Abstract
An ensemble approach, based on the combination of multiple linear regressions in subspace and variable clustering and therefore named VCS-MLR, was proposed for near-infrared spectroscopy (NIRS) calibration. By an experiment involving the determination of five components in tobacco samples, it was shown that VCS-MLR improved the performance by 61.4, 23.3, 10.2, 20.5, and 18, respectively, with respect to partial least squares regression (PLSR). The results confirmed that VCS-MLR can result in a more accurate calibration model but without the increase of computational burden. Moreover, the superiority of VCS-MLR was highlighted for small sample problems.
Acknowledgments
This work was supported by Sichuan Province Science Foundation for Youths (09ZQ026-066) and the Scientific Research Startup Fund for Doctors, Yibin University (2008B06).
Notes
Note. The values in brackets denote the optimal number of latent variables for PLSR and the optimal subspace size for both RS-MLR and VCS-MLR; the percentages of the last row denote the relative RMSEP reduction with respect to PLSR.