PROSAC as a selection tool for SO-PLS regression: A strategy for multi-block data fusion

Area of Science:

• Chemometrics
• Data Science
• Spectroscopy

Background:

• Multi-block fusion chemometric models like SO-PLS integrate spectral data for improved sample quality prediction.
• Pre-processing is crucial to reduce noise but selecting methods and blocks is complex with many data sources.
• Efficiently handling pre-processing, block selection, and ordering is vital for SO-PLS model performance.

Purpose of the Study:

• To address the challenge of efficient pre-processing, selection, and ordering of data blocks for targeted SO-PLS applications.
• To introduce a novel methodology that automates and optimizes these complex steps.
• To improve the accuracy and efficiency of chemometric models in handling large, multi-source spectral datasets.

Main Methods:

• Introduction of the PROSAC-SO-PLS methodology, utilizing pre-processing ensembles with response-oriented sequential alternation calibration (PROSAC).
• Implementation of a stepwise forward selection strategy, aided by the Gram-Schmidt process, to prioritize data blocks.
• Blocks are selected based on their effectiveness in minimizing prediction error, indicated by reduced prediction residuals.

Main Results:

• PROSAC-SO-PLS successfully identified optimal pre-processed data blocks and their sequential order for SO-PLS.
• Empirical validation on three near-infrared (NIR) datasets demonstrated consistent superiority over single-block PLS and PROSAC-only methods.
• Significant reduction in prediction errors, with RMSEP decreasing by 5-25% for seven out of eight response variables analyzed.

Conclusions:

• PROSAC-SO-PLS provides a versatile and efficient approach for ensemble pre-processing in NIR data modeling.
• It simplifies the application of SO-PLS by mitigating concerns regarding pre-processing sequence and block order.
• This methodology streamlines data pre-processing and model building, enhancing the accuracy and efficiency of chemometric analyses.