Linear or non-linear multivariate calibration models? That is the question.

Area of Science:

• Analytical Chemistry
• Data Science
• Chemometrics

Background:

• Data science, machine learning, deep learning, and artificial neural networks are increasingly used across disciplines to analyze complex, non-linear data.
• Chemometrics offers statistical tools for analytical chemistry but often results in overly complex models that do not improve predictive power, violating the principle of parsimony.

Purpose of the Study:

• To demonstrate that easily interpretable modified linear models can efficiently solve non-linear analytical data sets.
• To introduce a local sample selection method for model building in chemometrics.

Main Methods:

• Utilized modified linear models based on local sample selection prior to model building.
• Selected subsets of samples from the neighborhood of unknown samples in latent variable space.
• Applied the method to two near-infrared spectroscopy (NIRS) datasets for food sample analysis.

Main Results:

• Non-linear analytical data were solved with equal or better efficiency using the proposed models.
• Local regression achieved analytical performance comparable to more complex chemometric models.
• The local regression approach resulted in a considerably lower computational burden.

Conclusions:

• Modified linear models with local sample selection provide an efficient and interpretable alternative for analyzing complex analytical data.
• This method achieves high analytical performance without the computational complexity of advanced chemometric techniques.
• The findings support the use of parsimonious models in chemometrics, particularly in applications like NIRS analysis of food properties.