Scholarlink

<p>In multi-view clustering, the assumption of complete views often fails in practice due to missing data, motivating the study of incomplete multi-view clustering (IMC). While many IMC methods have achieved promising results to some extent, most struggle with high computational costs, which limits their scalability to large datasets. Graph-based IMC approaches mitigate this issue by using anchor graphs to approximate full similarity matrices. However, current methods that rely on static anchor graph construction still face scalability issues. Their fusion mechanisms often overlook the structure of anchor graphs, failing to accelerate the spectral clustering process effectively when dealing with three or more views. To address these issues, we introduce a novel large-scale incomplete multi-view clustering based on unified structure-preserving anchor graph LIMC-USAG. First, LIMC-USAG introduces a novel fusion strategy that preserves the structure of individual anchor graphs and can naturally scale to scenarios involving more than two views. Furthermore, we adopt k-means algorithm to select a small set of representative anchors from each view and compute sample-anchor similarities using a gaussian kernel function, enabling efficient and time-saving similarity computation. Recognizing the impact of the Gaussian bandwidth parameter on clustering performance, we further analyze and tune it based on the characteristics of dataset, improving both adaptability and robustness across different scenarios. Extensive experiments demonstrate that LIMC-USAG achieves consistently competitive performance in both accuracy and efficiency, suggesting its potential for large-scale IMC applications.</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

Large-scale incomplete multi-view clustering based on unified structure-preserving anchor graph

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Xiaolan Liu,
  • Gaojian Luo,
  • Mengying Xie,
  • Pei Huang

摘要

In multi-view clustering, the assumption of complete views often fails in practice due to missing data, motivating the study of incomplete multi-view clustering (IMC). While many IMC methods have achieved promising results to some extent, most struggle with high computational costs, which limits their scalability to large datasets. Graph-based IMC approaches mitigate this issue by using anchor graphs to approximate full similarity matrices. However, current methods that rely on static anchor graph construction still face scalability issues. Their fusion mechanisms often overlook the structure of anchor graphs, failing to accelerate the spectral clustering process effectively when dealing with three or more views. To address these issues, we introduce a novel large-scale incomplete multi-view clustering based on unified structure-preserving anchor graph LIMC-USAG. First, LIMC-USAG introduces a novel fusion strategy that preserves the structure of individual anchor graphs and can naturally scale to scenarios involving more than two views. Furthermore, we adopt k-means algorithm to select a small set of representative anchors from each view and compute sample-anchor similarities using a gaussian kernel function, enabling efficient and time-saving similarity computation. Recognizing the impact of the Gaussian bandwidth parameter on clustering performance, we further analyze and tune it based on the characteristics of dataset, improving both adaptability and robustness across different scenarios. Extensive experiments demonstrate that LIMC-USAG achieves consistently competitive performance in both accuracy and efficiency, suggesting its potential for large-scale IMC applications.