Scholarlink

<p>Multi-view anomaly detection leverages diverse information from multiple perspectives to identify anomalies, garnering widespread attention. However, existing methods often struggle to effectively address the complexity of capturing multiple anomaly types using a single scoring criterion. To tackle this challenge, we propose a novel approach, Multi-view Outlier Detection via Tensor Decomposition and Information Decoupling (MOD-TDID). MOD-TDID constructs similarity matrices and enhanced k-nearest neighbour graphs for each view, employing Newton-based Laplacian Optimization (NBLO) to capture the global and local structural features of multi-view data. Through an adaptive weighting strategy, shared and view-specific information is decoupled, enabling our method to flexibly identify commonalities and uniqueness within multi-view data. Moreover, high-order tensor regularization based on Tensor Singular Value Decomposition (t-SVD) enhances robustness and precision in complex data environments. Finally, we introduce a Dual Inconsistency Score (DIS) to comprehensively detect categorical, attribute-based, and hybrid anomalies. Extensive experiments conducted on multiple benchmark datasets demonstrate that MOD-TDID excels in identifying diverse anomalies and capturing intricate patterns, showcasing its potential in multi-view anomaly detection. The source code for this study is available at <a href="https://github.com/YF-W/MOD-TDID">https://github.com/YF-W/MOD-TDID</a>.</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 outlier detection via tensor decomposition and information decoupling

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • Yuefei Wang,
  • Jiaying Mei,
  • Zixu Wang,
  • Jinyan Cao,
  • Xinyuan Wang,
  • Siyi Qiu,
  • Binxiong Li,
  • Xi Yu

摘要

Multi-view anomaly detection leverages diverse information from multiple perspectives to identify anomalies, garnering widespread attention. However, existing methods often struggle to effectively address the complexity of capturing multiple anomaly types using a single scoring criterion. To tackle this challenge, we propose a novel approach, Multi-view Outlier Detection via Tensor Decomposition and Information Decoupling (MOD-TDID). MOD-TDID constructs similarity matrices and enhanced k-nearest neighbour graphs for each view, employing Newton-based Laplacian Optimization (NBLO) to capture the global and local structural features of multi-view data. Through an adaptive weighting strategy, shared and view-specific information is decoupled, enabling our method to flexibly identify commonalities and uniqueness within multi-view data. Moreover, high-order tensor regularization based on Tensor Singular Value Decomposition (t-SVD) enhances robustness and precision in complex data environments. Finally, we introduce a Dual Inconsistency Score (DIS) to comprehensively detect categorical, attribute-based, and hybrid anomalies. Extensive experiments conducted on multiple benchmark datasets demonstrate that MOD-TDID excels in identifying diverse anomalies and capturing intricate patterns, showcasing its potential in multi-view anomaly detection. The source code for this study is available at https://github.com/YF-W/MOD-TDID.