Scholarlink

<p>Partial multi-label learning (PML) tackles the challenge of learning from multi-label data with noisy labels. In addition to the interference from noisy labels, multi-label data also often suffers from the "curse of dimensionality" induced by high-dimensional features, which hinders the performance of existing PML algorithms. To simultaneously eliminate the adverse effects induced by noisy labels and redundant features, this paper proposes a novel partial multi-label feature selection based on feature-label collaboration, namely PMFS-FLC. Specifically, PMFS-FLC first employs a regression model to learn the latent label distributions from feature space, and applies an <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(l_1\)</EquationSource> </InlineEquation>-norm-based loss function to measure the discrepancy between the label distribution and the candidate label set, thereby enhancing robustness against noise. Then, a feature-label collaboration regularization is formulated, leveraging negative information between features and labels to accurately recover label distributions. Finally, by imposing an <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(l_{2,1}\)</EquationSource> </InlineEquation>-norm constraint on the weight matrix, PMFS-FLC performs feature selection within an explicit optimization framework. Extensive experiments on multiple datasets verify that PMFS-FLC outperforms state-of-the-art counterparts. The codes and datasets can be downloaded from <a href="https://github.com/zhenzhenSun-FZU/PMFS-FLC">https://github.com/zhenzhenSun-FZU/PMFS-FLC</a>.</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

PMFS-FLC: a partial multi-label feature selection with feature and label collaboration

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Zhenzhen Sun,
  • Zexiang Chen,
  • Jinghua Liu

摘要

Partial multi-label learning (PML) tackles the challenge of learning from multi-label data with noisy labels. In addition to the interference from noisy labels, multi-label data also often suffers from the "curse of dimensionality" induced by high-dimensional features, which hinders the performance of existing PML algorithms. To simultaneously eliminate the adverse effects induced by noisy labels and redundant features, this paper proposes a novel partial multi-label feature selection based on feature-label collaboration, namely PMFS-FLC. Specifically, PMFS-FLC first employs a regression model to learn the latent label distributions from feature space, and applies an \(l_1\) -norm-based loss function to measure the discrepancy between the label distribution and the candidate label set, thereby enhancing robustness against noise. Then, a feature-label collaboration regularization is formulated, leveraging negative information between features and labels to accurately recover label distributions. Finally, by imposing an \(l_{2,1}\) -norm constraint on the weight matrix, PMFS-FLC performs feature selection within an explicit optimization framework. Extensive experiments on multiple datasets verify that PMFS-FLC outperforms state-of-the-art counterparts. The codes and datasets can be downloaded from https://github.com/zhenzhenSun-FZU/PMFS-FLC.