Stacked ensemble extreme learning machine coupled with Partial Least Squares-based weighting strategy for nonlinear...

Area of Science:

• Chemometrics
• Machine Learning
• Spectroscopic Data Analysis

Background:

• Extreme Learning Machine (ELM) is effective for nonlinear multivariate calibration.
• Enhancing ELM's generalization and robustness is crucial for complex datasets.

Purpose of the Study:

• To introduce a modified ELM model, Stacked Ensemble ELM (SE-ELM), for improved chemometric analysis.
• To evaluate different weighting strategies for combining ELM sub-models.

Main Methods:

• Developed SE-ELM by creating sub-models on spectral sub-regions and combining predictions.
• Investigated three weighting strategies: Winner-takes-all (WTA), Constraint Non-negative Least Squares (CNNLS), and Partial Least Squares (PLS).
• Validated models on six spectroscopic datasets, comparing against ELM, BPNN, and RBFNN using Wilcoxon signed rank test.

Main Results:

• All SE-ELM models demonstrated superior robustness and accuracy over traditional ELM.
• The PLS-based weighting strategy outperformed WTA and CNNLS weighting strategies.
• SE-ELM with PLS weighting showed statistically comparable or superior results to BPNN and RBFNN.

Conclusions:

• SE-ELM offers a significant advancement in nonlinear multivariate calibration.
• PLS weighting is recommended as the optimal strategy for SE-ELM due to its ability to handle multi-collinearity.
• The proposed SE-ELM approach provides a more accurate and robust alternative for spectroscopic data analysis.