On Multivariate Calibration Problems

Multivariate calibration is a classic problem in the analytical chemistry field and frequently solved by partial least squares method in the previous work. Unfortunately there are so many redundant features in that problem, that feature selection are often performed before modeling by partial least squares method and the features not selected are usually discarded. In this paper, the redundant information is, however, reused in the learning of partial least squares method within the frame of multitask learning. Results on three multivariate calibration data sets show that multitask learning can greatly improve the accuracy of partial least squares method.