Automated support vector regression

Multivariate calibration is an important procedure for analytical chemistry. Automated or self-configuring methods can be used by scientists who lack expertise, may be embedded into data processing pipelines, and are less prone to user bias; however, the development of such algorithms is often neglected by the chemometrics community. Support vector regression (SVR) is a powerful method for accommodating megavariate data. SVR offers the advantage of fast calibration and flexibility in a variety of loss functions (ie, minimization of the residual error). By embedding bootstrapped Latin partitions (BLPs) into the calibration, the key parameter, the cost C, can be optimized to furnish an automated method. The methods are termed super SVR (sSVR). The BLP predictions of the calibration set accurately model the external prediction error of the entire calibration set. Prediction rates for super partial least squares (sPLS) are compared with sSVRs using three loss functions, Gauss, Laplace, and Huber. For linear data with uniformly distributed noise, sPLS is faster and gave better predictions. However, for data with outliers or a real data set of single-beam near infrared spectra of bovine plasma, gasoline, and wheat, the sSVRs performed better than sPLS, and generally, the Huber loss function gave the best results. Support vector regression (SVR) offers the advantage of fast calibration and flexibility with a variety of loss functions. By embedding bootstrapped Latin partitions (BLPs) into the calibration, the cost C can be optimized to automate the method, super SVR (sSVR). Prediction rates for super partial least squares (sPLS) are compared to sSVRs using the Gauss, Laplace, and Huber loss functions. For plasma, gasoline, and wheat NIR standard data sets, the sSVR performed better than sPLS, although sPLS was faster.