Scholarlink

<p>In partial multi-label learning (PML), instances are associated with candidate label sets containing both partial ground-truth labels and noisy labels. This introduces two key challenges: 1) weakened label-instance associations, which act as incorrect supervision and negatively impact classification performance; 2) label correlations can effectively enhance classifier robustness, yet label correlations extracted directly from noisy label information are inherently biased and imprecise. This paper proposes PML-LRIG, a novel method based on local reconstruction and indirect guidance to address these challenges. Our approach focuses on restoring local label-instance associations and extracting reliable label correlations to guide the classifier. We integrate instance clustering with category and geometric correlations to recover labels and generate label distributions. To handle complex distributions, we design a framework for high-order label correlation learning in PML. Additionally, we implement a dual-module classifier where one module learns sample-specific features while the other captures high-order label correlations, indirectly enhancing the classifier’s generalization capabilities. Extensive experiments show that PML-LRIG outperforms state-of-the-art methods.</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

Partial multi–label learning with local reconstruction and indirect guidance

  • 期刊论文

  • 发表日期：

    2026年4月29日

查看全文
  • Yanan Wu,
  • Yu Chen,
  • Hui Chen,
  • Yu Wang,
  • Xiaozhao Fang,
  • Guoxu Zhou,
  • Zhouqiang Qiu,
  • Junqiu Fan

摘要

In partial multi-label learning (PML), instances are associated with candidate label sets containing both partial ground-truth labels and noisy labels. This introduces two key challenges: 1) weakened label-instance associations, which act as incorrect supervision and negatively impact classification performance; 2) label correlations can effectively enhance classifier robustness, yet label correlations extracted directly from noisy label information are inherently biased and imprecise. This paper proposes PML-LRIG, a novel method based on local reconstruction and indirect guidance to address these challenges. Our approach focuses on restoring local label-instance associations and extracting reliable label correlations to guide the classifier. We integrate instance clustering with category and geometric correlations to recover labels and generate label distributions. To handle complex distributions, we design a framework for high-order label correlation learning in PML. Additionally, we implement a dual-module classifier where one module learns sample-specific features while the other captures high-order label correlations, indirectly enhancing the classifier’s generalization capabilities. Extensive experiments show that PML-LRIG outperforms state-of-the-art methods.