Standardization of near infrared spectra based on multi-task learning

binfeng yu,haibo ji,yu kang
DOI: https://doi.org/10.1080/00387010.2015.1055770
2016-01-01
Spectroscopy Letters
Abstract:In order to model the near infrared spectral difference between two instruments, this paper presents an approach based on multi-task learning for multivariate instrument standardization. A multi-task learning approach using trace norm regularization is employed to estimate the transformation matrix in direct standardization, and then is extended to the construction of the nonlinear transformation. The proposed approach is compared with the piecewise direct standardization (PDS) on two real data sets. Experimental results show that the proposed approach sometimes outperforms the conventional PDS method, and the multi-task learning method can be a promising way to overcome the over-fitting problem existing in direct standardization.
What problem does this paper attempt to address?