Hierarchical extreme learning machine for dimensionality reduction in near-infrared spectral analysis

Dimensionality reduction (DR) is a crucial step in the analysis of high-dimensional near-infrared (NIR) spectroscopy. However, traditional spectral DR techniques are often challenged by selection difficulties and limited effectiveness in improving model accuracy. A novel Hierarchical Extreme Learning Machine (HELM) for DR was proposed in this paper. The mutually orthogonal weight vectors are employed between the input data and each hidden node in the HELM. The efficiency of feature learning is enhanced through the reduction of redundancy and the improvement of independence among the extracted features. The Hierarchical Extreme Learning Machine-Partial Least Squares (HELM-PLS) model was developed by integrating HELM with PLS. The findings demonstrated that the HELM-PLS model showed excellent performance regarding R(Zahir et al., 2022²⁾, RMSE and RPD across three public datasets. Compared with other DR methods, the HELM-PLS model showed better prediction accuracy, robustness and generalization capability. These outcomes demonstrate the superiority of the HELM method in spectral data DR and feature extraction. The HELM-PLS model provides an efficient and accurate predictive tool for handling complex spectral data.