Recently, a research team from the Institute of Intelligent Machines, the Hefei Institutes of Physical Science of the Chinese Academy of Sciences, proposed a novel variable selection algorithm for neural network, which they named VSNN. This algorithm aims to improve spectral data analysis in nonlinear models.
The related research findings have been published in Analytica Chimica Acta.
Spectral analysis is widely used in many fields, but extracting useful information from complex data is a challenge. Traditional methods like partial least squares are not effective for nonlinear data, but neural networks work well in such cases.
In this study, the research team designed evaluation vectors for assessing the importance of neural network variables based on neural network interpretability studies.
The researchers created evaluation vectors—sensitivity vectors, gradient vectors, and weight gradient vectors—to assess which variables are most important in a neural network model. The VSNN algorithm works across different neural network models and helps to identify key information from spectral data.
Tests on datasets for corn, tablets, and meat showed that VSNN outperforms other variable selection methods and improves the accuracy of predictions. This makes the algorithm a valuable tool for spectral data analysis.
This study offered a promising tool for improving model accuracy and performance as machine learning technology continues to evolve.
Researchers proposed a novel variable selection algorithm for neural network. (Image by ZHANG Pengfei)
