Robust boosting neural networks with random weights for multivariate calibration of complex samples

Area of Science:

• Chemometrics
• Machine Learning
• Spectroscopy

Background:

• Neural networks with random weights (NNRW) show promise for regression tasks.
• However, NNRW is susceptible to outliers and can be unstable with complex real-world data.

Purpose of the Study:

• To introduce a novel method, robust boosting NNRW (RBNNRW), to enhance the performance of NNRW.
• To address the limitations of NNRW concerning outlier sensitivity and stability.

Main Methods:

• RBNNRW integrates a robust boosting technique with NNRW.
• It sequentially builds NNRW sub-models using robust reweighted sampling.
• Predictions are aggregated using a weighted median approach.

Main Results:

• RBNNRW demonstrated enhanced stability and accuracy compared to conventional PLS, NNRW, and boosting NNRW (BNNRW).
• The proposed method showed superior performance, particularly in the presence of outliers in spectral datasets (wheat, light gas oil, diesel fuel).

Conclusions:

• The integration of robust boosting significantly improves the stability and accuracy of NNRW.
• RBNNRW offers a more reliable approach for regression analysis with complex and potentially outlier-prone datasets.