Finding features - variable extraction strategies for dimensionality reduction and marker compounds identification in...

Area of Science:

• Analytical Chemistry
• Chemometrics
• Food Science

Background:

• Gas Chromatography-Ion Mobility Spectrometry (GC-IMS) is a powerful technique for analyzing volatile organic compounds (VOCs).
• Non-target screening (NTS) with GC-IMS generates large datasets requiring dimensionality reduction for machine learning.
• Variable reduction is crucial to prevent overfitting, reduce training times, and simplify models.

Purpose of the Study:

• To compare Principal Component Analysis (PCA) and Partial Least Squares (PLS) for dimensionality reduction in GC-IMS data.
• To evaluate their effectiveness in identifying marker compounds and as preprocessing steps for supervised learning.
• To assess interpretability and performance in analyzing a honey botanical origin dataset.

Main Methods:

• Applied PCA and PLS to a GC-IMS dataset from honey samples.
• Explained data formatting for high-dimensional GC-IMS data.
• Integrated PCA and PLS as preprocessing steps for four supervised learning algorithms.

Main Results:

• Both PCA and PLS offer per-variable visualizations aiding interpretability and marker discovery.
• PLS demonstrated superior performance as a preprocessing step for supervised learning, enhancing accuracy.
• PCA proved effective in identifying preprocessing weaknesses, such as data misalignments.

Conclusions:

• Dimensionality reduction techniques like PCA and PLS are essential for GC-IMS data analysis.
• PLS is recommended for improving accuracy in supervised learning workflows.
• PCA is valuable for diagnostic analysis of data quality and preprocessing steps.