Selection and transformation of input variables for RVM based on MI–PCA–MI and 4‐CBA concentration model

Considering that chemical process is a high nonlinear system with high-dimensional and complex variables, relevance vector machine (RVM) with the novel selection and transformation of the candidate input variables, which combines principal component analysis (PCA) with mutual information (MI), is proposed and named as MI–PCA–MI–RVM. First, to omit the irrelevant input variables, the selected input variables are determined according to their correlation degree with output variable via MI. Second, to eliminate the multicollinearity among the selected input variables, PCA is applied for them to be principal components (PCs). Third, the PCs are reordered according to their correlation degree with output variable via MI and denoted as MI–PCs, and the optimal MI–PCs are determined as the input variables for RVM according to the predicting performance. Further, MI–PCA–MI–RVM is employed to develop the 4-carboxybenzaldehyde concentration model. The result shows that MI–PCA–MI–RVM model has better predicting and robust performance than PCA–RVM and RVM models. © 2012 Curtin University of Technology and John Wiley & Sons, Ltd.