Comparing support vector machines to PLS for spectral regression applications

In order to on-line control the quality of industrial products, often spectroscopic methods are used in combination with regression tools. Partial Least Squares (PLS) is the most used regression technique for this task whereas Support Vector Machines (SVMs) are hardly known and used in chemometrics. Theoretically, regression by SVMs (SVR) can be very useful due to its ability to find nonlinear, global solutions and its ability to work with high dimensional input vectors. This paper compares the use and the performance of PLS and SVR for two spectral regression applications. rst application is the use of both high-resolution Raman spectra and low-resolution Raman spectra (which are cheaper to measure) for the determination of two monomer masses during a copolymerisation reaction. In the second application near-infrared (NIR) spectra are used to determine ethanol, water, and iso-propanol mole fractions in a ternary mixture. The NIR spectra used suffer from nonlinear temperature-induced variation which can affect the predictions. Clearly, for both applications, SVR outperformed PLS. With SVR, the usage of the cheaper low-resolution Raman spectra becomes more feasible in industrial applications. Furthermore, regression by SVR appears to be more robust with respect to nonlinear effects induced by variations in temperature.