Scholarlink

<p>In Multi-View Partial Multi-label Learning (MVPML), each instance is represented by feature vectors from multiple views and is associated with multiple labels, among which only a subset is valid. Since multi-view data typically exhibits higher dimensionality than single-view data, extracting discriminative features becomes more challenging. However, existing feature selection methods either focus solely on single-view settings or assume that all labels are accurately annotated. To overcome these limitations, we propose the first Multi-View Partial Multi-label Feature Selection method based on label disambiguation and shared subspace (MVPMFS). Specifically, we first perform label disambiguation by fully leveraging the fused similarity, near and far neighbors, and negative labels. Next, we model a latent subspace shared by the feature spaces of multiple views and the label space, along with corresponding latent mappings. We then capture high-order relationships between features and labels through structural alignment. Finally, we propose the concept of feature-specific labels (FSL) and develop a weight adjustment strategy based on FSL to further eliminate invalid feature weights. Experimental results demonstrate that MVPMFS achieves superior performance on multiple synthetic MVPML 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

Multi-view partial multi-label feature selection based on label disambiguation and shared subspace

  • 期刊论文

  • 发表日期：

    2026年4月14日

查看全文
  • Yemin Han,
  • Jing Chai,
  • Fa Zhu,
  • Xingchi Chen,
  • David Camacho

摘要

In Multi-View Partial Multi-label Learning (MVPML), each instance is represented by feature vectors from multiple views and is associated with multiple labels, among which only a subset is valid. Since multi-view data typically exhibits higher dimensionality than single-view data, extracting discriminative features becomes more challenging. However, existing feature selection methods either focus solely on single-view settings or assume that all labels are accurately annotated. To overcome these limitations, we propose the first Multi-View Partial Multi-label Feature Selection method based on label disambiguation and shared subspace (MVPMFS). Specifically, we first perform label disambiguation by fully leveraging the fused similarity, near and far neighbors, and negative labels. Next, we model a latent subspace shared by the feature spaces of multiple views and the label space, along with corresponding latent mappings. We then capture high-order relationships between features and labels through structural alignment. Finally, we propose the concept of feature-specific labels (FSL) and develop a weight adjustment strategy based on FSL to further eliminate invalid feature weights. Experimental results demonstrate that MVPMFS achieves superior performance on multiple synthetic MVPML datasets.