Feature selection of infrared spectra analysis with convolutional neural network

Area of Science:

• Chemometrics
• Machine Learning
• Spectroscopy

Background:

• Convolutional Neural Networks (CNNs) excel in data analysis but often lack interpretability, hindering understanding of their decision-making processes.
• Interpreting complex models is crucial for validating results and identifying key features in scientific applications.

Purpose of the Study:

• To develop an interpretable CNN model for infrared data analysis.
• To extract and visualize informative features from CNNs using an ascending stepwise linear regression (ASLR) approach.
• To compare the performance and interpretability of the proposed CNN method with Partial Least Squares Discriminant Analysis (PLS-DA).

Main Methods:

• Implemented an interpretable CNN model for infrared spectral data.
• Utilized an ascending stepwise linear regression (ASLR) to identify informative neurons in the CNN's flatten layer.
• Employed CNN characteristics to visualize active variables linked to extracted neurons.
• Applied Partial Least Squares Discriminant Analysis (PLS-DA) for comparative analysis.

Main Results:

• CNN models achieved high test set accuracies: 93.27% (Tablet), 97.50% (meat), and 96.65% (juice).
• PLS-DA models showed comparable accuracies: 95.19% (Tablet), 95.50% (meat), and 98.17% (juice).
• Both methods demonstrated stable patterns in active variables, and Monte-Carlo cross-validation confirmed the CNN model's repeatability.

Conclusions:

• The proposed interpretable CNN approach effectively analyzes infrared data, providing insights into feature importance.
• The method offers a valuable alternative for understanding complex spectral data, complementing traditional chemometric techniques.
• The study validates the robustness and repeatability of the interpretable CNN strategy for scientific data analysis.