Near-infrared calibration transfer via support vector machine and transfer learning

Abstract

Support Vector Regression (SVR) has been widely used as a nonlinear regression method in multivariate calibration. This paper investigates the transfer of SVR models between different environmental conditions based on a transfer learning method. In this method, the difference of the SVR regression vectors between old and new conditions is penalized by a L₂ norm penalty, which can prevent the new SVR model from changing too much. By utilizing the commonality of SVR models between different conditions, this method reduces the requirement of training samples for building a new SVR model and thus can update the old SVR model when a new instrument or new conditions are encountered. The new method was compared with several traditional calibration transfer methods including the Piecewise Direct Standardization (PDS) method and two kinds of model updating methods. Experimental results on three benchmark near-infrared datasets show that the transfer results obtained by the proposed approach can be better than those obtained by the existing calibration transfer methods, especially when the standardization samples are selected from the support vectors used in the primary model.