Scholarlink

<p>A core challenge in deep multi-view clustering is the conflict between preserving view-specific features and enforcing global consistency, which risks representation degeneration. We propose a “protect-then-fuse” framework to resolve this conflict.Our method employs a staged, dual-contrastive architecture: a view-specific loss (L<sub>VSCL</sub>) first enhances individual view representations, which then guide a similarity-based loss (L<sub>SGCL</sub>) to learn a coherent global structure. Experiments show our method outperforms state-of-the-art baselines, with significant gains on complex datasets. Ablation studies further confirm the critical contributions of both the staged training and the dual-loss design. Our framework thus provides an effective solution to the consistency-specificity dilemma, successfully integrating diverse views while mitigating feature suppression. This validates the “protect-then-fuse” strategy as a robust paradigm for the field.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Similarity-guided contrastive learning for deep multi-view clustering

  • 期刊论文

  • 发表日期：

    2025年10月7日

查看全文
  • Guanzheng Jiang,
  • Changming Zhu

摘要

A core challenge in deep multi-view clustering is the conflict between preserving view-specific features and enforcing global consistency, which risks representation degeneration. We propose a “protect-then-fuse” framework to resolve this conflict.Our method employs a staged, dual-contrastive architecture: a view-specific loss (LVSCL) first enhances individual view representations, which then guide a similarity-based loss (LSGCL) to learn a coherent global structure. Experiments show our method outperforms state-of-the-art baselines, with significant gains on complex datasets. Ablation studies further confirm the critical contributions of both the staged training and the dual-loss design. Our framework thus provides an effective solution to the consistency-specificity dilemma, successfully integrating diverse views while mitigating feature suppression. This validates the “protect-then-fuse” strategy as a robust paradigm for the field.