Scholarlink

<p>Real-world multi-view data are often incomplete and noisy, which poses significant challenges for graph-based incomplete multi-view clustering (IMC) methods. This paper is motivated by two key challenges. First, the pairwise graphs constructed in the original feature space are highly sensitive to noise and can be easily corrupted. Second, existing methods often fail to fully exploit discriminative partition-level information across views, leading to suboptimal recovery of missing relationships. To address these issues, we propose a robust inference framework that simultaneously estimates missing graph entries and explicitly separates noise through a dedicated noise matrix. At the same time, our model incorporates partition-level guidance to enhance the stability and accuracy of pairwise similarity recovery. Extensive experiments on several benchmark datasets demonstrate that our method achieves competitive clustering performance under varying levels of missing data and noise. We also conduct a diagnostic study to illustrate the limitations of representative baselines, which helps clarify the motivation behind our framework and highlights its practical advantages. Our code is available at <a href="https://github.com/ycyanglx-alt/RIMGEIMCcode.">https://github.com/ycyanglx-alt/RIMGEIMC<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\_\)</EquationSource></InlineEquation>code.</a></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

Robust inference of missing graph entries for incomplete multi-view clustering

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Chunyu Yang,
  • Hongyun Yue

摘要

Real-world multi-view data are often incomplete and noisy, which poses significant challenges for graph-based incomplete multi-view clustering (IMC) methods. This paper is motivated by two key challenges. First, the pairwise graphs constructed in the original feature space are highly sensitive to noise and can be easily corrupted. Second, existing methods often fail to fully exploit discriminative partition-level information across views, leading to suboptimal recovery of missing relationships. To address these issues, we propose a robust inference framework that simultaneously estimates missing graph entries and explicitly separates noise through a dedicated noise matrix. At the same time, our model incorporates partition-level guidance to enhance the stability and accuracy of pairwise similarity recovery. Extensive experiments on several benchmark datasets demonstrate that our method achieves competitive clustering performance under varying levels of missing data and noise. We also conduct a diagnostic study to illustrate the limitations of representative baselines, which helps clarify the motivation behind our framework and highlights its practical advantages. Our code is available at https://github.com/ycyanglx-alt/RIMGEIMC \(\_\)code.