Application of subspace ensemble radical basis function networks to quantitative analysis of near-infrared and mid-infrared spectroscopy

In spectral analysis, one of the challenges is to develop a reliable calibration model. The accuracy of model is of crucial, and even directly determines the availability of analytical methods. For the same dataset, different calibration strategies may have different performance. In analytical chemistry, one of the most important tasks is to develop effective calibration algorithms. Inspired by the advantages of ensemble learning, a calibration algorithm based on subspace and radical basis function (RBF) networks, abbreviated as "SERBF", is proposed for multivariate calibration. The algorithm mainly uses the random changes of the width of Gaussian function in the hidden layer, feature subspace and weight initialization to ensure the diversity required by the ensemble framework. Classic partial least squares (PLS), single-model RBF and non-subspace ensemble RBF algorithms (ERBF) are also considered as references. Three real datasets concerning tablet, wine and soil analysis are chosen for experiment. Compared with PLS, the proposed SERBF algorithm can improve the performance by 16.4%, 43.7% and 32.3% on the test set, respectively. Hence the results confirm that the proposed SERBF can be a feasible method for constructing quantitative model for molecular spectroscopy-based applications.