Support vector machine classification and regression based hybrid modeling method and its application in Raman spectral analysis

Ruan Hua,LianKui Dai

2010-01-01

Abstract:In multivariate calibration, the model performance depends not only on model structure and parameters, but also the training sample distribution. In practical application, training samples often distribute unevenly in space, Therefore the model performance based on whole training sample set degrades. Aiming at this problem, a new hybrid modeling method based on support vector classification and regression is proposed in this paper. A classification decision tree with binary tree form is firstly built using least-squares support vector classifier; then least-squares support vector regression is used to construct the regression model for each class. For an unknown sample, the established classification decision tree is applied to determine its class and then corresponding regression model is selected for quantitative analysis. This method was applied to Raman spectral analysis of gasoline octane number; and the standard prediction error is 0.22. However, the standard prediction error from the calibration based on the whole data set is 0.54, which is approximately 2.5 times larger. Analysis result shows that the proposed method has greatly improved the model performance and thus demonstrates its potential for general purpose analysis.