A novel variable selection algorithm based on neural network for near-infrared spectral modeling.

Area of Science:

• Spectroscopy
• Chemometrics
• Machine Learning

Background:

• Partial Least Squares (PLS) is a standard for spectral data modeling, with many variable selection algorithms developed to improve its predictive ability and interpretability.
• Advancements in neural network technology have led to their increasing application in spectral data modeling, though current research often overlooks variable selection in favor of network architecture.
• Existing neural network approaches for spectral data modeling often prioritize network structure over effective variable selection, limiting model interpretability and predictive performance.

Purpose of the Study:

• To introduce a novel neural network-based variable selection algorithm, VSNN (Variable Selection based on Neural Network), for spectral data.
• To evaluate the performance of VSNN by analyzing the impact of different neural network types, activation functions, and variable importance measures.
• To compare the predictive performance of VSNN against established methods such as Partial Least Squares (PLS), standard Neural Networks (NN), and Joint Mutual Information Maximisation (JMIM).

Main Methods:

• Developed VSNN, a neural network-based variable selection algorithm that iteratively removes unimportant variables using an exponentially decreasing function (EDF).
• Integrated various neural network types and activation functions within the VSNN framework.
• Tested VSNN on four diverse spectral datasets: corn moisture, corn oil, tablets, and meat.

Main Results:

• VSNN significantly enhanced model predictive ability across all tested datasets compared to PLS, NN, and JMIM.
• Non-linear activation functions notably improved VSNN performance on non-linear spectral data, as demonstrated with the meat dataset.
• Root Mean Square Error of Prediction (RMSEP) values were substantially reduced by VSNN, e.g., for the meat dataset from 3.2 to 0.36, and for corn moisture from 0.0409 to 0.002, showcasing improved accuracy.

Conclusions:

• VSNN offers a flexible framework for improving variable selection, modeling, and predictive performance in spectral data analysis.
• The algorithm's adaptability to different neural network architectures and importance evaluation metrics positions it for increased effectiveness with advancing machine learning technologies.
• VSNN demonstrates significant potential as a powerful algorithmic tool for variable selection in spectroscopy, enhancing both model accuracy and interpretability.