Scholarlink

<p>While multi-view unsupervised feature selection (MUFS) has achieved substantial progress, real-world unbalanced and incomplete multi-view data with uneven cross-view missing proportions poses significant challenges. Non-uniform missingness undermines cross-view complementarity and triggers feature selection bias, wherein highly complete views dominate feature evaluation, leading to suboptimal feature subsets, degraded convergence performance, and excessive computational overhead. To tackle this problem, we propose WTUIMUFS, a two stage MUFS method integrating fractal wave-optics optimization (FWOO) and tensor anchor graph joint learning. In the first stage, a Fractal WOO (FWOO) is developed for reliability-aware feature space exploration. By integrating fractal phase-shift perturbation with intensity-guided adaptive updating, FWOO enhances global search capability and alleviates premature convergence under heterogeneous view completeness. In the second stage, a Tensor Anchor Graph Joint Learning model is constructed to simultaneously perform view-aware missing value completion and tensor-guided latent representation modeling, thereby restoring high-order cross-view correlations disrupted by incomplete observations. Additionally, a view-completeness-aware dynamic reweighting mechanism is embedded throughout both stages to adaptively balance view contributions based on observation ratios and reconstruction reliability, mitigating dominance from highly complete views while preventing over-reliance on heavily imputed ones. Extensive experiments on eight benchmark datasets demonstrate that WTUIMUFS consistently outperforms eight state-of-the-art methods in terms of five evaluation metrics, convergence behavior, and runtime efficiency. Under varying missing ratios, average improvements over the second-best method reach 3.85% (ACC), 3.45% (NMI), 4.28% (Purity), 3.21% (F-score), and 2.78% (ARI), confirming the effectiveness, robustness, and computational efficiency of the proposed method.</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

Wave optics optimization with tensor anchor graph joint learning for unbalanced incomplete multi-view unsupervised feature selection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月11日

查看全文
  • Yanli Cai,
  • Fuyin Dai,
  • Zhigang Zhang,
  • Kaifan Zhang

摘要

While multi-view unsupervised feature selection (MUFS) has achieved substantial progress, real-world unbalanced and incomplete multi-view data with uneven cross-view missing proportions poses significant challenges. Non-uniform missingness undermines cross-view complementarity and triggers feature selection bias, wherein highly complete views dominate feature evaluation, leading to suboptimal feature subsets, degraded convergence performance, and excessive computational overhead. To tackle this problem, we propose WTUIMUFS, a two stage MUFS method integrating fractal wave-optics optimization (FWOO) and tensor anchor graph joint learning. In the first stage, a Fractal WOO (FWOO) is developed for reliability-aware feature space exploration. By integrating fractal phase-shift perturbation with intensity-guided adaptive updating, FWOO enhances global search capability and alleviates premature convergence under heterogeneous view completeness. In the second stage, a Tensor Anchor Graph Joint Learning model is constructed to simultaneously perform view-aware missing value completion and tensor-guided latent representation modeling, thereby restoring high-order cross-view correlations disrupted by incomplete observations. Additionally, a view-completeness-aware dynamic reweighting mechanism is embedded throughout both stages to adaptively balance view contributions based on observation ratios and reconstruction reliability, mitigating dominance from highly complete views while preventing over-reliance on heavily imputed ones. Extensive experiments on eight benchmark datasets demonstrate that WTUIMUFS consistently outperforms eight state-of-the-art methods in terms of five evaluation metrics, convergence behavior, and runtime efficiency. Under varying missing ratios, average improvements over the second-best method reach 3.85% (ACC), 3.45% (NMI), 4.28% (Purity), 3.21% (F-score), and 2.78% (ARI), confirming the effectiveness, robustness, and computational efficiency of the proposed method.