Iterative Robust Least Square Support Vector Machine for Spectral Analysis

The aim of this study is to develop a novel robust regression algorithm: robust least squares support vector machine (RLS-SVM), to overcome the limitation of the existing support vector machine at high percent of contamination for spectral analysis. In the algorithm, firstly a subset is selected randomly from the original data set to build regression model and the robust estimates of the residuals for the whole set are generated; then the confidence interval of the residuals distribution is applied iteratively to detect outliers. Finally, the LS-SVM estimates are created from the regression model being trained with the selected subset without outliers. The proposed algorithm is applied in the near infrared spectral analysis of gasoline samples in order to predict their octane number with some outliers. Compared with other support vector machine algorithms, the test results show the breakdown point value for the algorithm can be over 45 %. The results also show its priority in predicted precision.