Bagging for robust non-linear multivariate calibration of spectroscopy

This paper presents the application of the bagging technique for non-linear regression models to obtain more accurate and robust calibration of spectroscopy. Bagging refers to the combination of multiple models obtained by bootstrap re-sampling with replacement into an ensemble model to reduce prediction errors. It is well suited to "non-robust" models, such as the non-linear calibration methods of artificial neural network (ANN) and Gaussian process regression (GPR), in which small changes in data or model parameters can result in significant change in model predictions. A specific variant of bagging, based on sub-sampling without replacement and named subagging, is also investigated, since it has been reported to possess similar prediction capability to bagging but requires less computation. However, this work shows that the calibration performance of subagging is sensitive to the amount of sub-sampled data, which needs to be determined by computationally intensive cross-validation. Therefore, we suggest that bagging is preferred to subagging in practice. Application study on two near infrared datasets demonstrates the effectiveness of the presented approach.