Scholarlink

<p>In multi-view partial multi-label (MVPML) learning, each instance is characterized by multiple heterogeneous feature representations and is associated with a set of candidate labels that often include redundancies. The primary goal of MVPML is to identify the relevant labels from the candidate labels accurately. Existing methods focus on capturing the topological structures among samples from different views in the feature space to guide label disambiguation. However, they often overlook the presence of view-specific bias and redundant information within the topologies, which ultimately hinders the classification performance. To overcome these limitations, we propose a novel MVPML framework, called <Emphasis Type="BoldItalic">I</Emphasis><i>ntegrating</i> <Emphasis Type="BoldItalic">C</Emphasis><i>ross-</i><Emphasis Type="BoldItalic">G</Emphasis><i>raph</i> <Emphasis Type="BoldItalic">C</Emphasis><i>onsensus Constraints for Label Disambiguation in Multi-View Partial Multi-Label Learning</i> (ICGC), which explicitly explores both inter-view consistency and view-specific uniqueness by employing cross-graph constraints. Specifically, a sparsity constraint on individuality is imposed across graphs to construct more refined and consistent topological structures among samples from different views, which are subsequently integrated into a unified weight matrix. Furthermore, graph Laplacian regularization is utilized to facilitate the transfer of consensus information from the feature space to the label space, thereby improving the precision of label disambiguation. Comprehensive experimental evaluations on real-world datasets validate the superior performance of our model.</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

Integrating cross-graph consensus constraints for label disambiguation in multi-view partial multi-label learning

  • 期刊论文

  • 发表日期：

    2025年10月21日

查看全文
  • Yutong Wu,
  • Zheming Xu,
  • Congyan Lang,
  • Songhe Feng

摘要

In multi-view partial multi-label (MVPML) learning, each instance is characterized by multiple heterogeneous feature representations and is associated with a set of candidate labels that often include redundancies. The primary goal of MVPML is to identify the relevant labels from the candidate labels accurately. Existing methods focus on capturing the topological structures among samples from different views in the feature space to guide label disambiguation. However, they often overlook the presence of view-specific bias and redundant information within the topologies, which ultimately hinders the classification performance. To overcome these limitations, we propose a novel MVPML framework, called Integrating Cross-Graph Consensus Constraints for Label Disambiguation in Multi-View Partial Multi-Label Learning (ICGC), which explicitly explores both inter-view consistency and view-specific uniqueness by employing cross-graph constraints. Specifically, a sparsity constraint on individuality is imposed across graphs to construct more refined and consistent topological structures among samples from different views, which are subsequently integrated into a unified weight matrix. Furthermore, graph Laplacian regularization is utilized to facilitate the transfer of consensus information from the feature space to the label space, thereby improving the precision of label disambiguation. Comprehensive experimental evaluations on real-world datasets validate the superior performance of our model.