Predictive modelling of colossal ATR-FTIR spectral data using PLS-DA: empirical differences between PLS1-DA and...

Area of Science:

• Chemometrics and Analytical Spectroscopy
• Machine Learning for Spectral Data Analysis

Background:

• Partial Least Squares-Discriminant Analysis (PLS-DA) is widely used for analyzing high-dimensional, multicollinear data.
• For multi-class problems (K>2), PLS-DA can be implemented as K one-versus-all PLS1-DA models or a single PLS2-DA model.
• Empirical differences between these two PLS-DA approaches require investigation for large spectral datasets.

Purpose of the Study:

• To empirically compare the performance of PLS1-DA and PLS2-DA algorithms.
• To evaluate differences in model accuracy, stability, and fitting for colossal ATR-FTIR spectral data.
• To assess model parsimony and overfitting tendencies of both PLS-DA algorithms.

Main Methods:

• A large dataset of blue gel pen ink ATR-FTIR spectra was utilized.
• Four sub-datasets were created: raw and Asymmetric Least Squares (AsLS) preprocessed, considering global and local spectral regions.
• Multiple PLS-DA models (up to 50 PLS components) were built and evaluated using cross-validation, autoprediction, and external testing.

Main Results:

• Both PLS1-DA and PLS2-DA algorithms demonstrated satisfactory model accuracy and stability.
• PLS1-DA models exhibited significantly higher accuracy rates compared to PLS2-DA models.
• PLS2-DA models showed greater stability and were less prone to overfitting, indicating superior parsimony.

Conclusions:

• PLS1-DA achieves higher accuracy at the expense of reduced parsimony and stability, with an increased risk of overfitting.
• PLS2-DA offers better model stability and parsimony but with a trade-off in accuracy.
• The choice between PLS1-DA and PLS2-DA depends on the specific requirements for accuracy versus stability and parsimony in spectral data modeling.