Performance evaluation of relevance vector machines as a nonlinear regression method in real-world chemical spectroscopic data

Conventional multivariate calibration methods have been developed in chemometrics, using linear regression techniques as principal component regression (PCR) and partial least squares (PLS). Nevertheless, nonlinear methods such as neural networks have been also introduced, and more recently support vector (SVR) based methods. This paper presents the application of relevance vector machines regression method (RVMR) as an alternative regression technique based on the Bayesian theory, for the prediction of physical-chemical properties from chemical spectroscopic data of different instrumental sources. In terms of measuring the real effectiveness and generalization capability of this approach, a comparison study of its performance with other known regression techniques are presented. The good results obtained in terms of root mean square error of prediction (RMSEP) in the prediction of properties of interest, combined with the high sparseness capability exhibited, make this approach a good alternative to solve multivariate regression problems in practice.