Scholarlink

<p>In real-world scenarios, the prevalence of incomplete multi-view data has spurred interest in incomplete multi-view clustering (IMVC). Scholars have used information entropy to quantify incomplete view information, addressing the prediction and recovery of missing data in IMVC. However, current IMVC methods based on information theory face three challenges: 1) Conventional methods are constrained by the theoretical limitations of mutual information and conditional entropy, which restrict their applicability to dual-view scenarios and lack extensibility for multi-view frameworks. 2) Current data recovery approaches predominantly rely on single- or dual-view information, failing to fully exploit the comprehensive semantics of multi-modal data. 3) Traditional clustering methods neglect the quality discrepancies across views after data recovery, resulting in suboptimal discriminative power of fused representations. To address these challenges, this paper proposes a <Emphasis Type="BoldUnderline">Glo</Emphasis>bal-driven Information Theo<Emphasis Type="BoldUnderline">ry</Emphasis> for <Emphasis Type="BoldUnderline">I</Emphasis>ncomplete <Emphasis Type="BoldUnderline">M</Emphasis>ulti-<Emphasis Type="BoldUnderline">V</Emphasis>iew <Emphasis Type="BoldUnderline">C</Emphasis>lustering (GloryIMVC). Specifically, we extend the information-theoretic framework to multi-view settings by introducing global cross-view consistency learning, which leverages unified global representations to predict missing data, thereby addressing the scalability limitations of existing frameworks. Additionally, a complementary learning strategy driven by global representations is proposed, which integrates complementary information from multiple views to significantly enhance the quality of recovered data. Finally, we design an adaptive fusion mechanism with modulation factors to dynamically balance the reliability of predicted information, thereby generating a more discriminative global representation for clustering tasks. Extensive experiments on six datasets validate the effectiveness of GloryIMVC compared to 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

GloryIMVC: global-driven information theory for incomplete multi-view clustering

  • 期刊论文

  • 发表日期：

    2026年1月21日

查看全文
  • Yong Zhang,
  • Li Jiang,
  • Wenzhe Liu,
  • Hongwei Yin

摘要

In real-world scenarios, the prevalence of incomplete multi-view data has spurred interest in incomplete multi-view clustering (IMVC). Scholars have used information entropy to quantify incomplete view information, addressing the prediction and recovery of missing data in IMVC. However, current IMVC methods based on information theory face three challenges: 1) Conventional methods are constrained by the theoretical limitations of mutual information and conditional entropy, which restrict their applicability to dual-view scenarios and lack extensibility for multi-view frameworks. 2) Current data recovery approaches predominantly rely on single- or dual-view information, failing to fully exploit the comprehensive semantics of multi-modal data. 3) Traditional clustering methods neglect the quality discrepancies across views after data recovery, resulting in suboptimal discriminative power of fused representations. To address these challenges, this paper proposes a Global-driven Information Theory for Incomplete Multi-View Clustering (GloryIMVC). Specifically, we extend the information-theoretic framework to multi-view settings by introducing global cross-view consistency learning, which leverages unified global representations to predict missing data, thereby addressing the scalability limitations of existing frameworks. Additionally, a complementary learning strategy driven by global representations is proposed, which integrates complementary information from multiple views to significantly enhance the quality of recovered data. Finally, we design an adaptive fusion mechanism with modulation factors to dynamically balance the reliability of predicted information, thereby generating a more discriminative global representation for clustering tasks. Extensive experiments on six datasets validate the effectiveness of GloryIMVC compared to state-of-the-art methods.