Scholarlink

<p>As the dimensionality of multi-label data continues to grow, multi-label feature selection has garnered increasing attention. While many existing approaches leverage latent semantic indexing to embed the label space into a lower-dimensional form, they often underutilize label information and ignore crucial properties. To overcome these challenges, we present a unified framework for Multi-View Orthogonal Feature Selection (MVOFS). MVOFS improves label representation and alleviates label imbalance by learning orthogonal subspaces and jointly optimizing feature components. Specifically, it integrates three complementary mechanisms: (1) a latent orthogonal basis sharing strategy to construct a redundancy-free latent sub-label space, (2) a dual-feature collaboration scheme that models both label-shared and label-specific features to balance discriminability and diversity, and (3) a dynamic graph regularization module that adaptively updates similarity structures during optimization to preserve consistency in the latent space. Extensive experiments on 20 benchmark datasets demonstrate that MVOFS achieves consistently competitive performance compared to state-of-the-art multi-label feature selection methods, validating the effectiveness and robustness of the proposed framework.</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

Redundancy-free label space and dual-feature collaboration for multi-label feature selection

  • 期刊论文

  • 发表日期：

    2025年11月10日

查看全文
  • Yaqing Liu,
  • Ying Li,
  • Ying Wang,
  • Haoqian Wu,
  • Haoyu Guo

摘要

As the dimensionality of multi-label data continues to grow, multi-label feature selection has garnered increasing attention. While many existing approaches leverage latent semantic indexing to embed the label space into a lower-dimensional form, they often underutilize label information and ignore crucial properties. To overcome these challenges, we present a unified framework for Multi-View Orthogonal Feature Selection (MVOFS). MVOFS improves label representation and alleviates label imbalance by learning orthogonal subspaces and jointly optimizing feature components. Specifically, it integrates three complementary mechanisms: (1) a latent orthogonal basis sharing strategy to construct a redundancy-free latent sub-label space, (2) a dual-feature collaboration scheme that models both label-shared and label-specific features to balance discriminability and diversity, and (3) a dynamic graph regularization module that adaptively updates similarity structures during optimization to preserve consistency in the latent space. Extensive experiments on 20 benchmark datasets demonstrate that MVOFS achieves consistently competitive performance compared to state-of-the-art multi-label feature selection methods, validating the effectiveness and robustness of the proposed framework.