A comparative analysis of deep learning and chemometric approaches for spectral data modeling

Area of Science:

• Chemometrics
• Spectroscopic Data Analysis
• Machine Learning

Background:

• Five distinct modeling approaches for spectroscopic data analysis were evaluated.
• These included Partial Least Squares (PLS), interval PLS (iPLS), LASSO, and Convolutional Neural Networks (CNNs).
• Various pre-processing techniques, including classical methods and wavelet transforms, were combined with these models.

Purpose of the Study:

• To conduct a comprehensive comparison of different pre-processing methods and modeling techniques for spectroscopic data.
• To identify optimal combinations for regression and classification tasks in low-data scenarios.
• To evaluate the effectiveness of wavelet transforms as a pre-processing alternative.

Main Methods:

• Evaluated 5 modeling approaches: PLS, iPLS, LASSO, and CNNs.
• Utilized classical pre-processing and wavelet transforms.
• Tested models on beer dataset (regression) and waste lubricant oil dataset (classification).

Main Results:

• Interval PLS (iPLS) variants demonstrated superior performance in the regression case study.
• Wavelet transforms enhanced model performance and maintained interpretability for both linear and CNN models.
• CNNs showed good performance on raw spectra for classification, with further improvements observed after pre-processing.

Conclusions:

• No single pre-processing and model combination is universally optimal, especially in low-data settings.
• Interval PLS is effective for regression, while CNNs offer potential for classification, benefiting from pre-processing.
• Wavelet transforms are a valuable alternative to classical pre-processing for spectroscopic data analysis.