Supervised Manifold Learning for NIR Modeling of Cigarette Brand Identification

For NIR data has the character of high dimension, nonlinear, and high noise, we often confront the problem of dimensionality reduction when building the classification model on Near-Infrared spectra data. Traditional classification methods and linear dimensionality reduction techniques are difficult to improve the model performance. In this paper, a novel nonlinear modeling for NIR spectra analysis was proposed by combining S-Isomap and KNN. S-Isomap is a supervised manifold learning method which can effectively find out the intrinsic low dimensional structure and extract important information from the raw data. Compared with KLLE, KPCA, and other classification methods such as SVM or LDA, the results show that S-Isomap-KNN method performs the best on the modeling of cigarette brand identification. The method is also a good technique for visualization.