Scholarlink

<p>Change-point detection (CPD) receives extensive studies due to its wide applications in various fields. However, CPD remains a challenging problem for complex data with medium or high dimensions, high correlations, outliers or heavy-tailed distribution. This article proposes an integrated change-point detection method called PCA-uCPD, which utilizes principal components analysis (PCA) to project the original data series into uncorrelated principal components (PCs). Subsequently, we apply existing univariate change-point detection methods to the mapped PCs, followed by a proposed refining technique to obtain the ultimate change-point estimates for the original data sequences. The proposed method admits a flexible architecture that is thus capable of dealing with complex data. Theoretical justifications have been provided to guarantee the feasibility of the proposed methods. Moreover, we conduct simulations to assess performance across various data-generating scenarios. The efficacy of PCA-uCPD is further demonstrated through applications in both genetic and financial 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

PCA-uCPD: an ensemble method for multiple change-point detection in moderately high-dimensional data

  • 期刊论文

  • 发表日期：

    2025年1月16日

查看全文
  • Shanshan Qin,
  • Zhenni Tan,
  • Dongwei Wei,
  • Yuehua Wu

摘要

Change-point detection (CPD) receives extensive studies due to its wide applications in various fields. However, CPD remains a challenging problem for complex data with medium or high dimensions, high correlations, outliers or heavy-tailed distribution. This article proposes an integrated change-point detection method called PCA-uCPD, which utilizes principal components analysis (PCA) to project the original data series into uncorrelated principal components (PCs). Subsequently, we apply existing univariate change-point detection methods to the mapped PCs, followed by a proposed refining technique to obtain the ultimate change-point estimates for the original data sequences. The proposed method admits a flexible architecture that is thus capable of dealing with complex data. Theoretical justifications have been provided to guarantee the feasibility of the proposed methods. Moreover, we conduct simulations to assess performance across various data-generating scenarios. The efficacy of PCA-uCPD is further demonstrated through applications in both genetic and financial datasets.