Optimized sample-weighted partial least squares.

Area of Science:

• Chemometrics
• Multivariate data analysis
• Machine learning

Background:

• Traditional multivariate calibration methods assume equal sample contribution, ignoring representativeness.
• Partial Least Squares (PLS) is a common method, but its performance can be limited by sample heterogeneity.

Purpose of the Study:

• To introduce a novel multivariate regression method, Optimized Sample-Weighted PLS (OSWPLS).
• To improve prediction accuracy in multivariate calibration by accounting for sample representativeness.

Main Methods:

• Incorporation of weighted sampling concept into PLS.
• Utilizing Particle Swarm Optimization (PSO) to determine optimal sample weights.
• Application and comparison on meat and fuel datasets.

Main Results:

• OSWPLS significantly reduced the root mean squared error of prediction (RMSEP) for meat data from 3.03 to 2.35.
• OSWPLS showed comparable or slightly better performance than PLS for fuel data analytes.
• The method demonstrated stability and efficiency within moderate PSO cycles.

Conclusions:

• OSWPLS offers improved prediction ability over standard PLS by optimizing sample weighting.
• The approach effectively handles sample representativeness in calibration sets.
• OSWPLS is a stable and efficient method for multivariate regression problems.