Monotone Support Vector Regression And Its Application

In order to realize monotone regression, a modified monotone support vector regression (MonSVR) was proposed. In the MonSVR, monotone knowledge is described by first order differences inequality constraints of the kernel expansion, which is directly added into SVR dual formulation to maintain a convex optimization problem. The analysis result of the number of the added constraints shows that monotonicity of the regression curve is easily obtained by MonSVR on the condition that a small finite set of constraints on some discrete points are added in the SVR dual formulation.To model the multiple regression, additive support vector regression is employed. In additive support vector regression, the influence of independent variables on the dependent variable is realized through an additive structure of kernels and monotone knowledge on the independent variables is realized through adding extra constraints to the formulation of the additive support vector regression. And the method is referred as AddMonSVR.Further, MonSVR is employed to develop the true boiling point curve of crude oil. The results show that the monotonicity is obtained by MonSVR and that the whole predicting performance of MonSVR is better than some references’ methods. And the method AddMonSVR is employed in modeling the side burning reaction of the industrial p-xylene oxidation. Through long tern operation experience, the monotone knowledge between the independent and dependent variables has been known. And by using AddMonSVR, they can easily be incorporated into the data-model. The knowledge incorporated data-driven model shows better prediction accuracy than general data-driven models and easily gives a qualitatively data-analytic visualization of burning side effects.