Scholarlink

<p>Multi-label feature selection has garnered significant attention in the fields of machine learning and data mining due to its pivotal role in handling high-dimensional data and multi-label classification problems. Pseudo-label learning techniques are frequently employed to mitigate the incompatibility between the binary nature of real labels and linear mapping. However, existing methods often overlook the preservation of the geometric structure of the pseudo-label matrix. The geometric structure of the pseudo-label matrix, such as orthogonality, is crucial for maintaining the correlation and independence of feature vectors. Yet, current methods typically neglect this aspect, leading to insufficient discriminative power and consistency of pseudo-labels. To address these issues, this study proposes a novel multi-label feature selection method—Sparse Multi-label Feature Selection via Label Rotation Learning (LRSMFS). This method introduces a continuous pseudo-label matrix to simulate the distribution of real labels, capturing the complex relationships between features and labels. Simultaneously, it employs the <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\varvec{\ell }_{\varvec{2,0}}\)</EquationSource> </InlineEquation>-norm to control the sparsity of feature selection. By incorporating an orthogonal rotation matrix to optimize the pseudo-labels, the method preserves the geometric structure of feature vectors, thereby reducing the impact of noisy labels and enhancing the accuracy and consistency of pseudo-labels. Additionally, manifold learning is utilized to capture the local structural information of the data, constructing a comprehensive objective function. An alternating optimization strategy is adopted to update the feature selection matrix, pseudo-label matrix, and orthogonal rotation matrix, significantly improving the ability to model label correlations and the generalization capability of the model. The suggested approach shows significant performance improvements over comparable comparative approaches, according to comprehensive experiments carried out on eight multi-label 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

Sparse Multi–label feature selection via label rotation learning

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • Ruijia Li,
  • Yingcang Ma,
  • Hong Chen,
  • Qing Wan

摘要

Multi-label feature selection has garnered significant attention in the fields of machine learning and data mining due to its pivotal role in handling high-dimensional data and multi-label classification problems. Pseudo-label learning techniques are frequently employed to mitigate the incompatibility between the binary nature of real labels and linear mapping. However, existing methods often overlook the preservation of the geometric structure of the pseudo-label matrix. The geometric structure of the pseudo-label matrix, such as orthogonality, is crucial for maintaining the correlation and independence of feature vectors. Yet, current methods typically neglect this aspect, leading to insufficient discriminative power and consistency of pseudo-labels. To address these issues, this study proposes a novel multi-label feature selection method—Sparse Multi-label Feature Selection via Label Rotation Learning (LRSMFS). This method introduces a continuous pseudo-label matrix to simulate the distribution of real labels, capturing the complex relationships between features and labels. Simultaneously, it employs the \(\varvec{\ell }_{\varvec{2,0}}\) -norm to control the sparsity of feature selection. By incorporating an orthogonal rotation matrix to optimize the pseudo-labels, the method preserves the geometric structure of feature vectors, thereby reducing the impact of noisy labels and enhancing the accuracy and consistency of pseudo-labels. Additionally, manifold learning is utilized to capture the local structural information of the data, constructing a comprehensive objective function. An alternating optimization strategy is adopted to update the feature selection matrix, pseudo-label matrix, and orthogonal rotation matrix, significantly improving the ability to model label correlations and the generalization capability of the model. The suggested approach shows significant performance improvements over comparable comparative approaches, according to comprehensive experiments carried out on eight multi-label datasets.