PaperlayerView exact revision on arXiv

Source-linked AI summary

Adversarial Consistency-Guided Representation Learning for Multi-view Clustering

Yuchen Lin, Kunpeng Xu, Ying Fang, Lifei Chen

arXiv:2609.35212v1cs.LG

TL;DR

Shared representations in multi-view clustering may retain view-identifying information, potentially compromising cross-view clustering consistency. ACGRL reduces view identifiability adversarially and freezes the resulting references to guide view-specific learning. It outperforms representative methods across four benchmark datasets, with the reported ACC advantage varying by dataset.

  • Problem

    Shared representations can retain view-identifying information, potentially compromising the consistency of cross-view clustering structures.

    Source: §1 · p. 1

  • Method

    ACGRL uses gradient reversal to reduce view identifiability, then freezes the resulting references to guide view-specific learning and clustering.

    Source: §1 · p. 1

  • Results

    ACGRL achieves the highest ACC on all four datasets, exceeding the strongest baseline by 5.40 percentage points on NGs.

    Source: §4.3 · p. 3

  • Takeaways & Limitations

    Ablation and representation-separation analyses further support the effectiveness of ACGRL’s core designs.

    Source: §5 · p. 4

    Higher CCSI values indicate lower average similarity under the selected measures, but do not directly establish semantic disentanglement.

    Source: §4.5 · p. 4

Abstract

from arXiv · show

Multi-view clustering aims to capture cross-view consistency while exploiting view-specific information. However, shared representations learned to capture cross-view consistency may still retain view-identifying information, potentially compromising the consistency of cross-view clustering structures. To address this issue, we propose ACGRL, an adversarial consistency-guided representation learning framework for multi-view clustering. ACGRL employs a gradient-reversal view discriminator to reduce view identifiability and obtain invariant reference representations. These representations are then frozen to provide fixed references for disentangling view-specific information from cross-view common information in the subsequent learning stage. The fixed reference representations are concatenated with the learned view-specific representations for reconstruction and clustering, with cross-view cluster alignment encouraging consistent clustering assignments. Experiments on four benchmark datasets demonstrate the superior clustering performance of ACGRL compared with representative multi-view clustering methods.

1. INTRODUCTION

ACGRL addresses the possibility that shared representations retain view-identifying information, which may compromise cross-view clustering consistency. It reduces view identifiability adversarially, freezes the resulting references to guide view-specific learning, and outperforms representative methods on four benchmark datasets.

1. INTRODUCTION

  • ACGRL uses gradient reversal to reduce view identifiability in shared representations.The paper identifies retained view information as a potential threat to consistent cross-view clustering structures.Source: §1 · p. 1
  • Frozen shared representations guide view-specific learning and are concatenated with those representations for reconstruction and clustering, while cross-view alignment encourages consistent assignments.These fixed references are used to separate common from view-specific information.Source: §1 · p. 1
  • ACGRL outperforms representative multi-view clustering methods on four benchmark datasets.The introduction states the overall comparison without reporting individual dataset values.Source: §1 · p. 1

2. RELATED WORK

Prior multi-view clustering methods model cross-view relations through approaches such as subspace or graph learning, while representation disentanglement methods separate shared and view-specific information. ACGRL targets view-identifying information in the shared component and uses frozen representations to guide view-specific learning.

2.1. Multi-view Clustering

  • Prior multi-view clustering approaches model cross-view relations through subspace or graph learning, while deep methods use reconstruction, self-expression, and representation alignment.EPFMVC learns view-specific representations and fuses them according to inter-view relations; self-expressive approaches learn a shared affinity structure.Source: §2.1 · p. 1

2.2. Multi-view Representation Disentanglement

  • Shared-specific disentanglement methods include variational modeling and distilled disentangling, but separating the components does not necessarily remove view-identifying information from the shared component.ACGRL explicitly targets that information and uses frozen representations to guide view-specific learning.Source: §2.2 · p. 1

3. METHOD

ACGRL learns consistent and complementary representations for multi-view clustering, using adversarially learned, frozen view-invariant representations to supervise view-specific learning before integrating both for reconstruction and clustering.

3.1. Adversarial Consistency Learning

  • A gradient-reversal discriminator makes source-view identification difficult for the consistency extractors, encouraging view-invariant representations that are later frozen as references.The discriminator learns to identify the source view while the consistency extractors are optimized to make that identification difficult.Source: §3.1 · p. 2

3.2. Consistency-Supervised View-Specific Learning

  • Frozen shared representations anchor view-specific learning, while an auxiliary classifier assigns shared features to class 0 and view-specific features to their respective view classes.The view labels encourage view-specific representations to retain their own variations; an entropy term encourages confident assignments.Source: §3.2 · p. 2

3.3. Reconstruction and Clustering Objectives

  • Each view concatenates its invariant and view-specific representations for input reconstruction, while contrastive clustering treats same-cluster assignment vectors across views as positive pairs and different-cluster vectors as negative pairs.The view-specific assignment matrices are uniformly averaged to produce the consensus cluster assignment.Source: §3.3 · p. 3

4. EXPERIMENTS

ACGRL achieves the highest ACC on all four benchmark datasets, with gains over the strongest baseline ranging from 0.15 to 12.88 percentage points. Ablations, CCSI comparisons, and t-SNE visualizations examine its components and learned representations.

4.1. Datasets and Baseline

  • The evaluation uses NGs, BBCSport, Cora, and Hdigit, and compares ACGRL with nine representative multi-view clustering methods.The datasets contain 500, 544, 2,708, and 10,000 samples, respectively.Source: §4.1 · p. 3

4.2. Experimental Setup

  • ACGRL is optimized with Adam and cosine annealing; Stage I and Stage II are trained for 300 and 400 epochs, respectively.The initial learning rate is 10^-4 except for the clustering module, which uses 10^-3.Source: §4.2 · p. 3

4.3. Experimental Results and Analysis

  • ACGRL achieves the highest ACC on all four datasets, beating the strongest baseline by 5.40 points on NGs, 0.15 on Hdigit, 12.88 on BBCSport, and 9.34 on Cora.The comparison is reported in Table 1.Source: §4.3 · p. 3, Table 1 · p. 4

4.4. Ablation Study

  • Removing any of the four objectives reduces ACC, ARI, and NMI on Cora and BBCSport within the complete two-stage framework.Table 2 compares loss-function combinations.Source: §4.4 · p. 3, Table 2 · p. 4
  • Full ACGRL outperforms the MSE- and MI-based Stage I variants and the variant without fixed-reference guidance on NGs and BBCSport.Figure 2 presents the component analysis.Source: §4.4 · p. 3, Figure 2 · p. 3

4.5. Representation Separation Analysis

  • ACGRL has higher CCSI than DMVCS and DFL-NET for within-view common–specific and across-view specific-representation comparisons on all four datasets.Higher CCSI indicates lower average similarity under CKA and CCA, not semantic disentanglement; Figure 3 presents the comparison.Source: §4.5 · p. 4, Figure 3 · p. 4

4.6. Visualization Analysis

  • On Cora and Hdigit, t-SNE shows learned representations with clearer cluster structures than raw features, including improved separation on Cora and more compact, distinguishable clusters on Hdigit.Figure 4 places raw features on the left and learned representations on the right.Source: §4.6 · p. 4, Figure 4 · p. 4

5. CONCLUSION

ACGRL is a two-stage framework that uses adversarial training to suppress view-identifying information in shared representations, then freezes them as guidance for separating shared and view-specific information. Four-dataset experiments and supporting analyses show consistent gains over representative baselines.

5. CONCLUSION

  • ACGRL freezes shared representations after adversarial suppression of view-identifying information, using them as fixed guidance to separate shared and view-specific information in stage two.The framework targets disentangling consistent and complementary representations for multi-view clustering.Source: §5 · p. 4
  • Across four benchmark datasets, ACGRL consistently outperforms representative baselines.Source: §5 · p. 4
  • Ablation and representation-separation analyses further support the effectiveness of ACGRL’s core designs.Source: §5 · p. 4

Compliance with Ethical Standards

The study uses publicly available benchmark datasets and involves no human or animal subjects; therefore, ethical approval was not required.

Compliance with Ethical Standards

Loading 2609.35212v1…