Response to letter to the editor from Y. Takefuji on “Beyond principal component analysis: Enhancing feature reduction in electronic noses through robust statistical methods”

Abstract

Background

Principal Component Analysis (PCA) is extensively utilized in Electronic Nose (E-nose) research for dimensionality reduction, allowing simplification of high-dimensional data and enhancing computational efficiency. However, its dependency on linear assumptions and sensitivity to outliers pose significant challenges, particularly when faced with nonlinear or overlapping datasets.

Scope and approach

This paper explores advanced methods such as Kernel PCA (KPCA), t-Distributed Stochastic Neighbor Embedding (t-SNE), and Uniform Manifold Approximation and Projection (UMAP), which are more adept at managing nonlinear data, as well as nonparametric methods like Spearman's correlation and Kendall's tau to illuminate sensor data relationships. Furthermore, we examine the necessity of Variance Inflation Factor (VIF) analysis in addressing multicollinearity, highlighting hybrid approaches like Random Forest-VIF and PCA-VIF that enhance model stability and interpretability.

Key findings and conclusions

The original article by Zheng et al. (2025) demonstrates the broad applicability of PCA in detecting alcoholic beverages using E-noses, yet emphasizes the requirement for further research into its limitations. While PCA is foundational, its shortcomings call for the integration of advanced methodologies that cater to the complexities of E-nose data. Future research should focus on refining preprocessing protocols, utilizing nonlinear techniques, and managing data variability to improve accuracy and robustness, ultimately expanding E-nose applications across various domains and ensuring reliable performance.