligned representations.

Complementary information is exploited through the subsequent enhancement and aggregation mechanisms. Specifically, same-granularity cross-view attention identifies relevant information from corresponding branches across views, while learnable gates adaptively regulate its contribution. In Eq. (10), the enhanced Raw representation serves as the semantic anchor, and attention selectively integrates information from the enhanced Main and Graph branches. The residual formulation preserves the Raw representation while incorporating additional global and local structural cues. Furthermore, the reconstruction objectives encourage information preservation throughout this process.

Therefore, contrastive learning establishes semantic correspondence, whereas attention, gating, and residual aggregation enable adaptive exploitation of different structural perspectives. Although these mechanisms do not explicitly maximize structural novelty, the complementary neighborhood relationships and incremental discriminative gains provide empirical support for their effectiveness. Please refer to Dpyw W1 for the detailed analysis.

Q4:

To evaluate LIVC beyond multi-omics datasets, we conduct additional experiments on the Yale face-image dataset, which differs substantially from the biological data used in the original evaluation. Yale contains 165 samples, 15 clusters, and 3 views, with feature dimensions of 4096, 3304, and 6750, respectively. This provides an additional high-dimensional multi-view setting with different data characteristics.

We compare LIVC with 11 representative multi-view clustering methods. As shown in the following table, results are reported as ACC/NMI/PUR (%) over five random seeds.

MethodACC (%)NMI (%)PUR (%)
GMAE47.03 ± 15.4752.30 ± 15.1448.00 ± 15.21
MCSF62.91 ± 0.5169.52 ± 0.7162.91 ± 0.51
MISCC27.52 ± 1.3139.59 ± 1.3930.30 ± 0.94
WAH-MVC18.79 ± 6.7319.35 ± 10.3018.91 ± 6.71
GMVC31.27 ± 1.8236.70 ± 2.6132.61 ± 1.93
SSLNMVC36.36 ± 4.2041.11 ± 7.0537.58 ± 4.75
AEMVC60.00 ± 4.9861.27 ± 4.7660.24 ± 4.97
DMVC-MIC29.82 ± 0.7937.89 ± 0.9832.36 ± 0.33
SCMVC44.12 ± 6.8149.46 ± 6.2045.21 ± 6.78
ACCMVC38.42 ± 4.7543.86 ± 2.9739.88 ± 4.23
SEM29.70 ± 4.7338.01 ± 3.9431.39 ± 4.87
LIVC63.76 ± 0.5968.92 ± 1.3563.76 ± 0.59

On Yale, LIVC achieves the highest ACC (63.76%) and PUR (63.76%), while its NMI (68.92%) is close to the best result of 69.52% obtained by MCSF. These results show that LIVC remains competitive on a high-dimensional face-image dataset with substantially different characteristics from multi-omics data, providing additional evidence of its applicability beyond the original benchmarks.

I appreciate the authors' efforts in addressing my concerns. Their clarification of the motivation behind intra-view complementarity, particularly the difficulty of simultaneously capturing heterogeneous structural relationships without explicit structural constraints, addresses my primary concern. The additional experimental evidence further supports their claims. Therefore, I am willing to raise my rating to 3.

Reviewer_dcSrOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes LIVC, a framework for high-dimensional multi-view clustering that models complementary structures within each view before integrating information across views. For each view, the framework constructs raw, PCA-reconstructed, and KNN-smoothed representations and coordinates them through a shared view-specific encoder and contrastive learning. Granularity-aligned attention and learnable gates exchange information between corresponding branches across views. Residual refinement and low-rank bilinear fusion then produce a unified representation for K-means clustering. Experiments on five TCGA multi-omics datasets report competitive clustering performance against comparison methods, along with component ablations and parameter sensitivity analyses.

Critical strengths
  1. The paper constructs three complementary representations for each view. This design retains raw information and captures both global principal structure and local neighborhood relationships before cross-view integration.
  2. The paper conducts extensive experiments and analyses, including ablations and parameter studies. These results provide useful evidence that LIVC is effective across several datasets.
Critical weaknesses
  1. The paper treats PCA and KNN branches as sources of global and local structure, but does not separately analyze their contributions to the final clustering performance.
  2. The method uses bilinear multiplication to fuse learned representations. However, the advantage of this mechanism still needs further explanation.
Questions for authors
  1. Why were the PCA-reconstructed main branch and KNN-smoothed graph branch selected, and how does each contribute to clustering? The paper uses them to capture global principal structure and local neighborhood relationships within each view, but does not separately analyze their contributions to final clustering performance. An ablation study that removes one branch at a time to assess its effectiveness would be helpful.
  2. How does the learnable gating mapping Gvb(⋅)\mathcal{G}_v^b(\cdot) learn which cross-view information to retain? The paper uses its output to scale transferred information and updates its parameters through the joint objective. However, it remains unclear how this objective guides the mapping to suppress irrelevant information.
  3. How sensitive is LIVC’s clustering performance to the low-rank interaction dimension rfr_f? The paper fixes rfr_f at 64 without analyzing how different values affect the model. A sensitivity analysis of rfr_f would be helpful.
3 comments
AuthorsOpenReview ↗

Reply to Reviewer dcSr (Part 1 of 2)

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

W1:

We thank the reviewer for this suggestion. To separately evaluate the contributions of global principal structure and local neighborhood structure, we conduct branch-combination ablation experiments. Please refer to Dpyw Q1 for the detailed results. Specifically, we compare Raw (R), Raw+PCA (R+P), Raw+KNN (R+K), and the full three-branch configuration (R+P+K).

On CESC, adding PCA or KNN to Raw improves ACC from 51.35% to 79.73% and 91.89%, respectively, while combining all three branches further increases ACC to 93.92%. On STAD, Raw+PCA achieves 55.31% ACC, compared with 51.77% for Raw+KNN and 42.12% for Raw alone. These results show that the two branches contribute differently to clustering performance, with their relative effectiveness varying across datasets.

W2:

We thank the reviewer for this comment. Unlike linear or additive fusion, which primarily combines representations through weighted summation, low-rank bilinear fusion explicitly models multiplicative interactions between the enhanced representation and its residual refinement. Such interactions allow the model to capture feature dependencies that may not be adequately represented by simple additive operations. Furthermore, low-rank factorization avoids the computational and parameter costs of full bilinear fusion, while the residual connection preserves the enhanced representation.

To evaluate its effectiveness, we compare linear, additive, and low-rank bilinear fusion. Please refer to Dpyw Q2 for the detailed results. Low-rank bilinear fusion achieves the highest ACC on four datasets and ties with linear fusion on CESC. For example, on LGG, ACC increases from 95.72% (linear) and 95.25% (additive) to 96.20%, while NMI reaches 84.90%, compared with 83.21% and 82.62%, respectively. These comparisons suggest that modeling multiplicative feature interactions can improve fusion performance, although the gains are not uniform across all metrics.

Q1:

We select PCA reconstruction and KNN smoothing to capture complementary structural characteristics within each view. Specifically, the PCA-reconstructed Main branch emphasizes dominant global variation while suppressing low-variance fluctuations, whereas the KNN-smoothed Graph branch aggregates information from neighboring samples to enhance local structural consistency. Meanwhile, the Raw branch retains the original feature information, providing a reference for potentially overlooked discriminative relationships.

To evaluate their individual contributions, we conduct branch-removal ablation experiments. Please refer to Dpyw Q1 for the detailed results. Removing PCA (R+K) decreases ACC across all five datasets compared with the full model, while removing KNN (R+P) decreases ACC on three datasets. For example, on CESC, removing PCA or KNN reduces ACC from 93.92% to 91.89% and 79.73%, respectively. On SARC, the corresponding ACC decreases from 64.95% to 59.28% and 54.64%.

Although their contributions vary across datasets, these results indicate that PCA and KNN provide different benefits to clustering, supporting the joint modeling of global principal structure and local neighborhood relationships.

Q2:

The learnable gate adaptively controls cross-view information transfer based on both the target representation and the attention-aggregated information. Specifically, as defined in Eqs. (8)–(9), the gate is computed as

\bar Z_v^b=Z_v^b+\lambda_b G_v^b\odot C_v^b.$$ During training, gradients from the view-level reconstruction loss $L_{\mathrm{rec}}$ and fusion-level reconstruction loss $L_{\mathrm{fus}}$ propagate through the enhanced representations to update the gating mapping. When transferred features help preserve information required for reconstruction, the optimization can increase their contributions; conversely, when they introduce harmful interference, their gate weights can be reduced. Meanwhile, the contrastive objectives $L_{\mathrm{intra}}$ and $L_{\mathrm{coh}}$ encourage semantic consistency in the representations before gating, indirectly improving the quality of cross-view information available for selection. Thus, the gate learns to regulate information transfer through **reconstruction-driven feedback rather than predefined relevance labels**. Although this mechanism can suppress unhelpful transfer, it does not guarantee the complete elimination of irrelevant information.
AuthorsOpenReview ↗

Reply to Reviewer dcSr (Part 2 of 2)

Q3:

To investigate the effect of the low-rank interaction dimension, we conducted a sensitivity analysis with rf = 8, 16, 32, 64, and 128 across five datasets. As shown in the following table, results are reported as ACC/NMI/PUR (%).

Datasetrf=8rf=16rf=32rf=64rf=128
CESC90.54/43.44/90.5489.19/43.59/89.1991.22/48.68/91.2294.59/57.09/94.5992.57/52.69/92.57
LGG93.59/77.26/93.5993.35/75.37/93.3593.82/76.61/93.8295.96/84.20/95.9695.72/83.66/95.72
SARC64.43/40.25/65.9859.79/38.35/67.0163.92/40.17/66.4964.95/39.84/64.9563.40/40.24/67.53
STAD53.70/40.11/75.2452.09/39.30/76.8552.41/39.10/75.2454.34/42.13/76.5355.31/42.73/77.81
UCEC65.19/39.52/66.6763.70/38.77/66.6764.20/38.31/66.4265.43/41.14/67.4163.70/38.60/66.17

The results show that rf=64 achieves the best ACC, NMI, and PUR on CESC, LGG, and UCEC, while rf=128 performs best on STAD. On SARC, the optimal dimension varies across evaluation metrics. Overall, rf=64 provides a favorable balance between clustering performance and interaction dimensionality, supporting its selection as the default setting.

The authors have adequately addressed my concerns and clarified the issues I raised. Their responses have answered my questions, so I have increased my overall rating to 4.

Reviewer_DpywOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LIVC for high-dimensional multi-view clustering. It constructs raw, PCA-reconstructed, and neighborhood-smoothed representations within each view, performs gated cross-view interaction between corresponding branches, and combines the resulting representations through residual low-rank bilinear fusion. The framework is trained using reconstruction and contrastive objectives, followed by K-means clustering. Experiments on five TCGA multi-omics datasets demonstrate competitive clustering performance.

Critical strengths
  1. The paper explores intra-view structural complementarity, a potentially useful perspective for high-dimensional, small-sample clustering.
  2. LIVC achieves the best mean results on 14 of 15 dataset–metric combinations over five seeds.
Critical weaknesses
  1. The paper does not clearly demonstrate which clustering structures are lost by a single representation or establish the complementarity among the three branches. It is therefore difficult to determine whether the improvements arise from structural complementarity or from denoising and smoothing.
  2. The connection between the central problem and the proposed same-granularity cross-view interaction and low-rank bilinear fusion is insufficiently established. How contrastive alignment preserves complementary structures while promoting consistency also lacks sufficient explanation and validation.
Questions for authors
  1. What evidence shows that a single representation loses clustering-relevant structures and that the three branches recover complementary information, beyond the benefits of denoising and smoothing?
  2. Why are same-granularity cross-view interaction and low-rank bilinear fusion necessary to address the identified problem? How does contrastive alignment preserve complementary structures while encouraging consistency?
2 comments
AuthorsOpenReview ↗

Reply to Reviewer Dpyw (Part 1 of 2)

Thank you for your careful review. We address your concerns below.

W1:

A single learned representation may overlook certain clustering-relevant relationships. Raw may miss same-class neighborhoods, PCA reconstruction may suppress low-variance discriminative information, and KNN smoothing may blur cluster boundaries. Thus, the three branches emphasize different structural properties.

To examine their complementarity, we analyze same-class directed top-10 KNN edges across five datasets, five seeds, and four views per dataset. The first three columns report edges absent from Raw but identified only by Main, only by Graph, or by both. The last two report their retention in the fused representation (Same Ret. and Diff. Ret. for newly identified same-class and different-class edges).

DatasetMain OnlyGraph OnlyBothSame Ret. (%)Diff. Ret. (%)
CESC56.90466.5521.0518.069.40
LGG204.601466.8560.4515.543.32
SARC47.15426.3021.9024.2115.74
STAD98.35685.3028.1017.0010.28
UCEC150.65823.2043.7014.755.93

On LGG, Main and Graph identify 204.60 and 1466.85 distinct same-class edges absent from the learned Raw representation, with another 60.45 edges shared by both. Across five datasets, their newly identified edges overlap by only 3.57%–4.50%. Moreover, the fused representation retains a higher proportion of newly identified same-class edges than different-class edges.

We further assess their discriminative value through frozen-representation prediction experiments. The following table reports reductions in held-out log-loss (nats), averaged over five seeds. Positive values indicate improvements. R, P, and K denote Raw, PCA, and KNN, respectively, while All denotes R+P+K.

DatasetR → R+PR → R+KR+P → AllR → All
CESC+0.0002+0.0069+0.0059+0.0061
LGG+0.0009+0.0050+0.0026+0.0034
SARC+0.0050-0.0049-0.0010+0.0040
STAD+0.0041+0.0139+0.0184+0.0225
UCEC+0.0012-0.0024+0.0026+0.0038

On STAD, adding KNN to Raw+PCA reduces held-out log-loss by 0.0184 ± 0.0033 nats, with positive gains across all five seeds. Although the gains vary across datasets, these results support structural complementarity and additional discriminative value beyond Raw alone.

W2:

Raw, Main, and Graph emphasize original information, global principal structure, and local neighborhoods, respectively. Directly mixing these representations across views may introduce interference between different structural roles. We therefore adopt same-granularity cross-view interaction (Raw–Raw, Main–Main, and Graph–Graph), followed by gated information transfer and Raw-centered integration (Eq. 10). This ordering aims to exploit complementary structures while limiting premature interference.

Linear or additive fusion may not explicitly capture multiplicative feature dependencies. LIVC uses low-rank bilinear fusion to model second-order interactions between enhanced and residually refined representations without constructing a full outer product. The residual connection preserves the primary information pathway. For comparisons with linear and additive fusion, please refer to Dpyw Q2.

Eqs. (5) and (6) encourage sample-level semantic consistency, rather than identical structural representations. Reconstruction losses encourage information preservation, while residual connections and learnable gates help regulate information transfer. Thus, contrastive alignment promotes semantic correspondence without explicitly requiring the branches to become identical, although it does not guarantee preservation of complementarity.

AuthorsOpenReview ↗

Reply to Reviewer Dpyw (Part 2 of 2)

Q1:

To evaluate the contributions of different structural branches to clustering performance, we conduct a branch-combination ablation study. The neighborhood complementarity analysis and frozen-representation prediction experiments are presented in DpywW1 .

The following table compares Raw (R), Raw+PCA (R+P), Raw+KNN (R+K), and the full three-branch configuration (R+P+K). Results are reported as ACC/NMI/PUR (%).

DatasetRR+PR+KR+P+K
CESC51.35/5.13/85.8179.73/24.72/85.8191.89/47.29/91.8993.92/52.07/93.92
LGG95.96/83.76/95.9695.49/82.94/95.4994.06/77.49/94.0695.96/84.20/95.96
SARC54.64/42.07/70.1054.64/41.04/70.1059.28/38.94/68.5664.95/39.84/64.95
STAD42.12/30.39/69.7755.31/44.77/79.4251.77/39.73/75.5654.34/42.13/76.53
UCEC63.95/42.88/65.1967.41/45.62/70.3763.70/36.22/66.1765.43/41.14/67.41

On CESC, combining all three branches improves ACC from 51.35% (Raw only) to 93.92%, outperforming both two-branch variants. On SARC, the full model achieves 64.95% ACC, compared with 54.64% for Raw+PCA and 59.28% for Raw+KNN.

While PCA reconstruction and KNN smoothing provide denoising and smoothing benefits, the observed non-overlapping same-class relationships and incremental prediction gains further support the complementary value of combining these structural perspectives.

Q2:

Raw, Main, and Graph capture different structural information. Directly mixing these representations during cross-view interaction may introduce interference due to their different structural roles. We therefore adopt same-granularity cross-view interaction (Raw–Raw, Main–Main, and Graph–Graph), allowing corresponding branches to be enhanced separately across views. Learnable gates regulate transferred information, while Eq. (10) integrates the enhanced Main and Graph representations around the Raw semantic anchor. This same-granularity enhancement followed by cross-granularity integration aims to exploit complementary structures while reducing premature mixing.

After structural enhancement, linear or additive fusion may not adequately capture multiplicative dependencies among latent features. LIVC therefore employs low-rank bilinear fusion to model second-order interactions between the enhanced representation and its residual refinement. Low-rank factorization avoids constructing a full outer product, while the residual connection preserves the enhanced representation. We compare three fusion strategies, as shown below. Results are reported as ACC/NMI/PUR (%).

DatasetLinearAdditiveLow-Rank Bilinear
CESC88.51/42.09/88.5187.84/40.67/87.8488.51/42.09/88.51
LGG95.72/83.21/95.7295.25/82.62/95.2596.20/84.90/96.20
SARC61.86/38.90/65.4658.25/39.34/69.5963.40/40.96/66.49
STAD51.45/38.45/73.9552.09/39.41/74.6052.41/39.52/74.60
UCEC57.53/41.10/65.1961.73/35.94/65.9362.72/36.87/66.42

Low-rank bilinear fusion achieves the highest or tied-highest scores in 13 of 15 comparisons. The exceptions are SARC PUR and UCEC NMI, where additive and linear fusion perform better, respectively.

Regarding contrastive alignment, Eqs. (5) and (6) encourage sample-level semantic consistency without forcing identical structural relationships. View-level and fusion-level reconstruction losses encourage information preservation, while residual pathways and learnable gates help limit the loss of complementary information. Thus, contrastive alignment promotes semantic correspondence, while the other mechanisms support structural complementarity. Although alignment alone cannot guarantee structural preservation, distinct clustering-relevant relationships remain observable in the learned representations. Please refer to Dpyw W1 for details.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.