Scholarlink

<p>In non-targeted chemical analysis, quantifying the uncertainty associated with classification confidence remains challenging. Guided by metrological standards, this study introduces an adaptive Monte Carlo method to evaluate the uncertainty in principal component analysis (PCA) and <i>K</i>-means clustering. This approach addresses the limitations of traditional methods in handling uncertainty propagation through high-dimensional data reduction and complex classification algorithms. Using proton nuclear magnetic resonance spectra of natural and artificial cream samples, eight confidence-level calculation methods were implemented and compared. The results identified the SIMCA (Soft Independent Modeling of Class Analogy) weighted averaging method as the optimal approach, followed by the Mahalanobis distance method. To ensure computational stability during extensive Monte Carlo simulations (over 10000 iterations), three novel techniques were introduced: standardization of PCA direction, standardization of cluster numbering (the “minimum-index sorting” method), and a method for handling and reporting the misclassification rate. The relationship between computational precision (defined by the significant figures of standard deviation) and efficiency was investigated, providing practical guidance for balancing accuracy and computational cost. This optimal ranking was further validated across different sample matrices (cream and oil) and classification complexities (2- and 3-cluster), confirming the general robustness of the framework. This work establishes a robust framework for uncertainty evaluation in non-targeted chemical classification, offering&#xa0;a new tool for chemical metrology.</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

An adaptive Monte Carlo method for evaluating uncertainty in confidence level of non-targeted principal component analysis classification

  • 期刊论文

  • 发表日期：

    2026年2月17日

查看全文
  • Ting Huang,
  • Wei Li,
  • Wei Zhang,
  • Huan Yao

摘要

In non-targeted chemical analysis, quantifying the uncertainty associated with classification confidence remains challenging. Guided by metrological standards, this study introduces an adaptive Monte Carlo method to evaluate the uncertainty in principal component analysis (PCA) and K-means clustering. This approach addresses the limitations of traditional methods in handling uncertainty propagation through high-dimensional data reduction and complex classification algorithms. Using proton nuclear magnetic resonance spectra of natural and artificial cream samples, eight confidence-level calculation methods were implemented and compared. The results identified the SIMCA (Soft Independent Modeling of Class Analogy) weighted averaging method as the optimal approach, followed by the Mahalanobis distance method. To ensure computational stability during extensive Monte Carlo simulations (over 10000 iterations), three novel techniques were introduced: standardization of PCA direction, standardization of cluster numbering (the “minimum-index sorting” method), and a method for handling and reporting the misclassification rate. The relationship between computational precision (defined by the significant figures of standard deviation) and efficiency was investigated, providing practical guidance for balancing accuracy and computational cost. This optimal ranking was further validated across different sample matrices (cream and oil) and classification complexities (2- and 3-cluster), confirming the general robustness of the framework. This work establishes a robust framework for uncertainty evaluation in non-targeted chemical classification, offering a new tool for chemical metrology.