Scholarlink

<p>Multiple kernel clustering (MKC) excels at integrating information from multiple kernels for effective data clustering. Multiple kernel subspace clustering (MKSC) further enhances this by incorporating subspace learning, achieving significant performance improvements. However, most of existing MKSC methods often learn the affinity graphs directly and overlook the rich high-order correlations between different kernels, leading to suboptimal clustering results. This paper introduces MKSC-DT, a novel approach that addresses this limitation by utilizing dual tensors to capture high-order correlations in both feature space and semantic space. MKSC-DT learns multiple new kernels in original feature spaces and projects them onto a clean subspace to generate candidate affinity graphs. These new kernels and candidate graphs are then stacked into two third-order tensors, enabling the exploration of high-order relationships. An efficient alternating optimization algorithm is proposed to solve the resulting objective function. Extensive experiments on benchmark datasets demonstrate the superiority of MKSC-DT compared to state-of-the-art MKC methods, showcasing its effectiveness in leveraging high-order correlations for improved clustering performance.</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

Multiple kernel subspace clustering with dual tensors learning

  • 期刊论文

  • 发表日期：

    2025年6月19日

查看全文
  • Yinsong Liao,
  • Zhenwen Ren,
  • Bin Wu,
  • Chunyu Zhao

摘要

Multiple kernel clustering (MKC) excels at integrating information from multiple kernels for effective data clustering. Multiple kernel subspace clustering (MKSC) further enhances this by incorporating subspace learning, achieving significant performance improvements. However, most of existing MKSC methods often learn the affinity graphs directly and overlook the rich high-order correlations between different kernels, leading to suboptimal clustering results. This paper introduces MKSC-DT, a novel approach that addresses this limitation by utilizing dual tensors to capture high-order correlations in both feature space and semantic space. MKSC-DT learns multiple new kernels in original feature spaces and projects them onto a clean subspace to generate candidate affinity graphs. These new kernels and candidate graphs are then stacked into two third-order tensors, enabling the exploration of high-order relationships. An efficient alternating optimization algorithm is proposed to solve the resulting objective function. Extensive experiments on benchmark datasets demonstrate the superiority of MKSC-DT compared to state-of-the-art MKC methods, showcasing its effectiveness in leveraging high-order correlations for improved clustering performance.