Scholarlink

<p>Spectral clustering (SC) is a graph-based clustering algorithm widely applied in data mining and image processing. However, existing SC methods fail to deal with large-scale datasets due to the time-consuming construction of graphs and eigenvalue decomposition, and usually require additional <i>k</i>-means clustering to generate discrete label matrix, leading to information loss and effectiveness reduction. To address these issues, an innovative non-parametric discrete graph clustering model called Fast Discrete Spectral Clustering with Harmonic Min-Max Cut on Anchor Similarity Graph (FDSC-HMCA) is proposed in this letter. Specifically, the proposed method has the following features: (1) a novel approach, based on the anchor graph and the second-order similarity technique, is developed to construct the anchor-based adjacency matrix, making the model applicable to large-scale datasets; (2) a harmonic min-max cut framework with the anchor-based adjacency matrix is proposed, which significantly enhances robustness to “worst-case cluster” boundaries; (3) a coordinate descent (CD) algorithm is employed to solve the non-convex optimization problem of FDSC-HMCA, and the discrete label matrix is directly learned without extra <i>k</i>-means. Moreover, the computational complexity of the proposed method is analyzed. Experimental clustering results demonstrate that FDSC-HMCA exhibits good effectiveness and efficiency on six real-world datasets.</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

Fast discrete spectral clustering with Harmonic min-max cut on Anchor Similarity Graph

  • 期刊论文

  • 发表日期：

    2026年4月25日

查看全文
  • Bin Li,
  • Xiaojun Yang,
  • Weihao Zhao,
  • Jing Wang,
  • Feiping Nie

摘要

Spectral clustering (SC) is a graph-based clustering algorithm widely applied in data mining and image processing. However, existing SC methods fail to deal with large-scale datasets due to the time-consuming construction of graphs and eigenvalue decomposition, and usually require additional k-means clustering to generate discrete label matrix, leading to information loss and effectiveness reduction. To address these issues, an innovative non-parametric discrete graph clustering model called Fast Discrete Spectral Clustering with Harmonic Min-Max Cut on Anchor Similarity Graph (FDSC-HMCA) is proposed in this letter. Specifically, the proposed method has the following features: (1) a novel approach, based on the anchor graph and the second-order similarity technique, is developed to construct the anchor-based adjacency matrix, making the model applicable to large-scale datasets; (2) a harmonic min-max cut framework with the anchor-based adjacency matrix is proposed, which significantly enhances robustness to “worst-case cluster” boundaries; (3) a coordinate descent (CD) algorithm is employed to solve the non-convex optimization problem of FDSC-HMCA, and the discrete label matrix is directly learned without extra k-means. Moreover, the computational complexity of the proposed method is analyzed. Experimental clustering results demonstrate that FDSC-HMCA exhibits good effectiveness and efficiency on six real-world datasets.